Partial Differential Equations
Introduction
General Features of Partial Differential Equations
Classification of Partial Differential Equations
Classification of Physical Problems
Elliptic Partial Differential Equations
Parabolic Partial Differential Equations
Hyperbolic Partial Differential Equations
The Convection-Diffusion Equation
Initial Values and Boundary Conditions
Well-Posed Problems
Summary
lL1_ INTRODUCTION
Partial differential equations (PDEs) atise in all fields of engineering and science. Most
real physical processes are governed by partial differential equations. In many cases,
simplifying approximations are made to reduce the governing PDEs to ordinary differ-
ential equations (ODEs) or even to algebraic equations. However, because of the ever
increasing requirement for more accurate modeling of physical processes, engineers and
scientists are more and more required to solve the actual PDEs that govern the physical
problem being investigated, Part IM is devoted to the solution of partial differential
equations by finite difference methods.
For simplicity of notation, the phrase partial differential equation frequently will be
replaced by the acronym PDE in Part IIL. This replacement cenerally makes the text flow502 Part lil
The objectives of Part Il are:
1. To present the general features of partial differential equations
2. To discuss the relationship between the type of physical problem being solved,
the classification of the corresponding governing partial differential equation,
and the type of numerical method required
3. To present examples to illustrate these concepts.
1.2 GENERAL FEATURES OF PARTIAL DIFFERENTIAL EQUATIONS
A partial differential equation (PDE) is an equation stating a relationship between a
function of two or more independent variables and the partial derivatives of this function
with respect to these independent variables. The dependent variable f is used as a generic
dependent variable throughout Part III, In most problems in engineering and science, the
independent variables are either space (x, y, z) or space and time (x, y, z, t). The dependent
variable depends on the physical problem being modeled. Examples of three simple partial
differential equations having two independent variables are presented below:
(iII.1)
(11.2)
(mm.3)
Equation (111.1) is the two-dimensional Laplace equation, Eq. (U11.2) is the one-dimen-
sional diffusion equation, and Eq, (IIl,3) is the one-dimensional wave equation. For
simplicity of notation, Eqs. (111.1) to (11.3) usually will be written as
Sx thy =9 (IL4)
f= fe i)
Sa= Che (IIL.6)
where the subscripts denote partial differentiation.
The solution of a partial differential equation is that particular function, f(x, y) or
(x, which satisfies the PDE in the domain of interest, D(x, y) or D(x, 1), respectively,
and satisfies the initial and/or boundary conditions specified on the boundaries of the
domain of interest. In a very few special cases, the solution of a PDE can be expressed inPartial Differential Equations 503
where V? is the Laplacian operator, which in Cartesian coordinates is
z ez ez
ae tae
Equation (II1.5), which is the one-dimensional diffusion equation, in four independent
variables is
=U Sec thy +hz) =4VF (am.9)
The parameter at is the diffusion coefficient. Equation (III.6), which is the one-dimensional
wave equation, in four independent variables is
Sn =O fee thy the) =O VF (uL.10)
The parameter c is the wave propagation speed. Problems in two, three, and four
independent variables occur throughout engineering and science.
Equations (II1.4) to (III.10) are all second-order partial differential equations. The
order of a PDE is determined by the highest-order derivative appearing in the equation. A
large number of physical problems are governed by second-order PDEs. Some physical
problems are governed by a first-order PDE of the form
of, + bf, =0 (1.11)
where a and b are constants. Other physical problems are govemed by fourth-order PDEs
such as
Sreox +hesy + hay = 9 (11.12)
Equations (II1.4) to (111.12) are all Jinear partial differential equations. A linear PDE
is one in which all of the partial derivatives appear in linear form and none of the
coefficients depends on the dependent variable. The coefficients may be functions of the
independent variables, in which case the PDE is a linear, variable coefficient, PDE. For
example,
af, + bxf, =0 (11.13)
where a and b are constants, is a variable coefficient linear PDE, whereas Eqs. (IIL.4) to
(111.12) are all linear PDEs. If the coefficients depend on the dependent variable, or the
derivatives appear in a nonlinear form, then the PDE is nonlinear. For example,
the + of, =0 (11.14)
af + bf, =0 (1WL.15)
are nonlinear PDEs.
Equations (III.4) to (111.15) are all homogeneous partial differential equations. An
Peet
V= (11.8)
SPE eee eee ee ee eee eee eee504 Part Ill
Equations (II1.4) to (111.16) are all examples of a single partial differential equation
governing one dependent variable. Many physical problems are governed by a system of
PDEs involving several dependent variables, For example, the two PDEs
af, +g, =0 (i11.17a)
Ag, + Bf, =0 (101.176)
comprise a system of two coupled partial differential equations in two independent
variables (x and f) for determining the two dependent variables, f(x, ¢) and g(x, ¢). Systems
containing several PDEs occur frequently, and systems containing higher-order PDEs
occur occasionally. Systems of PDEs are generally more difficult to solve numerically than
a single PDE.
As illustrated in the preceding discussion, a wide variety of partial differential
equations exists. Each problem has its own special governing equation or equations and its
own peculiarities which must be considered individually. However, useful insights into the
general features of PDEs can be obtained by studying three special cases. The first special
case is the general quasilinear (i.e., linear in the highest-order derivative) second-order
nonhomogeneous PDE in two independent variables, which is
Afec + Bhy + Shy + Df. + Ef, + Ff =G (111.18)
where the coefficients A to C may depend on x, y, f,, and f,, the coefficients D to F may
depend on x, y, and f, and the nonhomogeneous term G may depend on x and y. The
second special case is the general quasilinear first-order nonhomogeneous PDE in two
independent variables, which is
af, + bf, = (111.19)
where a, b, and c may depend on x, f, and f. The third special case is the system of two
general quasilinear first-order nonhomogeneous PDEs in two independent variables, which
can be written as
af, + bf, + cg, + dg, =e (111.20a)
Af, + Bf, + Ca, + De, = E (11.20b)
where the coefficients a to d and A to D and the nonhomogeneous terms e and E may
depend on x, t, f, and g. The general features of these three special cases are similar to the
general features of all the PDEs discussed in this book. Consequently, these three specialPartial Differential Equations 505
are governed by fourth-order PDEs. The classification of PDEs is most easily explained for
a single second-order PDE. Consequently, in the following discussion, the general
quasilinear (i.e., linear in the highest-order derivative) second-order nonhomogeneous
PDE in two independent variables [i.e., Eq. (I11.18)] is classified first. The classification of
the general first-order nonhomogeneous PDE in two independent variables [i.e., Eq.
(111.19)] is studied next. Finally, the classification of the system of two quasilinear first-
order nonhomogeneous PDEs [i.e., Eq. (Ill.20)] is studied. The classification of higher-
order PDEs, larger systems of PDEs, and PDEs having more than two independent
variables is considerably more complicated.
The general quasilinear second-order nonhomogeneous partial differential equation
in two independent variables is [see Eq. (I1.18)]
Aes + Bh + Gy + Dh + Ef, + Ff = GE (L21)
The classification of Eq. (111.21) depends on the sign of the discriminant, B? ~ 4AC, as
follows:
B’-44C Classification
Negative Elliptic
Zero Parabolic
Positive Hyperbolic
The terminology elliptic, parabolic, and hyperbolic chosen to classify PDEs reflects
the analogy between the form of the discriminant, B* — 4C, for PDEs and the form of the
discriminant, B? — 44C, which classifies conic sections. Conic sections are described by
the general second-order algebraic equation
Ax + Bry + OC? + Dx + By +F =0 (11.22)
The type of curve represented by Eq. ([1.22) depends on the sign of the discriminant,
B’ ~ 4AC, as follows:
B—4AC Type of curve
Negative Ellipse
Zero Parabola
Positive Hyperbola
The analogy to the classification of PDEs is obvious. There is no other significance to the506 Part Ill
have some very special features. The prefix hyper is used to denote spaces of more than
three dimensions, that is, xyzt spaces, and curves and surfaces within those spaces. In two-
dimensional space, which is the case considered here, characteristics are paths (curved, in
general) in the solution domain along which information propagates. In other words,
information propagates throughout the solution domain along the characteristics paths.
Discontimuities in the derivatives of the dependent variable (if they exist) also propagate
along the characteristics paths. If a PDE possesses real characteristics, then information
propagates along these characteristics. If no real characteristics exist, then there are no
preferred paths of information propagation. Consequently, the presence or absence of
characteristics has a significant impact on the solution of a PDE (by both analytical and
numerical methods).
A simple physical example can be used to illustrate the physical significance of
characteristic paths. Convection is the process in which a physical property is propagated
(ie., convected) through space by the motion of the medium occupying the space. Fluid
flow is a common example of convection. The convection of a property f of a fluid particle
in one dimension is governed by the convection equation
ftuf, =0 (11.23)
where u is the convection velocity. A moving fluid particle carries (convects) its mass,
momentum, and energy with it as it moves through space. The location x(t) of the fluid
particle is related to its velocity u(¢) by the relationship
dx
== TT.24)
aan (11.24)
The path of the fluid particle, calléd its pathline, is given by
7
X=Xp +| u(t) dt (111.25)
to
The pathline is illustrated in Figure Illa.
Along the pathline, the convection equation fi.e., Eq. (II1.23)] can be written as
fru =f Eh=
which can be integrated to yield f = constant. Consequently, the fluid property f is
convected along the pathline, which is the characteristic path associated with the
convection equation. Equation (111.24), which is generally called the characteristic
equation, is the differential equation of the characteristic path. The physical significance
of the pathline (ie., the characteristic path) as the path of propagation of the fluid property
' ae 7 w| |
(111.26)Partial Differential Equations 507
Jf is quite apparent for fluid convection. Equation (111.26), which is generally called the
compatibility equation, is the differential equation which applies along the characteristic
path,
To illustrate further the property of a characteristic path as the path of propagation in
a convection problem, consider the triangular property distribution illustrated in Figure
IIl.1b. As the fiuid particles move to the right at the constant convection velocity u, each
particle carries with it its value of the property f. Consequently, the triangular property
distribution simply moves (i.e., convects) to the right at the constant convection velocity u,
unchanged in magnitude and shape. The apex of the triangle, which is a point of
discontinuous slope in the property distribution, convects as a discontinuity in slope at
the convection velocity u. This simple convection example illustrates the significance of
characteristic paths.
Let’s return to the classification of Eq. (IlI.21). Several procedures exist for
determining the characteristics, and hence the classification, of PDEs. Because discon-
tinuities in the derivatives of the solution, if they exist, must propagate along the
characteristics, one approach is to answer the following question: Are there any paths in
the solution domain D(x, y) passing through a general point P along which the second
derivatives of f(x, y), that is, fir, fy, and f,,, are multivalued or discontinuous? Such paths,
if they exist, are the paths of information propagation, that is, the characteristics.
One relationship for determining the three second derivatives of f(x, y) is given by
the partial differential equation itself, Eq. (111.21). Two more relationships are obtained by
applying the chain rule to determine the total derivatives of f, and f,, which are themselves
functions of x and y. Thus,
Uf) = fee Ce + fy dy (11.27)
Uf) =fy thy Y (1.27b)
Equations (111.21) and (111.27) can be written in matrix form as follows:
AB CVT fa -Df, - Ef, -F +G
dx dy 0 fy |= | aK) (111.28)
0 dx dy|| fy df)
Equation (111.28) can be solved by Cramer's rule to yield unique finite values of f.,, fy, and
Sy, unless the determinant of the coefficient matrix vanishes. In that case, the second
derivatives of f(x,y) are either infinite, which is physically meaningless, or they are
indeterminate, and thus multivalued or discontinuous.
Setting the determinant of the coefficient matrix of Eq. (III.28) equal to zero yields
A(dyy — B(dx\(dy) + C(dxy = 0 (11.29)
Equation (111.29) is the characteristic equation corresponding to Eq, (111.21). Equation
SRI a ae508 Part Ii
S(x,y) may be multivalued or discontinuous. These two families of curves, if they exist, are
the characteristic paths of the original PDE, Eq. (II.21).
The two families of characteristic curves may be complex, real and repeated, or real
and distinct, according to whether the discriminant, B’ — 4AC, is negative, zero, or
positive, respectively. Accordingly, Eq. (III.21) is classified as follows:
B’—44C Characteristic curves _ Classification
Negative Complex Elliptic
Zero Real and repeated Parabolic
Positive Real and distinct Hyperbolic
Consequently, elliptic PDEs have no real characteristic paths, parabolic PDEs have one
real repeated characteristic path, and hyperbolic PDEs have two real distinct characteristic
paths.
The presence of characteristic paths in the solution domain leads to the concepts of
domain of dependence and range of influence. Consider a point P in the solution domain,
D(x, y). The domain of dependence of point P is defined as the region of the solution
domain upon which the solution at point P, f(x), ¥,), depends. In other words, /(x,, ¥,)
depends on everything that has happened in the domain of dependence. The range of
influence of point P is defined as the region of the solution domain in which the solution
S(x,y) is influenced by the solution at point P. In other words, f(x,, yp) influences the
solution at all points in the range of influence.
Recall that parabolic and hyperbolic PDEs have real characteristic paths. Conse-
guently, they have specific domains of dependence and ranges of influence. Elliptic PDEs,
on the other hand, do not have real characteristic paths. Consequently, they have no
specific domains of dependence or ranges of influence. In effect, the entire solution
domain of an elliptic PDE is both the domain of dependence and the range of influence of
every point in the solution domain. Figure III.2 illustrates the concepts of domain of
dependence and range of influence for elliptic, parabolic, and hyperbolic PDEs.
Unlike the second-order PDE just discussed, a single first-order PDE is always
hyperbolic. Consider the classification of the single general quasilinear first-order non-
homogeneous PDE, Eq, (III.19):
af, + bf, =e (11.31)
og WT NU]Partial Differential Equations 509
The characteristic paths, if they exist, are determined by answering the following question:
Are there any paths in the solution domain D(x, £) passing through a general point P along
which the first derivatives of f(x, t) may be discontinuous? Such paths, if they exist, are the
characteristics of Eq. (11131).
One relationship for determining f; and f, is given by Eq. (1131). Another
relationship is given by the total derivative of f(t, x):
df =fdttf, de (111.32)
Equations (111.31) and (111.32) can be written in matrix form as
[4 al ~ [s| (1.33)
As before, the partial derivatives f, and f, are uniquely determined unless the determinant
of the coefficient matrix of Eq. (III.33) is zero. Setting that determinant equal to zero gives
the characteristic equation, which is
adx—bdt=0 (1.34)
Solving Eq. (111.34) for dx/dt gives
2 (135)
Equation (III.35) is the differential equation for a family of paths in the solution domain
along which f, and f, may be discontinuous, or multivalued. Since a and b are real
functions, the characteristic paths always exist. Consequently, a single quasilinear first-
order PDE is always hyperbolic. The convection equation, Eq. (III.23), is an example of
such a PDE.
As a third example, consider the classification of the system of two general coupled
quasilinear first-order nonhomogeneous partial differential equations, Eq. (II.20):
af, + bf, + cg, + dg, =e (UL.36a)
Af, + Bf, + Cg, + Dg, = E (I11.36b)
The characteristic paths, if they exist, are determined by answering the following question:
Are there any paths in the solution domain D(x, 1) passing through a general point P along510 Part Ill
Equations (111.36) and (111.37), which comprise a system of four equations for determining
Su fe & and g,, can be written in matrix form as follows:
abe aff e
ABpcobI||f£|_le
a dk 0 Offs, ||] af (01:38)
00 a d&dlg. dg
As before, the partial derivatives are uniquely determined unless the determinant of the
coefficient matrix of Eq. (III.38) is zero. Setting that determinant equal to zero yields the
characteristic equation, which is
(aC — Ac)(dx)” — (aD — Ad + bC ~ Be\(dx)(dt) + (6D ~ Bdy(dt? =0 —(IIL.39)
Equation (111.39), which is a quadratic equation in dx/dt, may be written as
A(dx)? — B(dx\(at) + C(aty’ = 0 (11.40)
where A = (aC — Ac), B = (aD — Ad + bC — Bc), and C = (6D — Bd). Equation (111.40)
can be solved by the quadratic formula to yield
(1.41)
dt 24
Equation (111.41) is the differential equation for two families of curves in the xt
plane, corresponding to the + signs. Along these two families of curves, the first
derivatives of f(x, 1) and g(x,¢) may be multivalued. These two families of curves, if
they exist, are the characteristic paths of the original system of PDEs, Eq. (11.36). The
slopes of the two families of characteristic paths may be complex, real and repeated, or real
and distinct, according to whether the discriminant, B’ — 44, is negative, zero, or
positive, respectively. Accordingly, Eq. (III.36) is classified as follows:
B-44T Classification
Negative Elliptic
Zero Parabolic
Positive Hyperbolic
In summary, the physical interpretation of the classification of a partial differential
equation can be explained in terms of its characteristics. If real characteristics exist,
preferred paths of information propagation exist. The speed of propagation of information
Pee ee eee eee eee eee eee eee eee ree eee eeePartial Differential Equations 511
solution at all the other points. Since there are no curves along which the derivatives may
be discontinuous, the solution throughout the entire solution domain must be continuous.
Physical problems governed by PDEs that have complex characteristics are equilibrium
problems. Thus, elliptic PDEs govern equilibrium problems. These concepts are related to
the classification of physical problems in the next section.
The significance of the classification of a PDE as it relates to the numerical
approximation of the PDE is as follows. For an elliptic PDE which contains only
second-order spatial derivatives, there are no preferred physical information propagation
paths. Consequently, all points are dependent on all other points and all points influence all
other points. This physical behavior should be accounted for in the numerical approxima-
tion of the PDE.
For a parabolic PDE which contains only second-order spatial derivatives, the
preferred physical information propagation paths are lines (or surfaces). of constant time
(or constant timelike variable). In other words, at each time (or timelike variable) level, all
points are dependent on all other points and all points influence all other points. This
physical behavior should be accounted for in the numerical approximation of the PDE.
For a hyperbolic PDE which contains only first-order spatial derivatives, distinct
physical information propagation paths exist. Physical information propagates along these
distinct physical propagation paths. This physical behavior should be accounted for in the
numerical approximation of the PDE.
Elliptic and parabolic PDEs exist which contain first-order spatial derivatives in
addition to second-order spatial derivatives, In such cases, the physical behavior associated
with the second-order spatial derivatives is the same as before, but the physical behavior
associated with the first-order spatial derivatives acts similarly to the behavior of the first-
order spatial derivatives in a hyperbolic PDE. This physical behavior should be accounted
for in the numerical approximation of such PDEs.
I.4_ CLASSIFICATION OF PHYSICAL PROBLEMS
Physical problems fall into one of the following three general classifications:
1, Equilibrium problems
2. Propagation problems
3. Eigenproblems
Each of these three types of physical problems has its own special features, its own
particular type of governing partial differential equation, and its own special numerical
solution method. A clear understanding of these concepts is essential if meaningful
numerical solutions are to be obtained.
1.4.1 Equilibrium Problems
PRP e Cee eee ee eee ete eee eae ae eae512 Part Ill
Domain of dependence
and
Range of influence
Figure III.3. Solution domain for an equilibrium problem.
points. Figure IIL.3 illustrates the closed solution domain D(x, y) and its boundary B.
Consequently, equilibrium problems are solved numerically by relaxation methods.
A classical example of an equilibrium problem governed by an elliptic PDE is steady
heat diffusion (i.e., conduction) in a solid (see Section IIL5). The governing PDE is the
Laplace equation
VT =0 (0.42)
where 7’ is the temperature of the solid. In two dimensions, Eq. (II1.42) is
Tz + Dy =0 (1.43)
Along the boundary B, the temperature 7(x, y) is subject to the boundary condition
aT +bT, =c¢ (11.44)
at each point on the boundary, where 7,, denotes the derivative normal to the boundary.
Equilibrium problems arise in all fields of engineering and science. Equilibrium
problems in partial differential equations are analogous to boundary-value problems in
ordinary differential equations, which are considered in Chapter 8.
L.4.2 Propagation Problems
Propagation problems are initial-value problems in open domains (open with respect to
one of the independent variables) in which the solution f(x, r) in the domain of interest
D(x, #) is marched forward from the initial state, guided and modified by boundary
conditions. Propagation problems are governed by parabolic or hyperbolic PDEs. Propa-
gation problems in PDEs are analogous to initial-value problems in ODEs, which are
considered in Chapter 7.
The _tajority | of propagation problems are unsteady problems. The diffusion
eee eeePartial Differential Equations 513
in which the initial property distribution at location yo, f(x, ¥) = F(@), is marched forward
in space in the y direction. The general features of these two PDEs, Eqs. (IIl.45a) and
(IIL45b), are identical, with the space coordinate y in Eq. (III.45b) taking on the character
of the time coordinate ¢ in the diffusion equation. Consequently, the marching direction
in a steady-state space propagation problem is called the timelike direction, and the
corresponding coordinate is called the timelike coordinate. The space direction in which
diffusion occurs [i.e., the x direction in Eqs. (I11.45a) and (III.45b)] is called the spacelike
direction, and the corresponding coordinate is called the spacelike coordinate. In the
present discussion, unsteady and steady propagation problems are considered simulta-
neously by considering the time coordinate in the diffusion equation, Eq. (III.45a), to be a
timelike coordinate, so that Eq. (III.45a) models both unsteady and steady propagation
problems.
The solution of a propagation problem is subject to initial conditions specified at a
particular value of the timelike coordinate and boundary conditions specified at each point
on the spacelike boundary. The domain of interest D(x, f) is open in the direction of the
timelike coordinate. Figure III.4 illustrates the open solution domain D(x, t) and its
boundary B which is composed of the initial time boundary and the two physical
boundaries. Propagation problems are initial-value problems, which are solved by
marching methods.
A classical example of a propagation problem governed by a parabolic PDE is
unsteady heat diffusion in a solid (see Section III.6). The governing PDE is the diffusion
equation:
T,=0VT (uL.46)
where T is the temperature and « is the thermal diffusivity of the solid. In one space
dimension, Eq. (111.46) is
T, = 0D y (11.47)
Since Eq. (111.47) is first order in time, values of 7 must be specified along the initial time
boundary. Since Eq. (111.47) is second order in space, values of T must be specified along
both space boundaries,
Parabolic PDEs have real repeated characteristics. As shown in Section III.6 for the
diffusion equation, parabolic PDEs have specific domains of dependence and ranges of
influence and infinite information propagation speed. Thus, the solution at each point in
tf Open boundary
March514 Part Ill
the solution domain depends on a specific domain of dependence and influences the
solution in a specific range of influence.
In two variables (e.g., space x and time f), parabolic PDEs have two real repeated
families of characteristics. As illustrated in Figure IIL.5, both families of characteristics
have zero slope in the xt plane, which corresponds to an infinite information propagation
speed. Consequently, parabolic PDEs behave like hyperbolic PDEs in the limit where the
information propagation speed is infinite. Thus, the solution at point P depends on the
entire solution domain upstream of and including the horizontal line through point P itself.
The solution at point P influences the entire solution domain downstream of and including
the horizontal line through point P itself. However, the solution at point P does not depend
on the solution downstream of the horizontal line through point P, nor does the solution at
point P influence the solution upstream of the horizontal line through point P. Numerical
methods for solving propagation problems governed by parabolic PDEs must take the
infinite information propagation speed into account.
A classical example of a propagation problem governed by a hyperbolic PDE is
acoustic wave propagation (see Section III.7). The governing PDE is the wave equation
=a VP’ (IIL.48)
where P’ is the acoustic pressure (ie., the pressure disturbance) and a is the speed of
propagation of small disturbances (i.e., the speed of sound), In one space dimension,
Eq. (111.48) is
Pi, = &P%, (111.49)
Since Eq. (III.49) is second order in time, initial values of both P’ and P; must be specified
along the initial time boundary. Since Eq. (III.49) is second order in space, values of P’
must be specified along both space boundaries.
Hyperbolic PDEs have real distinct characteristics. As shown in Section III.7 for the
wave equation, hyperbolic PDEs have finite domains of dependence and ranges of
influence and finite information propagation speed. Thus, the solution at each point in
the solution domain depends only on the solution in a finite domain of dependence and
influences the solution only in a finite range of influence.
In two variables (e.g., space x and time 1), hyperbolic PDEs have two real and
distinct families of characteristics. As illustrated in Figure III.6, both families of
March } Open
boundaryPartial Differential Equations 515
characteristics have finite slope in the x’ plane, which corresponds to a finite information
speed. For acoustic fields, these two real families of characteristics are the right-running
(ie., in the positive x direction) and left-running (.e., in the negative x direction) acoustic
waves. These characteristics are illustrated in Figure 111.6 at a particular point P. The
characteristics have finite information propagation speed, thus giving rise to a finite
domain of dependence and a finite range of influence for each point in the solution
domain. The solution at point P depends only on the solution within the domain of
dependence defined by the characteristics from the upstream portion of the solution
domain. The solution at point P influences only the solution within the range of influence
defined by the downstream propagating characteristics. The portion of the solution domain
outside of the domain of dependence and the range of influence of point P neither
influences the solution at.point P nor depends on the solution at point P. Numerical
methods for solving propagation problems governed by hyperbolic PDEs must take the
finite information propagation speed into account.
From the above discussion, it is seen that propagation problems are governed by
either a parabolic or a hyperbolic PDE. These two types of PDEs exhibit many similarities
(e.g., an open boundary, initial data, boundary data, domains of dependence, and ranges of
influence). Both types of problems are solved numerically by marching methods. However,
there are significant differences in propagation problems governed by parabolic PDEs and
hyperbolic PDEs, due to the infinite information propagation speed associated with
parabolic PDEs and the finite information propagation speed associated with hyperbolic
PDEs. These differences must be accounted for when applying marching methods to these
two types of partial differential equations.
Propagation problems arise in all fields of engineering and science. Propagation
problems governed by hyperbolic PDEs are somewhat analogous to initial-value problems
in ODEs, while propagation problems governed by parabolic PDEs share some of the
features of both initial-value and boundary-value problems in ODEs. Table III.1
summarizes the general features of PDEs as presented in this section.
IL4.3 Eigenproblems
Eigenproblems are special problems in which the solution exists only for special values
(ie., eigenvalues) of a parameter of the problem. The eigenvalues are to be determined in
March Open
boundary516 Part tit
Table IIL.1 General Features of Partial Differential Equations
Type of physical problem
Equilibrium Propagation
Mathematical. Elliptic Parabolic Hyperbolic
classification
of the PDE
Characteristics Complex Real and repeated Real and distinct
Information ‘Undefined Infinite Finite
propagation
speed
Domain of Entire Present and Past solution
dependence solution entire past domain between,
domain solution domain characteristics
Range of Entire Present and Future solution
influence solution entire future domain between
domain solution domain characteristics
Type of Relaxation Marching Marching
numerical
method
addition to the corresponding configuration of the system. Eigenproblems for PDEs are
analogous to cigenproblems for ODEs, which are considered in Section 8.9, Eigen-
problems for PDEs are not considered in this book.
1.5 ELLIPTIC PARTIAL DIFFERENTIAL EQUATIONS
A classical example of an elliptic PDE is the Laplace equation:
Vr=0 (111.50)
The Laplace equation applies to problems in ideal fluid flow, mass diffusion, heat
diffusion, electrostatics, etc. In the following discussion, the general features of the
Laplace equation are illustrated for the problem of steady two-dimensional heat diffusion
in a solid.
Consider the differential cube of solid material illustrated in Figure 111.7. Heat flow
in a solid is governed by Fourier’s law of conduction, which states that
aT
“6 (u1.s1)518 Part Ill
The characteristics associated with Eq. (III.58) are determined by performing a
characteristic analysis. In this case, Eq. (111.28) becomes
10 I4f Te 0
ax dy 0 |} tT, |=| az (111.60)
o & wl{7,| Lar)
The characteristic equation corresponding to Eq. (III.58) is determined by setting the
determinant of the coefficient matrix of Eq, (111.60) equal to zero and solving the resulting
equation for the slopes of the characteristic paths. Thus,
(dy? + (de)? = 0 (111.61)
ee tV-1 {IIL.62)
Equation (111.62) shows that there are no real characteristics associated with the steady
two-dimensional heat conduction equation. Physically, this implies that there are no
preferred paths of information propagation, and that the domain of dependence and
range of influence of every point is the entire solution domain. The temperature at every
point depends on the temperature at all the other points, including the boundaries of the
solution domain, and the temperature at each point influences the temperature at all the
other points. The temperature distribution is continuous throughout the solution domain
because there are no paths along which the derivative of temperature may be discontin-
uous. The domain of dependence and the range of influence of point P are illustrated
schematically in Figure TIL.3.
Another classical example of an elliptic PDE is the Poisson equation, which is the
nonhomogeneous Laplace equation. Consider the problem of steady heat conduction in a
solid with internal energy generation # (J/s) given by
E=Q@,y,.2aV (111.63)
where Q is the energy generation rate per unit volume (J/m°-s). For steady heat flow, the
sum of the energy transferred to the solid by conduction and the internal energy generation
must equal zero. Thus, Eq. (111.56) becomes
a(,7 a (aT a(,aT oy
—lk— —lk— —[k— = ut
a's) +5 (5) “e(hz) 2-8 amen
When the thermal conductivity & is constant (Le., neither a function of temperature nor
location), Eq. (111.64) becomes
eePartial Differential Equations 519
In summary, steady heat conduction is an equilibrium problem and must be solved
by relaxation methods. The PDE governing steady heat conduction is a classical example
of an elliptic PDE.
1.6 PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS
A classical example of a parabolic PDE is the diffusion equation:
=p (U.66)
The diffusion equation applies to problems in mass diffusion, momentum diffusion, heat
diffusion, etc. The general features of the diffusion equation are illustrated for the problem
of unsteady one-dimensional heat diffusion in a solid.
Consider the heat diffusion analysis presented in Section III.5. The net flow of heat
in the x, y, and z directions is given by Eqs. (1IL.53) to (IIT.55), respectively. For steady-
state heat flow, there is no net change in the amount of energy stored in the solid, so the
sum of the net heat flow components is zero. In an unsteady situation, however, there can
be a net change with time in the amount of energy stored in the solid. The energy £ (J)
stored in the solid mass dim (kg) is given by
Egos = dm CT = (o aV)CT = (pCT) aV .67)
where p is the density of the solid material (kg/m’), dV is the differential volume (m*), 7
is the temperature (K), and C is the specific heat (I/kg-K), which is a physical property of
the solid material. The sum of the net heat flow components must equal the time rate of
change of the stored energy. Thus,
A&pCT) _ a oT a or a or
My a ae(Eae) M(B) ba Fe) ames)
Equation (111.68) governs the unsteady diffusion of heat in a solid. When the thermal
conductivity 4, density p, and specific heat C are all constant (i.e., neither functions of
temperature or position), Eq. (IIL.68) simplifies to
T= (Ty + Ty + Tz) =aVT (11.69)
where «=/pC is the thermal diffusivity (m?/s). Equation (111.69) is the diffusion
equation.
For unsteady one-dimensional heat diffusion, Eq, (111.69) becomes520 Part lil
The characteristics associated with Eq. (III.70) are determined by performing a
characteristic analysis. In this case, Eq. (11.28) becomes
« 0 OV[T. L
ds at 04) T, | =| aT) (0.72)
0 a dt{| T, a(T,)
The characteristic equation corresponding to Eq. (III.70) is determined by setting the
determinant of the coefficient matrix of Eq. (II.72) equal to zero and solving for the slopes
of the characteristic paths. In the present case, this yields
adr? =0 (1.73)
dt = £0 (1.74)
t = constant (1.75)
Equation (II1.74) shows that there are two real repeated roots associated with the
characteristic equation, and Eq. (IIJ.75) shows that the characteristics are lines of constant
time. The speed of propagation of information along these characteristic paths is
dx dx
e=S=iyate (1.76)
Consequently, information propagates at an infinite speed along lines of constant time.
This situation is illustrated schematically in Figure IIIS. The information at point P
propagates at an infinite speed in both directions. Consequently, the temperature at point P
depends on the temperature at all other points in physical space at all times preceding and
including the current time, and the temperature at point P influences the temperature at all
other points in physical space at all times after and including the current time. In other
words, the domain of dependence of point P is the finite region ahead of and including the
current time line. The range of influence of point P is the semi-infinite region after and
including the current time line. In this regard, the diffusion equation behaves somewhat
like an elliptic PDE at each time level.
In summary, unsteady heat diffusion is a propagation problem which must be solved
by marching methods. The PDE governing unsteady heat diffusion is a classical example
of a parabolic PDE.
U.7 | HYPERBOLIC PARTIAL DIFFERENTIAL EQUATIONS
A classical example of a hyperbolic PDE is the wave equation:Partial Differential Equations 521
McDonald (1985) or Zucrow and Hoffman (1976)], those basic physical laws yield the
following system of quasi-linear first-order PDEs:
Pp, +V(pV) =0 (1.78)
pV, + p(VV)V + VP =0 (mL.79)
P,+V-VP —a(p,+ V-Vp) =0 (II.80)
where p is the fluid density (kg/m’), V is the fluid velocity vector (m/s), P is the static
pressure (N/m*), and a is the speed of propagation of small disturbances (m/s) (i.e. the
speed of sound). Equations (II1.78) to (1.80) are restricted to the flow of a pure substance
with no body forces or transport phenomena (i.e., no mass, momentum, or energy
diffusion). For unsteady one-dimensional flow, Eqs. (11.78) to (111.80) yield:
Py + Ply + Mp, = 0 (mL81)
pu, + pau, + Py =0 (111.82)
P,+uP,—a@(p, + up,) =0 (IIL83)
Equations (111.81) to (111.83) are more general examples of the simple one-dimen-
sional convection equation
ftw, =0 (uLs4)
where the property fis being convected by the velocity through the solution domain
D(x, 1). Equation (111.84) in three independent variables is
rush swe apever= 2 =o aites)
where uw, v, and w are the velocity components in the x, y, and z directions, respectively, and
the vector onentoe D/Dt is called the substantial derivative:
D
Dt stunt vptegagtvT (111.86)
Equations (111.81) and (111.83) are frequently combined to eliminate the derivatives
of density. Thus,
P+ uP, + plu, =0 (01.87)
Equations (III.81) to (11.83), or Eqs. (111.82) and (IIL.87), are classical examples of a
system of nonlinear first-order PDEs.
Acoustics is the science devoted to the study of the motion of small amplitude
disturbances in a fluid medium. Consider the classical case of infinitesimally small
perturbations in velocity, presure, and density in a stagnant fluid. In that case,
u=mytu=u P=Py+P p=pto' dy+a’ (11.88)522 Part Ill
Equations ({1J.89) and (II1.90) can be combined to solve explicitly for cither the pressure
perturbation P’ or the velocity perturbation w. Differentiating Eq. (111.89) with respect to x
and Eq. (111.90) with respect to f and combining the results to eliminate 1, yields the wave
equation for the pressure perturbation, P’:
Pi, = aP (1.91)
Differentiating Eq. (IIL89) with respect to ¢ and Eq. (11.90) with respect to x and
combining the results to eliminate P,, yields the wave equation for the velocity perturba-
tion u:
ul, = apul, (ail.92)
Equations (111.91) and (111.92) show that the properties of a linearized acoustic field
are governed by the wave equation. In terms of the general second-order PDE defined by
Eq, (IlL.21), A = 1, B~0, and C = —a3. The discriminant, B? — 4C, is
B’ —44C = 0 — 4(1)(-a3) = 4ag > 0 (111.93)
Consequently, Eqs. (II1.91) and (11.92) are hyperbolic PDEs.
Since Eqs. (H1.91) and (101.92) both involve the same differential operators [i.c.,
( Je = @3( Jue], they have the same characteristics. Consequently, it is necessary to study
only one of them, so Eq, (111.91) is chosen, The characteristics associated with Eq. (111.91)
are determined by performing a characteristics analysis. In this case, Eq. (111.28) becomes
1 0 -& VPP, 0
ade (0 |) Py |=] dP) (mm.94)
0 dt dk Py aPy)
The characteristic equation corresponding to Eq. (IIL91) is determined by setting the
determinant of the coefficient matrix of Eq. (111.94) to zero and solving for the slopes of
the characteristic paths. This yields
(axy — ajar = 0 (111.95)
Equation (111.95) is a quadratic equation for d¢/dr. Solving for dx/dr gives
dx
—= 1.
G7 8 (111.96)
X= 4 agt 97)
Equation (111.96) shows ‘that there are two real distinct roots associated with the
characteristic equation, and Eq. (III.97) shows that the characteristic paths are straightPartial Differential Equations 523
point P depends only upon the solution within the finite domain of dependence illustrated
in Figure IIL.6. Likewise, the perturbation pressure at point P influences the solution only
within the finite range of influence illustrated in Figure 116. The finite speed of
propagation of information and the finite domain of dependence and range of influence
must be accounted for when solving hyperbolic PDEs.
Equations (111.89) and (111.90) are examples of a system of two coupled first-order
convection equations of the general form:
f+ ag, =0 (ILL.99a)
8, + af, =0 (UL.99b)
Differentiating Eq, (111.99) with respect to 1, differentiating Eq. (IH.99b) with respect to x
and multiplying by a, and subtracting yields the wave equation:
Sa = Oh, (1.100)
Consequently, the second-order wave equation can be interpreted as a system of two
coupled first-order convection equations.
In summary, unsteady wave motion is a propagation problem which must be solved
by marching methods. The wave equation governing unsteady wave motion is a classical
example of a hyperbolic PDE.
U8 THE CONVECTION-DIFFUSION EQUATION
The Laplace equation and the Poisson equation gover steady diffusion processes. The
diffusion equation governs unsteady diffusion processes, The convection equation and the
wave equation govern unsteady convection processes. When convection and diffusion are
both present in a physical process, the process is governed by the convection-diffusion
equation, The unsteady convection-diffusion equation is given by
AVN =a VS (1.101)
and the steady convection-diffusion equation is given by
VU =a VF (41.102)
Equations (iI.101) and (II1.102) are both second-order PDEs. Consider the unsteady
one-dimensional convection-diffusion equation:
SH Uf, = Oh (III.103)524 Part lll
This characteristic path is not found in a classical characteristic analysis of the unsteady
convection-diffusion equation.
Consider the steady two-dimensional convection-diffusion equation:
uf, + Uf, = A Sor thy) (U.105)
The discriminant of Eq, (111.105) is B? — 44C = —4. Consequently, Eq. (II1.105) is an
elliptic PDE, with no real characteristic paths. However, the terms uf, and »f, model
physical convection, which has distinct information propagation paths. These information
propagation paths are not found in a classical characteristic analysis of the steady
convection-diffusion equation.
The significance of the above discussion is as follows. When numerically approx-
imating PDEs which contain both first-order and second-order spatial derivatives, the
different physical behavior associated with the different spatial derivatives should be taken
into account.
LQ INITIAL VALUES AND BOUNDARY CONDITIONS
A differential equation governs a family of solutions. A particular member of the family of
solutions is specified by the auxiliary conditions imposed on the differential equation.
For steady-state equilibrium problems, the auxiliary conditions consist of boundary
conditions on the entire boundary of the closed solution domain. Three types of boundary
conditions can be imposed:
1, Dirichlet boundary condition: The value of the function is specified.
Ff is specified on the boundary. (111.106)
2. Neumann boundary condition: The value of the derivative normal to the
boundary is specified.
¥ is specified on the boundary. (uL.107)
3. Mixed boundary condition: A combination of the function and its normal
derivative is specified on the boundary.
af + b= is specified on the boundary. (aL. 108)
One of the above types of boundary conditions must be specified at each point on the
boundary of the closed solution domain. Different types of boundary conditions can be
specified on different portions of the boundary.Partial Differential Equations 525
For a PDE containing a second-order time (or timelike) derivative, two initial conditions
are required along the time (or timelike) boundary:
f(x. y, 2,0) = F(x. y, z) on the time boundary (HL110a)
Si y, 2, 0) = G(x, y, z) on the time boundary (UT. 110b)
‘The required boundary conditions on the physical boundaries of the solution domain can
be of the Dirichlet type, Eq. (Ill. 106), the Neumann type, Eq, (II1.107), or the mixed type,
Eq, (IlI.108). Different types of boundary conditions can be specified on different portions
of the boundary.
Proper specifications of the type and number of auxiliary conditions is a necessary
condition to obtain a well-posed problem, as discussed in Section II.10.
1.10 WELL-POSED PROBLEMS
The general features of partial differential equations are discussed in the preceding
sections. Elliptic PDEs govern equilibrium problems in closed domains. No real char-
acteristics exist. Parabolic PDEs govern propagation problems in open domains. Real
repeated characteristics cxist. Hyperbolic PDEs govern propagation problems in open
domains. Real distinct characteristics exist. In all three cases, auxiliary conditions (1.¢.,
initial values and boundary conditions) are required to specify a particular solution of a
PDE. The interrelationship between the type of PDE, the auxiliary data, and whether or not
a solution exists and is unique gives rise to the concept of a well-posed problem,
Hadamard (1923) states that a physical problem is well posed if its solution exists, is
unique, and depends continuously on the boundary and/or initial data.
For an elliptic PDE, the solution domain D(x, y) must be closed, and continuous
boundary conditions must be specified along the entire physical boundary B. The boundary
conditions may be of three types: (a) Dirichlet boundary conditions, (b) Neumann
boundary conditions, or (c) mixed boundary conditions.
For a parabolic PDE, the solution domain D(x, 2) must be open in the time (or
timelike) direction, initial data must be specified along the time (or timelike) boundary, and
continuous boundary conditions must be specified along the physical boundaries of the
solution domain. The boundary conditions can be of the Dirichlet type, the Neumann type,
or the mixed type.
For a hyperbolic PDE, the solution domain D(x, t) must be open in the time (or
timelike) direction, initial data must be specified along the time (or timelike) boundary, and
continuous boundary conditions must be specified along the physical boundaries of the
solution domain. The boundary conditions can be of the Dirichlet type, the Neumann type,526 Part tit
{1.11 SUMMARY
The general features of partial differential equations have been presented, and the concept
of characteristics has been introduced. Characteristics are the physical paths along which
information propagates. Partial differential equations are classified as elliptic, parabolic, or
hyperbolic, according to whether there are no real characteristics, real repeated character-
istics, or real distinct characteristics, respectively. Examples of several partial differential
equations that arise in engineering and science have been presented.
The Laplace equation is a classical example of an elliptic PDE, which must be
solved by relaxation methods:
vr=0 (L111)
Chapter 9 is devoted to the solution of the Laplace equation, The diffusion equation is a
classical example of a parabolic PDE, which must be solved by marching methods:
f=aVye (1.112)
Chapter 10 is devoted to the solution of the diffusion equation. The convection equation is
a classical example of a hyperbolic PDE, which must be solved by marching methods:
FAVS =0 (1.113)
Chapter 11 is devoted to the solution of the convection equation. When convection and
diffusion are both present, the process is governed by the convection-diffusion equation:
HANNS =o VP (0.114)
The convection-diffusion equation is a more complicated example of a parabolic PDE,
which must be solved by marching methods. Section 10.9 is devoted to the solution of the
convection-diffusion equation. Some physica! problems are govemed by a system of
convection equations. In some cases, they can be recast as the wave equation:
tn =O VG (11.120)
The wave equation is a more complicated example of a hyperbolic PDE, which must be
solved by marching methods. Section 11.8 is devoted to the solution of the wave equation