Chapter 2

Numerical Methods for Parabolic PDE

2.1 Introduction

− Parabolic PDE arising in scientific and engineering problems are often of

the form
u t = L (u )
… (1)
nd
where L (u ) is a 2 order PD operator (linear or non-linear).

− Physical examples are: diffusion in an isotropic medium, unsteady state

heat conduction, boundary layer flow over a flat plate.

− The above examples are modelled by the parabolic equation

u t = div [ f grad u ]
= ∇ ⋅ [ f ∇u ]
… (2)

− We now turn to the development of finite difference methods for

parabolic PDE.

− To explain the basic ideas of numerical methods, let's consider the

simple diffusion equation
u t = u xx
… (3)

− Numerical methods for the parabolic PDE may be divided into

Explicit Methods Implicit Methods

• Provides for a non-iterative • Involves iterative

marching process for simultaneous calculations of
obtaining the solution at each many present values in terms
present point in terms of of known preceding and
known preceding and boundary values.
boundary values. • Stability problems are not as
• Stability problems are critical difficulty as in explicit
in this method. methods.
2.2 Explicit Methods

− Let us consider the dimensionless problem

u t = u xx 0 < x < 1, 0 < t <T
… (4)
I.C. u (x , 0) = f (x ) 0 < x <1
… (5)
B.C.'s u (0, t ) = g (t ) 0 < t ≤T
u (1, t ) = h (t ) 0 < t ≤T
… (6)

− Development of a finite difference analogue for this problem necessitates

the introduction of a net whose mesh points are denoted by
x i = i h , ti = j k
where i = 0, 1, 2, …, M; j = 0, 1, 2, …, N
with h = Δx and k = Δt .

− The boundaries are specified by i = 0 & i = M, and the initial line is

denoted by j = 0.

− If an approximation solution Ui,j is assumed to be known at all mesh

points up to time j, a method should be specified to advance the solution
to time j+1.

− The value of Ui,j+1 at x = 0 and x = 1 should be selected as those

boundary conditions specified on u in equations (6), i.e.

U 0, j +1 = g (t j +1 ) & U M , j +1 = h (t j +1 )
… (7)

− At other points, 0 < i < M, the PDE will be replaced by same difference
equation, by approximating the space derivative by the 2nd order central
difference and the time derivative by a forward difference at (x i , t j ) .

− The resulting equation is

1 1
⎡⎣U i , j +1 −U i , j ⎤⎦ = 2 ⎡⎣U i +1, j − 2U i , j + U i −1, j ⎤⎦
k h
i = 1, 2, ..., M-1
… (8)
− Solve equation (8) for U i , j +1 , we obtain the explicit equation for
"marching" ahead in time
U i , j +1 = rU i −1, j + (1 − 2r )U i , j + rU i +1, j
… (9)
Δt k
Where r = = 2
Δx 2
h
 − At t = 0, the solution is prescribed by the initial condition. Advance to
time t = Δt = k is carried out by employing equations (9) and (7).

− The steps are repeated to advance to time t = 2k and so forth…

− The explicit relation (equation (9)) is often called the forward difference
equation.

− It can be proved that the solution of the finite difference analog

converges to the solution of the differential equation as
Δx → 0 & Δt → 0 .

− The stability condition for this scheme is that

1 Δt
0 < r ≤ where r =
2 (Δx ) 2
… (10)
− The above implies that Δt should be small enough to ensure stability.

2.3 Implicit Methods

Δt 1
− The requirement that < places a severe restriction on the
( Δx )
2
2
interval size in the t direction with a resultant increase in computation.

− An implicit formula is one in which two or more unknown values in the

j + 1 row are specified in terms of known values in the j row (and
j − 1 , j − 2 , … if necessary) by a single application of the expansion.

t
j +1
j
i −1 i i +1

x
− The simplest implicit method is that suggested first by O'Brien et al, by
approximately the derivative u xx in the j + 1 row instead of in the j
row. We obtain
1 1
⎡⎣U i , j +1 − U i , j ⎤⎦ = 2 ⎡⎣U i +1, j +1 − 2U i , j +1 + U i −1, j +1 ⎤⎦
k h
… (11)
i.e. in terms of the unknowns in the j + 1 row
− rU i −1, j +1 + (1 + 2r )U i , j +1 − rU i +1, j +1 = U i , j
… (12)
where r = k 2
h
− Crank and Nicolson used an average of approximations in the j and
j + 1 row instead of equation (11).

− More generally, one can introduce a weighting factor λ and replace

equation (11) by
⎧⎪λ ⎣⎡U i −1, j +1 − 2U i , j +1 + U i +1, j +1 ⎦⎤ ⎪⎫
U i , j +1 − U i , j = r ⎨ ⎬
⎪⎩+ (1 − λ ) ⎡⎣U i −1, j − 2U i , j + U i +1, j ⎤⎦ ⎭⎪
… (13)

1
− With λ = 1 , we get O'Brien formulation, and with λ = , we get Crank-
2
Nicolson formula.

− Example:

u t = u xx
… (14)
I.C.'s u ( x , 0) = 1
… (15)
∂u
B.C.'s u ( 0, t ) = 0 , (1, t ) = 1
∂x
… (16)
1 1 k 3 2
Let h = , k = , so r = 2 = and λ = .
3 6 h 2 3

• The finite difference approximation for (14) is equation (13).

∂u
• To discretize the B.C. (1, t ) = 1 , a false boundary is
∂x
introduced, i.e. U 4, j −U 2, j = 1.2h = 2 .
3
… (16)
• Upon successively setting i = 1, 2, and 3, we obtain the matrix
form of the resulting three equations as
⎡0 1 0 ⎤⎡
⎡ 3 −1 0 ⎤ ⎡U 1, j +1 ⎤ ⎢ 2 ⎥ U 1, j ⎤ ⎡ 0 ⎤
⎢ −1 3 −1⎥ U ⎢ ⎥ ⎢1 1 ⎥ ⎢U ⎥ + ⎢ 0 ⎥
⎢ ⎥ ⎢ 2, j +1 ⎥ = ⎢ 2 0 2 ⎥ ⎢ 2, j ⎥ ⎢ ⎥
⎢⎣ 0 −2 3 ⎦⎥ ⎣⎢U 3, j +1 ⎦⎥ ⎢ 0 1 0 ⎥ ⎣⎢U 3, j ⎦⎥ ⎣⎢1 ⎦⎥
⎣ ⎦
… (17)
which can be reduced to
⎡U 1, j +1 ⎤ ⎡1 3 1 ⎤ ⎡U 1, j ⎤ ⎡1⎤
⎢ ⎥ 1 ⎢ ⎥ ⎢ ⎥ 1 ⎢ ⎥
⎢U 2, j +1 ⎥ = 12 ⎢ 3 3 3 ⎥ ⎢U 2, j ⎥ + 18 ⎢3⎥
⎢U 3, j +1 ⎥ ⎢⎣ 2 6 2 ⎥⎦ ⎢⎣U 3, j ⎥⎦ ⎢⎣8⎥⎦
⎣ ⎦
… (18)
And then U 1, j +1 , U 2, j +1 and U 3, j +1 can be determined.
 − In case the points in the field of integration increases leading to very
large system of equations, it would difficult to solve by matrix inversion,
and other solution technique are to be used.

2.4 An Unconditionally Unstable Difference Equation

− If 2nd order central difference is used for discretizing the time derivative,
such as
U i , j +1 −U i , j −1 1
= 2 ⎡⎣U i +1, j − 2U i , j + U i −1, j ⎤⎦
2k h
… (19)
the resulting difference equation is unconditionally unstable.

2.5 Numerical Methods for Non-Linear Problems

− Let's consider the non-linear equation

Lu = f (u )
… (20)
where L is a linear partial differential operator

− In order to deal with non-linear systems, we have to linearize the PDE.

− There are basically 2 methods of linearization.

Picard Method Newton Method

− In Picard method, we introduce a sequence of functions u ( n ) which

satisfy the boundary conditions specified for u , and the linearized
equation is written as
Lu ( n +1) = f ⎡⎣u ( n ) ⎤⎦
… (21)
− When the sequence u converges, the convergence is linear, i.e.
n

u ( n +1) − u =○ ⎡⎣u ( n ) − u ⎤⎦ as n → ∞
… (22)
− As for Newton linearization, one replaces the R.H.S. of equation (21) by
f (u ( n ) ) + (u ( n +1) − u ( n ) ) f ′ (u ( n ) )
∴The linearized PDE becomes
Lu ( n +1) = f (u ( n ) ) + (u ( n +1) − u ( n ) ) f ′ (u ( n ) )
i.e. Lu ( n +1) − f ′ (u ( n ) ) u ( n +1) = f (u ( n ) ) − u ( n ) f ′ (u ( n ) )
… (23)
− This sequence, when convergent, is usually quadratically convergent, i.e.
u ( n +1) − u =○ ⎡⎢(u ( n ) − u ) ⎤⎥ as n → ∞
2

⎣ ⎦
… (24)
− From the above, one can see that the solution of the non-linear problem
is two level,