Numerical Methods in

Fluid Flow and Heat Transfer

Prof. M. G. Mousa
2021
The Method of Modelling and Mathematical Modeling The
process of modeling of physical systems in the real world should
generally follow the path illustrated schematically in the chart
below:
The mathematical formulation of the problem is the reduction
of the physical problem to a set of either algebraic or differential
equations subject to certain assumptions.
INTRODUCTION
Write the Application of Navier-Stokes Equations •
The equations are nonlinear partial differential equations
• No full analytical solution exists
• The equations can be solved for several simple flow conditions
• Numerical solutions to Navier-Stokes equations are increasingly being used
to describe complex flows
Boundary layer
example
example
At y=0.0 u/U=0

At y= u/U=1

At y= du/dy=0
Initial Value Problems Of
Ordinary Differential
Equations

9.1 INTRODUCTION
Problems of solving ordinary differential equations (ODE) are classified into initial
value problems and boundary value problems. Many initial value problems are time-
dependent, in which all the conditions for the solution are specified at the initial time.
The numerical methods for initial value problems are significantly different from
those for boundary value problems. Therefore, the present chapter discussed the
numerical solution methods for the former type only. And chapter 10 describes the
numerical methods for the latter.
The initial value problem of a first-order ODE may be written in the form

Y (t) = ƒ(y, t), y (0) = yo

Where ƒ(y, t) is a function of y and t, and the second equation is initial condition. In
Eq. (9.1.1), the first derivative of y is given by a known function of y and t, and we
desire to compute the unknown function y by numerically integrating ƒ(y, t). If ƒ
grations discussed in chapter 4. However, the fact that ƒ is a function of the unknown
function y makes the integration different.
The initial condition is always a part of the problem definition because the
solution of an initial value problem can be uniquely determined only if an initial
condition is given.

More examples for initial value problems of a firs-order ordinary differential

equations follow:

(a) y ' (t ) = 3 y + 5, y (0) = 1

(b) y ' (t ) = t y + 1, y (0) = 0

1
(c) y ' (t ) = − y ( 0) = 1
1 + y 2'

(d ) y ' = z , z ' = − y, y (0) = 1, z (0) = 0

See Rieder and Busby for examples in engineering.

Numerical methods for ordinary differential equations calculate the
solution on the points, tn = tn-1 + h, where h is the step size (or time interval).
Three types of numerical integration methods for initial value problems are
described in this chapter. Euler methods, Rung-kutta methods, and predictor-
corrector methods. Major aspects of the methods are summarized in table 9. 1.
Table 9.1 summary of the methods for initial value problems of ODEs

Euler other
Name of methods relevant formula local global featuresª

Nonstiff equation:

Euler methods
Forward forward difference 0(h²) 0(h) SS, EC
Modified trapezoidal rule 0(h³) 0(h²) SS, EC, NL
Backward back ward difference 0(h²) 0(h) SS, EC, NL
…………………………………………………………………………………………..
Rang-Kutta
Second-order trapezoidal rule 0(h³) 0(h²) SS, EC, NL
4
³) SS, EC, N ) 0(h 0(h Third-order Simpson's 1/3
4 5
) SS, EC, NL ) 0(h 0(h Forth-order Simpson's 1/3 or 3/8
…………………………………………………………………………………………

Predictor-corrector
Second-order (identical with second-order Range-Kutta) SS, EC
4
³) NS, DC ) 0(h 0(h Third-order Newton backward
) NS, DC 4 ) 0(h 5 0(h Forth-order Newton backward
…………………………………………………………………………………………..

Stiff Equations:

Implicit methods backward difference; SS/ NS

Gear method

Exponential exponential
Transformation transformation SS

ª SS: self-starting capability.

NS: no self-starting capability.
EC: step size can be changed easily in the middle of solution.
DC: difficult to change the step size.
NL: solution of nonlinear equations may be necessary each step.
 Once we learn the numerical methods to solve first-order differential
equations, they can be easily applied to higher-order ODEs becomes a higher-order
ODE can be decomposed to a set of first-order differential equations. For example,
consider
y'''' + a y''' + b y'' + c y' + e y = g (9.1.2)
Where a, b, c, e, and g are constants or known functions of t. the initial conditions are
given as

y (0) = y0 , y' (0) = y'0 ,

y' ' (0) = y' '0 , y' ' ' (0) = y' ' '0

And where yo, y'o, y''o, and y'''o are prescribed values. By defining u, v, and w as

u= y' , v = y' ' , w = y' ' '

Eq. (9.1.2) can be written as

W' + aw + bv + cu + ey = g (9.1.3)
So, Eq. (9.1.2) is equivalent to the set of four first-order ordinary differential
equations:

y' = u , y (0) = y 0

u' = v , u (0) = y ' 0

v' = w , v (0) = y' ' 0
w' = g − aw − bv − cu − ey , w (0) = y' ' ' 0

The
numerical methods for first-order ordinary differential equations are then applicable to the
foregoing set.
The numerical methods may be3 applied to integro-differential equations, too.
For example, consider the equation given by

y ' ' + ay +  y ( s) ds = g , y (0) = y 0 , y ' (0) = y ' 0 (9.1.4)

by defining u and v as

t
u = y' , v= 0
y ( s ) ds

Eq. (9.1.4) becomes

u' = − a y − v + g , u (0) = y '0

u' = y , v (0) = 0 (9.1.5)

y' = u , y (0) = y0

The foregoing set of first-order ordinary differential equations can be solved by a numerical
method.

9.2 EULER METHODS

We set our study with the Euler methods, which are suitable for a quick programming
because of their great simplicity. It should be points out that, as the system of equations
becomes more complicated, the Euler methods are more often used. Indeed, a large fraction
of numerical methods for parabolic and hyperbolic partial differential equations, which are
far more complicated than ordinary differential equations, arte based on Euler methods
rather than the Range-kutta or predictor-corrector, methods.

Euler methods consist of three versions: (a) forward Euler, (b) modified Euler, and
(c) backward Euler methods.
9.2.1 Forward Euler method

The forward Euler method for y' = f (y, t) is obtained by rewriting the forward difference
approximation.

y n+ 1 − y n
 y' n
h
(9.2.2)

To

y n +1 = y n + h f ( yn , t n ) (9.2.3)

Where y'n = f (yn, tn) is used. Using Eq. (9.2.3), yn is recursively calculated as

y1 = y0 + h y' 0 = y0 + h f ( y0 , t 0 )

y 2 = y1 + h f ( y1 , t 1 )

y3 = y2 + h f ( y2 , t 2 )


y n = yn − 1 + h f ( yn − 1 , t n − 1 )

Example 9.1

(a) Solve y' = -20y + 7exp (-0.5t), y (0) = 5, using the forward Euler method with h =
0.001 for 0 < t < 0.02. Do this part by hand calculation
.
(b) Repeat the same for h = 0.01, 0.001, and 0.0001 on a computer for analytical
solution given by

y = 5 e − 20 t + ( 7 / 19.5 ) ( e − 0.5 t
−e − 20 t
)

(Solution)
(a) The first few steps of calculations with h = 0.1 are shown next:

t 0 = 0, y 0 = y ( 0) = 5

t 1 = 0.01, y 1 = y 0 + h y ' 0 = 5 + (0.01) ( − 20(5 ) + 7 exp (0) )

= 4.07

t 2 = 0.02 , y 2 = y 1 + h y ' 1 = 4.07 + (0.01) ( − 20 (4.07) + 7 exp ( − 0.005) )

= 3.32565

tn = n h , yn = y n −1 + h y'n − 1

The computational results for selected values of t with three values of times interval (grid
spacing) are shown in table 9.2.

Table 9.2 forward Euler method

t h = 0.01 h = 0.001 h = 0.0001

0.01 4.07000 (8.693) 4.14924 (0.769) 4.15617 (0.076)ª
0.02 3.32565 (14.072) 3.45379 (1.259) 3.46513 (0.124)
0.03 2.72982 (17.085) 2.88524 (1.544) 2.89915 (0.153)
0.04 2.25282 (18.440) 2.42037 (1.684) 2.43554 (0.167)
0.05 1.87087 (18.658) 2.04023 (1.722) 2.05574 (0.171)
0.06 1.56497 (18.125) 1.72932 (1.690) 1.74454 (0.168)
0.07 1.31990 (17.119) 1.47496 (1.613) 1.48949 (0.160)
0.08 1.12352 (15.839) 1.26683 (1.507) 1.28041 (0.150)
0.09 0.96607 (14.427) 1.09646 (1.387) 1.10895 (0.138)
0.10 0.83977 (12.979) 0.95696 (1.261) 0.96831 (0.126)

ª (error) × 100

Comments: Accuracy of the forward Euler method increases with a decreases in time
interval h. in effect, magnitudes of errors are approximately proportional to h. however,
further reduction of h without using double precision is not advantageous because it
increases numerical error caused by round-off (see chapter 1).

Although the forward Euler method is simple, it has to be used carefully for two kinds
of errors. The first is the truncation errors that we have already seen in example 9.1. The
second is a potential of instability, which occurs when the time constant of the equation is
negative (solution approaches zero if there is no source term), unless time interval h is
sufficiently small. A typical equation for a diminishing solution is y' = - α y, with y (0) =
y0 > 0, where α > 0. The exact solution is y = y oe (- αt). The forward Euler method for
this problem becomes

y n +1 = (1 −  h ) y n

If αh < 1, the numerical solution is diminishing and positive, but if αh > 1, the sign of the
solution alternates. Furthermore, if αh > 2, the magnitude of the solution increases after
each step, and the solution oscillates. This is the instability.
The forward Euler method is applicable to a set of first-order ODEs. Consider a set
of first-order ODEs given by

y ' = f ( y, z , t ) , y ( 0 ) = y 0
(9.2.4)
 z ' = g ( y, z , t ) , z ( 0 ) = z 0

The Euler method for equation (9.2.4) is written as

y n +1 = y n + h y' n = y n + h f ( yn , z n , t n )
(9.2.5)
z n +1 = z n + h z' n = z n + h f ( yn , z n , t n )

A higher-order ordinary differential equation may be broken into a set of coupled first-
order differential equations as mentioned earlier.

Example 9.2

Using the forward Euler method with h = 0.5, find the values of y (1) and y (1)
for

y'' (t) – 0.5 y' (t) + 0.15 y (t) = 0, y' (0) = 0, y (0) = 1

(Solution)

Let y' = z, then the second-order ODE becomes

y' = z, y (0) = 1
z' = 0.05 z – 0.15 y, z (0) = 0
We will denote yn = y (nh) and Zn = z (nh). The initial conditions are expressed as yo = y
(0) = 1 and zo = y' (0) = 0. Using the forward Euler method, y and z at n = 1 and n = 2
become:

t = 0.5:
y' 0 = z 0 = 0

z ' 0 = 0.05 z 0 − 0.15 y 0 = − 0.15

y 1 = y 0 + h y ' 0 = 1 + ( 0 .5 ) ( 0 ) = 1

z 1 = z0 + h z' 0 = 0 + ( 0.5 ) ( − 0.15 ) = − 0.075

t = 1:

y ' 0 = z 0 = − 0.075

z ' 0 = 0.05 z 1 − 0.15 y 1 = ( 0. 05)(−0.075) − ( 0.15)(1) = − 0.15375

y 2 = y 1 + h y ' 1 = 1 + (0.5 ) (− 0.075) = 0.96250

z 2 = z1 + h z ' 1 = − 0.075 + ( 0.5 ) ( − 0.15375) = − 0.15187

Therefore

y (1 ) = y 2 = 0.96250

y ' (1) = z (1) = z 2 = − 0.15187

Example 9.3
 Solve the following set of first-order ODEs by the forward Euler method with h =
0.005Π:

(A) y' = z, y (0) = 1

z' = - y, z (0) = 0

(Solution)
The calculations for the first few steps few steps with h = 0.0005Π are shown next.
t0 = 0: y0 = 1
z 0 = 0

t 1 = 0.0005 : y1 = y 0 + h z 0 = 1 + ( 0.0005 ) (0) = 1.0

z1 = z 0 − h y 0 = 0 − (0.0005 ) (1) = − 0.00157

t 2 = 0.001 : y 2 = y 1 + h z 1 = 1 + ( 0.0005 ) (−0.00157) = 0.99999

z 2 = z 1 − h y 1 = − 0.00157 − (0.0005 ) (1) = − 0.00314

In table 9.3 the results of the present calculations for selected values of t are compared to
the exact solution, y = cos (t) and z = - sin (t).
Table 9.3

Exact h = 0.005Π h = 0.0005Π

t y = cos (t) z = - sin (t) y z y z
0.5 Π 0 -1 1.32E -4 -1.01241 2.62E-6 -1.00123
Π -1 0 -1.02497 -2.67E-4 -1.00247 -5.25E-6
1.5 Π 0 1 -4.01E-4 1.03770 -7.88E-6 1.00371
2Π 1 0 1.05058 5.48E-4 1.00495 1.05E-5
3Π -1 0 -1.07682 -8.43E-4 -1.00743 -1.58E-5
6Π 1 0 1.15954 1.82E-3 1.01491 3.19E-5
8Π 1 0 1.21819 2.54E-3 1.01994 4.27E-5

IT IS OBSERVED IN THE RESULTS SHOWN IN TABLE 9.3 THAT THE ERROR INCREASES WITH
INCREASE IN T, AND PROPORATION TO H. (SEE THE Y VALUES FOR T = Π, 2Π, 3Π, 6Π AND 8Π:
Z VALUES FOR THESE T VALUES DO NOT FOLLOE THE SAME TREND BECAUSE, WHEN Z IS
CLOSED TO ZERO, THE ERRORS OF Z ARE SIGNIFICANTLY AFFECTED BY PHASE SHIFT).

9.2.3 MODIFIED EULAR METHOD

The motivation for the modified Euler method is twofold. First, the modified Euler method
is more accurate than the forward Euler method. Second, it is more stable then the forward
Euler method.
The modified Euler method is derived by applying the trapezoidal rule to
integrating y' = f (y, x):

h
y n +1 = yn + [ f (y n +1 ,t n +1 ) + f ( y n , t n )]
2 (9.2.6)
If ƒ is linear in y, both Eqs. (9.2.6) may easily solve for yn+1. For example, if the ODE is
given by

y' = ay + cos (t)

Eq. (9.2.6) becomes

h
y n +1 = y n + [a y n +1 + cos ( t n +1 )+ ay n + cos ( t n )]
2

Therefore, solving for yn+1 yields

1 + ah / 2 h/2
y n +1 = y + [ cos ( t n +1 ) + cos ( t n )] (9.2.7)
1 − ah / 2 1 − ah / 2
n

If ƒ is a nonlinear function of y, Eq. (9.2.6) becomes a nonlinear function of yn+1, so an

algorithm to solve nonlinear equation is necessary. A widely used method for solving
nonlinear equations is the successive substitution method (section 3.6):


(k ) h ( k − 1)
y − yn = f(y , t n +1 ) + f ( y n ,t n )]
n +1 2 n +1 (9.2.8)

(k ) (k )
Is an initial guess y n +1
is the kth iterative approximation for yn+1, and y n +1
Where

(k ) ( k −1)
Becomes less than y − y for yn+1. The above iteration is terminated when
n +1 n +1

A prescribed tolerance. The initial guess is set to yn. Then. The first iteration step becomes
identical with the forward Euler method. If only one more iteration step is used, the scheme
becomes the second-corrector method. But, in the modified Euler method, iteration is
continued until the tolerance of convergence is satisfied.
Example 9.4 shows an application of the modified Euler method to a nonlinear
first-order ODE.

Example 9.4

By the modified Euler method with h = 0.1, solve

y ' = − y 1.5 + 1, y (0) = 10

For 0 ≤ t ≤ 1. Print the results up to t = 1.

(Solution)

The modified Euler method becomes

h
y n +1 = y n + [ − ( y n +1 ) 1.5 − ( y n ) 1.5 + 2 ] (A)
2

For n = 0

h
y1 = y 0 + [ − ( y 1 ) 1.5 − ( y 0 ) 1.5 + 2 ]
2

The best estimate for y1 on the right side is yo. By introducing y1≈yo to the right side, the
equation becomes
 0.1
y 1  10 + [ − (10) 1.5 − (10 ) 1.5 + 2 ] = 6.93772
2

Introducing the above result to y1 of Eq. (A) again yields

0.1
y 1  10 + [ − ( 6.93772) 1.5 − (10 ) 1.5 + 2 ] = 7.60517
2

Repeating the substitution a few more times, we get

0.1
y 1  10 + [ − ( 7.60517) 1.5 − (10 ) 1.5 + 2 ] = 7.47020
2

0.1
y 1  10 + [ − ( 7.47020) 1.5 − (10 ) 1.5 + 2 ] = 7.49799
2

0.1
y 1  10 + [ − ( 7.49799) 1.5 − (10 ) 1.5 + 2 ] = 7.49229
2



0.1
y 1  10 + [ − ( 7.49326) 1.5 − (10 ) 1.5 + 2 ] = 7.49326
2

The computed results for ten time steps follow:

t y
0.0 10.0
0.1 7.4932
0.2 5.8586
0.3 4.7345
0.4 3.9298
0.5 3.3357
0.6 2.8859
0.7 2.5386
0.8 2.2658
0.9 2.0487
1.0 1.8738

Why is the accuracy of the modified Euler method higher than that of the forward
Euler method? To explain the reason analytically. Let us consider a test equation, y' = xy.
Equation (9.2.6) for this problem may then be written as

h 9.2.9))
y n +1 = y n + ( y n +1 + y n )
2

Or equivalently, solving for yn+1,

  h   h  −1
y n + 1 = 1 +  1 −  yn
 2  2 

Expanding the coefficient of the foregoing equation yields

 1 1 
y n +1 = 1 +  h + ( h) 2 + ( h ) 3 + ......  y n
 2 4 

Comparing this expansion to the Taylor expansion of the exact solution y (tn+1) = exp (αh)
yn, it is found that Eq. (9.2.9) is accurate to the second-order term. Thus, the modified
Euler method is a second-order (accurate) method. On the other hand, a similar analysis
fort the forward Euler method indicates that the forward Euler method is first-order
accurate.
The local error (error generated in each step) of the forward Euler method is
proportional to h² and its global error is proportional to h, whereas the local error of the
modified Euler method is proportional to h³ and its global error is proportional to h². The
order of errors of the backward Euler method is the same as in the forward Euler method.
If the modified Euler method is applied to asset of ODEs, the whole equations must
be solved simultaneously or "implicitly." However, the advantage of the implicit solution
is that the method is more stable than the forward Euler method and thus allows a larger
time step.

9.2.4 Backward Euler Method

The backward Euler method is based on the backward difference approximation and is
written as

y n + 1 = y n + h f ( yn +1 , t n + 1 ) (9.2.10)

The accuracy of this method is the same as that of the forward Euler method. Besides if ƒ
is a nonlinear function of y, an iterative scheme has to be used in each step just as in the
modified Euler method. However, the advantages are (a) the method is stable for stiff
problems, and (b) positively of solution is guaranteed when the exact solution is positive.
See applications of the backward Euler method in section 9.5 and chapter 12.

SUMMARY OF THIS SECTION

(a) The forward Euler method is based on the forward difference approximation. Its error
in one interval is proportional to h² and its global error to h. the forward Euler method may
become unstable if the ODE has a negative time constant unless a small h is used.

(b) The modified Euler method is based on the trapezoidal rule. If the ODE is not linear,
an iterative method is necessary for each interval. Its error in one interval is proportional
to h³ and its global error to h².

(c) The backward Euler method is based on the backward difference approximation. Its
error is comparable to those of the forward Euler method. The method is stable so it is used
to solve stiff problems that are difficult to solve by other methods.

9.3 RUNGE-KUTTA METHODS

A major drawback of the Euler methods is that the orders of accuracy are low. This
disadvantage is twofold. To maintain a high accuracy requires very small h, which
increases computational time and causes round-off errors.
In Runge-kutta methods, the order of accuracy is increased by using intermediate
points in each step interval. A higher accuracy also implicates that errors decrease more
quickly than in lower-order accuracy methods when h is reduced.
Consider an ordinary differential equation

(9.3.1) y' f ( y , t ) , y (0) = y 0

To calculate yn+1 at tn+1 h with a known value of yn, we integrate Eq. (9.3.1) over the
interval [tn, tn+1] as

t n +1
y n +1 = y n +  f ( y , t ) dt (9.3.2)
tn

Runge-kutta methods are derived by applying a numerical integration method to the

integral on the right side of Eq. (9.3.2) [fox/Mayers]. In the remainder of this section, the
second, third-, and fourth-order Runge-Kutta methods are explained.

1
 f ( y, t ) dt  h [ f ( y n , t n ) + f ( y n +1 , t
2
n +1 )] (9.3.3)

9.3.1. Second-order Runge-Kutta method

Here we examine an application of the trapezoidal rule to the right side of Eq. (9.3.2):

In Eq. (9.3.3) yn+1 is not known, so the second term is approximated by ƒ (y¯ n+1, tn+1),
where y¯n+1 is the first estimate for yn+1 calculated by the forward Euler method. The scheme
derived here is called the second-order Runge-Kutta method and summarized as

y n + 1 = y n + h f (y n , t n )

h
y n +1 = y n + [ f ( y n , t n ) + f ( y n +1 , t n +1 )]
2
Or in a more standard form as

k1 = h f (y n ,t n )

k 2 = h f ( y n + k 1 , t n +1 )
(9.3.4)
1
y n + 1 = y n + [k1 + k 2 ]
2

Cond-order Runge-Kutta method is identical to the Euler predictor-corrector

method, which is the simplest predictor-corrector method (see section 9.4.). It is also
equivalent to the modified Euler method, having only two iteration steps.

Example 9.5

The circuit shown in figure E9.5 has a self-inductance of L =50H, a resistance of R =

20 ohm, and a voltage source of V = 10 volt. If the switch is closed at t = 0, the
current /(t) satisfies

d
L /(t ) + R /(t ) = E , /(0) = 0 ( A)
dt

Find value of the current for 0

<t ≤ 10 sec by using the second-order Runge- kutta method with h = 0.1.

‹Solution›
We first rewrite Eq. (a) as

d R E
I =− I +  f ( I , t)
dt L L
Then, the second-order .Runge-kutta method becomes

 R E
k 1 = h − I +
L 
n
 L

 R E
k 2 = h  − (I n + k 1 ) + 
 L L

1
I n +1 =I n (k 1 +k 2 )
2

The calculations for the first two steps are shown next:
N = 0 (t=0.1): k1 =0.1[(-0.4) (0) +0.2] = 0.02

K2 =0.1[(-0.4) (0 +0.02] = 0.0192

1 1
I1 = I 0 (k1 + k 2 ) = 0 + (0.02 + 0.0192) = 0.0196
2 2

n = 1(t=0.2): k1 =0.1[(-0.4) (0.0196) +0.2] = 0.019216

k2 =0.1[(-0.4) (0.0196 +0.019216) + 0.2] = 0.018447

1
I 2 = I1 ( k1 + k 2 )
2

1
= 0.0196 + ( 0.019216+ 0.018447 ) = 0.038431
2

The final result of computation (at multiples of 10 steps)is as follows:

t (sec) t (amp)
0 0
1 0.1648
2 0.2752
3 0.3493
4 0.3990
5 0.4332
6 0.4546
7 0.4695
8 0.4796
9 0.4863
10 0.4908
(∞) (0.5000)

The accuracy of the second-order Runge-Kutta method may be analyzed by using

test equation y' = αy as described toward the end of section 9.2. However, to show a more
formula approach, we consider here a generic form. y' = f (y, x). We first expand the exact
value of yn+1 in the Taylor series:

h2
y n +1 = yn +h f + [ f t + fy f ]
2

h3
+ [f u +2 f ty f + fyy f 2
+ f t f y +f 2
y f ] + 0 (h 4 ) (9.3.6)
6

Where all the derivatives of y are expressed in terms of f and the partial derivatives of f at
tn.

Next, we expand the second equation in Eq. (9.3.4) in a Taylor series:

h2 h3
y n +1 = yn +h f + [ f t + fy f ] + [f u +2 f ty f + fyy f 2
] + 0 (h 4 ) (9.3.6)
2 4

By comparing Eq. (9.3.6) to Eq. (9.3.5), Eq. (9.3.4) is found to be accurate to the order of
h² and the discrepancy (error generated in one step) is proportional to h³. Notice that the
second-order Runge-Kutta method is identical to the modified Euler method given by Eq.
(9.3.8) with two iteration steps. However, the order of accuracy of the former is identical
to that of the latter, which requires iterative convergence. This indicates that the iteration
in the modified Euler method increases accuracy only a little. (Indeed, using the second-
order Runge Kutta method with a smaller h is far more effective in improving accuracy
than using the modified Euler method with strict iterative convergence.) y' = αy, but this
approach is left as a student's exercise.
 Application of the second-order Runge Kutta method to a higher-order ordinary
differential equation is easy. For illustration, we consider the second-order differential
equation:

y''(t) + ay'(t) + by(t) = q(t), y(0) = 1, y'(0) = 0 (9.3.7)

Where a and b are coefficients and q(t) is a known function, and two initial conditions are
given. By defining

z (t) = y'(t) (9.3.8)

Eq. (9.3.7) can be reduced to couple first-order differential equations:

Y' = ƒ(y, z, t) ≡ z, y (0) = 1

(9.3.9)
Z' = g(y, z, t) ≡ -az – by + q, z (0) = 0

The second-order Runge Kutta method for the foregoing equations is written as

k 1 = h f ( y n , z n ,t n ) = hz n

l 1 = h g ( y n , zn , tn ) = h ( − a z n − b yn + qn )

k 2 = h f ( y n + k1 , z n + l1 , t n + 1 ) = h ( z n + l1 )

l2 = h g ( y n + k1 , z n + l1 , t n + 1 ) = h ( − a( z n + l1 ) − b ( y n + k1 ) + q n + 1 ) (9.3.10)

1
yn + 1 = yn + ( k1 + k1 )
2

1
z n +1 = z n + (l1 + l 2 )
2

Example 9.6
 A cubic material of mass M = 0.5 kg is fixed to the lower end of a missies spring.
The upper end of the spring is fixed to a structure at rest. The cube receives resistance R =
- B dy/dt from the air, where B is a damping constant.

(See figure E9.6). The equation of motion is

d2 d
M 2 y +B y + k y = 0, y ( 0 ) = 1, y ' ( 0 ) = 0 (A)
dt dt

Where y is the displacement from the static position, k is the spring constant equal to 100
kg/s², and B = 10 kg/s.

(a) Calculate y (t) for 0< t < 0.05 using the second-order Runge Kutta method with h =
0.025 by hand calculations.

(b) Calculate y (t) for 0 < t < 10 sec using the second-order Runge Kutta method with h =
0.001.

(c) Repeat the calculation for B = 0.

(Solution)

Equation (A) may be written as

y' = z ≡ ƒ (y, z, t) y (0) = 1

(B)
B k
z' = − z− y  g ( y, z , t ) , z (0) = 0
M M
By setting a = B/M = 20 and b = kM = 200 and g = 0, the second-order Runge Kutta method
for Eq. (A) becomes the form of Eq. (9.3.9)

(a) For n = 1: t = 0.025

k1 = h f ( y 0 , z 0 , t 0 ) = h z 0 = 0.025(0) = 0

l1 = h g ( y 0 , z 0 , t 0 ) = h ( − 20 z 0 − 200 y 0 ) = 0.025 ( − 20 (0) − 200 (1) ) = − 5

k 2 = h f ( y 0 + k1 , z 0 + l1 , t 0 ) = h ( z 0 + l1 ) = 0.025(0 − 5 ) = − 0.125

l 2 = h g ( y 0 + k1 , z 0 + l1 , t 0 ) = h ( − 20 ( z 0 + l1 − 200 ( y 0 + k1 ) )

= 0.025 [ − 20 (0 − 5) − 200 (1 + 0) ] = − 2.5

1
y1 = y0 + ( 0 − 1.25) = 0.9375
2

1
z1 = z0 + ( − 5 − 2.5 ) = − 3.75
2

For n = 2: t = 0.05

k1 = h f ( y1 , z1 , t1 ) = h z1 = 0.025(−3.75) = − 0.09375

l1 = h g ( y1 , z1 , t1 ) = h ( − 20 z1 − 200 y1 )

= 0.025 ( − 20 (−3.75) − 200 (0.9375) ) = − 2.8125

k 2 = h f ( y1 + k1 , z1 + l1 , t1 ) = h ( z1 + l1 )

= 0.025(−3.75 − 2.8125) = 0.1640625

l 2 = h g ( y1 + k1 , z1 + l1 , t1 ) = h ( − 20 ( z1 + l1 ) − 200 ( y1 + k1 ) )

= 0.025 ( − 20 (−3.75 − 2.8125) − 200 (0.9375 − 0.093750) )

= − 0.9375
1
y 2 = y1 + ( − 0.09375 − 0.1640625) = 0.80859
2

1
z 2 = z1 + ( − 2.8125 − 0.9375) = − 5.625
2

(b) And (c) this part of the computations was performed by using PROGRAM 9 – 1. The
computational results after every 50 steps up to 0.75 sec are shown bellow:

(b) (c)
t (sec) y (meter) y (meter)
(B = 10) (B = 0)
0 1.000 1.000
0.05 0.823 0.760
0.1 0.508 0.155
0.15 0.238 -0.523
0.2 0.066 -0.951
0.25 -0.016 -0.923
0.3 -0.042 -0.45
0.35 -0.038 0.235
0.4 -0.025 0.810
0.45 -0.013 0.996
0.5 -0.004 0.705
0.55 0.000 0.075
0.6 0.001 -0.590
0.65 0.001 -0.973
0.7 0.001 -0.889
0.75 0.000 -0.378

9.3.2 Third-order Runge Kutta Method

A Runge Kutta that is more accurate than the second-order Runge Kutta method may be
derived by using a higher-order numerical integration scheme for the second term of Eq.
(9.3.2). Using the Simpson's 1/3 rule, Eq. (9.3.2) becomes

h
yn + 1 = yn + [ f ( yn , t n ) + 4 f ( y n + 1 , t 1 ) + f ( y n +1 , t n +1 ) ] (9.3.11)
6 2 n+
2

where y n + 1 and y n + 1 are estimates because y 1 and y n + 1 are not know.

2 n +
2

the estimate y n + 1 is obtained by the forward Euler method as

2

h
y n + 1 = yn + f ( yn , tn ) (9.3.12)
2 2
 The estimate y n + 1 may be obtained by

y n + 1 = yn + h f ( yn , tn )

or
y n + 1 = yn + h f ( y n +
1 t 1 )
2 n+
2

or a linear combination of both

y n + 1 = y n h [  f ( y n , t n ) + (1 −  ) f ( y n + 1 , t 1 )] (9.3.13)
2 n+
2

Here θ is an undetermined parameter, which will be determined to maximize accuracy of

the numerical method. With Eq. (9.3.13), the whole scheme is written in the following form:

k1 = h f ( y n , t n )

1 h
k 2 = h f ( yn + k1 , t n + )
2 2
(9.3.14)

k 3 = h f ( y n +  k1 + (1 −  ) k 2 , t n + h )

1
yn + 1 = yn + ( k1 + 4 k 2 + k 3 )
6

To optimize θ, we expand k1, K2 and k3 in Taylor series as

k1 = h f ( 9.3.15 a )

1 2 1
k2 = h f + h ( f1 + f y f ) + h 3 ( f u + 2 f t y f + f y y f 2 ) ( 9.3.15 b)
2 8

1 3
k3 = h f + h 2 ( f t + f y f ) + h [ fu + 2 ft y f
2

+ fyy f 2
+ (1 −  ) ( f t + f y f ) f y ] ( 9.3.15 c)

Where ƒ and its derivatives are evaluated at tn. By introducing Eq. (9.3.15) into Eq. (9.3.14)
and comparing it to Eq. (9.3.5), we find that θ = - 1 is the optimum because Eq. (9.3.14)
then agrees with Eq. (9.3.5) to the third-order term.
The forgoing derivative may be more easily understood if it is applied to the test
equation y' = αy.
In summary, the third-order-accurate Runge Kutta method is written as

k1 = h f ( yn , t n )

 1 h
k 2 = h f  yn + k1 , t n + 
 2 2
(9.3.16)
k3 = h f ( yn − k1 + 2 k 2 , t n + h )

1
yn + 1 = yn + ( k1 + 4 k 2 + k3 )
6
Example 9. 8

Solve

y' = t y + 1 , y (0) 0

Using the fourth-order Runge-Kutta method, Eq. (9.3.17), with h = 0.2, 0.1. And 0.05,
respectively, and evaluate the error for each h at t = 1, 2, 3, 4, and 5

(Solution)

Computations for this example were performed by using PROGRAM 9 – 2. The results
are as shown below:

h = 0.2 h = 0.1 h = 0.05

t y p.e y p.e y p.e
1 1.41067 (0.00) 1.41069 (0.00) 1.41068 (0.00)
2 8.83839 (0.01) 8.83937 (0.00) 8.83943 (0.00)
3 112.394 (0.11) 112.506 (0.01) 112.514 (0.00)
4 3716.42 (0.52) 3734.23 (0.04) 3735.72 (0.00)
5 330549 (1.71) 335798 (0.15) 336273 (0.01)

ª p.e...Percentage error

A comparison of these results to those of Euler methods reveals that the error of the fourth-
order Runge-Kutta method with h = 0.1 is comparable to that of the modified Euler method
with h = 0.001 also, the fourth-order Runge-Kutta method with h = 0.2 is comparable to
the forward Euler method with h = 0.001.

Application of the fourth-order Runge-Kutta method to a set of ordinary differential

equations is very similar to that of the second-order Runge-Kutta method.
For simplicity of explanation, we consider a set of two equations.
 y' = f (y, z, t)
(9.3.19)
z' = g (y, z, t)

The fourth-order Runge-Kutta method for the set of two equations becomes

k1 = h f ( y n , z n , t n )

l1 = h g ( y n , z n , t n )
 k l h
k2 = h f  yn + 1 , zn + 1 , tn + 
 2 2 2

 k l h
l2 = h g  y n + 1 , z n + 1 , t n + 
 2 2 2
(9.3.20)
 k l h
k3 = h f  yn + 2 , z n + 2 , t n + 
 2 2 2

 k l h
l3 = h g  y n + 2 , z n + 2 , t n + 
 2 2 2

k4 = h f ( y n + k 3 , z n + l3 , t n + h )

l4 = h g ( y n + k 3 , z n + l3 , t n + h )

1
yn + 1 = yn + [ k1 + 2 k 2 + 2 k 3 + k 4 ]
6 (9.3.21)

1
zn + 1 = zn + [ l1 + 2 l 2 + 2 l3 + l 4 ] (9.3.22)
6
 Even when the number of equations in a set is greater than two, the derivation of
the fourth-order Runge-Kutta method is essentially the same. A program to solve a set of
equations using the fourth-order Runge-Kutta method is given as PROGRAM 9 – 3.

Example 9.9

Repeat the problem in example 9.3 by using the fourth-order Runge-Kutta method
with h = 0.2Π, and h = 0.5Π.

(Solution)

PROGRAM 9 – 3 is used to produce the results shown below:

Exact h = 0.2 Π h = 0.05Π

t y = cos (t) z = -sin(t) y z y z

0.5Π 0 -1 1.23E-4 -0.99997 1.32E-6 -0.99999

Π -1 0 -0.99993 -2.48E-4 -0.99999 -2.65E-6
1.5 Π 0 1 -3.72E-4 0.99990 -3.96E-6 0.99999
2Π 1 0 0.99987 4.95E-4 0.99999 5.29E-6
3Π -1 0 -0.99989 -7.43E-4 -0.99999 -7.94E-6
6Π 1 0 0.99960 1.49E-3 0.99999 1.57E-5
8Π 1 0 0.99947 1.98E-3 0.99999 2.11E-5

Comparing these values to the results of the forward Euler solution in example 9.3, the
accuracy of the fourth-order Runge-Kutta method even with h = 0.2Π is significantly better
than the forward Euler method with h = 0.01Π.
9.3.5 Error, stability, and Grid Interval optimization

The Range-Kutta methods are subject to two kinds of error—truncation error and
instability. As discussed earlier, the truncation error is due to the discrepancy between the
Taylor expansion of the numerical method and the Taylor expansion of the exact solution.
The amount of error decreases as the order of the method becomes higher. On the other
hand, instability is an accumulated effect of the local error such that the error of the solution
grows unboundedly as the time steps are advanced.
To analyze the instability of a Runge-Kutta method, let us consider the test equation

y' = αy (9.3.23)

Where α < 0. For a given value of yn the exact value for yn+1 is analytically given as

Yn+1 = exp (αh) yn (9.3.24)

Notice that because α < 0. | yn+1 | decreases as n (or time) increases.

The numerical solution of Eq. (9.3.23) by the fourth-order Runge-Kutta method
becomes

k1 =  h y n

 k   1 
k 2 =  h  yn + 1  =  h  1 +  h  yn
 2  2 

 k   1  1  (9.3.25)
k3 =  h  yn + 2  =  h  1 +  h  1 +  h   yn
 2  2  2 

 
k 4 =  h ( yn + k3 ) =  h  1 +  h  1 + 1 
 h 1+
1
 h    y n
  2  2 

 1 1 1 
yn +1 =  1 +  h + (  h ) 2 + ( h ) 3 + ( h) 4  yn (9.3.26)
 2 6 24 
Equation (9.3.26) equals the first five terms of the Taylor expansion for the right side of
Eq. (9.3.24) about tn. The factor

1 1 1
 =1 +h + (  h) 2 + ( h ) 3 + (  h) 4 (9.3.27)
2 6 24

In Eq. (9.3.26) is approximating exp (αh) of Eq. (9.3.24), so the truncation error and
instability of Eq. (9.3.26) both originate in this approximation.
Equation (9.3.27) and exp (αh) are plotted together in figure 9.1 for comparison. The figure
indicates that if α < 0 and the modulus (absolute value) of αh increase, the deviation of γ
from exp (αh) increases, so that the error of the Runge-Kutta method increases. Particularly,
if αh ≤ -2.785, the method becomes unstable because the modulus of the numerical solution
grows in each step whereas the modulus of the true solution decreases by a factor, exp (αh),
in each step.

In practical applications of the Runge-Kutta method, the size of an optimal grid

interval can be determined in the following way. For illustration purposes, suppose we
desire to keep there local error of the third-order Runge-Kutta method less than ξ. The local
error of the third-order Runge-Kutta method for a test interval h is proportion . . So we
express the local error in the form h 4 to

Eh = B h4 (9.3.28)

Where B a constant that depends on the given problem. If we apply the same Runge-Kutta
method in two steps with h/2 as the time interval, the error becomes . Times 2Where
4
the factor 2 is due to the accumulation (h / 2) preoperational to
Two steps thus, it becomes of error in

4
h 1
2 E h/2 = 2 B   = B h4 (9.3.29)
2 8

By subtracting Eq. (9.3.29) from Eq. (9.3.28), we get

 1 7 (9.3.30)
Eh − 2 E h / 2 = B h 4 − B h4 = B h4
8 8

The left side of the forgoing equation may be evaluated by a numerical experiment—
that is, by running the scheme twice starting from the same initial value. In the first run,
only one time step is advanced using a trial value for h as the time interval. We denote the
result of this calculation as [y1] h. in the second run; [y2] h/2 is calculated in two time steps
using h/2 as the time interval. Using the results of those two calculations, the left side of
Eq. (9.3.30.) is evaluated as

Eh − 2 E h / 2 =  y1 h −  y2 h / 2 (9.3.31)

Introducing Eq. (9.3.31) into Eq. (9.3.30) and solving for B yields

(  y1 h −  y2 h / 2 ) / h 4
8
B = (9.3.32)
7
Once B is determined, the maximum (or optimum) h that satisfies the criterion Eh ≤ξ may
be found by introducing Eh = ξ into Eq. (9.3.28) and solving for h. as follows:

  
0.25

h = 
 B

The theory we just described is reminiscent of the Romberg integration explained in section
3.2.

Example 9.10

Assume that a fourth-order Runge-Kutta method is applied to

y
y' = − , y ( 0) = 1
1 +t2
Find an optimal step interval satisfying Eh ≤ 0.00001.

(Solution)

For the fourth-order Runge-Kutta method, the local error is expressed by

Eh = B h 5 (A)

The approach is very similar to Eqs (9.3.28) through (9.3.33) except that the order of
5
Error is five. The error accumulated in two steps using h/2 is 2Eh/2 = 2B (h/2)
The deference between the errors of one-step and two-step calculations, namely Eh – Eh/2,
is numerically evaluated by

2Eh – 2Eh/2 = [y1]h – [y2]h/2 . (B)

IN Eq. (B), [y1] h is result of the fourth-order rung-kutta method for only one step
With h, and [y2]h/2 is the result of the same for two steps with h/2. Introducing Eq. (A) into
Eq. (B) and solving for B, we have

16
B = ( [ y1 ]h − [ y1 ]h / 2 ) / h 5 (C)
15

Now we actually run the fourth order Runge-Kutta method for only one step with h
= 1 starting with the given initial condition. Then we run it for two steps with h/2 = 1/2.
The results are

[y1]1 = 0.4566667 (only interval only)

[y2]1/2 = 0.4559973 (only intervals)

From Eq (c), we obtain B as

16
B = ( 0.4566667 − 0.4559973) / (1) 5 = 6.310− 4 (D)
15
By introducing this into Eq. (a), the local error for any h is expressed by

Eh = 6.3  10 − 4 h 5

The maximum h that the given criterion Eh < 0.00001, is

h = ( 0.00001/ 6.3  10− 4 ) 1/ 5 = 0.44 (E)

SUMMARY OF THIS SECTION

(a) The Runge-Kutta methods are derived by integrating the first order ODE with
numerical integration methods. The second-order Runge-Kutta method is identical to the
modified Euler method with two iteration cycles as well as to the second order predictor-
corrector method.

(b) A higher-order ODE can be solved by a Runge-Kutta method after it is transformed to

a set of first-order ODEs.

(c) Each Runge-Kutta method becomes unstable if α is negative and |αh| exceeds a certain
criterion.

(d) The local error of a Runge-Kutta method can be found by running the same method
twice: the first time for one interval with a value of h, and the second time for two intervals
with h/2.

9.4 PREDICTOR-CORRECTOR METHODS

9.4.1 Third-Order Adams Predictor-corrector Method

A predictor-corrector method consists of a predictor step and a corrector step in each

interval. The predictor estimates the solution for the new point, and then the corrector
improves its accuracy. Predictor-corrector methods use the solutions for previous points
instead of using intermediate points in each interval.

to explain the methods, let us consider an equispaced time interval and assume that
the solution has been calculated up to time point n so that the values of y and y' on the
pervious time points may be used for the calculation of yn+1.
Both predictor and corrector formulas are derived by introducing appropriate
polynomial approximation for y'(t) into Eq. (9.3.2). The most primitive member of the
predictor-corrector methods is the second-order predictor-corrector method, which is
identical to the second-order Runge-Kutta method.
Let us derive a third-order predictor by approximating y' = ƒ (y, tr) with a
quadratic interpolation polynomial fitted of f'n, y'n-1 and y'n-2:

1
y' ( z ) = [ ( z + h ) ( z + 2 h ) y ' n − 2 z ( z + 2 h ) y ' n − 1 + z ( z + h ) y ' n − 2 ] + E ( Z ) (9.4.1)
2 h2

Where z is a local coordinate defined by

z = t - tn
And E (z) is the error (see section 2.3). Equation (9.4.1) is the Lagrange interpolation fitted
to the values y'n, y'n-1 and y'n-2. The error of the polynomial is

z ( z + h) ( z + 2 h ) y ( i v ) (  ) , t n − 2    t n + 1
1
E ( z)= (9.4.2)
3!

Here, the derivative in the error term is of the fourth order because a quadratic polynomial
is fitted to y'.
Equation (9.3.2) can be rewritten in terms of the local coordinate z = t – tn as

h
y n +1 = y n + 0
y' ( z ) d z
(9.4.3)

Introducing Eq. (9.4.1) into Eq. (9.4.3) yields

h
y n +1 = y n + ( 23 y 'n − 16 y 'n −1 + 5 y 'n − 2 ) + 0 ( h 4 ) (9.4.4)
12

Equation (9.4.4) is called the third-order Adams-Bash forth predictor formula. The error of
Eq. (9.4.4) is attributable to Eq. (9.4.2) and is evaluated by integrating Eq. (9.4.2) in [0, h],
as allows:

h y (  ) , tn − 2   tn + 1
3 4 (iv)
0( h4 ) =
8

The deriving Eq. (9.4.4), notice that Eq. (9.4.1) has been used as an extrapolation. As
pointed out in section 2.9, extrapolation is lees accurate than interpolation (see section 2.9
and Appendix A). Therefore, Eq. (9.4.4) is used only as a predictor and is written as

h
y n +1 = y n + ( 23 y 'n − 16 y 'n −1 + 5 y 'n − 2 ) + 0 ( h 4 )
12

Where the over bar indicates a predictor.

is y ' n +1 To derive a corrector formula, a predicted value of y'n+1 denoted by

Into y'(t) = ƒ(y, t) as y n +1 necessary, which is calculated by introducing

y'n + 1 = f ( y n + 1 , t n +1 )

, y'n and y'n-1 is written as y ' n +1 the quadratic polynomial fitted to

y' ( z ) =
1
2 h2
 
z ( z + h ) y ' n + 1 − 2 ( z − h) ( z + h) y ' n + z ( z − h) y ' n − 1 + E ( z ) (9.4.6)

Where z is the local coordinate defined after Eq. (9.4.1). The error of this equation is
 z ( z + h) ( z + 2 h ) y ( i v ) (  ) , t n − 1    t n + 1
1
E ( z)=
3!

Introducing Eq. (9.4.6) into Eq. (9.4.3) yields a corrector formula as

h
y n +1 = y n + ( 5 y ' n + 1 + 8 y ' n − y ' n −1 ) + 0 ( h 4 )
12 (9.4.7)

The error is

h y (  ) , t n −1    t n + 1
1 4 (iv)
0( h4 ) = −
24

Equation (9.4.7) is named the Adams-Moulton corrector formula of order 3. The set of Eqs.
(9.4.5) and (9.4.7) is called the third-order Adams predictor-corrector method.
As seen from the preceding derivation, numerous formulas can be derived by
changing the choice of the extrapolating and interpolating polynomials.
In discussing the predictor-corrector methods, we have assumed that the
solutions for previous points are available. The third-order predictor-corrector method
needs three previous values of y as explained earlier. therefore, to start up the method, the
solutions for n = 0, n = 1, and n = 2 are necessary, the first of which is given by an initial
condition, but the second and third should be provided by some other means than the
predictor-corrector method, such as a Runge-Kutta method.

9.4.2 Fourth-Order Dams Predictor-Corrector Method

The interpolation polynomial fitted to y' at points n, n-1, n-2, …, n-m may be written in the
Newton backward formula [see Eq. (2.4.14)] as

m
 s + k − 1 k
gm (t ) =  ( − 1 ) k    y' n −k
(9.4.8)
k =0  k 
 Where

t − tn
s=
h

By introducing Eq. (9.4.8) into Eq. (9.3.2), we obtain the Adams-Bashforth predictor
formula of order m + 1:


y n + 1 = yn + h b0 y 'n + b1  y 'n −1 + ......... + bm m y 'n − m  (9.4.9)

Where

1  s + k − 1 (9.4.10)
bk =  0

 k
 d s


The first few value of bk follows:

b0 = 1

1
b1 =
12

5
b2 =
12

3
b3 =
8

251
b4 =
720
if we set m = 2 in Eq. (9.4.9) for example, we obtain the third-order predictor given by Eq.
(9.4.4). By following the same procedure for m = 3, the fourth-order predictor formula is
derived as

9h
y n +1 = y n + ( 55 y 'n − 59 y 'n −1 + 37 y 'n − 2 − 9 y 'n − 3 ) + 0 (h 5 ) (9.4.11)
24
Where

h y (  ) , tn −3   tn + 1
251 5 ( v )
0 ( h5 ) =
720

The corrector formulas may be derived by using the polynomial fitted to y' at grid points,
n+1, n, n-1, …., n –m +1. The Newton backward interpolation formula fitted to y' at these
points (see section 2.5) is

m
 s + k − 2 k
gm (t )      y ' n + 1 −k
(9.4.12)
k =0  k 

Introducing this equation into Eq. (9.3.2) yields the Adams-Moulton corrector formula


y n + 1 = y n + h c0 y ' n + 1 + c1  y' n + ......... + cm m y' n − m 

Where

1  s + k − 2
ck =  0

 k
 d s


The first sew values of Ck follow:

By setting m = 3 in Eq. (9.4.13), we obtain the fourth-order Adams-Moulton corrector:

 h
y n +1 = y n + ( 9 y 'n + 1 +19 y 'n −5 y 'n − 1 + y 'n − 2 ) + 0 (h 5 ) (9.4.14)
24

Where y'n = f (yn, tn) and

h y (  ) , tn − 2   tn + 1
19 5 ( v )
0 ( h5 ) = −
720

The set of Eqs. (9.4.11) and (9.4.14) is called the fourth-order Adams predictor-corrector
method.

9.4.3 Advantage and Disadvantages of Predictor-Corrector Methods

An advantage of predictor-corrector methods is their computational efficiency: they use

information from previous steps. Indeed, the function ƒ(y, t) is evaluated only twice in each
step regardless of the order of the predictor-corrector method, whereas the fourth-order
Runge-Kutta method evaluates ƒ(y, t) four times in each interval. Another advantage is that
the local error can be detected at each step with a small additional computing effort. The
technique of detecting local error is discussed in subsection 9.4.4. On the other hand, there
are some disadvantages as follows:

(a) The method cannot be started by itself because of the use of previous points, until the
solutions for enough points are determined, another method such as Runge-Kutta method
must be used.

(b) Because previous points are used, changing the interval size in the middle of solution
is not easy. Although there predictor-corrector formulas may be derived on nonuniformly
spaced points, the coefficients of the formulas change for each interval, so programming
becomes very cumbersome.

(c) The predictor-corrector method cannot be used if y' becomes discontinuous. This can
happen when one of the coefficients of the differential equation changes discontinuously
in the middle of the domain.
However, the last two difficulties can be overcome as follows: because the predictor-
corrector program must contain a self-starting method such as a Runge-Kutta method
anyway, the computation can be restarted whenever the step interval has to be changed or
when y' becomes discontinuous.

9.4.4 Analysis of Local Error and Instability of Predictor-Corrector Methods

One advantage of the predictor-corrector methods is that local error may be evaluated
easily by observing the difference between the predictor and the corrector in each step. For
illustration of analysis, we consider the third-order Adams predictor-corrector method.
Equations (9.4.4) and (9.4.7) indicate that, assume that yn, yn-1, yn-2… are exact, the predictor
and corrector values become

3 4 (i v )
y n + 1 = yn + 1 , exact − h y ( )
8 (9.4.15)

1 4 (iv)
y n + 1 = y n + 1 , exact + h y (  ) (9.4.16)
24

If we assume further that the values of the fourth derivative in Eqs. (9.4.15) and (9.4.16)
take the same value, then subtracting Eq. (9.4.16) from Eq. (9.4.15) yields

10 4 ( i v ) (9.4.17)
y n + 1 − yn + 1 = − h y ( )
24

Back substituting Eq. (9.4.17) into Eq. (9.4.16) yields

y n + 1 , exact − yn + 1 =
1
10
(y n +1 − yn + 1 ) (9.4.18)
The right side of Eq. (9.4.18) is the local error of the corrector. Because it is expressed in
terms of the difference between the predictor and the corrector, the calculation is simple.
By using this algorithm at every step interval, the local error of the method can be
automatically monitored in a program.

Now we study the stability of a predictor-corrector method by considering again the

third-order Adams predictor-corrector method given by Eq. (9.4.5) and Eq. (9.4.7).
Suppose we apply the method to the test equation given by

y' = αy (9.4.19)

Introducing Eq. (9.4.19) into Eqs. (9.4.4) and (9.4.7) yields

h
y n +1 = y n + ( 23 y 'n − 16 y 'n −1 + 5 y 'n − 2 )
12

h
y n +1 = y n + ( 5 y 'n + 1 + 8 y 'n − y 'n − 1 )
12

In the foregoing two equations and reorganizing the terms lead to y n +1 Eliminating

yn + 1 = − a2 yn − a1 yn − 1 − a0 yn − 2 (9.4.20)

Where

a 2 = − (1 + 13 b + 115 b 2 )

a1 = b + 80 b 2

a 0 = − 25 b 2

h
b =
12
 Equation (9.4.20) may be considered as the initial value problem of a difference
equation, for which the analytical solution may be obtained in a similar way as for a third-
order linear ordinary differential equation. Indeed, the analytical solution of Eq. (9.4.20)
may be found in the form.

yn = c  n (9.4.21)

Where γ is a characteristic value and c is a constant. By introducing Eq. (9.4.21) into Eq.
(9.4.20), we get the characteristic equation:

 3 + a2  2 + a1  + a0 = 0 (9.4.22)

Equation (9.4.22) is a third-order polynomial equation, so it has three roots although two
of them can be complex values. We denoted the three roots by

 1 ,  2 , and  3

Because each of y1, y2, and y3 satisfies Eq. (9.4.20), a linear combination of all the solution
is also a solution of Eq. (9.4.20). The general solution of Eq. (9.4.20) may now be written
as

yn = c1 (  1 ) n + c2 (  2 )n + c3 (  3 )n (9.4.23)

Where c1, c2 and c3 are determined when the initial values of yo, y1 and y2 are given
(remember that the third-order predictor-corrector method needs three starting values).
The exact solution to the original problem, Eq. (9.4.19_), is given by
 yn = y ( 0 ) exp ( n h ) (9.4.24)

Where y (0) is the initial value of y (t). The question is each of the three terms in Eq. (9.4.23)
is related to Eq. (9.4.24). The answer is that one term in Eq. (9.4.23) is an approximation
for Eq. (9.4.24), but the other two are irrelevant to the true solution and constitute a part of
the error of the scheme. we assumed that the first term is the approximation and that the
second and the third are the errors. Instability of the method is then related to the second
and third are the error terms vanish as n increases, there is no instability behavior of the
numerical solution occurs. This is the instability, and it happens if

2  1 or 3  1 or both

When the predictor-corrector method is applied to Eq. (9.4.19), both α and h affect the
instability. However, because α and h always appear as a product [see Eq. (9.4.20), we can
consider αh as one parameter. the roots of Eq. (9.4.22) for various values of αh are shown
in table 9.2.

Table 9.2 Characteristic values of the third-order predictor-corrector method applied to y'(t) = αy (t)

Percentage
αh exp (αh) γ1 Error γ 2, γ 3 | γ 2|, | γ 3|
0.1 1.1051 1.1051 0 0.006±0.039j 0.040
0.2 1.2214 1.2214 0 0.014±0.074j 0.075
0.5 1.6487 1.6477 0.06 0.047±0.155j 0.162
1.0 2.7183 2.6668 1.90 0.108±0.231j 0.255
1.5 4.4816 4.1105 8.3 0.155±0.266j 0.308
2.0 7.3891 5.9811 23 0.190±0.283j 0.341
2.5 12.1825 8.2705 32 0.215±0.292j 0.362
-0.1 0.9048 0.9048 0 -0.003±0.043j 0.043
-0.2 0.8187 0.8189 0.02 -0.002±0.092j 0.092
-0.3 0.7408 0.7416 0.1 -0.003±0.145j 0.145
-0.4 0.6703 0.6732 0.43 -0.011±0.203j 0.203
-0.5 0.6065 0.6147 1.35 0.022±0.265j 0.266
-1.0 0.3678 0.4824 31.2 0.116±0.588j 0.600
-1.5 0.2231 0.4944 121. 0.338±0.821j 0.889
-2.0 0.1353 0.5650 419. 0.731±0.833j 1.109
j = −1

In table 9.2 γ1 is the root relevant to the true solution; indeed, it is an

approximation to exp (αh). The other two γs are irrelevant roots. The last column shows
the magnitude of the second and the third roots. it is seen that when αh > 0 (that is, α > 0),
the magnitudes of γ2 and γ3 are allows smaller than γ1. Therefore, as n increases, the
magnitude of the second and third terms relative to the first term diminishes; thus, there is
no instability if α > 0.
In the second half of table 9.2 where α < 0, the relevant root γ1 is always smaller
than 1 and it decreases as αh becomes more negative. when the magnitude of αh is very
small, the irrelevant roots are smaller than the relevant root, but the magnitude of the
irrelevant roots keeps increasing and exceeds the relevant roots before αh reaches -1. The
magnitude of the irrelevant root exceeds unity approximately when αh = - 1.8. if the
magnitude of the irrelevant root exceeds unity, the second and third terms of Eq. (9.4.20)
will show an erratic behavior. That is, while the first term approaches 0, the second and the
third terms diverge with oscillatory behavior.
Therefore, instability of the third-order predictor-corrector occurs if αh < -1.8.
Table 9.2 also provides important information on the accuracy of the third-order Adams
predictor-corrector method. As described previously, the first root γ1 in table 9.2
approximates exp (αh). The table shows that, for α > 0, the percentage of error is small as
|αh| increases, and for αh = -0.5. Which is still far from the instability domain, the
percentage of error is significant.

SUMMARY OF THIS SECTION

(a) A predictor-corrector method consists of s predictor and corrector.

(b) The predictors of the Adams predictor-corrector methods are named Adams-Bashforth
predictors. They are derived by integrating a polynomial extrapolation of y' for the previous
points.

(c) The correctors of the Adams predictor-corrector methods are named Adams-Moulton
predictors. And they are derived by integrating a polynomial interpolation of ( the
predicted value for the new point). y ' y' for the previous points plus

(d) The second-order predictor-corrector method is identical with the second-order Runge-
Kutta method.
(e) The third-and fourth-order predictor-corrector methods cannot be self-started. However,
once started, their computational efficiency is higher than that of the Runge-kutta method.
Error check in each interval is easier than for the Runge-Kutta method.

9.5 MORE APPLICATIONS

In this section, five applications of the numerical methods for initial value problems are
shown. Although the fourth-order Runge-Kutta method is used throughout this section, it
replaced by any other method for ordinary differential equations described in this chapter.

Example 9. 12

A metal piece of 0.1 kg mass and 200º C (or 473ºk) is suddenly placed in a room
of temperature 25º C, where it is subject to both natural convection cooling and radiation
heat transfer. Assuming that the temperature distribution in the metal is uniform, the
equation for the temperature may be written as

dT A
= [   ( 2974 − T 4 ) + hc (297 − T ) ] , T ( 0 ) = 473 ( A)
dt pcv

Where T is the temperature is degrees Kelvin, and we assume that the constants are given
by
p = 300 k g / m 3 ( density of the matel )

v = 0.001m 3 ( volume of the metal )

A = 0.25 m 3 ( surface area of the metal )

c = 900 J / k g K ( specific heat of the metal )

hc = 30 J / m 2 K ( heat transfer coefficient )

 = 0.8 ( emissivity of the metal )

 = 5.67  10 − 8 w / m 2 k 4 ( stefan− Boltzmann cons tan t )

(Solution)

This problem can be solved by modifying PROGRAM 9 – 2, which uses the

fourth-order Runge-Kutta method. Temperatures calculated by the fourth-order Runge-
Kutta method with h = 1 sec follow for selected values of t:

t (sec) T(ºK)
0 473
10 418.0
20 381.7
30 356.9
60 318.8
120 300.0
180 297.4

Example 9.13
 The electric current of the circuit shown in figure E9.13A satisfies the integro-
differential equation

di 1 t 1
L
dt
+ Ri +
C 0
i ( t' ) d t' +
C
q (0) = E (t ), t  0
(A)

Where the switch is closed at t = 0; I = i(t) is the current (amp); R is resistance (ohm); L,
C and E are given by

L = 200 Henry
C = 0.001 farad
E (t) = 1 volt for t > 0

Initial conditions are q (0) = 0 (capacitor's initial charge) and I (0) = 0. Calculate the current
for 0 ≤ t ≤ 5 sec after closing the switch (t = 0) for the following four values of R:

(a) R = 0 ohm
(b) R = 50 ohm
(c) R = 100 ohm
(d) R = 300 ohm

(Solution)
We first define

t
q (t ) = 0
i ( t' ) d t '
(B)
Differentiating Eq. (B) yields

d (C)
q (t ) = i (t ), q (0) = 0
dt

Introducing Eq. (B) into Eq. (A) and rewriting give

d R 1 1 E (t )
i ( t ) = − i ( t) − q (t ) + q(0) + , i ( 0) = 0 (D)
dt L LC LC L

thus, Eq. (A) is transformed to a set of two first-order ODEs Eqs (C) and (D) PROGRAM
9 N- 4 was modified for the problem in two respects (see note below). The result of the
computation is shown in a graphic form in figure E9.13b.

note: (a) to perform the calculations for all four cases in one run, four pairs of coupled
first-order ODEs are incorporated, the first pair corresponding to the first case, the section
pair to the second case, and so on. This is possible because not all equations in the program
have to be mathematically coupled.
(b) A graphic plotting routine is added, so all four cases are plotted on one graphic output.

Example 9.14

The three-mass system is shown in the figure below. The displacements of the three
masses satisfy the equations given by
 M 1 y' '1 + B1 y'1 + k1 y1 − B1 y' 2 − k 2 y2 = F1 ( t )

− B1 y '1 − k1 y1 M 2 y ' ' 2 + B1 y' 2 + ( k1 + k 2 ) y 2 − k 2 y3 = 0 (A)

− k 2 y 2 + M 3 y' '3 + B2 y '3 + ( k 2 + k 3 ) y3 = F3 ( t )

Constants and initial conditions are as follows:

k1 = k2 = k3 = 1 (b spring constants, kgm / s ²)

M 1 = M 2 = M 3 = 1 (Mass, kg )

F1 ( t ) = 1, F3 (t ) = 0 (Force, Newton)

B1 = B2 = 0.1 (Damping coefficients, kg / s)

y1 (0) = y '1 (0) = y 2 = y ' 2 (0) = y3 (0) = y '3 (0) = 0 (Initial conditions)

Solve the foregoing equations by using the fourth-order Runge-Kutta method for 0 ≤ t ≤
30 sec with h = 0.1

(Solution)

By defining
 y4 = y '1 , y5 = y'2 , and y6 = y '3 (B)

Eq. (A) is written as a set of six first-order ODEs as follows:

y'1 = y 4 (C 1)

y' 2 = y5
(C 2)

y'3 = y 6
(C 3)
1
y'4 = [ − B1 y 4 − k1 y1 + B1 y 5 + k 2 y 2 + F1 ] (C 4)
M2

1
y '5 = [ B1 y 4 + k1 y1 − B1 y 5 − ( k1 + k 2 ) y 2 + k 2 y 3 ] (C 5)
M2

1
y'6 = [ k 2 y 2 − B2 y 6 − ( k 2 k 3 ) y 3 + F3 ] (C 6)
M3

These equations are solved by modifying PROGRAM 9 – 3. The computational results are
shown in figure E9.14b.

Example 9.15
A rod 1.0 m long placed in a vacuum is heated by an electric current through the
rod. The temperature at both ends is fixed at 273º k. The heat is dissipated from the surface
by radiation heat transfer to the environment whose temperature is 273º k. Using the
following constants, determine the temperature distribution in the axial direction:
k = 60 W/ mk (thermal conductivity)

Q = 50 W / m (heat generation rate per unit length of the bar)

 = 5.6710 −8 W / m 2 k 4 (Stefan-Boltazmann constant)

A = 0.0001 m² (the cross section area)

p = 0.01 m (perimeter of the rod)

(Solution)

The heat conduction equation in the axial direction x is written as

d2
− Ak 2
T + p ( T 4 − 273 4 ) = Q 0  x  1.0 (A)
dx

With the boundary conditions

T (0) = T (1.0) = 273 K

Where T is temperature in degrees Kelvin.

The present problem is a boundary value problem (boundary conditions are
specified at x = 0 and x= 1), but it can be solved as an initial value problem on the trial-
and-error basis. By defining y1 and y2 as

y1 ( x ) = T ( x )

y 2 ( x) = T ( x)
Eq. (A) may be rewritten as a set of two first-order ODEs as

y '1 = y2 (B)

p Q
y '2 =  ( y 4 − 273 4 ) −
Ak kA

Only one initial condition, y1 (0) = 273, is known from the boundary conditions (but y2 (0)
is not known). So we solve Eq. (A) with trail values for y2 (0) until the boundary condition
for the right end, namely y1 (1) = 273, is satisfied. This approach is called the shooting
method [Rieder/ Busby].
For the present example, PROGRAM 9 – 3 is used with some modifications. The
results are directly plotted on a printer and shown in figure E9. 15. It is seen that y2 (0) =
1160 is too small as an initial guess, whereas y2 (0) = 1300 is too large. Some y2 (0) in
between these values should give the best result. After a few more trials, y2 (0) = 1200 is
found to satisfy almost exactly the right boundary condition.

Example 9.16

the temperature of a perfectly insulated iron bar 55 cm long is initially at 200 º C. the
temperature of the left edge is suddenly reduced and fixed to 0º C at t = 0 sec. calculate the
temperature distribution at even 100 sec is reached. the property constants are

k = 80.2 w / m k (thermal conductivity)

p = 7870 k g / m³ (density)

c = 447 k J / kg k (specific heat unit)

(Solution)

We first divide the rod into eleven control volumes as shown in figure E 9.16a. Denoting
the average temperature of control volume I by Ti (t). The heat balance equation for control
volume i is written as

p c  x A ( d T / dt ) = ( qi − 1 − qi ) A

In Eq. (A), q, is the heat flux (rate of condition of the heat transfer per unit cross-sectional
area) at the boundary of the control volumes I and I + 1, and written

INPUT PRINTING OR PLOTTING INTERVAL (P.I.)

103
INPUT NUMBER OF STEPS IN ONE P.I. OF X
5
INPUT MAXIMUM X TO STOP CALCULATION
1005
H= 20 T0 T1 T2 …. T10

t= 100 0 109 198 200 200 200 200 200 200

(Sec)
200 0 191 197 199 200 200 200 200

300 0 67 122 160 182 193 197 199 200 200 200

400 0 58 108 146 172 186 194 198 199 200 200

500 0 52 99 136 163 180 190 195 198 199 200

600 0 48 91 127 155 173 186 193 196 198 199

700 0 44 85 120 147 167 181 190 195 197 198

800 0 41 80 114 141 162 176 186 192 196 197

900 0 39 76 108 135 156 172 183 190 194 196

1000 0 37 72 104 130 152 168 179 187 192 194

Figure E9.16b Result of computations

By

k
qi = − ( Ti + 1 − Ti ) for i = 0, 1, 2, .......... , 9
x (B)

And

q10 = 0 (C)
introducing Eq. (B) into Eq. (A) and rearranging yield

d Ti k
= (Ti − 1 − 2 Ti + Ti + 1 ) (D – 1)
dt p c  x2

for I = 1, 2, 3,……., 9 and

d T10 k
= (T9 − T10 ) (D-
dt p c  x2 2)

Equation (D) may be considered as a set of first-order ODEs and solved by using
one of the Runge-Kutta methods. The set of equations is solved by PROGRAM 9 – 3 with
some modifications. The computed results are shown in figure E9. 16b.

Notes:

(a) Equation (D) may be viewed as a semi difference approximation for the heat conduction
equation (parabolic partial differential equation)
   T (x ,t )
k T (x , t) = p c
x x t

With the initial condition, T (x, 0) 200º C, and the boundary conditions, T (0, t) = T' (55, t)
= 0.

(b) The present solution technique for the partial differential equation using a numerical
method for ODEs is called the method of lines.

(c) space-dependant and time-dependant thermal conductivity can be implemented with a

minor change: that is to recalculate k for each boundary of the control volumes in each
time step.

(d) The author's study indicates that the computations using h = 50 sec agree well with that
of h = 1 sec, but the solution scheme becomes unstable with h = 100 sec.

9.6 STIFF ODEs

9.6.1 Why Stiff Equations Are Difficult

Stiffness refers to a very short time constant of an ODE. Consider, for example,

y' = −  y + s (t ) , y ( 0 ) = y0 (9.6.1)

Where α > 0. The solution of this equation is, if s = 0

y (t ) = y0 e −  t (9.6.2a)

And if s (t) ≠ 0, then

 t
y ( t ) = y0 e −  t 0
s (  ) e  d  (9.6.2b)

The response of the system to the initial condition as well as to the change of s (t) is
characterized by 1 / |α| that is called the time constant.

Solution of a stiff problem with a standard Runge-Kutta or predictor-corrector

method is difficult or, sometimes, impossible. For example, if the fourth-order Runge-
Kutta method is used for Eq. (9.6.2a and b), the computation becomes unstable unless h <
2.785 / |α| (sec Subsection 9.3.5). As the time constant becomes shorter, one has to use a
progressively smaller time step. For α = - 100000 sec ˉ¹ as an example, h must be smaller
than 2.785 / 100000 = 0.000002785 sec just to maintain stability. The predictor-corrector
methods discussed earlier in this chapter are subject to similar constraints.
When very fast transients of a system are computed, the necessity of small time
steps is understandable. On the other hand, when s (t) is slowly varying function or a
constant, the solution changes vary slowly, so naturally we desire to use larger time steps.
Nonetheless, the same time steps are necessary to assure stability of the numerical solution,
no matter how slow the actual change of the the solution is.
Stiffness is particularly serious for a set of ODEs [Gear (1971); Gear (1979); Hall
/ Watt; Fertziger; Kuo]. if the set of equation contains only one stiff equation, the stability
of a numerical method is governed by the short time constant of the stiffest equation. * For
example, in the case of two equations,

(9.6.3) y' = - y + z + 3

z = −10 7 z + y

The second equation has a significantly shorter time constant than the first.
A number of numerical methods that allow a large time step have been proposed
including the implicit Runge-Kutta method and the rational Runge-Kutta method. Two
such methods are introduced in the remainder of this section.

9.6.2 Implicit Methods

For simplicity, let us consider a set of two ODEs:

d
y = f ( y , z, t )
dt
(9.6.4)
d
z = g ( y , z , t)
dt

Using the backward difference approximation to the left side, we can write an implicit
scheme as

yn + 1 − yn = h f ( yn + 1 , zn + 1 , t n + 1 )  h fn + 1

(9.6.5)

zn + 1 − zn = h g ( yn + 1 , zn + 1 , tn + 1 )  h gn + 1

Where the ƒ and g terms on the right side have unknowns yn+1 and z n+1,
If ƒ and g are nonlinear functions, Eq. (9.6.5) cannot be solved in a closed form.
However, the iterative solution explained in Subsection 9.2.3 can be applied very easily
[Hall/Watt]. Indeed, the reader is encouraged to try it. Unfortunately, it is not
computationally efficient for a large system of ODEs. A more efficient approach is to
linearize the equations by Taylor expansions [Kubicek; Constantinites]. The Taylor
expansion of fk, n+1 about tn becomes

f n + 1 = f n + f y  y + f z  z + ft h

(9.6.6)
g n + 1 = g n + g y  y + g z  z + gt h

Where
  y = yn + 1 − yn ,  z = zn + 1 − zn (9.6.7)

Introducing Eq. (9.6.6) into Eq (9.6.5) and using Eq. (9.6.7) yields
.

1 − h f y − h fz   y  h fn h2 ft 
 −h g 1 − h g z   z  =   (9.6.8a)
 y   h g n h 2 gt 

or more compactly

( I − hJ )  y = RH S (‫ال‬8‫ز‬6‫ز‬9)

Where

R H S = vector on the right side of Eq. (9.6.8a)

J = the Jacobian matrix defined by

 fy fz 
J = 
g y g z 

I = identity matrix

 y = col (  y ,  z )

Equation (9.6.8) is solved by Gauss elimination. The implicit method is unconditionally

stable unless nonlinear effects cause instability.
The method has been extended to a larger set of coupled ODEs. The Gear methods
[Gear, 1971], which are available in NAG library [NAG], use higher-order backward
difference approximations with variable grid spacing.
9.6.3 Exponential Method

Exponential transformation and exponential fitting have been proposed and used by various
researchers to stiff ODEs. The subsection gives only a brief introduction of the basic ideas
in exponential methods.
To explain the principle, consider a single first0order ODE;

y' = f (y, t) (9.6.9)

Where, for simplicity of the discussions, we assume f does not include t explicitly.

To both terms of Eq. (9.6.9) yields

y ' + c y = f ( y, t ) + c y (9.6.10)

−ct
as an integrating factor, Eq. (9.6.10) is integrated e Where c is a constant. Using
 
As t n , t n + 1 in the interval

h
 [ f ( y ( t n +  ) , t n +  ) + c y ( t n +  ) ] e c (  − h ) d  (9.6.11)
−ch
y (t n +1 ) = yn e +
0

Where tn+1=tn+h. equation (9.6.11) is exact regardless of the choice of c.

Several different numerical schemes may be derived by introducing an
approximation for f+cy in the integrand. However, the accuracy of approximation
integration is then affected by the value of c. to find n appropriate value of c, we write y as

y (t ) = yn + y (t ) (9.6.12)

Introducing Eq. (9.6.9) yields

 y' = f ( y n + y)

= f n + ( f y ) n  y + 0( y 2 ) (9.6.13)
By ignoring the second-order error term, Eq. (9.6.13) can be equivalently written as

y' − ( f y ) n y = f n − ( f y ) n y n
(9.6.14)

Which is a linearized approximation for Eq.(9.6.9) about t=tn. if c in Eq. (9.6.10)is set to

c = − ( fy )n (9.6.15)

Then, Eq. (9.6.10) becomes identical with Eq. (9.6.14). By using

y' ' = f ' = f y y' = f y f (9.6.16)

Eq. (9.6.15) may also be expressed by

c = − ( f ' / f )n (9.6.17)

An explicit numerical scheme is obtained by setting the terms in the brackets of Eq.
(9.6.11) by

[ f ( y , tn +  ) + c y ( tn +  ) ]  f n + c yn (9.6.18)

Because the right side of Eq.(9.6.18) reduces to

 y n + 1 = yn e c h + ( 1 / c ) ( 1 − e − c h ) [ f n + c yn ]

 1 e −c h 
= yn + h f n   (9.6.19)
 ch 

Which is known as exponentially fitted method [Bui; Oran; Hetric; Fergason/hasen]. Not
only this method is unconditionally stable but also positivity of the solution is guaranteed
whenever the exact solution is expected to be positive.
The errors of Eq. (9.6.19) come from the approximation of Eq. (9.6.18). A more
accurate method using an iterative procedure is developed in the remainder of this
subsection. Based on Eq.(9.6.19), a predictor for y(t) for tn < t < tn+1 can be set

 1 − e − c 
y ( t ) = yn +   fn ,  = t − t n (9.6.20)
 c 

And for tn+1,

 1 − e −ch 
y n +1 = yn +   fn
 c 

By introducing Eq. (9.6.20) into Eq. (9.6.11) we obtain

h
y n + 1 = yn +  =0
[ f ( y ( t n +  ) , t n +  ) − f n + c y ( t n +  ) − cy n ] e c (  − h ) d  (9.6.21)

The second term of Eq. (9.6.21) is a correction of Eq. (9.6.20), and can be evaluated by
any one of the following:
(a) Analytical integration if it is possible.
(b) Approximating the terms in the brackets by a linear interpolation.
(c) Integrating by the trapezoidal rule.

Approach (a) is not easy unless ƒ is a simple function, so we do not consider it any
further. To pursue (b), the linear interpolation of the bracketed part is written as

[ f ( y , t n +  ) − f n + c y ( t n +  ) − c yn ]  B  (9.6.22)

Where

f n +1 − f n + c( y n +1 − y n )
B=
h

Introducing Eq. (9.6.22), the corrector becomes

Bh 2  1 − e − ch  (9.6.23)
yn +1 = y n +1 +  
ch  ch − 1 

If the trapezoidal rule is used, the corrector becomes

Bh 2
yn +1 = yn +1 + (9.6.24)
2

Which agrees with Eq. (9.6.23) in the limit of ch → 0?

The second term of Eq. (9.6.23) or Eq. (9.6.24) is a correction to Eq. (9.6.219). To
compute the second term, Eq. (9.6.19) is first evaluated, and then the second term is
computed.

An extension of the exponential method to a set of nonlinear equations is

straightforward, and the procedure is essentially the same as for a single equation. That
is, Eq. (9.6.19) is independently evaluated for all the equations. Once the predictors for
all the variables are obtained, then the second term of Eq. (9.6.23) or Eq. (9.6.24) is
evaluated.

SUMMARY OF THIS SECTION

(a) AN ODE becomes stiff if its time constant is short and ƒ' / ƒ < 0 (if there is no
homogeneous term, the solution approaches zero). if a standard numerical method such
as one of the Runge-Kutta or predictor-corrector methods is used, a very small time
step is required even when the solution is slowly changing.

(b) To alleviate the difficulty of the stiff ODEs, two methods, including an implicit
method and exponential method, are introduced.
11.1 Basic concepts and definitions

An equation containing the dependent and independent variables and one

or more partial derivatives of the dependent variable is called a partial differential

equation. In general it may be written in the form

F (x,y,.....,u,ux, uy,......,uxx, uyy,......)=0 (11.1)

involving several independent variables x,y,.....,an unknown function u of

these variables and the partial derivatives ux, uy,......,uxx, uxy, uyy ........of the function.

(11.1) is considered in a suitable subset D of Rn. For the sake of convenience we

confine our discussion for n=2. However extension of properties discussed here to

higher values of n is possible.

Here as in the case of ordinary differential equations, we define the order

of a partial differential equation to be the order of the derivative of highest order

occurring in the equation. The power of the highest order derivative in a differential

equation is called the degree of the partial differential equation.

u u
Example 11.1 (a) x +y =0 is a first-order equation in two variables with
x y

variable coefficients.

u u
(b) a +b =c; where x,y are independent variables, a and b are constants;
x y

is partial differential equation of first-order with constant coefficients.

u u
(c) + -(x+y) u=0 is a partial differential equation of first-order.
x y
  2u  2u  2u u u
(d) a(x) +2b(x) +c(x) =x+y+u+ + is a partial differential
x 2
xy y 2
x y

equation of second-order.

 2u  2u  2u u u
(e) a(x) +2b(x) +c(x) = f(x,y,u, , )
x 2
xy y 2
x y

where a(x), b(x) and c(x) are functions of x and f(..,..,.,.,.) is a function of x,y,u,

u u
and , is a partial differential equation of second order.
x y

 2u u
(f) u + =y is a partial differential equation of second-order.
xy x

 2u  2u  2u
(g) + 2y + 3x = 4 sin x is a partial differential equation of second-
x 2 xy y 2

order and degree one.

 2u  2u
(h) = is a partial differential of second-order.
x 2 y 2

2
 u 
2
 u 
(i)   +   = 1 is a partial differential equation of first-order and second
 x   y 

degree.

By a solution of a partial differential equation of the type

F (x,y,u, ux,uy,uxx,uyy, uxy) =0 (11.2)

we understand functions u=(x,y) which satisfy (11.2) identically in D, that is, if we

put values of quantities on the left hand side we get right hand side.

Example 11.2.

(i) Show that sin n(x+y), cosn(x+y) and ex+y are solutions of the partial

differential equation
   u   u 
  -   =0
  x   y 
(ii) Show that u(x,y)=(x+y)3 and u(x,y)=sin (x-y) are solutions of the partial

differential equation

 2u  2u
- =0
x 2  y 2
u
Solution (i) = n cos n (x+y) if u(x,y) = sin n(x+y)
x
u
= n cos n (x+y) if u(x,y) = sin n(x+y)
y
u u
L.H.S. of the equation is - = ncos n(x+y) – n cos n(x+y) =0=R.H.S.
x y
u
= -nsin n(x+y) if u(x,y) = cos n(x+y)
x

u
= -nsin n(x+y) if u(x,y) = cos n(x+y)
y
L.H.S. = [-nsin n(x+y)]-[-n sin n(x+y)]=0=R.H.S.

u
= ex+y if u(x,y) ex+y
x

u
= ex+y if u(x,y) = ex+y
y

L.H.S. = ex+y -ex+y =0 = R.H.S.

u  2u
(ii) For u(x,y) = (x+y)3, =3(x+y)2, =6 (x+y)
x x 2
u  2u
For u(x,y)=(x+y)3, =3(x+y)2, = 6(x+y)
y y 2
This implies that L.H.S. of the given partial differential is
 2u  2u
- = 6(x+y)-6(x+y)=0=R.H.S.
x 2  y 2
 u  2u
For u(x,y)=sin (x-y), =cos (x-y), = - sin (x-y)
x x 2
u  2u
= - cos (x-y), = - sin (x-y)
y y 2
L.H.S. of the partial differential equation is
 2u  2 y
- = - cos (x-y) + cos (x-y) = 0 = R.H.S.
x 2  y 2
Therefore (x+y)3 and sin (x-y) are solutions of
 2u  2u
- = 0.
x 2  y 2

A partial differential equation is said to be linear if the unknown function

u(.,.) and all its partial derivatives appear in an algebraically linear form, 'that is, of

the first degree. For example the equation

A uxx+2Buxy+Cuyy+Dux+Euy+Fu = f (11.3)

where the coefficients A,B,C,D.E and F and the function f are functions of x and y,

is a second-order linear partial differential equation in the unknown u(x,y).

Left hand side of (11.3) can be abbreviated by Lu, where u has continuous

partial derivatives of upto second order.

If u is a function having continuous partial derivatives of appropriate order,

say n then a partial derivative can be written as Lu=f where L is a differential

operator, that is, L carries u to the sum of scalar multiplications of its partial

derivatives of different order. An operator L is called linear differential operator if L

(u+v)= Lu+v where  and  are scalars and u and v are any functions with

continuous partial derivatives of appropriate order. A partial differential equation is

called homogeneous if Lu=0, that is, f on the right hand side of a partial differential
equation is zero, say f=0 in 11.3. The partial differential equation is called non-

homogeneous if f0.

(x+2y) ux +x2uy = sin (x2+y2) is a non-homogeneous partial differential

equation of first-order.

(x+2y) ux+x2uy=0 is a homogeneous linear partial differential equation of

first-order.

xuxx +yuxy+uyy=0 is a homogeneous linear partial differential equation of

second-order.

xuxx+y uxy+uyy=sin x is a non-homogeneous linear partial differential

equation of second-order.

The general solution of a linear partial differential equation is a linear

combination of all linearly independent solutions of the equation with as many

arbitrary functions as the order of the equation; a partial differential equation of

order 2 has 2 arbitrary functions. A particular solution of a differential equation is

one that does not contain arbitrary functions or constants. Homogeneous linear

partial differential equation has an interesting property that if u is its solution then

a scalar multiple of u, that is, cu, where c is a constant, is also its solution. Any

equation of the type F(x,y,u,c1,c2)=0, where c1 and c2 are arbitrary constants, which

is a solution of a partial differential equation of first-order is called a complete

solution or a complete integral of that equation. An equation F(,)=0 involving

arbitrary function. F connecting two known functions  and  of x, y and u, and

providing a solution of a first order differential equation is called a general solution

or general integral of that equation. It is clear that in some sense general solution
provides a much broader set of solutions than a complete solution. However a

general solution may be derived once a complete solution is known.

u u  2u
Very often ux = , uy = ,uxx = 2
x y x

 2u  2u
uxy = and uyy = are respectively denoted by p, q,r, s and t.
xy y 2

In this notation the general form of partial differential equation of first-order

is

F(x,y,u,p,q)=0 (11.4)

The general second-order partial differential equation is of the form

F(x,y,u,p,q,r,s,t)=0 (11.5)

A partial differential equation is said to be quasilinear if it is linear in all the

highest-order derivatives of the dependent variable. The most general form of a

quasi linear second- order equation is

A(x,y,u,p,q) uxx + B(x,y,u,p,q) uxy + C(x,y,u,p,q) uyy +f(x,y,u,p,q)=0 (11.6)

A partial differential equation of first-order is called semilinear if it is linear

u
in the principal part, namely the terms involving first derivatives: thus, for A +
x

u
B = C, these equations are defined to be such that the left hand side, which
y

contains all derivatives is linear in u in that A,B depend on x and y alone; however

C may depend non linearly on u. A semi linear partial differential equation of

second-order is of the form

 2u  2u  2u u u
A + 2B +C = f(x,y,u, , ) (11.7)
x 2
xy y 2
x y
where A,B,C are functions of x and y.

11.2. Classification of Partial Differential Equations

We have seen the classification of Partial Differential equations into linear,

quasilinear, semi linear, homogeneous and non-homogeneous categories in

Section 11.1. In this section we mainly focus on the classification of second order

equations into elliptic, hyperbolic and parabolic types. Notion of Cauchy data (initial

and boundary conditions) and characteristic for partial differential equations are

introduced.

11.2.1 Initial and Boundary Value Problems

A partial differential equation subject to certain conditions in the form of

initial or boundary condition is known as an initial-value or a boundary value

problem. The initial conditions, also known as Cauchy conditions, are the values

of the unknown function u(.,.) and an appropriate number of its derivatives at the

initial point.

Let us consider a second-order partial differential equation for the function

u(.,.) in the independent variables x and y, and suppose that this equation can be

solved explicitly for uyy, and hence can be represented in the form

uyy = F(x,y,u,ux,uy,uxx,uxy) (11.8)

For some value y=y0, we prescribe the initial values of the unknown function

u and of the derivative with respect to y

u(x,y0)=f(x) (11.9)

uy(x,y0)=g(x) (11.10)
 The problem of determining the solution of (11.8) satisfying initial conditions

(11.9)-(11.10) is known as the initial-value problem. Here initial-value usually

refer to the data assigned at y=y0. If initial values are prescribed along some curve

 in the (x,y) plane, that is, finding solution of equation (11.8) subject to prescribed

value of y on some curve  is called the Cauchy problem. These conditions are

called Cauchy data. Actually two names are synonymous.

Example 11.3 (a) ut = uxx 0<x<1, t>0

u(x,0)= cos x 0 x  l

is an initial-value problem.

(b) Suppose that is a curve in the (x,y) plane; we define Cauchy data to be the

prescription of u on . It is convenient to write this boundary condition in the

parametric form

x=x0(s), y=y0(s), u=u0(s), for s1 s  s2.

u u
A(x,y,u) +B (x,y,u) =C (11.10)
x y

subject to (11.10) is a Cauchy problem

(c) Let us consider the equation

A(x,y) uxx + B(x,y) uxy + C uyy = F(x,y,u,ux,uy). (11.11)

Let (x0,y0) denote points on a smooth curve  in the (x,y) plane. Also let the

parametric equations of this curve  be

x=x0 (), y0=y0 ()

where  is a parameter.
 We suppose that two functions f() and g() are prescribed along the curve

. The Cauchy problem is now one of determining the solution u(x,y) of Equation

(11.11) in the neighbourhood of the curve  satisfying the Cauchy conditions

u
u=f(), =g()
n

on the curve . n is the direction of the normal to  which lies to the left of  in the

counter clockwise direction of increasing arc length. The functions f() and g(()

are the Cauchy data.

The solution of the Cauchy problem is a surface, called an integral surface,

in the (x,y,u) space passing through  a curve having  as its projection in the (x,y)

u
plane and satisfying =g() which represents a tangent plane to the integral
n

surface along .

Types of Boundary Conditions

The boundary conditions on partial differential equation (11.6) fall into the

following three categories:

(i) Dirichlet boundary conditions (also known as boundary conditions of the

first kind), when the values of the unknown function u are prescribed at each

point of the boundary  of a given domain  on which (11.6) is defined.

(ii) Neumann boundary conditions (also known as boundary conditions of the

second kind), when the values of the normal derivatives of the unknown

function u are prescribed at each point of the boundary .

(iii) Robin boundary conditions (also known as boundary conditions of the third

kind, or mixed boundary conditions), when the values of a linear combination

 of the unknown function u and its normal derivative are prescribed at each

point of the boundary .

u  2u
Example 11.4 (i) = k 2 , 0<x<l,t>0
t x

u(x,o)=f(x)

u
(x,o)=g(x), 0<x<l
t

u(0,t) =T1(t)

u(l,t)=T2(t), t>0

It is a Dirichlet boundary value problem.

u  2u
(ii) =k , 0<x< l, t>o
t x 2

u
u(x,o)=f(x), (x,o)=g(x), 0<x< l
t

u u
(0,t) =T3(t), (l,t)=T4(t), t>0
n n

It is an example of Neumann boundary value problem.

u  2u
(iii) = k 2 , 0<x< l, t>0
t x

u
u(x,o)=f(x), (x,o)=g(x), 0<x< l,
t

u 
u(0, t) +  (0, t) = 0,
n 
t  0 .
u 
u(l, t) +  (l, t ) = 0
x 

It is an example of Robin boundary problem.

 It may be observed, a major part of scientific and technological studies are

devoted to initial and boundary value problems. Solutions of few important initial

and boundary value problems will be discussed in the next chapter.

11.2.2 Classification of Second-order partial differential Equations

For f=0 in Equation (11.3), the most general form of a second-order

homogeneous equation

A uxx + 2B uxy+C uyy + D ux+E uy +Fu=0 (11.12)

For a correspondence of this equation with an algebraic quadratic equation,

we replace ux by , uy by , uxx by 2, uxy by , and uyy by 2. The left hand side

of Equation (11.12) reduces to a second degree polynomial in  and :

P( ,)=A2+2B+C2+D+E+F=0 (11.13)

It is known from analytical geometry and algebra that the polynomial

equation P (,)=0 represents a hyperbola, parabola, or ellipse according as its

discriminant. B2-AC is positive, zero, or negative. Thus, the partial differential

equation (11.12) is classified as hyperbolic, parabolic, or elliptic according as

the quantity

B2-AC>0, B2-AC=0, or B2-AC<0.

The equation

A u2x+2B uxy + C u2y = 0 (14.14)

is called the characteristic equation of the partial differential equation (11.13).

Solutions of (11.14) are called the characteristics

Example 11.5 Examine whether the following partial differential equations are

hyperbolic, parabolic, or elliptic.

  2u  2u
(i) + x +4=0
x 2 y 2

 2u  2u
(ii) + y =0
x 2 y 2

2 2u  2 u
(iii) y - =0
x 2 y 2

(iv) uxx + x2 uyy = 0

(v) x uxx + 2x uxy + y uyy = 0

Solution (i) A = 1, C = x, B = 0

B2-AC = 0 –x <0 for x>0

Thus the equation is elliptic if x > 0, is hyperbolic if x < 0 and it is parabolic

if x = 0.

(ii) A=1, B=0, C=y

B2-AC=0-y >0 if y<0 and so the equation is hyperbolic if y<0. It is parabolic

if y=0 and it is elliptic if y>0.

(iii) A=y2, B=0, C = -1.

B2-AC=y2>0 for all y. Therefore the equation is hyperbolic.

(iv) A=1, B=0, C=x2

B2-AC=0-x2<0 for all x. The equation is elliptic

(v) A=x, B=x, C=y

B2-AC=x2-xy=x(x-y)>0 for x>0 x>y

In this case the equation is hyperbolic B2-AC=o if x=y. For this the equation

is parabolic. B2-AC <0 if x>y and x<0 or if x<y and x>0

In this case the equation is elliptic.

11.3 Solutions of Partial Differential Equations of First-order

11.3.1 Solution of Partial Differential Equations of first-order with constant

coefficients.

The most general form of linear partial differential equations of first order

with constant coefficients is

Aux+Buy+Ku=f(x,y) (11.15)

where A,B and K are constants

Let u(x,y) be a solution of (11.15) then

du=uxdx+uydy (11.16)

From (11.15) and (11.16) we get the auxiliary system of equations

(comparing coefficients of ux, uy and remaining terms).

dx dy du
= = (11.17)
A B f ( x, y ) − Ku

Bx − c
The solution of the left pair is Bx-Ay=c or y= , where c is an arbitrary
A
constant of integration
 dx dy
 A = B or Bdx-Ady=0 or Bx-Ay=c by integrating both sides of the


previous equation  .

The other pair

dx du
=
A f ( x, y ) - Ku

is reduced to an ordinary differential equation with u as the dependent variable and

x as the independent variable, namely

du f ( x, y ) - Ku
=
dx A
 Bx - c
f ( x, )
du Ku f ( x, y ) A
or + = =
dx A A A

The integrating factor of this differential equation is eKx/A. Making change of

variable by v=ueKx/A (11.15) takes the form

kx
Avx+Bvy = f(x,y)e A =g(x,y)

The substitution v=ueKy/B in (11.15) leads to Avx+Bvy=f(x,y) eKy/B. Thus, we

need to consider only the formal reduced form

Aux+Buy=f(x,y) (11.18)

The auxiliary system of equations for (11.18) is

dx dy du
= = (11.19)
A B f ( x, y )

dx dy
The solution of = is
A B

Bx-Ay=c, which gives

Ay + c
x=
B
Substituting this value in

dy du
= we get
B f ( x, y )

dy du
=
B Ay + c
f( , y)
B

Ay + c
f( , y)
or du=F(y,c) dy where F(y,c)= B
B

Solution of this equation is

 u=G(y,c)+c1, where Gy (y,c)= F(y,c).

Thus, the general solution is obtained by replacing c1 by  (c) and c by Bx-

Ay, thereby yielding

u(x,y)=G(y, Bx-Ay)+ ( Bx-Ay), and the solution of equation (11.15) is

u(x,y)=[G(y, Bx-Ay)+ ( Bx-Ay)]e-Kx/A (11.20)

Equations (11.17) are called the equations of the Characteristics. These

equations contain two independent equations, with two solutions of the form

F(x,y,u)=c1 and G(x,y,u)=c2. Each of these represents a family of surfaces. The

curves of intersection of these two families of surfaces are known as the

characteristics of the partial differential equation. The projections of these curves

in the (x,y)-plane are called the base characteristics. The general solution

represents a family of surfaces, and these surfaces are called integral surfaces.

Thus, the equation Bx-Ay=c represents a family of planes. The intersection

of any one of these planes with an integral surface is a curve whose projection in

the (x,y)-plane is again given by Bx-Ay=c, but this time this equation represents a

straight line and is the base characteristic. Therefore, the solution u on a base

characteristic Bx-Ay=c is given by u=G(y,c)+c1, and the general solution is the

same as above.

Example 11.6 Find the general solution of the first-order linear partial differential

equation with the constant coefficients:

4ux+uy=x2y

Solution: The auxiliary system of equations is

dx dy du
= = 2
4 1 x y
 From here we get

dx dy
= or dx-4dy=0. Integrating both sides
4 1

dx du
we get x-4y=c. Also = 2 or x2y dx=4du
4 x y

x-c
or x2 ( ) dx =4du or
4

1
(x3 – cx2) dx = du
16

Integrating both sides we get

3 x 4 - 4cx 3
u=c1+
192

3 x 4 - 4cx 3
= f(c)+
192

After replacing c by x-4y, we get the general solution

3 x 4 - 4( x - 4 y )x 3
u=f(x-4y)+
192

x4 x3y
=f(x-4y)- +
192 12

11.3.2. Lagrange's Method

The general form of first-order linear partial differential equations with

variable coefficients is

P(x,y)ux+Q(x,y)uy+f(x,y)u=R(x,y) (11.21)

We can eliminate the term in u from (11.21) by substituting u=ve-(x,y), where

(x,y) satisfies the equation

 P(x,y) x(x,y)+ Q (x,y) y(x,y)=f(x,y)

Hence, Eq (11.21) is reduced to

P(x,y)ux+Q (x,y) uy =R(x,y) (11.22)

where P,Q,R in (11.22) are not the same as in (11.21). The following theorem

provides a method for solving (11.22) often called Lagrange's Method.

Theorem 11.1 The general solution of the linear partial differential equation of first

order

Pp+Qq=R; (11.23)

u u
where p= , q = , P, Q and R are functions of x y and u
x y

is F(, ) = 0 (11.24)

where F is an arbitrary function and  (x,y,u) =c1 and  (x,y,u)=c2 form a solution

of the auxiliary system of equations

dx dy du
= = (11.25)
P Q R

Proof: Let  (x,y,u)=c1 and  (x,y,u)=c2 satisfy (11.25), then equations

xdx+y dy +udu=0

and

dx dy du
= =
P Q R

must be compatible, that is, we must have P x+Qy+Ru=0

Similarly we must have

Px+Qy+Ru=0

Solving these equations for P,Q, and R, we have

 P Q R
= = (11.26)
(,  ) / ( y, u) (,  ) / (u, x ) (,  ) / ( x, y )

where  (,)/(y,u)= yu- yu0 denotes the Jacobian.

Let F(,)=0. By differentiating this equation with respect to x and y,

respectively, we obtain the equations

F     F    
 + p +  + p = 0
  x u    x u 

F     F    
 + q +  + q = 0
  y u    y u 

F F
and if we now eliminate and from these equations, we obtain the
 
( ,  ) (,  ) ( ,  )
equation p +q = (11.27)
( y, u) (u, x ) ( x, y )

Substituting from equations (11.26) into equation (11.27), we see that

F(,)=0 is a general solution of (11.23). The solution can also be written as

 =g() or =h(),

Example 11.7 Find the general solution of the partial differential equation y2up +

x2uq = y2x

Solution: The auxiliary system of equations is

dx dy du
2
= 2 = 2 (11.28)
y u x u xy

Taking the first two members we have x2dx = y2dy which on integration

given x3-y3 = c1. Again taking the first and third members,

we have x dx = u du

which on integration given x2-u2 = c2

Hence, the general solution is

 F(x3-y3,x2-u2) = 0

11.3.3 Charpit's Method for solving nonlinear Partial Differential Equation of

First-Order

We present here a general method for solving non-linear partial differential

equations. This is known as Charpit's method.

Let

F(x,y,u, p.q)=0 (11.29)

be a general non linear partial differential equation of first-order. Since u

depends on x and y, we have

du=uxdx+uydy = pdx+qdy (11.30)

u u
where p=ux= , q = uy=
x y
If we can find another relation between x,y,u,p,q such that

f(x,y,u,p,q)=0 (11.31)

then we can solve (11.28) and (11.30) for p and q and substitute them in equation

(11.29). This will give the solution provided (11.29) is integrable.

To determine f, differentiate (11.28) and (11.30) w.r.t. x and y so that

F F F p F q
+ p+ + =0 (11.32)
x u p x q x

f f f p f q
+ p+ + =0 (11.33)
x u p x q x

F F F p F q
+ q+ + =0 (11.34)
y u p y q y

f f f p f q
+ q+ + =0 (11.35)
y u p y q y
 p q
Eliminating from, equations (11.31) and (11.32), and from equations
x y
(11.33) and (11.34) we obtain
 F f f F   F f f F   F f f F  q
 -  +  - p +  -  =0
 x p x p   u p u p   q p q p  dx

 F f f F   F f f F   F f f F  p
 -  +  - q +  -  =0
 y q y q   u q u q   p q p q  dy

Adding these two equations and using

q  2u p
= =
x xy y

and rearranging the terms, we get

 F  f  F  f  F F  f  F F  f
 -  +  -  +  - p - q  + +p 
 p  x  q  y  p q  u  x u  p
(11.36)
 F f  f
+  + q  =0
 y u  q

Following arguments in the proof of Theorem 11.1 we get the auxiliary

system of equations

dx dy du dp dq df
= = = = = (11.37)
- F - F F F F F F F 0
-p -q +p +q
p q p q x u y u

An Integral of these equations, involving. p or q or both, can be taken as

the required equation (11.30). p and q determined from (11.28) and (11.30) will

make (11.29) integerable.

Example 11.8 Find the general solution of the partial differential equation.

2 2
 u   u 
  x +   y - u = 0 (11.38)
 x   y 
 u u
Solution: Let p = ,q=
x y

The auxiliary system of equations is

dx dy du dp dq
= = = = (11.39)
2px 2qy 2(p x + q y ) p - p
2 2 2
q - q2

which we obtain from (11.36) by putting values of

F F F F F
= 2px , = 2qy, = p2, = - 1, = q2
p q x u y

and multiplying by -1 throughout the auxiliary system. From first and 4th expression

in (11.38) we get

p 2 dx + 2pxdp
dx = . From second and 5th expression
py

q 2 dy + 2qydq
dy=
qy

Using these values of dx and dy in (11.38) we get

p 2 dx + 2pxdp q 2 dy + 2qydq
=
p2 x q2 y

dx 2 dy 2dq
or + dp = +
x p y q

Taking integral of all terms we get

ln|x| + 2ln|p| = ln|y|+2ln|q|+lnc

or ln|x| p2 = ln|y|q2c

or p2x=cq2y, where c is an arbitrary constant. (11.40)

Solving (11.37) and (11.39) for p and q we get cq2y+q2y -u=0

(c+1)q2y=u
 1
 u  2
q=  
 (c + 1)y 

1
 cu  2
p=  
 (c + 1)x 

(11.29) takes the following form in this case

1 1
 cu  2  u  2
du=   dx +   dy
 (c + 1)x   (c + 1)y 

1 1 1
 1+ c  2 c 2 i 2
or   du =   dx +   dy
 u  x y

1 1 1
By integrating this equation we obtain ((1+ c )u) 2
= (cx) 2
+ ( y) 2
+ c1

This is a complete solution.

11.3.4 Solutions of special type of partial differential equations

(i) Equations containing p and q only

Let us consider a partial differential equation of the type

F(p,q)=0 (11.41)

The auxiliary system of equations of Charpit's method (Equation (11.36))

takes the form

dx dy du dp dq
= = = =
Fp Fq pFp + qFq 0 0

It is clear that p=c is a solution of these equations. Putting value of p in

(11.40) we have

F(c,q)=0 (11.42)

So that q=G(c) where c is a constant

 Then observing that

du=cdx+G(c) dy

we get the solution u=cx +G(c) y+c1,

where c1 is another constant.

Example 11.9 Solve p2+q2=1

Solution: The auxiliary system of equation is

dx dy du dp dq
- = = 2 2
= =
- 2p 2q - 2p - 2q 0 0

dx dy du dp dq
or = = 2 = =
p q p +q 2
0 0

Using dp =0, we get p=c and q= 1- c 2 , and these two combined with du

=pdx+qdy yield

u=cx+y 1- c 2 + c1 which is a complete solution.

dx dx
Using = p , we get du = where p= c
du c

x
Integrating the equation we get u = + c1
c

dy
Also du = , where q = 1- p 2 = 1- c 2
q

dy 1
or du = . Integrating this equation we get u = y +c2
2 2
1- c 1- c

This cu = x+cc1 and u 1 - c 2 = y + c2 1- c 2

Replacing cc1 and c2 1- c 2 by -  and - respectively, and eliminating c,

we get
 u2 = (x-)2 + (y-)2

This is another complete solution.

This is another complete solution.

(ii) Clairaut equations

An equation of the form

u=px+qy+f(p,q)

or

F=px+qy+f(p,q)-u=0 (11.43)

is known as Clairaut equation.

The auxiliary system of equations for Clairaut equation takes the form

dx dy du dp dq
= = = =
x + fp y + fq px + qy + pf p + qf q 0 0

From here we find that

dp=0, dq=o implying

p=c1, q=c2

If we put these values of p and q in Eq. (11.42), we get

u = c1 x +c2y +f (c1, c2)

Therefore, F(x,y,u,c1,c2) = c1x + c2y + f (c1,c2) -u=0 is a complete solution

of (11.42).

(iii) Equations not containing x and y

Consider a partial differential equation of the type

F(u,p,q) = 0 (11.44)

The auxiliary system of equations take the form

 dx dy du dp dq
= = = =
Fp Fq pFq + qFq - pFu - qFu

dp dq
The last two terms yield =
p q

i.e. p = a2q where a2 is an arbitrary constant

This equation together with 11.43 can be solved for p and q and we

proceed as in previous cases.

Example 11.10 Solve u2+pq – 4 = 0

Solution. The auxiliary system of equations is

dx dy du dp dq
= = = =
q p 2pq - 2up - 2uq

The last two equations yield p = a2q.

Substituting in u2+pq – 4 = 0 gives

1
q=  4 - u 2 and p = + a 4 - u2
a

Then du = pdx+qdy yields

 1 
du = + 4 - u2  adx + dy 
 a 

du 1
or = + adx + dy
4-u 2 a

u  1 
Integrating we get sin--1 = +  adx + y + c 
2  a 

 1 
or u = + 2 sin  ax + y + c 
 a 

which is the required complete solution.

(iv) Equations of the type

 f(x,p) = g(y,q)

Then each of these functions must be constant, that is

f(x, p) = g(y, q) = C

Solving for p and q, and using du=pdx+qdy we can obtain the solution

Example 11.11 Solve p2(1-x2)-q2(4-y2) = 0

Solution Let p2(1-x2) = q2 (4-y2) = a2

a a
This gives p = and q =
2
1- x 4 - y2

(neglecting the negative sign).

Substituting in du = pdx + q dy we have

a a
du = dx + dy
1- x 2 4 - y2

 y
Integration gives u = a  sin' x + sin'  + c.
 2

which is the required complete solution.

11.3.5. Geometric concepts related to Partial Differential Equations of First

order

We have discussed geometrical interpretation of a first order ordinary

differential equation in chapter. .........

The situation for a partial differential equation is some what complicated. In

u u
this case the values of p= , q= are not unique at a point (x,y,u). If an integral
x y

surface is g(x,y,u)=0, then p and q represent the slopes of the curves of

intersection of the surface with the planes y=constant and x=constant, respectively.
Moreover, p,q,-1 represent the direction ratios of the normals to the surface at the

point (x,y,u). The derivatives p and q are constrained by F(x,y,u,p,q)=0. Obviously,

at a fixed point, p and q can be represented by a single parameter. Hence, there

are infinitely many possible normals and consequently infinitely many integral

surfaces passing through any fixed point. So, unlike the case of ordinary differential

equations, we cannot determine a unique integral surface by making it pass

through a point.

Cauchy established that a unique integral surface can be obtained by

making it pass through a continuous twisted space curve, also known as an initial

curve, except when the curve is a characteristic of the differential equation.

The infinity of normals passing through a fixed point generates a cone

known as the normal cone. The corresponding tangent planes to the integral

surfaces envelope a cone known as the Monge cone. In the case of a linear or a

quasi linear equation, the normal cone degenerates into a plane since each normal

is perpendicular to a fixed line. Consider the equation ap+bq=c, where a,b, and c

are functions x,y, and u. Then the direction p,q,-1 is perpendicular to the direction

ratios a,b,c. This direction is fixed at a fixed point. The Monge cone then

degenerates into a coaxial set of planes known as the Monge pencil. The common

axis of the planes is the line through the fixed point with direction ratios a,b,c. This

line is known as the Monge axis.

11.4 Solutions of Linear Partial Differential Equation of Second Order with

Constant Coefficients

11.4.1 Homogeneous Equations

   i i i
Let Dx= , Dy = ,D x = i , D iy = i ,
x y x y

We are looking for solving equations of the type

 2u  2u  2u
+ k + k =0 (11.45)
x 2 xy y 2
1 2

where k1 and k2 are constants.

(11.44) can be written as

(D 2
x )
+ k 1D xD y + k 2D2y u = 0

or F(Dx, Dy) u=0 (11.46)

The auxiliary equation of (11.45) (compare with Section 5.5) is

D 2x + k 1D x D y + k 2D 2y = 0

Dy then equation (11.45) can be written as

Let the roots of this equation be m1 and m2, that is, Dx=m1Dy, Dx=m2Dy

(Dx-m1Dy) (Dx-m2Dy)u=0- (11.47)

This implies

(Dx-m2Dy) u=0 or p-m2q=0

The auxiliary system of equations for p-m2q=0 is of the type

dx dy du
= =
1 - m2 0

This gives us -m2dx=dy

or y+m2x=c

and u=c1= (c)

Thus, u=(y+m2x) is a solution of (11.44).

From (11.46) we also have (Dx-m1Dy) u=0

 or p-m1q=0

Its auxiliary system of equations is

dx dy du
= =
1 - m1 0

This gives –m1dx=dy or m1x+y=c1 and u=c2 and so u=(y+m1x) is a solution

of (11.44).

Therefore u= (y+m2x) +  (y+m1x) is the complete solution of (11.44)

If the roots are equal (m1 = m2) then Equation 11.44 is equivalent to

(Dx-m1Dy)2 u = 0

Putting (Dx-m1Dy) u = z, we get

(Dx-m1Dy) z=0 which gives

z= (y+m1x)

Substituting z in (Dx-m1Dy) u=z gives

(Dx-m1Dy) u =  (y+m1x)

or p-m1q =  (y+m1x)

Its auxiliary system of equations is

dx dy du
= =
1 - m1 ( y + m1x )

which gives y+m1x = a & u +  (a) x+b

The complete solution in this case is

u= x  (y+m1x) +  (y+m1x)

Example 11.12 Find the solution of the equation

 2u  2u
- =0
x 2 y 2
Solution: In the terminology introduced above this equation can be written as

(Dx2-Dy2) u = 0.

or (Dx-Dy) (Dx+Dy)u=0

Its auxiliary equation is

(Dx-Dy)(Dx+Dy)=0,

that is, Dx - Dy =0

or Dx= -Dy. that is,

p=q or p = - q

p-q = 0 or p+q=0

Auxiliary system of equations for p-q=0 is

dx dy du
= =
1 -1 0

This gives x+y = c.

The auxiliary system for p+q = 0 is

dx dy du
= =
1 1 0

This gives x-y =c1

The complete solution is

u=(x+y)+  (x-y) where  and  are arbitrary functions.

Non-homogeneous Partial Differential Equations of the second-order

Equations of the type

 2u  2u  2u
+ k + k =f(x,y) (11.48)
x 2 xy y 2
1 2
 are called non-homogeneous partial differential equations of the second-

order with constant coefficients.

Let uc be the general solution of

 2u  2u  2u
+ k + k =0 (11.49)
x 2 xy y 2
1 2

and let up be a particular solution of (11.47)

Then uc+up is the solution of (11.47)

We have discussed the method for finding the general solution

(complementary function) of (11.48). In Section 5.6 we described the method of

undetermined coefficients for ordinary differential equations. That method is

applicable in finding particular solution of partial differential equations of the type

(11.47) Let f(Dx,Dy) be a linear partial differential operator with constant coefficients,

then the corresponding inverse operator is defined

1
as
f (D x ,D y )

The following results hold

 1 
f(Dx,Dy)  ( x, y ) = ( x, y ) (11.50)
 f (D x ,D y ) 

1 1  1 
( x, y ) =  ( x, y ) (11.51)
f1 (D x ,D y )f2 (D x ,D y ) f1(D x ,D y )  f2 D x ,D y ) 

1  1 
=  ( x, y ) (11.52)
f2 (D x ,D y )  f1(D x ,D y ) 
 1
1( x, y) + 2 ( x, y) =  1 1( x, y)
(D x ,D y ) f (D x ,D y )
(11.53)
1
+  2 ( x, y )
f (D x ,D y )

1 1
e ax+by = = e ax+by , f (a,b)  0 (11.54)
f (D x ,D y ) f (a,b)

f(Dx,Dy)  (x,y) eax+by=eax+by f(Dx+a, Dy+b) (x,y)

1 1
( x, y )e ax+by = e ax+by ( x, y ) (11.55)
f (D x ,D y ) f (D x + a,D y + b)

1 1
= e ax e by ( x, y ) = e by e ax ( x, y ) (11.56)
f (D x + a,D y ) f (D x ,D y + b)

f (D 2x , D 2y ) cos (ax+by) = f(-a2,-b2) cos (ax+by)

1 1
2 2
cos (ax + by ) = cos (ax + by ) (11.57)
f (D , D y )
x f (-a , - b 2 )
2

f (D 2x , D 2y ) sin (ax+by) = f(-a2,-b2) sin (ax+by)

1 1
2 2
sin (ax + by ) = sin (ax + by ) (11.58)
f (D , D y )
x f (-a , - b 2 )
2

1
When (x,y) is any function of x and y, we resolve into partial
f (D x , D y )

fractions treating f(Dx, Dy) as a function of Dx alone and operate each partial

fraction on (x,y), remembering that

1
(x,y) =  ( x, c − mx )dx
D x − mD y

where c is replaced by y+mx after integration.

Example 11.13
 Find the particular solution of the following partial differential equations

 2u  2u u
(i) 3 2 +4 - = e x −3 y
x xy y

 2u u
(ii) 3 - = e x sin( x + y )
x 2 y

Solution: (i) The equation can be written as

(3D 2x + 4 D xD y - D y ) u = ex-3y

1
up = ex-3y
3D + 4 D x D y - D y
2
x

1
= ex-3y by (11.53)
3 + 4(-3) - (-3)

1 x-3y
= - e
6

(ii) The equation can be written as

(3D2x-Dy)u=ex sin (x+y)

1
up = 2
ex sin (x+y)
3D - D y
x

1
= ex sin (x+y)
(3(D x + 1) 2 - D y )

1
= ex sin(x+y)
(3D + 6 D x + 3 − D y )
2
x

1
= ex sin(x+y)
(3(-1) + 6D x + 3 - D y

 1  ( 6D x + D y )
= ex   sin (x+y) = ex sin(x+y)
 6D -D  36D 2x - D 2 y
 x y 
 7 cos ( x + y )
= ex
- 35

1 x
=- e cos(x+y).
5

Example 11.14 Solve the partial differential equation

 2u 2  2u
-c = e-xsin t
t 2 x 2

Solution: The equation can be written as

(D 2t -c2Dx2) u = e-xsin t

The particular solution is

1
up= 2
2 2
e − x sin t
D - c Dx
t

−x 1 1
= e 2
sin t = e − x sin t
D - (c(D x - 1)
t - 1- c 2

1
=- e − x sin t
c +1
2

By proceeding on the lines of the solution of Example 11.12 we get

uc =  (x-ct)+  (x+ct)

1
u(x,t)=  (x-ct)+  (x+ct) - e − x sin t
c +12

The solution uc is known as the d' Alembert's solution of the wave

equation

 2u 2  2u
-c =0.
t 2 x 2

11.5 Monge's Method for a special class of non linear Equations (quasi

linear Equations) of the Second order.

 Let u(x,y) be a function of two variables x and y

u u  2u  2u  2u
Let p = , q = ,r = , s= , t= 2
x y x xy y

Monge's method provides a technique for solving a special class of partial

differential equation of second order of the type

F(x,y,u,p,q,r,s,t)=0 (11.59)

Monge's method comprises in establishing one or two first integrals of the

form

= f() (11.60)

where  and  are known function of x,y,u, p and q and the function f is

arbitrary; that is, in finding relations of the type (11.59) such that equation (11.58)

can be derived from equation (11.59). The following equations are obtained from

it by partial differentiation.

x+up+pr+qs=f'() {x+up+pr+qs} (11.61)

y+uq+ps+qt=f'() {y+uq+ps+qt} (11.62)

It may be noted that every equation of the type (11.58) does not have a first

integral of the type (11.59). By eliminating f'() from equations (11.60) and (11.61),

we find that any second order partial differential equation which possesses a first

integral of the type (11.59) must be expressible in the form

R1r+S1s+T1t+U1(rt-s2)=V1 (11.63)

where R1, S1,T1,U1 and V1 are functions of x,y,u, p and q defined by the relations

(, ) (, ) (, ) (, )

R1 = +q , T1 = +p (11.64)
(p, y ) (p, u) ( x, q) (u, q)
 (, ) (, ) (, ) (, )
S1= +q − −p (11.65)
(q, y ) (q, u) (p, x ) (p, u)

(, ) (, ) (, ) (, )

U1= , V1 = q +p + (11.66)
(p, q) (u, x ) ( y, u) ( y, x )

The equation (11.62) reduces to the form

R1r+S1s+T1t=V1 (11.67)

if and only if the Jacobian pp- qp=0 identically. Equation (11.66) is a non-

linear equation because the coefficients R1, S1, T1, V1 are functions of p and q as

well as of x,y, and u. Infact it is a quasi linear equation. We explain here the method

of finding solution of the equation of the type (11.66), namely

Rr+Ss+Tt = V (11.68)

for which a first integral of the form (11.59) exists. For any function u of x

and y we have the relations dp =rdx+sdy, dq=sdx+tdy (11.69)

Eliminating r and t from this pair of equations and equation (11.67), we see

that any solution of (11.67) must satisfy the relation

Rdpdy+Tdqdx - Vdxdy=0 (11.70)

Rdy2 +Tdx2 –Sdxdy=0 (11.71)

The method of finding solutions of (11.69) and (11.70) is explained through

the following example:

Example: 11.15

2 2
 u   2 u  u  u   2 u  u   2 u
Solve the equation   - 2   +  =0
 y  x  x  y  xy  x  y
2 2

This equation is of the form (11.67) where

 2 2
 u   2u u u  2u  u   2u

R=    , r = , S = −2 , s = , T =   , and t = , V =0
 y  x x y xy  x  y 2
2

Therefore (11.69) and (11.70) become respectively

q2dpdy + p2dq dx=0 (11.72)

(pdx+qdy)2 = 0 (11.73)

By the equation du=pdx+qdy and (11.72) we get du=0, which gives integral

u=c1. From (11.71) and (11.72) we have qdp =pdq, which has solution

p=c2q. Thus, the first integral is

p=q f(u) (11.74)

where f(.) is arbitrary. We solve (11.73) by Lagrange's method. The auxiliary

system of equations (characteristic equations) are

dx dy du
= =
1 - f (u) 0

with integral u=c1, y+x f(c1)=c2 leading to the general solution

y+x f(u)=g(u)

where the functions f and g are arbitrary.

11.6 Exercises

Write down the order and degree of partial differential equations in problems

1-5.

u u
1. + = u2
x y

 2u u
2. =
x 2
t

3
 u  u
3.   + =0
 x  y

u u
4. + 100 =0
t x

2 3
 u   u 
5.   +   = 0
 x   y 

6. Verify that the functions u(x,y)=x2-y2 and u(x,y) = ex sin y are

solutions of the equation

 2u  2u
+ =0
x 2 y

7. Let u=f(x,y), where f is an arbitrary differentiable function. Show that

u satisfies the equation

x ux –y uy = 0

Examine whether cos (xy), exy and (xy)3 are solutions of this partial

differential equation.

Classify the partial differential equations as hyperbolic, parabolic, or elliptic.

8. 4 uxx-7 uxy + 3 uyy= 0

9. 4 uxx-8 uxy + 4 uyy= 0

 10. a2 uxx+2a uxy +uyy = 0, a0

 2u  2u  2u
11. 4 - 12 + 9 =0
t 2 xt x 2

 2u  2u  2u
12. 8 - 2 - 3 =0
x 2 xy y 2

For what values of x and y are the following partial differential equations

hyperbolic, parabolic, or elliptic?

13. uxx+2xuxy+(1-y2) uyy=0

14. (1+y2) uxx+(1+x2) uyy=0

15. uxx + x2 uyy = 0

16. uxx -2 sin x uxy – cos2x uy = 0

17. Find the general solution of 2 ux-3 uy = cos x

18. Solve ux+exuy=y, u(0,y) = 1+y

Find the complete solutions of the equations in problem 19-25

19. p=(u +qy)2

20. 2(u+xp+yq)=yp2

21. u2=pqxy

22. xp+3yq=2(u-x2q2)

23. pq=1

24. p2y(1+x2)=qx2

25. u=p2-q2

26. p2q2+x2y2=x2q2(x2-y2)

27. Discuss the method for finding a complete solution of the equation of

the type
 F(u,p,q)=0

Solve partial differential equations of problems 28 to 32.

 2u u
28. + 12 +2=0
x 2
x

 2u  2u  2u
29. 4 - 16 + 15 =0
x 2 y 2 y 2

 2u  2u u
30. 3 2 +4 - =0
x xy y

 2u u
31. 3 - = sin (ax+by)
x 2 y

 2 u u u  2u
32. 3 - 2 - 5 = 3x+y+ex-y
x 2 x y y 2

Solve equations in problems 33-36 using Monge's method

 2u  2u
33. =
x 2 y 2

 u   u   u  u u  u 
2 2
34.     - x  - 
2 
 x   y   x xy y x 

2
  2u   u 
2
u u  2u  2u  u   u   2u  u   2u
35.  2    - 2 -   =  - 
 x   y  x q xy y 2  x   x  y 2  y  xy
36.