School of Computing & Mathematical Sciences

Terence Etchells

Numerical methods for solving second order differential equations

of the form
y + p( x , y ) y + q ( x , y ) y = r ( x )

with initial values

We have so far developed methods for solving first order differential equations, namely
the Euler, Modified Euler and Runge Kutta order 4 (the last two being predictorcorrector methods).
In calculating the analytical solution of a second order differential equation, two
arbitrary constants appear due to integrating twice. These constants can be found
from the initial or boundary values that define the differential equation. If the
differential equation y + p( x , y ) y + q ( x , y ) y = r ( x ) has a solution y ( x ) then the
initial value values take the form y ( x 0 ) = y 0 and y ( x 0 ) = z 0 . So, we know the
starting co-ordinate of the solution (as we did in the first order case) and we also know
its initial gradient. This initial information is essential for solving second order
differential equations numerically.
We also mentioned boundary value problems, these differ from initial value problems
in that were we are not given a starting gradient. Instead we are given a second coordinate through which the solution passes. Boundary value problems need special
treatment, numerically, and we shall study some of these methods later.
Recasting
The first step in solving a second order initial value problem is to recast the second
order equation into two simultaneous first order differential equations. For example, if
the differential equation is defined as
y 3xy + sin x y = e x where y(0) = 0 and y (0) = 1
then we recast this equation by first setting
y = z which when differentiated with respect to x gives y = z
and replacing y and y in the differential equation with z and z , respectively.
y = z
z 3xz + sin x y = e x
finally we rearrange the equations such that y and z are written as functions of x, y,
and z.
y = z
z = 3xz sin x y + e x
Exercise
Recast these differential equation into sets of first order differential equations.
So we have

(a)

y sin( xy ) y = y e x

Liverpool John Moores University

(b)

xy (1 + x 2 ) y + y = e x

Page 1

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

Solving Simultaneous First Order ODEs

The methods of Euler and Ring Kutta for first order odes can be applied to
simultaneous first order differential equations. In the first instance we will use Eulers
method to demonstrate the technique.
We are given two first order differential equations
dy
= f ( x, y, z)
dx
dz
= g ( x , y , z ) , (notice 3 variables 2 equations)
dx
with initial values x 0 , y 0 , z 0 .
We will applying Eulers method to obtain estimates for
y1 and z1 ,
y1 = y0 + hf ( x0 , y0 , z0 )
z1 = z 0 + hg ( x 0 , y 0 , z 0 )
were h is the step size.
This of course leads us to the more general recurrence relations
yn+1 = yn + hf ( xn , yn , zn )
zn+1 = zn + hg ( xn , yn , zn )
DERIVE ACTIVITY 8A
(A)
Develop a function that will calculate the Euler method approximations to two
simultaneous first order ODEs, with: starting values x 0 , y 0 , z 0 ; step size h ; n
applications.
Author
#1:
f ( x , y , z ):=
#2:
g ( x , y , z ):=
#3:
[ x 0: =, y 0: =, z 0: = ]
#4:
[ x + h, y + hf ( x , y , z ), z + hg ( x , y , z )]
#5:
ITERATES (#4,[ x , y , z ],[ x 0, y 0, z0], n)
#6:
EULER _ SIM ( x , y , z , x 0, y 0, z 0, h, n): = #5
Remove #3 to #5 and Save as EULERSIM.MTH
(B)
Use the function EULER_SIM to approximate the solution of
dy
= 4 y + 2z
dx
dz
= 3y z
dx
with initial values [0,1,1] from x=0 to x=1 with step size 0.1

#4:
#5:
#6:

Author
f ( x , y , z ):= 4 y + 2 z
g ( x , y , z ):= 3 y z
EULER _ SIM ( x , y , z ,0,11
, ,0.110
, )
approX #6

to give
Liverpool John Moores University

Page 2

3/2/99

School of Computing & Mathematical Sciences

(C)

Terence Etchells

The second order differential equation y + 3 y + 2 y = e x into two simultaneous

ODEs. The ode has initial conditions y(0) = 0 and y (0) = 1, use Eulers
method to approximate the solution between x=0 and x=1, with step size 0.1.
Compare this approximate solution with the analytical solution, which can be
calculated using the function DSOLVE2_IV() found in the utility file
ODE2.MTH.
Recasting we get

#8:
#9:
#10:

y = z
z = e x 3z 2 y

Author
f ( x , y , z ):= z
g ( x , y , z ):= e x 3z 2 y
( Alt e for the exponential e)
EULER _ SIM ( x , y , z ,0,0,1,0110
., )
approX #10 to give

Liverpool John Moores University

Page 3

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

We can plot y against x by extracting the first two columns of this matrix, the
following function will do this for us.
#12: EXTRACT (m): = [m`sub1, m`sub2]`
#13: EXTRACT(#11)
approX #13 , Plot (Beside at 40), Centre , Zoom in (once: press F9).
we get

Transfer Load Utility ODE2.MTH

Liverpool John Moores University

Page 4

3/2/99

School of Computing & Mathematical Sciences

#15:

Terence Etchells

Author
DSOLVE 2 _ IV (3,2, e x , x ,0,0,1)
Simplify #15 and Plot #16

we get

Which shows that, as one would expect, the Euler method is peeling away from the
solution as x increases.
The Modified Euler Method
In the single first order case, the modified Euler method is defined as
h
y n +1 = y n + ( k 1 + k 2 ) where k 1 = f ( x , y ) and k 2 = f ( x + h, y + hk 1 ).
2
As in the Euler method, we have a third variable, z, which is also changing as x steps
along the x axis. So we adjust the modified Euler method to account for the third
variable thus:
Let f ( x, y , z ) = k 1 and g ( x , y, z ) = l1 (we use l as it is next to k in the alphabet)
and also let
k 2 = f ( x + h, y + hk 1 , z + hl1 ) and l2 = g ( x + h, y + hk1 , z + hl1 ).
We can now define the recurrence equations for y and z as,
h
h
y n +1 = y n + ( k 1 + k 2 ) and z n +1 = z n + ( l1 + l2 )
2
2
respectively.

Liverpool John Moores University

Page 5

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

Example 8A
Find an approximate value for y(11
. ) and z(11
. ) , where y ( x ) and z ( x ) satisfy the
simultaneous differential equations,
dy
dz
= x + yz and
= xz
dx
dx
with initial values x 0 = 1, y 0 = 0 and z 0 = 0 and a step size of h=0.1.
So k 1 = x + yz and l1 = x z . Using the initial values we have that
k1 = 1 + 0 * 0 = 1 and l1 = 1 0 = 1
So, as k 2 = f ( x + h, y + hk 1 , z + hl1 ) and l2 = g ( x + h, y + hk1 , z + hl1 ),
then k 2 = 11
. + (0 + 01
. * 1)(0 + 0.1* 1) = 111
. and l2 = 11
. (0 + 01
. * 1) = 1
Which leads to,
h
01
.
y (11
. ) = y1 = y 0 + ( k 1 + k 2 ) = 0 =
(1 + 11
. ) = 01055
.
and
2
2
h
01
.
z (11
. ) = z1 = z 0 + ( l1 + l2 ) = 0 + (1 + 1) = 01
..
2
2
Whilst not overcomplicated in its computation, one wouldnt wish to be doing this
1000 times!
DERIVE ACTIVITY 8B
(A)
Write a function that will automate the modified Euler method for two
dy
= f ( x , y , z ) and
simultaneous first order differential equations
dx
dz
= g ( x , y , z ) ,with initial values x 0 , y 0 , z 0 and step size h.
dx

#1:
#2:
#3:
#4:
#5:
#6:
#7:
#8:
#9:
#10:

#12:
#13:

Author
f ( x , y , z ):=
g ( x , y , z ):=
[ x 0: =, y 0: =, z 0: = ]
k 1: = f ( x , y, z )
l1: = g ( x , y , z )
k 2: = f ( x + h, y + hk 1, z + hl1)
l 2: = g ( x + h, y + hk 1, z + hl1)
h
y + ( k 1 + k 2)
2
h
z + ( l1 + l 2)
2
[ x + h,#8,#9]
Simplify #10
Author
ITERATES (#11,[ x , y , z ],[ x 0, y 0, z 0], n)
RK 2 _ SIM ( x , y , z , x 0, y 0, z 0, h, n): =#12
Remove #4 to #12
Transfer Save Derive as RK2_SIM.MTH

Liverpool John Moores University

Page 6

3/2/99

School of Computing & Mathematical Sciences

(B)

Terence Etchells

Use the RK2_SIM() function to find y(1) and z(1) for the differential equation
in example 8A with step size h=0.1.
#5:
#6:
#7:

f ( , y , z ):=
yz
x y
xz
RK SIM y z

Exercise 8A
(1)
approximate 2 first order simultaneous equations
using the Runge Kutta order 4 method.
Use the RK2_SIM() function for different step size h
y (0) = 1, y (0) = 0 to show that
equation + 3 + 2 =
the modified Euler method has global error order h 2 .

Using the RK function in the DERIVE utility file ODE_APPR.MTH

There is a utility file that is supplied with DERIVE called ODE_APPR.MTH that
contains a function that will perform Runge Kutta order 4 approximations for any
number of simultaneous first order differential equations. The function is of the form
RK (r , v , v 0 , h, n) where: r is a vector whose elements are the right hand side of each of
the differential equation (in the form y = r ( x , y , L ) ); v is a vector whose elements are
the variables; v 0 is a vector whose elements are the initial values for the variable; h is

Liverpool John Moores University

Page 7

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

the step size; n is the number of iterations. As ever, an example will describe best how
to use this function.
DERIVE ACTIVITY 8C
(A)
Use the RK4() function in ODE_APPR.MTH to solve the following system of
first order differential equations with the given initial values;
dw
dy
dz
= x w + y,
= wx + y + z ,
= wxyz
dx
dx
dx
where x 0 = 0, w0 = 1, y 0 = 2, z 0 = 3, h=0.1 and n=10.

#1:
#2:
#3:
#4:

(B)

So r is the vector [ x w + y , wx + y + z , wxyz] ; v is the vector [ x , w, y , z ]; v 0 is

the vector [0,1,2,3].
The order of elements in v and v 0 is very important! As the derivatives are
with respect to x, then x is the first element, as the first differential equation is
dw
= x w + y then w is the next element, etc. To get RK() to the desired
dx
produce the results:
Transfer Load Utility ODE_APPR.MTH
Author
[ x w + y , wx + y + z , wxyz]
[ x , w, y , z ]
[0,1,2,3]
RK(#1,#2,#3,0.110
, )
approX #4 to give

Use the RK() function in ODE_APPR.MTH to approximate the solution to the

second order differential equation y + 5 y + 6 y = e x sin x with initial values
y(0) = 0 and y (0) = 1, between x=0 and x=2.
First recast the second order differential equation into two first order
equations, we have

Liverpool John Moores University

Page 8

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

dy
=z
dx
dz
= e x sin x 5z 6 y
dx
Author
#6:
[ z , e x sin x 5z 6 y ]
[ x , y, z]
#7:
#8:
[0,0,1]
RK(#6, #7, #8, 01
. , 20)
#9:
approX #9
We now have a large matrix of points for x, y and y , if we wish to plot x
against y then we need to extract the first two columns.
When
ODE_APPR.MTH
is
loaded
a
function
called
EXTRACT_2_COLUMNS(m,C1,C2) is entered into DERIVEs function
memory, this extracts columns C1 and C2 from the matrix m. We can use this
to extract the first two columns of the matrix in expression 10 to produce a
plot of the solution.
Author
#11: Extract_2_Columns(#10,1,2)
approX #11
Plot, Beside, at 40, #12
This particular differential equation can, in fact, be solved analytically. We can
use the utility file ODE2.MTH which contains the function
DSOLVE 2 _ IV ( p, q , r , x , x 0, y 0, v 0) that solves second order differential
equations of the form y + p( x ) y + q ( x ) y = r ( x ) with initial values y ( x 0) = y 0
and y ( x 0) = v 0 (where possible). We shall use this function here!
Transfer Load Utility ODE2.MTH
#13: DSOLVE 2 _ IV (5, 6, e x sin x , x ,0,0,1)
Simplify #13
Plot #14 Scale x: 0.5 y: 0.05
The following screen dump demonstrates the DERIVE output

Liverpool John Moores University

Page 9

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

RK and DSOLVE2_IV in action

(line = analytic solution : dots = numeric solution)

The Shooting Method for Boundary Value Problems

In the previous examples the differential equations where initial value problems, so
called because an initial value for position and gradient are supplied to solve the
equations. However, it is quite common for a differential equation to be defined in
terms of Boundary Values, i.e. an initial value for the position and another value for
the position. In other words we are given two points through which the solution
passes instead of an initial position and gradient. Clearly this will cause problems for
solution numerically because the methods described previously are reliant on the
knowledge of an initial gradient.
An example of a boundary value differential equation is
y + 3 y + 2 y = 5cos x where y(0) = 0 and y(1) = 0 .
Of course this particular second order differential equation has an analytical solution,
hence the boundary value nature of the ODE does not cause a problem as the boundary
values can determine the two arbitrary constants that arise from the general solution.
However, lets imagine that we do not know the analytical solution and we wish to
solve the equation using say the Runge Kutta method order four.
The title Shooting Method is descriptive of the method, it is very much like shooting at
a target. We do not know the initial gradient at which to shoot at the target but we
know where the target is. So we make a guess (not very scientific but there it is!) and
Liverpool John Moores University

Page 10

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

see how close we get to the target, we then adjust the initial gradient until we hit the
target. There you have it, the Shooting Method as employed by the Artillery for
centuries. Lets see this method in action!
DERIVE ACTIVITY 8D
We will use the example y + 3 y + 2 y = 5cos x where y(0) = 0 and y(1) = 0 .
Recasting the second order differential equation into two simultaneous first order
differential equations we have:
y = z
z = 3z 2 y 5cos x
Transfer Load the Utility File ODE_APPR.MTH
The initial value for the position is (0,0)and we will take the first guess at the initial
value of the gradient to be 1, this will mean that we have an initial starting vector of
[0,0,1]. If we use a step size of h=0.1, we will need 10 iteration to reach (1,0)

#1:
#2:

Author
[ z ,3z 2 y 5 cos x ]
RK (#1,[ x , y , z ],[0,0,1],0.110
, )
approX #2

we get

Liverpool John Moores University

Page 11

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

From inspection, we can see that we have missed the mark but not by a great deal. A
picture would be helpful at this stage to see what happened. The function
EXTRACT_2_COLUMNS(m,C1,C2) was automatically loaded with the utility file
ODE_APPR.MTH. If we now employ this function on #3 we can the plot our attempt
at the solution.
Author
#4:
EXTRACT _ 2 _ COLUMNS (#3,1,2)
approX #4
Plot, Beside, at Column 40 #5 Scale x:0.25 y:0.25 Centre (0,0.5)
resulting in

Intuitively, the initial gradient needs to be much higher. Edit #3 by highlighting it,
Author F3 and adjust the starting vector to [0,0,3], using a starting gradient of 3
instead of 1. approX and extract the first two columns and plot (as before) to give

Liverpool John Moores University

Page 12

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

As we can see the initial gradient was still not high enough, so we repeat the whole
process with a starting vector [0,0,4].

Liverpool John Moores University

Page 13

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

As you can see we have now over shot the mark, so our starting gradient is in the
region [3,4]. Once we know an interval in which the initial gradient lies we can use
Linear Interpolation to find a more accurate starting gradient.
When we used 3 as a starting gradient we had -0.185722 as our value for y(1) and
when we used 4 as a starting gradient we had 0.0468179 as our value for y(1) . Using
the method of linear interpolation with approximate values

h
1
=
h 0.798710 . Which implies that a
0185772
.
0.185772 + 0.0468179
starting value for the gradient of 3.79871 should be very close to the mark!
we have

which is very close!!

Liverpool John Moores University

Page 14

3/2/99

School of Computing & Mathematical Sciences

Terence Etchells

Finally, we will solve this equation analytically (only because we can) a see what the
analytical solution looks like.

#n:

Transfer Load Utility File ODE2.MTH

DSOLVE 2 _ BV (3,2,5 cos x , x ,0,0,1,0)
Simplify #n and Plot #(n+1)

Of course, the luxury of such verification is not always available, but it demonstrates
the accuracy of the method.
Exercise
Solve the following boundary value second order differential equations.
1.
2.

y + xyy + 3xy = e 3 x when y(0) = 1 and y(1) = 1

y + e 5 xy y = 6 xy cos x when y(0) = 1 and y(3) = 0

Liverpool John Moores University

Page 15

3/2/99