These slides are a supplement to the book Numerical Methods with

Matlab: Implementations and Applications, by Gerald W. Recktenwald,

c 2000–2006, Prentice-Hall, Upper Saddle River, NJ. These slides are
Numerical Integration of Ordinary Differential Equations copyright  c 2000–2006 Gerald W. Recktenwald. The PDF version
of these slides may be downloaded or stored or printed only for
for Initial Value Problems noncommercial, educational use. The repackaging or sale of these
slides in any form, without written consent of the author, is prohibited.

The latest version of this PDF file, along with other supplemental material
Gerald Recktenwald for the book, can be found at www.prenhall.com/recktenwald or
Portland State University web.cecs.pdx.edu/~gerry/nmm/.
Department of Mechanical Engineering
[email protected] Version 0.92 August 22, 2006

page 1

Overview Application: Newton’s Law of Motion

Newton’s Law of Motion is

• Motivation: ODE’s arise as models of many applications
F = ma
• Euler’s method
 A low accuracy prototype for other methods Acceleration is the time derivative of velocity, so
 Development
 Implementation dv
=a
 Analysis dt
• Midpoint method and
dv F
• Heun’s method =
dt m
• Runge-Kutta method of order 4
• Matlab’s adaptive stepsize routines If F (t) and v(0) are known, we can (at least in principle) integrate the preceding
equation to find v(t)
• Systems of equations
• Higher order ODEs

NMM: Integration of ODEs page 2 NMM: Integration of ODEs page 3

 Application: Newton’s Law of Cooling Newton’s Law of Cooling

The cooling rate of an object immersed

in a flowing fluid is T∞ Apply an energy balance

Q = hA(Ts − T∞) dT
mc = −Q = −hA(Ts − T∞)
dt
m, c
where Q is the heat transfer rate, h is the
Assume material is highly conductive ⇒ Ts = T
heat transfer coefficient, A is the surface
area, Ts is the surface temperature, and dT
T∞ is the temperature of the fluid. mc = −hA(T − T∞)
dt
When the cooling rate is primarily
controlled by the convection from the or
dT hA
surface, the variation of the object’s =− (T − T∞)
dt mc
temperature with is described by an
ODE.

NMM: Integration of ODEs page 4 NMM: Integration of ODEs page 5

Example: Analytical Solution Numerical Integration of First Order ODEs (1)

The ODE
dy
= −y y(0) = y0 The generic form of a first order ODE is
dt
can be integrated directly: dy
= f (t, y); y(0) = y0
dy dt
= −dt
y where the right hand side f (t, y) is any single-valued function of t and y .
ln y = −t + C The approximate numerical solution is obtained at discrete values of t
ln y − ln C2 = −t
tj = t0 + jh
y
ln = −t
C2 where h is the “stepsize”
−t
y = C2 e
−t
y = y0 e

NMM: Integration of ODEs page 6 NMM: Integration of ODEs page 7

 Numerical Integration of ODEs (2) Nomenclature

Graphical Interpretation
f (t0,y0) = slope at (t0,y0)
numerical solution at t3 y(t) = exact solution
y(tj ) = exact solution evaluated at tj
yj = approximate solution at tj
f (tj , yj ) = approximate r.h.s. at tj
y0 exact solution y(t)

h h h

t0 t1 t2 t3

NMM: Integration of ODEs page 8 NMM: Integration of ODEs page 9

Euler’s Method (1) Euler’s Method (2)

Consider a Taylor series expansion in the neighborhood of t0 Given h = t1 − t0 and initial condition, y = y(t0), compute
˛ ˛
dy ˛˛ (t − t0)2 d2y ˛˛ y1 = y0 + h f (t0, y0)
y(t) = y(t0) + (t − t0) + ˛ + ...
dt ˛t0 2 dt2 ˛ y2 = y1 + h f (t1, y1)
t0
... ...
Retain only first derivative term and define
˛ yj+1 = yj + h f (tj , yj )
dy ˛˛
f (t0, y0) ≡
dt ˛t0 or
yj = yj−1 + h f (tj−1, yj−1)
to get
y(t) ≈ y(t0) + (t − t0)f (t0, y0)

NMM: Integration of ODEs page 10 NMM: Integration of ODEs page 11

 Example: Euler’s Method Reducing Stepsize Improves Accuracy (1)

Use Euler’s method to integrate

1
Use Euler’s method to integrate Exact
dy h = 0.2
= t − 2y y(0) = 1 0.9 h = 0.1
dt dy h = 0.05
= t − 2y; y(0) = 1 0.8
dt
The exact solution is
1h −2t
i
0.7
y= 2t − 1 + 5e for a sequence of smaller h (see
4 demoEuler). 0.6

y
For a given h, the largest error in
0.5
Euler Exact Error the numerical solution is the Global
j tj f (tj−1 , yj−1 ) yj = yj−1 + h f (tj−1 , yj−1 ) y(tj ) yj − y(tj ) Discretization Error or GDE. 0.4

0 0.0 NA (initial condition) 1.0000 1.0000 0 0.3

1 0.2 0 − (2)(1) = −2.000 1.0 + (0.2)(−2.0) = 0.6000 0.6879 −0.0879
2 0.4 0.2 − (2)(0.6) = −1.000 0.6 + (0.2)(−1.0) = 0.4000 0.5117 −0.1117 0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
3 0.6 0.4 − (2)(0.4) = −0.400 0.4 + (0.2)(−0.4) = 0.3200 0.4265 −0.1065 t

NMM: Integration of ODEs page 12 NMM: Integration of ODEs page 13

Reducing Stepsize Improves Accuracy (2) Implementation of Euler’s Method

function [t,y] = odeEuler(diffeq,tn,h,y0)

% odeEuler Euler’s method for integration of a single, first order ODE
Local error at any time step is Here are results for the sample problem %
% Synopsis: [t,y] = odeEuler(diffeq,tn,h,y0)
plotted on previous slide: %
ej = yj − y(tj ) % Input: diffeq = (string) name of the m-file that evaluates the right
dy/dt = t − 2y; y(0) = 1 % hand side of the ODE written in standard form
% tn = stopping value of the independent variable
where y(tj ) is the exact solution % h = stepsize for advancing the independent variable
evaluated at tj . h max(ej ) % y0 = initial condition for the dependent variable
%
0.200 0.1117 % Output: t = vector of independent variable values: t(j) = (j-1)*h
GDE = max(ej ), j = 1, . . . % y = vector of numerical solution values at the t(j)
0.100 0.0502
t = (0:h:tn)’; % Column vector of elements with spacing h
For Euler’s method, GDE decreases 0.050 0.0240 n = length(t); % Number of elements in the t vector
linearly with h. y = y0*ones(n,1); % Preallocate y for speed
0.025 0.0117
% Begin Euler scheme; j=1 for initial condition
for j=2:n
y(j) = y(j-1) + h*feval(diffeq,t(j-1),y(j-1));
end

NMM: Integration of ODEs page 14 NMM: Integration of ODEs page 15

 Analysis of Euler’s Method (1) Analysis of Euler’s Method (2)

Rewrite the discrete form of Euler’s method as Introduce a family of functions zj (t), which are the exact solutions to the ODE given the
approximiate solution produced by Euler’s method at step j .
yj − yj−1
= f (tj−1, yj−1) (discrete)
h dzj
= f (t, y); zj (tj−1) = yj−1
Compare with original ODE dt
dy
= f (t, y) (continuous) Due to truncation error, Euler’s method produces a value of yj+1 that is different from
dt zj (tj+1) even though by design, yj = z(tj ).
Substitute the exact solution into the discrete approximation to the ODE to get
In other words
y(tj ) − y(tj−1)
− f (tj−1, y(tj−1)) = 0 yj+1 − z(tj+1) = 0
h
because yj+1 contains truncation error.

NMM: Integration of ODEs page 16 NMM: Integration of ODEs page 17

Analysis of Euler’s Method (3) Analysis of Euler’s Method (4)

1
The local discretization error (LDE) is the residual obtained when the exact zj (t) is
Example: Solve 0.9

0.8
substituted into the discrete approximation to the ODE
dy
= t − 2y; y(0) = 1 0.7
y(t) z(tj ) − z(tj−1)
dt 0.6 exact τ (t, h) = − f (tj−1, zj (tj−1))
h
y and z

0.5
The plot shows the numerical solution
0.4
(•) obtained with h = 0.5. The z(t) Note:
0.3
curve starting at each of the numerical z 3 (t) • User chooses h, and this affects LDE
0.2
solution points is shown as a solid line.
0.1
z 1 (t)
z 2 (t) • LDE also depends on t, the position in the interval
0
0 0.5 1 1.5 2
t

NMM: Integration of ODEs page 18 NMM: Integration of ODEs page 19

 Analysis of Euler’s Method (5) Analysis of Euler’s Method (6)

Using a Taylor series expansion for zj (x) we find that Assume that zj(ξ) is bounded by M in the interval t0 ≤ t ≤ tN . Then

z(tj ) − z(tj−1) h  hM
− f (tj−1, zj (tj−1)) = zj (ξ) τ (t, h) ≤ LDE for Euler’s method
h 2 2

where tj−1 ≤ ξ ≤ tj and zj ≡ d2zj /dt2. Although M is unknown we can still compute the effect of reducing the stepsize by taking
the ratio of τ (t, h) for two different choices of h
Thus, for Euler’s method the local discretization error is

h  τ (t, h2) h2
τ (t, h) = z (ξ) =
2 j τ (t, h1) h1

Since ξ is not known, the value of zj(ξ) cannot be computed.

NMM: Integration of ODEs page 20 NMM: Integration of ODEs page 21

Global Discretization Error for Euler’s Method Summary of Euler’s Method

Development of Euler’s method has demonstrated the following general ideas

General application of Euler’s method requires several steps to compute the solution to • The numerical integration scheme is derived from a truncated Taylor series
the ODE in an interval, t0 ≤ t ≤ tN . The local truncation error at each step approximation of the ODE.
accumulates. The result is the global discretization error (GDE) • The local discretization error (LDE) accounts for the error at each time step.
The GDE for Euler’s method is O(h). Thus LDE = O(h )
p

GDE(h1) h1 where h is the stepsize and p is an integer p ≥ 1.

= • The global discretization error (GDE) includes the accumulated effect of the LDE when
GDE(h2) h2
the ODE integration scheme is applied to an interval using several steps of size h.
p
GDE = O(h )

• The implementation separates the logic of the ODE integration scheme from the
evaluation of the right hand side, f (t, y). A general purpose ODE solver requires the
user to supply a small m-file for evaluating f (t, y).

NMM: Integration of ODEs page 22 NMM: Integration of ODEs page 23

 Higher Order Methods Midpoint Method (1)

We now commence a survey of one-step methods that are more accurate than Euler’s
Increase accuracy by evaluating slope twice in each step of size h
method.
• Not all methods are represented here k1 = f (tj , yj )
• Objective is a logical progression leading to RK-4
Compute a tentative value of y at the midpoint
• Sequence is in order of increasing accuracy and increasing computational efficiency
h
yj+1/2 = yj + f (tj , yj )
Methods with increasing accuracy, lower GDE 2
Method GDE re-evaluate the slope
h h
Euler O(h) k2 = f (tj +
, y j + k1 )
2
2 2
Midpoint O(h ) Compute final value of y at the end of the full interval
Heun O(h2)
yj+1 = yj + hk2
RK-4 O(h4)

Note that since h < 1, a GDE of O(h4) is much smaller than a GDE of O(h). LDE = GDE = O(h2)

NMM: Integration of ODEs page 24 NMM: Integration of ODEs page 25

function [t,y] = odeMidpt(diffeq,tn,h,y0)

Midpoint Method (2) % odeMidpt Midpoint method for integration of a single, first order ODE
%
% Synopsis: [t,y] = odeMidpt(diffeq,tn,h,y0)
%
% Input: diffeq = (string) name of the m-file that evaluates the right
yj from Euler’s method
% hand side of the ODE written in standard form
% tn = stopping value of the independent variable
% h = stepsize for advancing the independent variable
estimate of slope % y0 = initial condition for the dependent variable
at tj–1 + 0.5h true solution for the given yj–1 %
yj–1 + 0.5 h k 1 % Output: t = vector of independent variable values: t(j) = (j-1)*h
% y = vector of numerical solution values at the t(j)

t = (0:h:tn)’; % Column vector of elements with spacing h

yj–1 yj from midpoint method n = length(t); % Number of elements in the t vector
0.5h 0.5h y = y0*ones(n,1); % Preallocate y for speed
h2 = h/2; % Avoid repeated evaluation of this constant

tj–1 tj % Begin Midpoint scheme; j=1 for initial condition

for j=2:n
k1 = feval(diffeq,t(j-1),y(j-1));
k2 = feval(diffeq,t(j-1)+h2,y(j-1)+h2*k1);
y(j) = y(j-1) + h*k2;
end

NMM: Integration of ODEs page 26 NMM: Integration of ODEs page 27

 Midpoint Method (3) Comparison of Midpoint Method with Euler’s Method

Solve
dy
= −y; y(0) = 1; 0≤t≤1
Midpoint method requires twice as much work per time step. Does the extra effort pay dt
off? The exact solution is y = e−t.
Consider integration with Euler’s method and h = 0.1. Formal accuracy is O(0.1). >> compEM

h nrhsE errE nrhsM errM

Repeat calculations with the midpoint method and h = 0.1. Formal accuracy is 0.20000 6 4.02e-02 12 2.86e-03
O(0.01). 0.10000 11 1.92e-02 22 6.62e-04
0.05000 21 9.39e-03 42 1.59e-04
For Euler’s method to obtain the same accuracy, the stepsize would have to be reduced by 0.02500 41 4.65e-03 82 3.90e-05
0.01250 81 2.31e-03 162 9.67e-06
a factor of 10. The midpoint method, therefore, achieves the same (formal) accuracy with 0.00625 161 1.15e-03 322 2.41e-06
one fifth the work!
For comparable accuracy:
➣ Midpoint method with h = 0.2 evaluates the right hand side of the ODE 12 times,
and gives max error of 2.86 × 10−3
➣ Euler’s method with h = 0.0125 evaluates the right hand side of the ODE 81 times,
and gives max error of 2.31 × 10−3

NMM: Integration of ODEs page 28 NMM: Integration of ODEs page 29

Heun’s Method (1) Heun’s Method (2)

Compute the slope at the starting point

yj from Euler’s method
k1 = f (tj , yj )

Compute a tentative value of y at the endpoint estimate of slope at tj yj from Heun’s method
∗
yj = yj + hf (tj , yj )

re-evaluate the slope

true solution for the given yj–1
∗ yj–1
k2 = f (tj + h, yj ) = f (tj + j, yj + hk1) h

Compute final value of y with an average of the two slopes tj–1 tj

k1 + k2
yj+1 = yj + h
2

LDE = GDE = O(h2)

NMM: Integration of ODEs page 30 NMM: Integration of ODEs page 31

 Summary So Far Runge-Kutta Methods

Generalize the idea embodied in Heun’s method. Use a weighted average of the slope
➣ Euler’s method evaluates slope at beginning of the step
evaluated at multiple in the step
➣ Midpoint method evaluates slope at beginning and at midpoint of the step
X
➣ Heun’s method evaluates slope at beginning and at end of step yj+1 = yj + h γmkm

Can we continue to get more accurate schemes by evaluating the slope at more points in where γm are weighting coefficients and km are slopes evaluated at points in the interval
the interval? Yes, but there is a limit beyond which additional evaluations of the slope tj ≤ t ≤ tj+1
increase in cost (increased flops) faster than the improve the accuracy.
In general, X
γm = 1

NMM: Integration of ODEs page 32 NMM: Integration of ODEs page 33

Fourth Order Runge-Kutta Fourth Order Runge-Kutta

Compute slope at four places within each step

2
4
k1 = f (tj , yj )
h h 1 3
k2 = f (tj + , y j + k1 ) true solution for the given yj –1
2 2 yj –1
h h
k3 = f (tj + , yj + k2)
2 2
0.5h 0.5h
k4 = f (tj + h, yj + hk3)
tj –1 tj
Use weighted average of slopes to obtain yj+1
„ «
k1 k2 k3 k4
yj+1 = yj + h + + +
6 3 3 6

LDE = GDE = O(h4)

NMM: Integration of ODEs page 34 NMM: Integration of ODEs page 35

 function [t,y] = odeRK4(diffeq,tn,h,y0)
% odeRK4 Fourth order Runge-Kutta method for a single, first order ODE
% Comparison of Euler, Midpoint and RK4 (1)
% Synopsis: [t,y] = odeRK4(fun,tn,h,y0)
%
% Input: diffeq = (string) name of the m-file that evaluates the right
% hand side of the ODE written in standard form
% tn = stopping value of the independent variable Solve
% h = stepsize for advancing the independent variable dy
% y0 = initial condition for the dependent variable = −y; y(0) = 1; 0≤t≤1
% dt
% Output: t = vector of independent variable values: t(j) = (j-1)*h
% y = vector of numerical solution values at the t(j)

t = (0:h:tn)’; % Column vector of elements with spacing h

>> compEMRK4
n = length(t); % Number of elements in the t vector
y = y0*ones(n,1); % Preallocate y for speed
h nrhsE errE nrhsM errM nrhsRK4 err4
h2 = h/2; h3 = h/3; h6 = h/6; % Avoid repeated evaluation of constants
0.20000 6 4.02e-02 12 2.86e-03 24 5.80e-06
0.10000 11 1.92e-02 22 6.62e-04 44 3.33e-07
% Begin RK4 integration; j=1 for initial condition
0.05000 21 9.39e-03 42 1.59e-04 84 2.00e-08
for j=2:n
0.02500 41 4.65e-03 82 3.90e-05 164 1.22e-09
k1 = feval(diffeq, t(j-1), y(j-1) );
0.01250 81 2.31e-03 162 9.67e-06 324 7.56e-11
k2 = feval(diffeq, t(j-1)+h2, y(j-1)+h2*k1 );
0.00625 161 1.15e-03 322 2.41e-06 644 4.70e-12
k3 = feval(diffeq, t(j-1)+h2, y(j-1)+h2*k2 );
k4 = feval(diffeq, t(j-1)+h, y(j-1)+h*k3 );
y(j) = y(j-1) + h6*(k1+k4) + h3*(k2+k3);
end

NMM: Integration of ODEs page 36 NMM: Integration of ODEs page 37

Comparison of Euler, Midpoint and RK4 (2) Summary: Accuracy of ODE Integration Schemes

RHS
Error step size evaluations
• GDE decreases as h decreases
Euler 1.2 × 10−3 0.00625 161
−3
• Need an upper limit on h to achieve a desired accuracy
Midpoint 2.9 × 10 0.2 12
• Example: Euler’s Method

Midpoint 2.4 × 10−6 0.00625 322 yj = yj−1 + h f (tj−1, yj−1)

−6
RK-4 5.8 × 10 0.2 24
when h and |f (tj , yj )| are large, the change in y is large
Conclusion: • The product, h f (tj , yj ), determines accuracy
➣ RK-4 is much more accurate (smaller GDE) than Midpoint or Euler
➣ Although RK-4 takes more flops per step, it can achieve comparable accuracy with
much larger time steps. The net effect is that RK-4 is more accurate and more
efficient

NMM: Integration of ODEs page 38 NMM: Integration of ODEs page 39

 Procedure for Using Algorithms Having Fixed Stepsize Adaptive Stepsize Algorithms

Let the solution algorithm determine h at each time step

• Develop an m-file to evaluate the right hand side
• Use a high order method, e.g. RK-4 • Set a tolerance on the error
• Compare solutions for a sequence of smaller h • When |f (tj−1, yj−1)| is decreases, increase h to increase efficiency and decrease
• When the change in the solution between successively smaller h is “small enough”, round-off
accept that as the h-independent solution. • When |f (tj−1, yj−1)| is increases, decrease h to maintain accuracy

The goal is to obtain a solution that does not depend (in a significant way) on h.

NMM: Integration of ODEs page 40 NMM: Integration of ODEs page 41

Adaptive Stepsize Algorithms (2) Embedded Runge-Kutta Methods (1)

There is a pair of RK methods that use the same six k values

How do we find the “error” at each step in order to judge whether the stepsize needs to
be reduced or increased? Fourth Order RK:

Two related strategies: yj+1 = yj + c1k1 + c2k2 + c3k3

4
➣ Use two h values at each step: + c4k4 + c5k5 + c6k6 + O(h )
1. Advance the solution with h = h1
2. Advance the solution with two steps of size h2 = h/2 Fifth Order RK:
3. If solutions are close enough, accept the h1 solution, stop ∗ ∗ ∗ ∗
yj+1 = yj + c1 k1 + c2 k2 + c3 k3
4. Otherwise, replace h1 = h2, go back to step 2
∗ ∗ ∗ 5
+ c4 k4 + c5 k5 + c6 k6 + O(h )
➣ Use embedded Runge-Kutta methods
Therefore, at each step an estimate of the truncation error is
∗
Δ = yj+1 − yj+1

NMM: Integration of ODEs page 42 NMM: Integration of ODEs page 43

 Embedded Runge-Kutta Methods (2) Matlab ode45 Function

Possible outcomes
• User supplies error tolerance, not stepsize
• If Δ is smaller than tolerance, accept the yj+1 solution. • Simultaneously compute 4th and 5th order Runge-Kutta solution
• If Δ is much smaller than tolerance, accept the yj+1 solution, and try increasing the • Compare two solutions to determine accuracy
stepsize. • Adjust step-size so that error tolerance is maintained
• If Δ is larger than tolerance, reduce h and try again.

NMM: Integration of ODEs page 44 NMM: Integration of ODEs page 45

Matlab ode45 Function (2) Using ode45

Error tolerances are Syntax:

τ̃ < max(RelTol × |yj |, AbsTol)
[t,Y] = ode45(diffeq,tn,y0)
where τ̃ is an estimate of the local truncation error, and RelTol and AbsTol are the error [t,Y] = ode45(diffeq,[t0 tn],y0)
tolerances, which have the default values of [t,Y] = ode45(diffeq,[t0 tn],y0,options)
[t,Y] = ode45(diffeq,[t0 tn],y0,options,arg1,arg2,...)
−3 −6
RelTol = 1 × 10 AbsTol = 1 × 10
User writes the diffeq m-file to evaluate the right hand side of the ODE.
Solution is controlled with the odeset function
options = odeset(’parameterName’,value,...)
[y,t] = ode45(’rhsFun’,[t0 tN],y0,options,...)

Syntax for ode23 and other solvers is the same.

NMM: Integration of ODEs page 46 NMM: Integration of ODEs page 47

 Matlab’s Built-in ODE Routines Interpolation Refinement by ode45 (1)

Function Description
The ode45 function attempts to obtain the solution to within the user-specified error
ode113 Variable order solution to nonstiff systems of ODEs. ode113 uses an
tolerances. In some situations the solution can be obtained within the tolerance by taking
explicit predictor-corrector method with variable order from 1 to 13.
so few time steps that the solution appears to be unsmooth. To compensate for this,
ode15s Variable order, multistep method for solution to stiff systems of ode45 automatically interpolates the solution between points that are obtained from the
ODEs. ode15s uses an implicit multistep method with variable solver.
order from 1 to 5.
Consider
ode23 Lower order adaptive stepsize routine for non-stiff systems of ODEs. dy
= cos(t), y(0) = 0
ode23 uses Runge-Kutta schemes of order 2 and 3. dt
The following statements obtain the solution with the default parameters.
ode23s Lower order adaptive stepsize routine for moderately stiff systems of
ODEs. ode23 uses Runge-Kutta schemes of order 2 and 3. >> rhs = inline(’cos(t)’,’t’,’y’);
>> [t,Y] = ode45(rhs,[0 2*pi],0);
ode45 Higher order adaptive stepsize routine for non-stiff systems of ODEs. >> plot(t,Y,’o’)
ode45 uses Runge-Kutta schemes of order 4 and 5.
(See plot on next slide)

NMM: Integration of ODEs page 48 NMM: Integration of ODEs page 49

Interpolation Refinement by ode45 (2) Interpolation Refinement by ode45 (3)

1 1
>> rhs = inline(’cos(t)’,’t’,’y’); Repeat without interpolation:
0.8 0.8
>> [t,Y] = ode45(rhs,[0 2*pi],0); >> options = odeset(’Refine’,1)
>> plot(t,Y,’o’) 0.6 >> [t2,Y2] = ode45(rhs,[0 2*pi],0,options);
0.6

0.4
>> hold on 0.4
>> plot(t2,Y2,’rs-’)
0.2 0.2

The two solutions are identical at

0 0
the points obtained from the Runge-
-0.2 -0.2
Kutta algorithm
-0.4 -0.4
>> max(Y2 - Y(1:4:end))
-0.6 ans = -0.6

0
-0.8 -0.8

-1 -1
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

NMM: Integration of ODEs page 50 NMM: Integration of ODEs page 51

 Coupled ODEs (1) Coupled ODEs (2)

Apply the 4th Order Runge-Kutta Scheme:

` ´
k1,1 = f1 tj , yj,1 , yj,2
Consider ` ´
k1,2 = f2 tj , yj,1 , yj,2
dy1 „ «
= f1(t, y1, y2) h h h
dt k2,1 = f1 tj + , yj,1 + k1,1 , yj,2 + k1,2
2 2 2
dy2 „ «
= f2(t, y1, y2) h h h
dt k2,2 = f2 tj + , yj,1 + k1,1 , yj,2 + k1,2
2 2 2
„ «
These equations must be advanced simultaneously. h h h
k3,1 = f1 tj + , yj,1 + k2,1 , yj,2 + k2,2
2 2 2
„ «
h h h
k3,2 = f2 tj + , yj,1 + k2,1 , yj,2 + k2,2
2 2 2
` ´
k4,1 = f1 tj + h, yj,1 + hk3,1 , yj,2 + hk3,2
` ´
k4,2 = f2 tj + h, yj,1 + hk3,1 , yj,2 + hk3,2

NMM: Integration of ODEs page 52 NMM: Integration of ODEs page 53

Coupled ODEs (3) Example: Predator-Prey Equations

Update y1 and y2 only after all slopes are computed

„ « dp1
k1,1 k2,1 k3,1 k4,1
yj+1,1 = yj,1 + h + + + = α1p1 − δ1p1p2 (prey)
6 3 3 6 dt
„ « dp2
k1,2 k2,2 k3,2 k4,2 = α2p1p2 − δ2p2 (preditor)
yj+1,2 = yj,2 + h + + + dt
6 3 3 6

NMM: Integration of ODEs page 54 NMM: Integration of ODEs page 55

 Evaluate RHS of Preditor-Prey Model Preditor-Prey Results
alpha(1) = 2.000000 delta(1) = 0.020000
5500

5000

Prey population
function dpdt = rhsPop2(t,p,flag,alpha,delta)
4500
% rhsPop2 Right hand sides of coupled ODEs for 2 species predator-prey system
%
4000
% Synopis: dpdt = rhsPop2(t,p,flag,alpha,delta)
% 3500
% Input: t = time. Not used in this m-file, but needed by ode45
% p = vector (length 2) of populations of species 1 and 2 3000
0 5 10 15 20 25 30
% flag = (not used) placeholder for compatibility with ode45
% alpha = vector (length 2) of growth coefficients
alpha(2) = 0.000200 delta(2) = 0.800000
% delta = vector (length 2) of mortality coefficients 120
%
115
% Output: dpdt = vector of dp/dt values

Predator population
dpdt = [ alpha(1)*p(1) - delta(1)*p(1)*p(2); 110
alpha(2)*p(1)*p(2) - delta(2)*p(2); ]; 105

100

95

90

85
0 5 10 15 20 25 30
time (arbitrary units)

NMM: Integration of ODEs page 56 NMM: Integration of ODEs page 57

Example: Second Order Mechanical System Second Order Mechanical System

F(t)
Governing equation is a second order ODE
m
2 F
x ẍ + 2ζωnẋ + ωnx =
k c m

c
X ζ≡ √
2 km
F = ma
q
Forces acting on the mass are
ωn ≡ k/m
Fspring = −kx
ζ and ωn are the only (dimensionless) parameters
Fdamper = −cẋ
F (t) − kx − cẋ = mẍ

NMM: Integration of ODEs page 58 NMM: Integration of ODEs page 59

 Equivalent Coupled First Order Systems Solve Second Order System with ODE45
function demoSmd(zeta,omegan,tstop)
% demoSmd Second order system of ODEs for a spring-mass-damper system
%
Define % Synopsis: smdsys(zeta,omegan,tstop)
y1 ≡ x y2 ≡ ẋ %
% Input: zeta = (optional) damping ratio; Default: zeta = 0.1
then % omegan = (optional) natural frequency; Default: omegan = 35
% tstop = (optional) stopping time; Default: tstop = 1.5
dy1 %
= ẋ = y2 % Output: plot of displacement and velocity versus time
dt
dy2 if nargin<1, zeta = 0.1; end
= ẍ if nargin<2, omegan = 35; end
dt if nargin<3, tstop = 1.5; end
F 2
= − 2ζωnẋ − ωnx y0 = [0; 0]; a0 = 9.8; % Initial conditions and one g force/mass
m [t,y] = ode45(’rhssmd’,tstop,y0,[],zeta,omegan,a0);

F 2 subplot(2,1,1);
= − 2ζωny2 − ωny1 plot(t,y(:,1)); ylabel(’Displacement’); grid;
m title(sprintf(’zeta = %5.3f omegan = %5.1f’,zeta,omegan));
subplot(2,1,2);
plot(t,y(:,2)); xlabel(’Time (s)’); ylabel(’Velocity’); grid;

NMM: Integration of ODEs page 60 NMM: Integration of ODEs page 61

Solve Second Order System with ODE45 Response of Second Order System to a Step Input
zeta = 0.100 omegan = 35.0
0.015
function dydt = rhsSmd(t,y,flag,zeta,omegan,a0)

Displacement
% rhsSmd Right-hand sides of coupled ODEs for a spring-mass-damper system
0.01
%
% Synopis: dydt = rhsSmd(t,y,flag,zeta,omegan,a0)
% 0.005
% Input: t = time, the independent variable
% y = vector (length 2) of dependent variables
% y(1) = displacement and y(2) = velocity 0
% flag = dummy argument for compatibility with ode45 0 0.5 1 1.5
% zeta = damping ratio (dimensionless)
% omegan = natural frequency (rad/s) 0.6
% a0 = input force per unit mass
% 0.4
% Output: dydt = column vector of dy(i)/dt values

Velocity
0.2
if t<=0, fonm = 0.0;
else, fonm = a0; % Force/mass (acceleration)
end 0

dydt = [ y(2); fonm - 2*zeta*omegan*y(2) - omegan*omegan*y(1)]; -0.2

0 0.5 1 1.5
time (seconds)

NMM: Integration of ODEs page 62 NMM: Integration of ODEs page 63

 General Procedure for Higher Order ODEs

Given The transformation is

dnu
= f (t, u) Define yi ODE for yi
dtn
dy1
y1 = u = y2
dt
du dy2
y2 = = y3
dt dt
d2u dy3
y3 = = y4
dt2 dt
... ...

dn−1u dyn
yn = = f (t, u)
dtn−1 dt

NMM: Integration of ODEs page 64