ECE 759: Mathematics 1 Workshop 1

Due on September 30, 2015

Amodei 2:30

Contents

Gonzalo

Introduction																																											3
Finite Difference Method Spatial Discretization	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	3 4
. Temporal Discretization	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	4
															1

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

Stability																																																						4
Problem 1																																																						5
(1.) .	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	6
(2.) .	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	6
(3.) .	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	6
Problem 2																																																						7
(1.) .	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	7
(2.) .	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	7
(3.) .	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	.	7

Page 2 of 8

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

Introduction

The first workshops of the subject Mathematics 1 focuses on the numerical solution of the one dimensional heat equation. The form of the heat equation is the following:

∂u ∂t ^{= µ}

∂ ² u _{∂x 2} + f (x, t)

(1)

Where u is the variable to solve, µ is the diffusion coefficient, f represents a source term, t is the time variable and x is the spatial coordinate. For a 1D finite line of length L, the heat equation must be combined with the following boundary and initial conditions

u(0, t)	=	u g (x,	t)	(2)
u(L, t)	=	u d (x,	t)	(3)
u(x, 0) = φ(x)				(4)
				(5)

The heat equation is a well known equation in partial derivatives and is capable of modeling numerous physical phenomena such as: heat transfer in stationary continuous mediums or specific laminar flows under certain conditions. The aim of this workshops is to solver this one dimensional heat equation using the finite difference method using different schemes for the temporal discretization and using a centered approach for the diffusive terms.

Finite Difference Method

The finite difference approximations for derivatives are one of the simplest and of the oldest methods to solve differential equations. It was already known by L. Euler (1707-1783) ca. 1768, in one dimension of space and was probably extended to dimension two by C. Runge (1856-1927) ca. 1908. The advent of finite difference techniques in numerical applications began in the early 1950s and their development was stimulated by the emergence of computers that offered a convenient framework for dealing with complex problems of science and technology. Theoretical results have been obtained during the last five decades regarding the accuracy, stability and convergence of the finite difference method for partial differential equations. The principle of finite difference methods is close to the numerical schemes used to solve ordinary dif- ferential equations (cf. Appendix C). It consists in approximating the differential operator by replacing the derivatives in the equation using differential quotients. The domain is partitioned in space and in time and approximations of the solution are computed at the space or time points. The error between the numerical solution and the exact solution is determined by the error that is commited by going from a differential operator to a difference operator. This error is called the discretization error or truncation error. The term truncation error reflects the fact that a finite part of a Taylor series is used in the approximation. Generally one can write an EDP in the following form:

∂u _∂t = F (u, x, t)

(6)

Where F is an spatial operator. For the example, for the heat equation F is from the form:

∂ ² u F (u, x, t) = µ _{∂x 2} + f (x, t)

(7)

For a fixed x _j the general form of the EDP can be written as follow:

Page 3 of 8

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

du

_dt = F (u, x _j , t)

(8)

Spatial Discretization Temporal Discretization

The temporal derivative of the variable u is usually aproximated as:

∂u

u

n+1

j

− u

n

j

∂t ^∼

∆t

Introducing this approximation in the equation 8 yields:

u

n+1

j

− u

n

j

And thus we can solve the equation for u

∆t

:

n+1

j

= F (u, x _j , t)

u

n+1

j

− u

n

j

t

=

t

n

n+1

F (u, x _j , t)dt

(9)

(10)

(11)

The different time schemes, modify the way the integral of the operator F (defined in 6) is defined. For the Euler explicit schemes, the integral is evaluated using data from the previous timestep F (u, x _j , t) = F(u ⁿ , x _j , t _n )

(12)

without having to solve a linear system of equations, although

they have same stability limits. For the Euler implicit scheme, the integral is evaluated using the variable to be obtained F (u, x _j , t) = F(u ⁿ⁺¹ , x _j , t _n+1 )

u

n+1

j

− u

j

n+1

n

=

F(u ⁿ , x _j , t _n )∆t + o(∆t)

This explicit schemes allow to obtain u

j

(13)

The implicit schemes generally need of a bigger computational cost because they generate a linear system

. Usually, the implicit schemes present

u

n+1

j

− u

n

j

F(u ⁿ⁺¹ , x _j , t _n+1 )∆t + o(∆t)

n+1

=

of equation which must be solve in order to obtain the variable u

j

more stability than the explicit schemes

Stability

In	this section a study of the stability of the homogeneous one dimensional heat equation will be presented.
A	scheme is stable if roundoff errors are not amplified in the calculations. The Fourier method can be used
to	check if a scheme is stable, using the following approximation:

(14)

The Von Neumann stability condition states that the following condition for the amplification factor G(w) must be fulfilled in order to ensure the stability of a discretized EDP:

n

u j

_a ⁿ _e ^iwj∆x

₌

|G(w)| =

_a

n+1

a

ⁿ

<= 1

(15)

Page 4 of 8

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

Introducing the Fourier approach (Equation 14) into the discretized equation and operating the expression yields:

|G(w)| = 1 + 2

µ∆t

∆x ²

(cos(w∆x) − 1) <= 1

(16)

Which gives a relation between the three parameters of the problem (∆t, ∆x and µ) in order to maintain the stability of the solution:

µ∆t ∆x ^{2 <=}

1

₂

(17)

And thus we can reach the Courant-Friedrichs-Lewy (CFL) condition for the heat equation using Euler explicit scheme and central differential scheme:

2µ∆t

CFL =

<= 1

(18)

∆x ² In cause of using an Euler implicit scheme for the temporal integration, the amplification function is of the form:

|G(w)| =

		1
1 +	µ∆t	(2 − cos(w∆x))
	∆x 2

<= 1

(19)

µ∆t

and there is no CFL condition, the

∆x ² solution is stable for all the combinations of ∆t, ∆x and µ. Although, one must take into account that the

bigger the temporal and spatial increments are, the bigger the error in the solution.

This condition is fulfilled for every value of the parameter

Problem 1

Listing ?? shows a MATLAB script that was used to generate Figure ??.

Listing 1: MATLAB Function Used to Generate Figure ??

function homework_example

%HOMEWORK_EXAMPLE Example

MATLAB function for

LaTeX homework templates.

%

This

function produces

a

figure

by

using

a

number

of

MATLAB commands.

It

%

should help

generate

the

power

of

the \matlabscript macro for including

5

%

MATLAB functions and

scripts within code.

%

%

The

figure

generated

by

this function will

be used

to

demonstrate how

to

%

use

\scalefig.

10

%

Generate 100

random data

points

with a normal distribution

X = randn(1, 100);

%

Generate a histogram

in

N

at 100

centers specified

at

Xcenters

[N, Xcenters]

= hist (X,

length(X)/10);

15

%

Plot

the sample CDF of the data

clf ;

subplot(211);

Problem 1 continued on next

Page 5 of 8

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

Problem 1 (continued)

cdf_datafig

=

s t a ir s ( sort(X), linspace(0,1,length(X)) );

20

%

Plot

an

expected CDF

on

top

of

it

hold on;

xlims

=

xlim; linspace(xlims(1), xlims(2), 1000);

expX =

25

cdf_expfig =

plot( expX, normcdf(expX), ’r--’ );

%

Label

the

figure

legend( [cdf_datafig, cdf_expfig], ’Data Distribution’,

’Expected Distribution’,

4

);

30

xlabel(’x’); ylabel(’F(x) = P(X \leq x)’);

t

i t l e (’Normal

Cummulative

Distribution

Function’);

%%% Notice the use of

TeX

in

the

above

ylabel.

In MATLAB

7.0,

LaTeX

can

35

%%% also

be

used if the interpreter

on

the text

object

is

changed from

%%% ’tex"

to

’latex’.

%

Plot

the sample pdf

of

the

data

subplot(212);

40

pdf_datafig

=

bar( Xcenters, N/length(X)/(Xcenters(2)-Xcenters(1)) );

xlim( xlims );

%

Plot

an

expected pdf

on

top

of

it

hold on;

45

pdf_expfig =

plot( expX, normpdf(expX), ’r--’ );

%

Label

the

figure

legend( [pdf_datafig, pdf_expfig], ’Data Density’,

’Expected Density’,

1

);

50

xlabel(’x’);

ylabel(’dF(x)/dx’);

t

i t l e (’Normal

Probability

Density

Function’);

(1.)

Still no Latin.

(2.)

It can be demonstrated that the heat equation ( is of the form

u(x, t) = A ₀ +

n=∞

n=1 A _n cos(nπ/L)

+

B _n sin( ^nπ

L

_{) e} −µ(nπL) ² t

₍₂₀₎

And thus when t → ∞ the solution tends tou(x, t) = A ₀

(3.)

As seen in the section and the equation , the explicit scheme used in this problem needs a value of CF L <= 1 in order to ensure the stability of the solution. Using a value of CF L = 1.1 makes the solution blow up in a few number of iterations.

Page 6 of 8

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

Problem 1

E

Problem 2

You were expecting lorem ipsum here, weren’t you?

(1.)

Still no Latin.

(2.)

It can be demonstrated that the heat equation ( is of the form

u(x, t) = A ₀ +

n=∞

n=1

A _n cos(nπ/L) + B _n sin( ^nπ

L

_{) e} −µ(nπL) ² t

And thus when t → ∞ the solution tends to

(3.)

₍₂₁₎

In this section, we will investigate the influence of the CFL number in the time discretization er- ror, and also the influence of the mesh spacing ∆x in the spatial discretization error, comparing our numerical solution u with the analytical solution u _ex , using an Euler explicit temporal scheme combined with a second order central scheme for the diffusive terms. As previously stated we expect a temporal discretization error of order O(∆T ) and an spatial discretization error of order O(∆x ² ). We will define the error for a given mesh as a function of the CFL number for a given time with the following expression:

E(t, CF L) =

_i=0 max (u

i=t

n

j

−

u _ex (x _j , t ⁿ )) ²

κ

(22)

Where t is the time where the error is evaluated and κ is the number of iterations needed to reach up to time t with a given CFL. The figure represents the error previously defined as a function of the CFL number for a given time of t = 1, using logaritmic scale for both axes, as one might expect that the error is proportional to the CFL (or timestep) E = k ∗ CF L. As we can observe in the figure, the error obtained is practically proportional to the CFL, or in other words, the temporal discretization error is of order O(∆t). This result coincides with the theoretical value of the temporal discretization error.

10

10

10

Time Discretization Error

Data

−

4

Slope

O(

∆ t)

Slope

O(

∆ t 2 )

−

5

− 6

10 − 2

10 − 1

10 0

10 1

CFL

Problem 2 [(3.)] continued on next

Page 7 of 8

Gonzalo

ECE 759 (Amodei 2:30): Mathematics 1

Problem 2

E

We will define the error for a given mesh as a function of the mesh spacing ∆ x for a given value of CF L = 0.1 with the following expression:

E(t, ∆x) =

i=t

i=0 ^{max (u}

n

j

−

u _ex (x _j , t ⁿ )) ²

κ

(23)

Where t is the time where the error is evaluated and κ is the number of iterations needed to reach up to time t with a given CFL. The figure represents the error previously defined as a function of the mesh spacing ∆x number for a given time of t = 1. In this case, the error obtained is practically proportional to the square of the mesh spacing , or in other words, the spatial discretization error is of order O(∆x ² ). This result coincides with the theoretical value of the spatial discretization error.

Mesh Discretization Error

10

10

10

10

10

−

3

−

4

−

5

Data

−

6

Slope

O( ∆ x

)

Slope

O( ∆ x

2 )

− 7

10 − 3

10 − 2

10 − 1

∆

x

Page 8 of 8