Lenguajes de Programación

Diplomatura de Especialización en Dinámica de Fluidos Computacional (CFD)

APPLICATIONS - MATLAB

MSc. Sebastian Valencia

MSc. Luis Angeles
 Solution of Equation Systems

• The solution of a system of equations is necessary in most engineering problems. Examples:

• interpolation and curve fitting problems;
• solution of differential equations – simulation of engineering problems.

• Most of the time of a simulation (using finite differences, finite volumes, finite elements or another
numerical method) is invested in the solution of the system of equations obtained with the
discretization.

• Discretization is a mathematical process through which approximate results of the differential equation
of the problem are obtained.

• There is a need for robust and fast methods.

Sebastian Valencia – Luis Angeles 2

 Solution of Equation Systems

• System of n equations and n unknows.

𝑎11 𝑥1 + 𝑎12 𝑥2 + ⋯ + 𝑎1𝑛 𝑥𝑛 = 𝑏1

𝑎21 𝑥1 + 𝑎22 𝑥2 + ⋯ + 𝑎2𝑛 𝑥𝑛 = 𝑏2
⋮
𝑎𝑛1 𝑥1 + 𝑎𝑛2 𝑥2 + ⋯ + 𝑎𝑛𝑛 𝑥𝑛 = 𝑏𝑛

𝑎11 𝑎12 ⋯ 𝑎1𝑛−1 𝑎1𝑛 𝑥1 𝑏1

𝑎11 𝑎12 ⋯ 𝑎1𝑛−1 𝑎1𝑛 𝑥2 𝑏2
𝑎11 𝑎12 ⋯ 𝑎1𝑛−1 𝑎1𝑛 𝑥3 = 𝑏3 𝑨𝒙 = 𝒃
⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮
𝑎11 𝑎12 ⋯ 𝑎1𝑛−1 𝑎1𝑛 𝑥𝑛 𝑏𝑛

Sebastian Valencia – Luis Angeles 3

 Solution of Equation Systems

• Uniqueness and existence of solutions:

• The rank of a matrix (𝑟𝑎𝑛𝑘 𝐴 ) is equal to the number of linearly independent rows or columns.

• if 𝑟𝑎𝑛𝑘 𝐴 = 𝑟𝑎𝑛𝑘 𝐴|𝑏 = 𝑛 (matrix dimension), then the solution of the system 𝑨𝒙 = 𝒃 is unique.
• if 𝑟𝑎𝑛𝑘 𝐴 = 𝑟𝑎𝑛𝑘 𝐴|𝑏 = 𝑟 < 𝑛, then there are infinite solutions.
• if 𝑟𝑎𝑛𝑘 𝐴 ≠ 𝑟𝑎𝑛𝑘 𝐴|𝑏 , then there is solution,

In MATLAB:

rank(A) gives the rank of the matrix A

Sebastian Valencia – Luis Angeles 4

 Solution of Equation Systems

• Example
>> rank(A)

>> A = [1 5 7; 4 2 -1; 10 ans =

4 7]
3
A= 𝑟𝑎𝑛𝑘 𝐴 = 𝑟𝑎𝑛𝑘 𝐴|𝑏 = 3, then the
>> rank([A,b]) solution of the system 𝑨𝒙 = 𝒃 is
1 5 7
4 2 -1 ans =
unique.
10 4 7
3
>> b = [3;5;11]
>> [A,b]
b=
ans =
3
5 1 5 7 3
11 4 2 -1 5
Sebastian Valencia – Luis Angeles 5
10 4 7 11
 Solution of Equation Systems

• Direct Methods
• The exact solution (if there are no truncation errors* of the
computer) is determined from a finite number of operations. Gaussian elimination
LU factorization
• They require a large storage capacity;
Cholesky
• They are robust;
Band Matrix
• They are fast.

• Iterative Methods
• It provides a sequence of approximate solutions that converge Jacobi
as the number of iterations approaches infinity. Gauss Seidel
• They require less storage capacity than direct methods; SOR
• There may be convergence problems.

Sebastian Valencia – Luis Angeles 6

 Solution of Equation Systems

• Direct Methods – LU factorization

• Decomposing A matrix (no singular, 𝐴 ≠ 0, 𝑟𝑎𝑛𝑘 = 𝑛)

𝑳𝑼𝒙 = 𝒃
Progressive
𝑳𝒚 = 𝒃 lower triangular system
substitution for L
𝑼𝒙 = 𝒚 upper triangular system matrix

Sebastian Valencia – Luis Angeles 7

 Solution of Equation Systems
Forward substitution algorithm
function x = forward_sub(L, b)
% x = forward_sub(L, b) is the solution
to L x = b
% L must be a lower-triangular matrix
% b must be a vector of the same
leading dimension as L

n = size(L, 1);
x = zeros(n, 1);
for i = 1:n
tmp = b(i);
for j = 1:(i-1)
tmp = tmp - L(i, j) * x(j);
end
x(i) = tmp / L(i, i);
end Sebastian Valencia – Luis Angeles 8

end
 Solution of Equation Systems
Back substitution algorithm
function x = back_sub(U, b)
% x = back_sub(U, b) is the solution to
U x = b
% U must be an upper-triangular matrix
% b must be a vector of the same
leading dimension as U

n = size(U, 1);
x = zeros(n, 1);
for i = n:-1:1
tmp = b(i);
for j = i+1:n
tmp = tmp - U(i, j) * x(j);
end
x(i) = tmp / U(i, i);
end Sebastian Valencia – Luis Angeles 9

end
 Solution of Equation Systems

• The LU decomposition algorithm

• Given a matrix 𝐴 there are many different algorithms to find the matrices 𝐿 and 𝑈 for the LU
decomposition. Here we will use the recursive leading-row-column LU algorithm. This algorithm is
based on writing 𝐴=𝐿𝑈 in block form as:

• Comparing the left- and right-hand side entries of the above block matrix equation we see that:

Sebastian Valencia – Luis Angeles 10

 Solution of Equation Systems

• The LU decomposition algorithm

• These four equations can be rearranged to solve for the components of the 𝐿 and 𝑈 matrices as:

• The first three equations above can be immediately evaluated to give the first row and column
of 𝐿 and 𝑈. The last equation can then have its right-hand-side evaluated, which gives the Schur
complement 𝑆22 of 𝐴. We thus have the equation 𝐿22 𝑈22 = 𝑆22 , which is an (𝑛 − 1) × (𝑛 − 1) LU
decomposition problem which we can recursively solve.

Sebastian Valencia – Luis Angeles 11

 Solution of Equation Systems

• The LU decomposition algorithm

function [L, U] = lu_decomp(A) L11 = 1;
% [L, U] = lu_decomp(A) is the LU U11 = A11;
decomposition A = L U
% A is any matrix L12 = zeros(1, n-1);
% L will be a lower-triangular matrix with 1 U12 = A12;
on the diagonal, the same shape as A
% U will be an upper-triangular matrix, the L21 = A21 / U11;
same shape as A U21 = zeros(n-1, 1);
[n, ~] = size(A); S22 = A22 - L21 * U12;
if n == 1 [L22, U22] = lu_decomp(S22);
L = 1;
U = A; L = [L11, L12; L21, L22];
return; U = [U11, U12; U21, U22];
end end
A11 = A(1,1);
A12 = A(1,2:end);
A21 = A(2:end,1);
A22 = A(2:end,2:end);
Sebastian Valencia – Luis Angeles 12
 Solution of Equation Systems

• The LU decomposition algorithm

• Python

Sebastian Valencia – Luis Angeles 13

 A = [1,2,2;4,4,2;4,6,4];
• The LU decomposition algorithm
b = [10;12;5];
rank(A)
A= 𝑨𝒙 = 𝒃 rank([A,b])
1 2 2
4 4 2 % LU decomposition
4 6 4 [L,U] = lu_decomp(A)
y = forward_sub(L,b)
b= x = back_sub(U,y)
10
12 % Matlab operators
5 y = L\b
x2 = U\y
x=
x3 = A\b
17.0000
-24.5000
21.0000 Sebastian Valencia – Luis Angeles 14
 Solution of Equation Systems

• Direct Methods – MATLAB tools

1) If 𝐴 is a square and non-singular matrix (det A ≠ 0) , the solution for the linear equations can be
obtained by:
𝑥 = 𝑖𝑛𝑣(𝐴) × 𝑏
However, it is recommended to use the multiplication by an “inverse”, by the left of a matrix, using the
forward slash “\”:
𝑥 = 𝐴\b
This gives the same result as the command 𝑖𝑛𝑣, but the inverse of 𝐴 is computed differently (using LU
factorization). Also, if 𝐴 has a special structure, such as symmetric, sparse (zero large number of
elements), or triangular, the “\” command provides an efficient algorithm for solving linear equations.
2) Function lu
𝐿, 𝑈 = 𝑙𝑢 𝐴
[𝐿, 𝑈, 𝑃] = 𝑙𝑢(𝐴)

Sebastian Valencia – Luis Angeles 15

 Solution of Equation Systems

• Direct Methods – MATLAB tools

The "\" operator has implemented different algorithms such that, when A is square:

• If A is a triangular matrix, use the backward or forward substitution method to solve the system.

• If A is symmetric try to compute a Cholesky decomposition of A.

• If A is not symmetric or the Cholesky decomposition fails, a Gaussian algorithm with partial pivot is
used.

• If A is singular gives an error.

Sebastian Valencia – Luis Angeles 16

 Solution of Equation Systems

• Direct Methods – LU factorization – Example

• Consider the following system of linear equations:

• Solve in MATLAB using a) LU factorization and b) Cholesky factorization

Sebastian Valencia – Luis Angeles 17

 Solution of Equation Systems

• Direct Methods – LU factorization – Example

• Consider the following system of linear equations:

LU factorization: Cholesky factorization:

>> A = [4 -1 0; -1 4 -1; 0 -1 4] If A is a symmetric and positive definite

matrix:
>> b = [2;8;-6]
𝐴 = 𝐿𝐿𝑇 ; 𝑈 = 𝐿𝑇
>> [L,U,P] = lu(A)
>> U = chol(A)
>> y = L\(P*b)
>> y = U'\b
>> x = U\y
>> x = U\y
x= x=
1.0000 1.0000
2.0000 2.0000 Sebastian Valencia – Luis Angeles 18
-1.0000 -1.0000
 Solution of Equation Systems

• Iterative Methods
• They provide a sequence of approximate solutions that converge as the number of
iterations, tends to infinity, that is:

𝑥 0 →𝑥 1 →𝑥 2 →⋯→𝑥 𝑘

Initial value lim 𝑥 (𝑘)

𝑘→∞

• Characteristics:
• It is used for sparse matrices (high-dimensional matrix, in which most of its elements are
zeros).
• Less storage memory required and are usually faster than direct methods.
• It is not "invasive" (Gaussian elimination "invades" the matrix).
• There may be some convergence problems.
Sebastian Valencia – Luis Angeles 19
 Solution of Equation Systems

• Iterative Methods – Jacobi Method

• In the system 𝐴𝑥 = 𝑏 , it is true that 𝑎𝑖𝑖 ≠ 0; 𝑖 = 1, 2, ⋯ , 𝑛 of the ith equation,
𝑎11 𝑥1 + 𝑎12 𝑥2 + ⋯ + 𝑎1𝑛 𝑥𝑛 = 𝑏1 𝑛
𝑎21 𝑥1 + 𝑎22 𝑥2 + ⋯ + 𝑎2𝑛 𝑥𝑛 = 𝑏2 for 𝑖 = 1,2, … , 𝑛
⋮ → ෍ 𝑎𝑖𝑗 𝑥𝑗 = 𝑏𝑖
𝑗=1
𝑎𝑛1 𝑥1 + 𝑎𝑛2 𝑥2 + ⋯ + 𝑎𝑛𝑛 𝑥𝑛 = 𝑏𝑛

• Solving for 𝑥𝑖 it is obtained:

𝑛 𝑛

𝑎𝑖𝑖 𝑥𝑖 + ෍ 𝑎𝑖𝑗 𝑥𝑗 = 𝑏𝑖 → 𝑥𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 + 𝑏𝑖 /𝑎𝑖𝑖 for 𝑖 = 1,2, … , 𝑛

𝑗=1,𝑗≠𝑖 𝑗=1,𝑗≠𝑖

𝑛 for 𝑖 = 1,2, … , 𝑛
Jacobi Method (𝒌+𝟏) (𝒌)
𝑥𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 + 𝑏𝑖 /𝑎𝑖𝑖 𝑘 = 0,1,2, …
initial solution 𝑥0 𝑗=1,𝑗≠𝑖
Sebastian Valencia – Luis Angeles 20
 Solution of Equation Systems

• Iterative Methods – Jacobi Method

• Solve

• Initial value

• Error : 1𝑒 − 4

Sebastian Valencia – Luis Angeles 21

 Solution of Equation Systems

• Iterative Methods – Jacobi Method

function [x,X,Actual_error]=jacobiv1(A,b,x_0,error,nm_iter)
% x_0 : x inicial
x = x_0; while actual_error > error && k < nm_iter
n = size(A,1); k = k+1;
actual_error = 1; x_ant = x; % x anterior
Actual_error = []; for i=1:n
X = x; sum = 0;
k = 0; for j=[1:i-1,i+1:n]
sum = sum + A(i,j)*x_ant(j);
% Matriz B de Jacobi end
B_J = -diag(diag(A))\(triu(A,1)+tril(A,-1)); x(i) = (-sum + b(i))/A(i,i);
% spectral radius condition end
rho = max(abs(eig(B_J))); actual_error = norm(b-A*x');
disp('Radio Espectral'); X = [X;x];
disp(rho); Actual_error = [Actual_error;actual_error];
if rho >= 1 end
disp('No converge'); array2table(X)
end end

Sebastian Valencia – Luis Angeles 22

 Solution of Equation Systems

• Iterative Methods – Gauss Seidel

• Considerando Jacobi para i=1,...n
𝑛 for 𝑖 = 1,2, … , 𝑛
(𝒌+𝟏) (𝒌)
𝑥𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 + 𝑏𝑖 /𝑎𝑖𝑖
𝑘 = 0,1,2, …
𝑗=1,𝑗≠𝑖

• To compute 𝑥2𝑘+1 , we use 𝑥1𝑘 , , but 𝑥1𝑘+1 has already been computed. Then, if the
method is convergent, it is better to use 𝑥1𝑘+1 instead of 𝑥1𝑘

𝑖−1 𝑛
(𝒌+𝟏) (𝒌+1) (𝒌) for 𝑖 = 1,2, … , 𝑛
𝑥𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 − ෍ 𝑎𝑖𝑗 𝑥𝑗 + 𝑏𝑖 /𝑎𝑖𝑖
𝑗=1, 𝑗=𝑖+1 𝑘 = 0,1,2, …

Sebastian Valencia – Luis Angeles 23

 Solution of Equation Systems

• Iterative Methods – Gauss Seidel

• Solve

• Initial value

• Error : 1𝑒 − 4

Sebastian Valencia – Luis Angeles 24

 Solution of Equation Systems

• Iterative Methods – Gauss Seidel

function [x,X,Actual_error]=gaussv1(A,b,x_0,error,nm_iter)
% x_0 : x inicial
x = x_0;
n = size(A,1); while actual_error > error && k < nm_iter
actual_error = 1; k = k+1;
Actual_error = []; for i=1:n
X = x; sum = 0;
k = 0; for j=[1:i-1,i+1:n]
sum = sum + A(i,j)*x(j);
% Matriz B de Gauss-Seidel end
B_GS = -tril(A)\triu(A,1); x(i) = (-sum + b(i))/A(i,i);
% spectral radius condition end
rho = max(abs(eig(B_GS))); actual_error = norm(b-A*x');
disp('Radio Espectral'); X = [X;x];
disp(rho); Actual_error = [Actual_error;actual_error];
if rho >= 1 end
disp('No converge'); array2table(X)
end end

Sebastian Valencia – Luis Angeles 25

 Solution of Equation Systems

• Iterative Methods – SOR Method (succesive over relaxation)

• It is a modification of the Gauss-Seidel Method to speed up convergence:
𝑖−1 𝑛
(𝒌+𝟏) (𝒌+1) (𝒌)
𝑥𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 − ෍ 𝑎𝑖𝑗 𝑥𝑗 + 𝑏𝑖 /𝑎𝑖𝑖
𝑗=1, 𝑗=𝑖+1

𝑖−1 𝑛 𝒌
(𝒌+𝟏) 𝒌+1 𝒌 𝒌 𝒂𝒊𝒊 𝒙𝒊
𝑥𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 − ෍ 𝑎𝑖𝑗 𝑥𝑗 − 𝒂𝒊𝒊 𝒙𝒊 + 𝑏𝑖 /𝑎𝑖𝑖 +
𝑎𝑖𝑖
𝑗=1, 𝑗=𝑖+1

(𝒌+𝟏) (𝒌) (𝒌)

𝑥𝑖 = 𝑥𝑖 + 𝑟𝑖 , where:
𝑖−1 𝑛
(𝒌) 𝒌+1 𝒌
𝑟𝑖 = − ෍ 𝑎𝑖𝑗 𝑥𝑗 − ෍ 𝑎𝑖𝑗 𝑥𝑗 + 𝑏𝑖 /𝑎𝑖𝑖
𝑗=1, 𝒋=𝒊
𝟎<𝝎<𝟐
(𝒌+𝟏) (𝒌) (𝒌)
SOR → 𝑥𝑖 = 𝑥𝑖 + 𝝎𝑟𝑖 𝝎: Relaxation parameter Sebastian Valencia – Luis Angeles 26
 Solution of Equation Systems

• Iterative Methods – Advantages and Disadvantages

• In general, there are no storage problems with iterative methods.

• For large dimensional systems of equations (1000x1000, or larger), efficient iterative

methods require less computational cost than direct methods.

• They are relatively easy to adapt to parallel calculation processes.

• The number of iterations required in iterative methods is not known a priori.

• For certain ill-conditioned problems, the convergence rate of direct methods can be
very slow.

Sebastian Valencia – Luis Angeles 27

 Solution of Equation Systems

• Problem 1 – Solution of a linear system by the iterative method

• Write an algorithm in MatLab for the solution of a linear system using the Jacobi method
and another one using the Gauss-Seidel method.
• A main program must call the respective algorithms (they can be functions) developed for
the solution of the system 𝐴𝑥 = 𝑏; where:
1
• 𝑎𝑖𝑗 = 𝑖+𝑗−1 and 𝑏 is obtained in such a way that the exact solution of the system is 𝑥𝑖 = 1.
• In the main program, the size of the n × n square matrix must be an input.
• Execute the program considering 𝑛 = 2, 10 and 15 for the two iterative methods
implemented.
• Explain what happens.

Sebastian Valencia – Luis Angeles 28

 Solution of Equation Systems

• Problem 1 – Solution of a linear system by the iterative method

Sebastian Valencia – Luis Angeles 29

 Solution of Equation Systems

• Problem 2 – Temperature Distribution

• The steady state temperature distribution of a flat plate can be approximated at 9
internal points by applying the discretized Laplace equation at each point. If the sides
of a square plate are kept at constant temperatures of 0 °C and 100 °C, the
temperatures at the 9 points can be obtained as the solution of the following system:

• Obtain the temperature at the different points of the plate and make a graph of the
temperature distribution. Sebastian Valencia – Luis Angeles 30
 Solution of Equation Systems

• Problem 2 – Temperature Distribution

Sebastian Valencia – Luis Angeles 31

 Solution of Equation Systems

• Project 01 – Solving Equation System

• Solve the system 𝑨𝒙 = 𝒃 using the following methods: a) Jacobi; b) Gauss-Seidel; c)

SOR with ω = 0.8 and 1.5. Compare the results in a big plot.

• Make a figure of number of iterations versus ω for SOR. Which value of ω is giving the
lowest number of iterations?

Sebastian Valencia – Luis Angeles 32

MSc. Sebastian Valencia – MSc. Luis Angeles
Mechanical Engineering Section
Department of Engineering
Pontificia Universidad Católica del Perú
Av. Universitaria 1801
San Miguel, Lima 32
Peru

[email protected] , [email protected].

Sebastian Valencia – Luis Angeles 33