ESCI 386 Scientific Programming,
Analysis and Visualization with
Python
Lesson 18 - Linear Algebra
Matrix Operations
In many instances, numpy arrays can be
thought of as matrices.
In the next slides we explore some matrix
operations on numpy arrays
Determinants
The determinant of an array is found by
using the det() function from the scipy.linalg
module.
>>> import scipy.linalg as slin
>>> a
array([[ 3, -5, 8],
[-1, 2, 3],
[-5, -6, 2]])
>>> slin.det(a)
259.0
3
Trace
The trace of an array is found by using the
trace() function from numpy.
>>> import numpy
>>> a
array([[ 3, -5,
[-1, 2,
[-5, -6,
>>> np.trace(a)
7
as np
8],
3],
2]])
Trace (cont)
Offset traces can also be computed.
>>> a
array([[ 3, -5, 8],
[-1, 2, 3],
[-5, -6, 2]])
>>> np.trace(a,-1)
-7
>>> np.trace(a,1)
-2
5
Inverses
Inverse of a matrix is computed from
scipy.linalg.inv() function.
>>> a
array([[ 3, -5, 8],
[-1, 2, 3],
[-5, -6, 2]])
>>> slin.inv(a)
array([[ 0.08494208, -0.14671815, -0.11969112],
[-0.05019305, 0.17760618, -0.06563707],
[ 0.06177606, 0.16602317, 0.003861 ]])
Inverses
Transpose of a matrix is computed from
numpy.transpose() function.
a
array([[ 3, -5, 8],
[-1, 2, 3],
[-5, -6, 2]])
>>> np.transpose(a)
array([[ 3, -1, -5],
[-5, 2, -6],
[ 8, 3, 2]])
7
NumPy Matrix Objects
NumPy also has matrix objects that are an
extension of arrays.
These matrix objects have built in methods
for determinant and inverse.
NumPy Matrix Objects
a
matrix([[ 3, -5, 8],
[-1, 2, 3],
[-5, -6, 2]])
>>> a.T
matrix([[ 3, -1, -5],
[-5, 2, -6],
[ 8, 3, 2]])
>>> a.I
matrix([[ 0.08494208, -0.14671815, -0.11969112],
[-0.05019305, 0.17760618, -0.06563707],
[ 0.06177606, 0.16602317, 0.003861 ]])
Matrix Objects Support Matrix
Multiplication
>>> a
matrix([[ 3, -5,
[-1, 2,
[-5, -6,
>>> b
matrix([[ 3],
[ 4],
[-1]])
>>> a*b
matrix([[-19],
[ 2],
[-41]])
8],
3],
2]])
10
Solving Systems of Equations
A system of linear, algebraic equations can be
written in matrix form.
a1 x b1 y c1 z d1
a2 x b2 y c2 z d 2
a3 x b3 y c3 z d3
a1
a2
a
3
b1
b2
b3
c1 x d1
c2 y d 2
d
c3
z
3
The 33 matrix is called the coefficient matrix
The right-hand side is a vector.
11
Methods/Functions for Solving Matrix Equations
Solving matrix equations is computationally
intensive.
We will discuss several methods for solving
these equations.
These methods are all from the scipy.linalg
module.
12
Methods/Functions for Solving Matrix Equations
We will illustrate these methods for the simple
system of equations:
4 x 5 y 8z 4
2x 8 y 7z 0
5 x 8 y 5
13
scipy.linalg.solve()
This method takes the coefficient matrix and the
right-hand side vector as arguments and return a
vector with the solutions.
>>> cm = np.array([[4, -5, 8], [2, -8, 7], [-5, 8, 0]])
>>> rhs = np.array([4,0,-5])
>>> soln = slin.solve(cm,rhs)
>>> soln
array([ 1.53112033, 0.33195021, -0.05809129])
14
Coefficient Matrix Must be Nonsingular!
If the coefficient matrix is singular (has
determinant of zero) then an error results.
>>> cm = np.array([[4, -5, 8], [8, -10, 16], [-5, 8, 0]])
>>> rhs = np.array([4,8,-5])
>>> soln = slin.solve(cm,rhs)
Traceback (most recent call last):
File "<pyshell#45>", line 1, in <module>
soln = slin.solve(cm,rhs)
File "C:\Python27\lib\site-packages\scipy\linalg\basic.py", line 68, in solve
raise LinAlgError("singular matrix")
LinAlgError: singular matrix
15
Very Large Systems of Equations
Very large systems of equations are very
computationally intensive to solve.
There are several specialized methods to
efficiently solve large systems of equation.
We will discuss some of these.
16
LU Decomposition
If the right-hand side vector changes, but the
coefficient matrix doesnt change, then the coefficient
matrix can be decomposed using LU decomposition.
This LU decomposition can then be used to solve the
system for any different right-hand side.
This saves time because the decomposition is the
single biggest drain on resources. So, by not having to
redo it every time, we save computational time.
17
LU Decomposition
To use LU decompositions:
First, use the slin.lu_factor() method on the coefficient
matrix, and assign the result to a new variable.
Then, use the slin.lu_solve() function with the
decomposition and the rhs as arguments.
>>> cm = np.array([[4, -5, 8], [2, -8, 7], [-5, 8, 0]])
>>> rhs = np.array([4,0,-5])
>>> soln = slin.solve(cm,rhs)
>>> soln
array([ 1.53112033, 0.33195021, -0.05809129])
18
LU Decomposition Example
>>> cm
array([[ 4, -5, 8],
[ 2, -8, 7],
[-5, 8, 0]])
>>> rhs
array([ 4, 0, -5])
>>> lu = slin.lu_factor(cm)
>>> soln = slin.lu_solve(lu,rhs)
>>> soln
array([ 1.53112033, 0.33195021, -0.05809129])
19
LU Decomposition Example
Once the LU decomposition is accomplished we
can use any right-hand side we want without
redoing the decomposition.
>>> soln = slin.lu_solve(lu,[4, -3, 9])
>>> soln
array([ 0.64315353, 1.52697095, 1.13278008])
>>> soln = slin.lu_solve(lu,[-2, -3, -12])
>>> soln
array([ 1.77593361, -0.39004149, -1.38174274])
20
Banded Matrices
Many large matrices in the sciences and
engineering are of a banded nature, meaning that
their non-zero values are along diagonals.
Methods for efficiently solving these type of
matrices have been developed.
Banded matrices also dont require as much
memory to store, since many of the values are
zero
21
Representing Banded Matrices
The banded matrix on the left is represented
as the non-square matrix shown on the right.
1 5 3 0
3 2 3 5
0 2 1 5
0 0 9 1
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0
0
9
5
2
2
0
0
0 0 0
0 0 0 0 0 3 5
4 0 0 0 5 3 5
1 2 1 1
3 2 0
0 1 6 3 2 9 0
3 2 7
0 9 1
0
2 6
5 3 1 7
2 0 2 1
2 3 9 0
22
Representing Banded Matrices
The upper diagonals have leading zeros.
Lower diagonals have trailing zeros.
1 5 3 0
3 2 3 5
0 2 1 5
0 0 9 1
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0
0
9
5
2
2
0
0
0 0 0
0 0 0 0 0 3 5
4 0 0 0 5 3 5
1 2 1 1
3 2 0
0 1 6 3 2 9 0
3 2 7
0 9 1
0
2 6
5 3 1 7
2 0 2 1
2 3 9 0
23
Solving Banded Matrix Equations
To solve a set of equations with a banded coefficient
matrix we use the scipy.linalg.solve_banded() function.
The format for this function is
slin.solve_banded((l,u), cm, rhs)
(l, u) is a tuple where l is the number of nonzero lower
diagonals, and u is the number of nonzero upper
diagonals.
cm is the coefficient matrix in banded form, and rhs is the
right-hand side vector.
24
In-class Exercise
Solve the set of matrix equations below using
solve_banded().
1 5 3 0
3 2 3 5
0 2 1 5
0 0 9 1
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0
0
9
5
2
2
0
0
0 a 5
0 0 0 b 3
0 0 0 c 3
4 0 0 d 2
0
3 2 0
e
0 1 6 f 7
3 2 7 g 8
0 9 1 h 1
0
25
In-class Results
a = 200.639937
b = -25.491352
c = -107.698899
d = -174.206761
e = 102.750000
f = 70.833333
g = -3.500000
h = 32.500000
26
