005background-linear algebra and probability
TRANSCRIPT
-
8/9/2019 005Background-Linear Algebra and Probability
1/46
Introduction to Linear Algebra
-
8/9/2019 005Background-Linear Algebra and Probability
2/46
Matrix
representation
of size n X d
-
8/9/2019 005Background-Linear Algebra and Probability
3/46
Symmetric Matrix
Diagonal Matrix
-
8/9/2019 005Background-Linear Algebra and Probability
4/46
Matrix multiplication with a vector
gives a transformed vector
-
8/9/2019 005Background-Linear Algebra and Probability
5/46
Inner Product
Norm of a
vector
Cosine of the angle
between 2 vectors
Cauchy Schwartz
Inequality
-
8/9/2019 005Background-Linear Algebra and Probability
6/46
Linear Independence of vectors / Basis
vectors
-
8/9/2019 005Background-Linear Algebra and Probability
7/46
Outer Product of 2 vectors
-
8/9/2019 005Background-Linear Algebra and Probability
8/46
Rank of a matrix
Number of independent rows / columns of a
matrix
Outer product between 2 vectors gives a
matrix of rank 1.
-
8/9/2019 005Background-Linear Algebra and Probability
9/46
Gradient of a scalar function
-
8/9/2019 005Background-Linear Algebra and Probability
10/46
Gradient of a vector-valued
function (Jacobian)
-
8/9/2019 005Background-Linear Algebra and Probability
11/46
Matrix / Vector Derivatives
Gradient of
Vector function
Gradient of scalar
function
-
8/9/2019 005Background-Linear Algebra and Probability
12/46
Eigen values and vectors
Diagonalization property
1VDVM
V Matrix, whose columns correspond toEigenvectors of
D Diagonal matrix, corresponding toeigenvalues of
M
M
-
8/9/2019 005Background-Linear Algebra and Probability
13/46
Rank deficient square matrices
A matrix Aof rank r will have r independentrows / columns, and is said to be rank-deficient.
There will be r independent eigenvectorscorresponding to r non- zero eigen
values
r eigen values are zero.
-
8/9/2019 005Background-Linear Algebra and Probability
14/46
Matrix Inverse
For square matrix M
-
8/9/2019 005Background-Linear Algebra and Probability
15/46
Orthogonal Matrix
IQQT
ji
jiqq
j
T
i 0
1
},.....,{ 21 Mqqq M vectors
Qcolumns of denote
Mqqq .....21
is of size M XM and is said to be
an orthogonal matrix
Q
-
8/9/2019 005Background-Linear Algebra and Probability
16/46
Positive Definite Matrix
Positive definite square
matrix A0Axx
T
Eigen values of positive definite matrix A are
non-negative.
x denotes a vector of size d x 1 .
A denotes a vector of size d x d .
-
8/9/2019 005Background-Linear Algebra and Probability
17/46
Real symmetric matrix
Elements of a matrix are real numbers
Matrix is symmetric
Eigen values of a real symmetric matrix arereal.
-
8/9/2019 005Background-Linear Algebra and Probability
18/46
Probability Theory
-
8/9/2019 005Background-Linear Algebra and Probability
19/46
Random variables
A random variable xdenotes a quantity that is
uncertain
May be result of experiment (flipping a coin) or a real
world measurements (measuring temperature) If observe several instances of xwe get different
values
Some values occur more than others and thisinformation is captured by a probability distribution
19
-
8/9/2019 005Background-Linear Algebra and Probability
20/46
Discrete Random Variables
20
-
8/9/2019 005Background-Linear Algebra and Probability
21/46
-
8/9/2019 005Background-Linear Algebra and Probability
22/46
-
8/9/2019 005Background-Linear Algebra and Probability
23/46
Continuous Random Variable
23
-
8/9/2019 005Background-Linear Algebra and Probability
24/46
Expectation/ Variance of continuous
random variable
-
8/9/2019 005Background-Linear Algebra and Probability
25/46
Joint Probability
Consider two random variables xand y
If we observe multiple paired instances, then some
combinations of outcomes are more likely than
others This is captured in the joint probability distribution
Written as P(x,y)
Can read P(x,y) as probability of xand y
25
-
8/9/2019 005Background-Linear Algebra and Probability
26/46
For 2 random variables x and y,
26
Marginalization property
-
8/9/2019 005Background-Linear Algebra and Probability
27/46
Marginalization
We can recover probability distribution of any variable in a joint distributionby integrating (or summing) over the other variables
27
-
8/9/2019 005Background-Linear Algebra and Probability
28/46
Marginalization
We can recover probability distribution of any variable in a joint distributionby integrating (or summing) over the other variables
28
-
8/9/2019 005Background-Linear Algebra and Probability
29/46
Marginalization
We can recover probability distribution of any variable in a joint distributionby integrating (or summing) over the other variables
Works in higher dimensions as wellleaves joint distribution betweenwhatever variables are left
29
-
8/9/2019 005Background-Linear Algebra and Probability
30/46
Conditional Probability
Conditional probability of xgiven that y=y1is relative
propensity of variable xto take different outcomes given that
y is fixed to be equal to y1.
Written as Pr(x|y=y1)
30
-
8/9/2019 005Background-Linear Algebra and Probability
31/46
Conditional Probability
Conditional probability can be extracted from joint probability Extract appropriate slice and normalize
31
-
8/9/2019 005Background-Linear Algebra and Probability
32/46
Conditional Probability
More usually written in compact form
Can be re-arranged to give
32
-
8/9/2019 005Background-Linear Algebra and Probability
33/46
Conditional Probability
This idea can be extended to more than two
variables
33
-
8/9/2019 005Background-Linear Algebra and Probability
34/46
Bayes Rule
From before:
Combining:
Re-arranging:
34
-
8/9/2019 005Background-Linear Algebra and Probability
35/46
Bayes Rule Terminology
Posteriorwhat we
know about yafter
seeingx
Priorwhat we knowabout ybefore seeing x
Likelihoodpropensity forobserving a certain value of
xgiven a certain value of y
Evidencea constant to
ensure that the left hand
side is a valid distribution35
-
8/9/2019 005Background-Linear Algebra and Probability
36/46
Independence
If two variables xand yare independent then variable xtellsus nothing about variable y(and vice-versa)
36
-
8/9/2019 005Background-Linear Algebra and Probability
37/46
Independence
When variables are independent, the joint factorizes into aproduct of the marginals:
37
-
8/9/2019 005Background-Linear Algebra and Probability
38/46
Discrete Random vectors
-
8/9/2019 005Background-Linear Algebra and Probability
39/46
Continuous Random vectors
-
8/9/2019 005Background-Linear Algebra and Probability
40/46
Continuous random vectors
-
8/9/2019 005Background-Linear Algebra and Probability
41/46
Discrete random vectors
-
8/9/2019 005Background-Linear Algebra and Probability
42/46
-
8/9/2019 005Background-Linear Algebra and Probability
43/46
Properties of covariance matrix
-
8/9/2019 005Background-Linear Algebra and Probability
44/46
1 dimensional Gaussian
-
8/9/2019 005Background-Linear Algebra and Probability
45/46
2 dimensional Gaussian
-
8/9/2019 005Background-Linear Algebra and Probability
46/46
Multi dimensional Gaussian