Basic principles of probability theory

Download Report

Transcript Basic principles of probability theory

Short review of probabilistic concepts
Probability theory plays very important role in statistics. This
lecture will give the short review of basic concepts of the
probability theory.
Basic principles and definitions
Conditional probabilities and independence
Bayes’s theorem and postulate
Random variables and probability distributions
Bayes’s theorem and likelihood
Expectations and moments
Random experiment
Random experiment satisfies the following conditions:
All possible distinct outcomes are known in advance
In any particular experiment outcome is not known in advance
Experiment can be repeated under identical conditions
The outcome space  is the set of the possible outcomes.
Example 1. Tossing a coin is a random experiment. The outcome space is {H,T} –
head and tail.
Example 2. Rolling a die. The outcome space is {1,2,3,4,5,6}
Example 3. Drawing from an urn with N balls, M of them is red and N-M is white.
The outcome space is {R,W} – red and white
Example 5. Measuring temperature (in C or in K): What is the outcome space?
Something that might or might not happen depending on outcome of the experiment
is called an event. An event is a subset of the outcome space
Example: Rolling a die. {1,2,3} or {2,4,6}
Example: Measuring temperature in Celsius. Give example of an event.
Classical definition of probability
If all the outcomes are equally likely then the probability of event A is the number of
outcomes in A (M(A)) divided by number of all outcomes (M):
P ( A) 
M ( A)
Example: If a coin is fair then probability of H is ½ and probability of T is ½
Example: If a die is fair then probability of {1} is 1/6
If the outcome space is real numbers or are in a space then probability is measured as
ratio of area of an event and area of outcome space:
P( A) 
M ( A)
M ( )
Where M is area.
Example: Outcome space is the interval [0,2]. What is probability of [0,1]?
Frequency definition of probability
Since random experiments can be repeated as many times as we wish under identical
conditions (in theory) we can measure the relative frequency of the occurrence of
an event. If number of trials is m and number of the occurrence of A is m(A) then
according to frequency definition probability of A is the limit:
P( A)  lim
m( A)
( m  )
According to the law of large numbers this limit exists. When the number of trials is
small then there might be strong fluctuations. As number of trials increases
fluctuations tend to decrease.
Other (subjective) definitions of probability
There are other definitions of probability also:
• Degree of belief. How much a person believes in an event. In that sense one
person’s probability would be different from another person’s. For example:
existence of “an extra-terrestrial life”.
• Degree of knowledge. In many cases exact value of an event exists but we do not
know it. By carrying out experiments we want to find this value. Since experiment
is prone to errors it is in general impossible to find exact value and we assign
probability for this. That is purpose of the most statistical procedures and
techniques. According to Jaynes if proper rules are designed then exactly same
information would produce exactly same probabilities. (See Jaynes, The
Probability theory: Logic of Science). This definition reflects our state of
knowledge about the event and can change as we update our knowledge.
Probability axioms
Probability is defined as a function from subsets of outcome space  to the real line
R that satisfies following conditions:
Non-negativity: P(A)  0
Additivity: if AB= then P(AB) = P(A) + P(B)
Probability of whole space is 1. P() = 1
All above definitions obey these rules. So any property that can be derived from
these axioms is valid for all definitions
Show that: P( )=0 (Hint:   = )
Show that: 0  P(A)  1 (Hint A and Ã=-A are not intersecting).
Conditional probability and independence
Let us consider if an event B has occurred or will occur and we want to know what is
the probability of A. Knowing B may influence our knowledge about A. Or
occurrence of B may influence of occurrence of A. The probability of A given B is
called conditional probability of A given B and is defined as (for P(B)>0):
P( A | B ) 
P( A  B )
P( B )
It is clear that now the event B has become new outcome space. Event A and B are
called independent if occurrence of B does not influence on probability of A.
P( A | B)  P( A) and P( B | A)  P( B)
It can also be written as:
P( A  B)  P( A) P( B)
Note that only one of the above equations is independent.
The Law of total probability
In many cases when direct calculation of probability is not known it is easier to
divide an event into smaller parts and calculate their probability and then take
weighted average of them. This can be done using the law of total probability.
Let B1, B2,,,Bn be partition of , I.e. they are mutually exclusive (BiBj=) and their
sum is  (1n Bi= ) then from the axioms of probability:
P( A)   P( A | Bi ) P( Bi )
i 1
Consider a box with N balls, M of them are red and N-M are white. We make two
draws. We don’t know what is the first ball. What is probability of the second ball
being red. (Hint: Use partition as ({R1} {W1}). Then use law of total probability
for ({R2}. Here subscript shows the first or the second draw.)
Bayes’s theorem
Bayes’s theorem is a tool that updates probability of an event in the light of an
evidence. It is written in various forms. All they are equivalent. Let us again
consider partition of outcome space – B1,B2,,,,Bn so that they are mutually exclusive
and sum of them is equal to . Then for one of these events (say j-th event) we can
P( A | B j ) P( B j )
P( A | B j ) P( B j )
P( B j | A) 
 P( A | Bi ) P( Bi )
P( A)
i 1
Usually P(Bj|A) is called posterior probability, P(Bj) is prior probability and P(A|Bj) is
likelihood. It is widely used in statistical inferences. We will come back to this
theorem again.
Example: A box contains four balls. There are two possibilities: a) all balls are white
(B1) b) two white and two red (B2). A ball is drawn and it is white (event A). What is
the probability that all balls are white. B1 (all white) and B2 (two white and two red)
are two possible outcomes with prior probabilities ½. If B1 is true then probability of
A is 1 and if B2 is true then probability of A is ½. Calculate P(B1|A). What is
probability P(B2|A)
Bayes’s postulate: If there is no prior information available then prior probabilities
should be assumed to be equal.
Random variables
Random variable is a function from outcome space to the real line
X:   R
Example: Consider random experiment of tossing the coin twice. Outcome space is:
Define random variable as
X((T,T)) = 0, X((H,T))=X((T,H)) = 1, X((H,H))=2
Example 2: Rolling a die. Outcome space {1,2,3,4,5,6). Define random variable
X(j) = j.
Probability distribution function
Discrete case (number of elements in outcome space is finite or countable infinite):
Probability function p assigns for each possible realisation x of a random variable X
the probability P(X=x). Obviously xp(x) = 1.
Example: The number of heads turning up in two tosses is random variable with
probability function p(1) = 1/4, p(0) =1/2, p(2) =1/4.
For continuous random variable it is not possible to define probability for each
realisation since their probability is usually 0. For them it is easy to define
distribution function:
F(x) = P(Xx)
i.e. probability that X is less than or equal to x. F(x) has following properties:
1) F(- ) = 0, 2) F(x) is monotonic and increasing function, 3) F(+ ) = 1.
This function is defined for discrete as well as continuous random variables. If
derivative of F(x) exists (it is usually defined for continuous random variables)
then it is called probability density function – f(x) = dF(x)/dx. Another relation
between them is:
F ( x) 
 f ( x )dx
Joint probability distributions
If there are more than one random variables then their joint probability distribution is
defined similarly. For discrete case:
p(x,y) = P((X,Y)=(x,y)) = P(X=x,Y=y)
Then xyp(x,y) = 1, p(x,y)0.
The marginal probability function p(x) is derived by summing over all possible values
of y
pX(x) = yp(x,y)
Conditional probability function of X given Y=y is:
p(x|y) = p(x,y)/pY(y)
Definition for the joint probability distribution for continous random variables is
F(x,y) = P(X  x,Yy). Probability density (f(x,y)) is derivative of probability function
with respect to its arguments. It has properties:
 
f ( x, y )  0,
  f ( x, y )dxdy  1
Marginal and conditional probability densities are defined similar to discrete random
variables by replacing summation with integration.
Joint probability distributions and independence
Random events {X=x} and {Y=y} are independent if
P(X=x, Y=y) = P(X=x)P(Y=y)
The random variables are independent if for all pairs (x,y) this relation holds. It can
also be written as
p(x,y) = pX(x)pY(y)
And then p(x|y) = pX(x) and p(y|x) = pY(y)
For continuous random variables definition is analogous. It can be defined by
replacing p with f everywhere.
f(x,y) = fX(x)fY(y), f(x|y) = fX(x), f(y|x) = fY(y)
Bayes’s theorem then becomes:
f(x|y) = fX(x) f(y|x)/fY(y)
Usually we will drop subscripts X and Y.
Where f(x|y) is posterior probability density, f(x) is prior probability density f(y|x) is
likelihood of y if x would be observed, f(y) can be considered as a normalisation
Expectation values. Moments
If X is a random variable and h(X) is its function then expectation value (discrete
case) is defined as:
E(h(X)) = xh(x)p(x)
If h(x) = x then it is called the first moment. If h(x) = xn then it is called n-th moment.
If h(x) = (x-E(X))n then it is called n-th central moment: The second central
moment is called variance of the random variable. First moment and second
central moment play important role in statistics and they have special symbols
   xp( x)  2  ( x   )2 p( x)
 - is also called as a standard deviation
When there are more than one random variable and their joint probability function is
known then their mixed moments also are defined. Most important of them is
covariance and correlation:
cov(x, y )   ( x  x )( y   y ) p( x, y ),  (x,y) 
cov(x, y )
x , y variables expectation values, moments,
x ycovariance and
For continuous random
correlation are defined similarly by replacing summation with integration. If
random variables are independent then their covariance is 0. Reverse is not true in
Entropy and maximum entropy
Entropy of a random variable is the average amount of information generated by
observing its value. The information generated by observing realisation of X=x is
defined as:
H(X=x) = ln(1/p(x)) = -ln(p(x))
Example: We have a coin. Probability of head is 0.9 and probability of tail is 0.1.
Define random variable. X(H) = 1, X(T) = 0. What is information generated by
observing head? What is information generated by observing tail?
Entropy is defined (discrete case) as
H(X) = -x p(x) ln(p(x))
It is Shannon entropy first derived in information theory by Claude Shannon. Since
ln(p(x)) can be considered as a random variable then entropy can be written as
H(X) = -E(ln(p(x))
Expected value of -ln(p(x)).
It is often used to generate prior probability distribution from given information.
Maximisation of this function under conditions gives the probability distribution.
Continuous case is defined similarly by changing summation with integration and p
with f.
Further reading
Berthold, M. and Hand, DJ (2003) “Intelligent data analysis”
Mardia, KV, Kent, JT and Bibby, JM (2003) “Mutlivariate analysis”
Jaynes, E. (2003) “The probability theory: Logic of science”