Friday, July 8, 2016
Unizor - Probability - Correlation
Notes to a video lecture on http://www.unizor.com
In this lecture we will talk about independent and dependent random variables and will introduce a numerical measure of dependency between random variables.
Assume a random variable ξtakes values
x1, x2,..., xM
p1, p2,..., pM.
Further, assume a random variable η takes values
y1, y2,..., yN
q1, q2,..., qN.
A short reminder about independence of these random variables.
If the conditional probability of a random variable ξ taking any one of its values under the condition that a random variable η took any one of its values equals to an unconditional probability of ξtaking that value, then a random variable ξ is independent of a random variable η.
In other words, ξ is independent of η if
P(ξ=xi | η=yj ) = P(ξ=xi )
where index i can take any value from 1 to M and index j can take values from 1 to N.
Simple consequences of this definition, as discussed in previous lectures are:
(a) independence is symmetrical, that is, if ξ is independent of η, then η is independent of ξ:
(b) for independent random variables the probability of them to take simultaneously some values equals to a product of their probabilities to take these values independently:
P(ξ=xi ∩ η=yj ) =
= P(ξ=xi ) · P(η=yj )
(c) Mathematical expectation of a product of two independent random variables equal to a product of their mathematical expectations:
E(ξ·η) = E(ξ)·E(η)
The last property of mathematical expectations for independent random variables is the basis of measuring the degree of dependency between any pair of random variables.
First of all, we introduce a concept of covariance of any two random variables:
Simple transformation by opening parenthesis converts it into an equivalent definition:
Cov(ξ,η) = E(ξ·η)−E(ξ)E(η)
Now we see that for independent random variables their covariance equals to zero (see property (c) above).
Incidentally, the covariance of a random variable with itself (kind of ultimate dependency) equal to its variance:
Cov(ξ,ξ) = E(ξ·ξ)−E(ξ)E(ξ) =
= E[(ξ−E(ξ))²] = Var(ξ)
Also notice that another example of very strong dependency, η = A·ξ, where A is a constant, leads to the following value of covariance:
= E(ξ·Aξ)−E(ξ)E(Aξ) =
= A·E[(ξ−E(ξ))²] = A·Var(ξ)
This shows that, when coefficient A is positive (that is, positive change of ξ causes positive change of η=A·ξ),covariance between them is positive as well and proportional to coefficient A. If A is negative (that is, positive change of ξ causes negative change of η=A·ξ), covariance between them is negative as well and still proportional to coefficient A.
One more example.
Consider "half-dependency" between ξ and η, defined as follows.
Let ξ' be an independent random variable, identically distributed with ξ.
Let η = (ξ + ξ')/2.
So, η "borrows" its randomness from two independent identically distributed random variables ξ and ξ'.
Then covariance between ξ andη is:
Cov(ξ,η) = Cov(ξ,(ξ+ξ')/2) =
Since ξ and ξ' are independent, expectation of their product equals to a product of their expectations.
So, our expression can be transformed further:=E(ξ²)/2+E(ξ)·E(ξ')/2 −
As we see, covariance between "half-dependent" random variables ξ and η=(ξ+ξ')/2, where ξ and ξ' are independent identically distributed random variables, equals to half of the variance of ξ.
All the above manipulations with covariance led us to some formulas where the variance plays a significant role. If we want a kind of measure that reflects the dependency between random variables not related to variances, but always scaled in the interval [−1, 1], we have to scale the covariance by a factor that depends on variances, thus forming a coefficient of correlation:R(ξ,η) =
Let's examine this coefficient of correlation in cases we considered above as examples.
For independent random variables ξ and η the correlation is zero because their covariance is zero.
Correlation between a random variable and itself equals to 1:R(ξ,ξ) = Cov(ξ,ξ)/Var(ξ,ξ) = 1
Correlation between a random variables ξ and Aξ equals to 1(for positive constant A) or −1(for negative A):
R(ξ,Aξ) = Cov(ξ,Aξ)/√(Var(ξ)·Var(Aξ) = A/|A|
which equals to 1 or −1, depending on a sign of A.
This seems to corresponds our intuitive understanding of rigid relationship between ξ and Aξ.
Correlation between "half-dependent" random variables, as introduced above, is:
R(ξ,(ξ+ξ')/2) = Cov(ξ,(ξ+ξ')/2)/√(Var(ξ)·Var((ξ+ξ')/2) = √2/2.
As we see, in all these examples the correlation is a number from an interval [−1,1] that is equal to zero for independent random variables, equals to 1 or−1 for rigidly dependent random variables and is inside this interval for partially dependent (like in our example of "half-dependent") random variables.
For those interested, it can be proved that this statement is true for any pair of random variables.
So, the coefficient of correlation is a good tool to measure the degree of dependency between two random variables.