1. Linear Algebra in Markov
Chains
Presented By : Luckshay Batra
luckybatra17@gmail.com
2. Introduction
Transition Matrix : A matrix in which all entries are
nonnegative and all columns sum to 1.
Thus, an n x n Transition matrix gives the probability of
movement from any one of n states to any of the other n
states.
Probability Vector : A vector whose coordinates are
nonnegative and sum to 1. Thus, an n-dimensional
probability vector gives the chances of occurrence of
each one of n events.
3. Markov Chain
If p0 is a probability vector and A is a transition matrix, then the sequence,
p0, p1, p2, p3, ... , where pk = (A^k)*p0 for k = 1, 2, 3, ..., is called a Markov
chain.
If A*p = p for some probability vector, then p is called a steady-state
vector.
A Stochastic model that describes the probabilities of transition among
the states of a system.
Change of state depends probabilistically only on the current state of the
system.
Independent of the past given that the present state is known.
The behavior depends on the structure of the transition matrix P.
4. Problem of Television Viewers
Suppose there are two regional news shows in the local
television viewing area, and we have conducted a survey of
viewers to determine which channel the viewers have been
watching. The first survey revealed that 40% of the viewers
watched station X and 60% watched station Y. Subsequent
surveys revealed that each week 15% of the X viewers
switched to station Y and 5% of the Y viewers switched to
station X.
We will use Transition matrices and Markov chains to make a
prediction about the future television market from this
information.
5. We assume that the 15%-5% switching trends will continue
indefinitely.
Let pk be a two-dimensional column vector whose entries give
the proportion of people who watch station X and the
proportion who watch station Y, in that order, during week k.
Thus, p0 = [0.4,0.6]t.
From the assumptions above, we see that the proportion of
people watching station X in week k = 1 will be 85% of the X
viewers in week k = 0 (which is 85% of 40%, or 34%) plus 5%
of the Y viewers in week k = 0 (which is 5% of 60%, or 3%) for
a total of 37%:
0.85(0.4) + 0.05(0.6) = 0.37.
6. Similarly,
0.15(0.4) + 0.95(0.6) = 0.63,
is the proportion of people watching station Y in week k = 1. So,
we have computed p1 = [0.37,0.63]t.
Similarly, we could compute the sequence of vectors p2, p3, ... ,
but this process would be quite tedious if we continued in the
manner above.
Andrei Markov (1856-1922) described a much more efficient
way of handling such a problem.
7. First, we think of the movement of viewers as being described by the
following array, which gives the weekly proportion of viewers who change
from one station to another:
From:
The matrix A of entries in the table is called the transition matrix for this
problem.
Looking carefully at the definition of matrix multiplication, we see that
p1=A*p0. Indeed, we see that, for k = 0, 1, 2, ..., and so on, pk+1=A*pk.
Each pk is called a probability vector, and the sequence, p0, p1, p2, p3, ..., is
called a Markov chain.
8. Because matrix multiplication is associative, we find
p2 = A*p1 = A(A*p0) = (A^2)*p0,
p3 = A*p2 = A((A^2)*p0) = (A^3)*p0,
and, in general,
pk = (A^k)*p0, for k = 1, 2, ...
Compute A^5. Then compute p5 directly from A^5 without
computing p1, ... , p4.
9. Suppose p is a probability vector with the property that A*p = p.
If this p describes the current viewers .
We can solve the equation A*p = p by rewriting it as A*p - p = 0 and factoring
out p.
We get the matrix equation (A - I)p = 0, where I is the 2x2 identity matrix.
Using our given transition matrix, solve this matrix equation for p.
From the solutions pick the one which is a probability vector (i.e., its entries
must sum to 1).
This vector is called the steady-state vector.
The fact that p is the steady-state vector does not imply that the Markov chain
will actually reach that state from p0.