Application to Markov Chains
Many natural phenomena progress through various stages and can be in a variety of states at each stage. For example, the weather in a given city progresses day by day and, on any given day, may be sunny or rainy. Here the states are “sun” and “rain,” and the weather progresses from one state to another in daily stages. Another example might be a football team: The stages of its evolution are the games it plays, and the possible states are “win,” “draw,” and “loss.”
The general setup is as follows: A “system” evolves through a series of “stages,” and at any stage it can be in any one of a finite number of “states.” At any given stage, the state to which it will go at the next stage depends on the past and present history of the system—that is, on the sequence of states it has occupied to date.
Markov Chain A Markov chain is such an evolving system wherein the state to which
it will go next depends only on its present state and does not depend on the earlier
history of the system.
(Note: The name honors Andrei Andreyevich Markov (1856–1922) who was a professor at the university in St. Petersburg, Russia.)
Even in the case of a Markov chain, the state the system will occupy at any stage is determined only in terms of probabilities. In other words, chance plays a role. For example, if a football team wins a particular game, we do not know whether it will win, draw, or lose the next game. On the other hand, we may know that the team tends to persist in winning streaks; for example, if it wins one game it may win the next game of the time, lose of the time, and draw of the time. These fractions are called the probabilities of these various possibilities. Similarly, if the team loses, it may lose the next game with probability (that is, half the time), win with probability , and draw with probability . The probabilities of the various outcomes after a drawn game will also be known.
We shall treat probabilities informally here: The probability that a given event will occur is the long-run proportion of the time that the event does indeed occur. Hence, all probabilities are numbers between and . A probability of means the event is impossible and never occurs; events with probability are certain to occur.
If a Markov chain is in a particular state, the probabilities that it goes to the various states at the next stage of its evolution are called the transition probabilities for the chain, and they are assumed to be known quantities. To motivate the general conditions that follow, consider the following simple example. Here the system is a man, the stages are his successive lunches, and the states are the two restaurants he chooses.
A man always eats lunch at one of two restaurants, and . He never eats at twice in
a row. However, if he eats at , he is three times as likely to eat at next time as at .
Initially, he is equally likely to eat at either restaurant.
- (a)
- What is the probability that he eats at on the third day after the initial one?
- (b)
- What proportion of his lunches does he eat at ?
The table of transition probabilities follows. The column indicates that if he eats at
on one day, he never eats there again on the next day and so is certain to go to
.
The column shows that, if he eats at on one day, he will eat there on the next day of the time and switches to only of the time.
The restaurant he visits on a given day is not determined. The most that we can expect is to know the probability that he will visit or on that day.
Let denote the state vector for day . Here denotes the probability that he eats at on day , and is the probability that he eats at on day . It is convenient to let correspond to the initial day. Because he is equally likely to eat at or on that initial day, and , so Now let
denote the transition matrix. We claim that the relationship holds for all integers . This will be derived later; for now, we use it as follows to successively compute . Hence, the probability that his third lunch (after the initial one) is at is approximately , whereas the probability that it is at is . If we carry these calculations on, the next state vectors are (to five figures): Moreover, as increases the entries of get closer and closer to the corresponding entries of . Hence, in the long run, he eats of his lunches at and at .Example 007199 incorporates most of the essential features of all Markov chains. The general model is as follows: The system evolves through various stages and at each stage can be in exactly one of distinct states. It progresses through a sequence of states as time goes on. If a Markov chain is in state at a particular stage of its development, the probability that it goes to state at the next stage is called the transition probability. The matrix is called the transition matrix for the Markov chain. The situation is depicted graphically in the diagram.
We make one important assumption about the transition matrix : It does not depend on which stage the process is in. This assumption means that the transition probabilities are independent of time—that is, they do not change as time goes on. It is this assumption that distinguishes Markov chains in the literature of this subject.
Suppose the transition matrix of a three-state Markov chain is If, for example, the
system is in state 2, then column 2 lists the probabilities of where it goes next. Thus,
the probability is that it goes from state 2 to state 1, and the probability is that it
goes from state 2 to state 2. The fact that means that it is impossible for it to go
from state 2 to state 3 at the next stage.
Consider the th column of the transition matrix .
If the system is in state at some stage of its evolution, the transition probabilities represent the fraction of the time that the system will move to state 1, state 2, …, state , respectively, at the next stage. We assume that it has to go to some state at each transition, so the sum of these probabilities is : Thus, the columns of all sum to and the entries of lie between and . Hence is called a stochastic matrix.As in Example 007199, we introduce the following notation: Let denote the probability that the system is in state after transitions. The matrices
are called the state vectors for the Markov chain. Note that the sum of the entries of must equal because the system must be in some state after transitions. The matrix is called the initial state vector for the Markov chain and is given as part of the data of the particular chain. For example, if the chain has only two states, then an initial vector means that it started in state 1. If it started in state 2, the initial vector would be . If , it is equally likely that the system started in state 1 or in state 2. Let be the transition matrix for an -state Markov chain. If is the state vector at
stage , then
for each .
- Heuristic Proof
- Suppose that the Markov chain has been run times, each time starting with the same initial state vector. Recall that is the proportion of the time the system goes from state at some stage to state at the next stage, whereas is the proportion of the time it is in state at stage . Hence is (approximately) the number of times the system is in state at stage . We are going to calculate this number another way. The system got to state at stage through some other state (say state ) at stage . The number of times it was in state at that stage is (approximately) , so the number of times it got to state via state is . Summing over gives the number of times the system is in state (at stage ). This is the number we calculated before, so Dividing by gives for each , and this can be expressed as the matrix equation .
If the initial probability vector and the transition matrix are given, Theorem 007270 gives , one after the other, as follows:
Hence, the state vector is completely determined for each by and . A wolf pack always hunts in one of three regions , , and . Its hunting habits are as
follows:
- (a)
- If it hunts in some region one day, it is as likely as not to hunt there again the next day.
- (b)
- If it hunts in , it never hunts in the next day.
- (c)
- If it hunts in or , it is equally likely to hunt in each of the other regions the next day.
If the pack hunts in on Monday, find the probability that it hunts there on Thursday.
The stages of this process are the successive days; the states are the three regions.
The transition matrix is determined as follows (see the table): The first habit asserts
that . Now column 1 displays what happens when the pack starts in : It never goes to
state 2, so and, because the column must sum to , . Column 2 describes what
happens if it starts in : and and are equal (by habit 3), so because the column sum
must equal . Column 3 is filled in a similar way. Now let Monday be the initial stage.
Then because the pack hunts in on that day. Then , , and describe Tuesday,
Wednesday, and Thursday, respectively, and we compute them using Theorem 007270.
Hence, the probability that the pack hunts in Region on Thursday is .
Steady State Vector
Another phenomenon that was observed in Example 007199 can be expressed in general terms. The state vectors were calculated in that example and were found to “approach” . This means that the first component of becomes and remains very close to as becomes large, whereas the second component gets close to as increases. When this is the case, we say that converges to . For large , then, there is very little error in taking , so the long-term probability that the system is in state 1 is , whereas the probability that it is in state 2 is . In Example 007199, enough state vectors were computed for the limiting vector to be apparent. However, there is a better way to do this that works in most cases.
Suppose is the transition matrix of a Markov chain, and assume that the state vectors converge to a limiting vector . Then is very close to for sufficiently large , so is also very close to . Thus, the equation from Theorem 007270 is closely approximated by
so it is not surprising that should be a solution to this matrix equation. Moreover, it is easily solved because it can be written as a system of homogeneous linear equations with the entries of as variables.In Example 007199, where , the general solution to is , where is a parameter. But if we insist that the entries of sum to (as must be true of all state vectors), we find and so as before.
All this is predicated on the existence of a limiting vector for the sequence of state vectors of the Markov chain, and such a vector may not always exist. However, it does exist in one commonly occurring situation. A stochastic matrix is called regular if some power of has every entry greater than zero. The matrix of Example 007199 is regular (in this case, each entry of is positive), and the general theorem is as follows:
Let be the transition matrix of a Markov chain and assume that is regular. Then
there is a unique column matrix satisfying the following conditions:
- (a)
- .
- (b)
- The entries of are positive and sum to .
Moreover, condition 1 can be written as
and so gives a homogeneous system of linear equations for . Finally, the sequence of state vectors converges to in the sense that if is large enough, each entry of is closely approximated by the corresponding entry of .This theorem will not be proved here. The interested reader can find an elementary proof in J. Kemeny, H. Mirkil, J. Snell, and G. Thompson, Finite Mathematical Structures (Englewood Cliffs, N.J.: Prentice-Hall, 1958).
If is the regular transition matrix of a Markov chain, the column satisfying conditions 1 and 2 of Theorem 007400 is called the steady-state vectorfor the Markov chain. The entries of are the long-term probabilities that the chain will be in each of the various states.
A man eats one of three soups—beef, chicken, and vegetable—each day.
He never eats the same soup two days in a row. If he eats beef soup on a
certain day, he is equally likely to eat each of the others the next day; if he
does not eat beef soup, he is twice as likely to eat it the next day as the
alternative.
- (a)
- If he has beef soup one day, what is the probability that he has it again two days later?
- (b)
- What are the long-run probabilities that he eats each of the three soups?
The states here are , , and , the three soups. The transition matrix is given in the
table. (Recall that, for each state, the corresponding column lists the probabilities for
the next state.) If he has beef soup initially, then the initial state vector is
Then two days later the state vector is . If is the transition matrix, then
so he eats beef soup two days later with probability . This answers (a.) and
also shows that he eats chicken and vegetable soup each with probability
.
To find the long-run probabilities, we must find the steady-state vector . Theorem 007400 applies because is regular ( has positive entries), so satisfies . That is, where
The solution is , where is a parameter, and we use because the entries of must sum to . Hence, in the long run, he eats beef soup of the time and eats chicken soup and vegetable soup each of the time.Practice Problems
In each case find the steady-state vector and, assuming that it starts in state 1, find
the probability that it is in state 2 after transitions.
A fox hunts in three territories , , and . He never hunts in the same territory on two
successive days. If he hunts in , then he hunts in the next day. If he hunts in or , he
is twice as likely to hunt in the next day as in the other territory.
Assume that there are three social classes—upper, middle, and lower—and that
social mobility behaves as follows:
- (a)
- Of the children of upper-class parents, remain upper-class, whereas become middle-class and become lower-class.
- (b)
- Of the children of middle-class parents, remain middle-class, whereas the others are evenly split between the upper class and the lower class.
- (c)
- For the children of lower-class parents, remain lower-class, whereas become middle-class and upper-class.
The prime minister says she will call an election. This gossip is passed from person
to person with a probability that the information is passed incorrectly at any stage.
Assume that when a person hears the gossip he or she passes it to one person who
does not know. Find the long-term probability that a person will hear that there is
going to be an election.
John makes it to work on time one Monday out of four. On other work days his
behaviour is as follows: If he is late one day, he is twice as likely to come to work on
time the next day as to be late. If he is on time one day, he is as likely to be late as
not the next day. Find the probability of his being late and that of his being on
time Wednesdays. Probability of being late , Probability of being on time
Suppose you have cent and match coins with a friend. At each match you
either win or lose cent with equal probability. If you go broke or ever get
cents, you quit. Assume your friend never quits. If the states are 0, 1, 2,
3, and 4 representing your wealth, show that the corresponding transition
matrix is not regular. Find the probability that you will go broke after
matches.
A mouse is put into a maze of compartments, as in the diagram. Assume that he
always leaves any compartment he enters and that he is equally likely to take any
tunnel entry.
If a stochastic matrix has a on its main diagonal, show that it cannot be regular.
Assume it is not .
If is the stage- state vector for a Markov chain, show that holds for all and (where
is the transition matrix).
A stochastic matrix is doubly stochastic if all the row sums also equal . Find the
steady-state vector for a doubly stochastic matrix.
Consider the stochastic matrix
, where and .
- (a)
- Show that is the steady-state vector for .
- (b)
- Show that converges to the matrix by first verifying inductively that for . (It can be shown that the sequence of powers of any regular transition matrix converges to the matrix each of whose columns equals the steady-state vector for .)
Text Source
This application was adapted from Section 2.9 of Keith Nicholson’s Linear Algebra with Applications. (CC-BY-NC-SA)
W. Keith Nicholson, Linear Algebra with Applications, Lyryx 2018, Open Edition, p. 134