Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Consider $A$ as a matrix, that when normalized represents an finite-state time-homogeneous Markov chain $M$ with entries $0\leq p_{i,j}\leq 1$, where each row sums up to $1$. We can also assume that $M$ is irreducible and aperiodic, hence it has an unique steady-state distribution $\pi$ ($\pi=\pi P$).

Now consider a parameter $0<\delta\leq 1$ is added to the diagonal elements of the matrix $A$ Markov chain. So $a'_{i,j}=a_{i,j}+\delta$, from where we get $A'$. All rows of $A'$ are then normalized to sum up to one. From it we get new transition matrix $P'$. We again compute the steady-state distribution of this newly constructed chain $M'$ and call it $\pi'$.

If we observe what is happening with the elements of $\pi'$ if $\delta$ is increased, we can see that each value in the distribution monitonically increases or decreases.

Now my question. How could I prove that this is true for all $M$ with above properties (or does anyone have an counter example?)?

It also seems that $\pi'$ converges to some $\pi_C$ if I set $\delta$ to some arbitrarily large number (not compute the limit). Is there a direct formula to compute such $\pi_C$?

For example, lets take this matrix $A$ for out initial matrix:

$$ A = \begin{bmatrix} 0.056&0.084&0.242&0.255\\ 0.071&0.056&0.249&0.210\\ 0.086&0.095&0.056&0.080\\ 0.115&0.102&0.101&0.056 \end{bmatrix} $$

Matrix $P$ for our Markov chain $M$, that we get by normalizing rows of $A$ (note that values are rounded to three decimal places):

$$ P = \begin{bmatrix} 0.087&0.132&0.380&0.401\\ 0.121&0.095&0.426&0.358\\ 0.271&0.300&0.175&0.254\\ 0.309&0.272&0.270&0.149 \end{bmatrix} $$

Steady state distribution for chain $M$ is $$\pi= \begin{bmatrix} 0.211&0.213&0.298&0.278 \end{bmatrix}$$

$P'$ for $\delta=0.5$ (diagonal elements of $A$ are increased by $0.5$ and $A'$ is normalized by rows)

$$ P' = \begin{bmatrix} 0.489&0.074&0.213&0.224\\ 0.065&0.512&0.229&0.193\\ 0.105&0.116&0.681&0.098\\ 0.132&0.117&0.116&0.636 \end{bmatrix} $$

Corresponding steady-state distribution for $M'$ (with $\delta=0.5$) is: $$\pi'=\begin{bmatrix}0.172 & 0.180 & 0.351 & 0.296\end{bmatrix}$$ And also I computed $\pi_C$ (with $\delta=100$) which is: $$\pi_C= \begin{bmatrix} 0.139&0.153&0.395&0.312 \end{bmatrix}$$

Below is a plot of steady-state distribution with increasing $\delta$. We can observe that each value in $\pi$ either monotonically increases or monotonically decreases.

plot of the steady-state distribution with increasing $\delta$

And a little background... I am constructing a new multi-criteria decision making method, which is in some sense simmilar to PageRank algorithm. We are trying to evaluate given alternatives, based on user preference. I discovered that increasing $\delta$ in some sense increases the separation between the alternatives.

-edit comment- Sorry for the not so clear explanation and the error that I made. Hope we can now understand eachother.

share|improve this question
    
How is $0<\delta\leq1$ compatible with $\delta\to\infty$? If $\delta\to\infty$, then $M'$ should converge to the identity matrix which does not have a unique stationary distribution. –  Eckhard Jan 3 '13 at 10:40
    
Yes it is not compatible. In my method I am using $0< \delta \leq 1$, but I noticed that if I increase $\delta$ to some larger values (e. g. 100), the values converge to above mentioned $\pi_C$ and not to $[0.25,0.25,0.25,0.25]$. Sorry for the confusion, I am a decision support grad students and mathematics was just a part of my undergrad program. –  Nejc Jan 3 '13 at 10:58
    
I'm a bit confused: if I'm not mistaken, the new matrix has to have the form $$ P_\delta := \frac{1}{1+\delta}(P+\delta I) $$ however, $\pi_\delta P_\delta = \pi_\delta$ if and only if $\pi_\delta P = \pi_\delta$ which implies that the stationary distribution is independent of $\delta$. I guess, it's not the case though. –  Ilya Jan 3 '13 at 11:44
    
@Ilya: That observation is actually compatible with a few short calculations I did. –  Eckhard Jan 3 '13 at 11:47
    
@Eckhard: I think, that using the average return time characterization of the steady-state distribution, one obtains that $\pi$ is independent on the diagonal elements and depends only on ratios of the off-diagonal ones. Thus, my previous guess was incorrect and indeed, $\pi_\delta \equiv \pi$ for $\delta>0$ whereas $\pi_\infty$ in undefined as $P_\delta \to I$. –  Ilya Jan 3 '13 at 11:53
show 5 more comments

1 Answer 1

up vote 1 down vote accepted

So far I can give only a partial answer. Let $A$ be the original matrix and let $$ \alpha(i):=\sum_j a_{ij} $$ be the sum of its rows. For simplicity assume that all $a_{ij}$ are positive - just to be sure that the correspondent stochastic matrix is irreducible and aperiodic. Define $$ K:=\mathrm{diag}(k(1),\dots,k(n)) $$ with $k(i) = 1/\alpha(i)$ to be the normalizing matrix, so that $P:=KA$ is a stochastic matrix. Denote the correspondent steady-state distribution by $\pi$. Now let $\delta>0$ be the perturbation of $A$ and define $$ A_\delta:=A+\delta I $$ where $I$ is the identity matrix, so $\alpha_\delta(i) = \alpha(i)+\delta$ and for the normalizing constants $$k_\delta(i) = \frac{k(i)}{1+\delta k(i)}. $$ The new stochastic matrix is then $P_\delta = K_\delta A_\delta$, so for the new steady-state distribution we have $$ \pi_\delta(K_\delta A_\delta) = \pi_\delta. $$ If we denote $\nu_\delta:=\pi_\delta - \pi$, by opening all the brackets we obtain the following equation on $\nu_\delta$: $$ \tag{1} \nu_\delta(P_\delta - I) = \pi(P-P_\delta). $$ Although equation $(1)$ does not have the unique solution as $P_\delta$ admits the eigenvalue $1$ and thus $(P_\delta - I)$ is not invertible, we also know that $\sum_i \nu_\delta(i) = 0$ since $\nu_\delta = \pi_\delta-\pi$. Using this last condition, we can obtain the desired solution of $(1)$, however I am not sure whether it is possible to have the solution in a neat form. As a special case, when $A$ is itself stochastic then $\nu_\delta\equiv 0$ and thus $\pi_\delta\equiv \pi$.

share|improve this answer
    
I suppose you meant $K_\delta$ instead of $N_\delta$... Anyway, i do not understand what you meant by $\sum \nu_\delta = 0$. Isn't $\nu_\delta$ fixed for some $\delta$. How can we sum them? –  Nejc Jan 4 '13 at 10:49
    
@Nejc: I updated the notation a bit. I meant, that the sum of elements of the vector $\nu_\delta$ is always zero as $\nu_\delta$ is a difference between two probability distributions. –  Ilya Jan 4 '13 at 11:00
    
OK, now I understand everything you have written. Can I now get the monotonicity per position in distribution (for instance for $i$-th element), if I rewrite the formula for index $i$? Instead of $\nu_\delta$ I would look at $\nu_\delta(i)$. –  Nejc Jan 4 '13 at 12:57
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.