and markov with (a’m,n)-matrix l. abolnikov
TRANSCRIPT
Journal ofApplied Mathematics and Simulation Vol. 1, Issue 1, 1987.
NECESSARY AND SUFFICIENT CONDITIONS FOR THE ERGODICITY OFMARKOV CHAINS WITH TRANSITION Am,n (A’m,n)-MATRIX
L. ABOLNIKOVLoyola Marymount University
Los Angeles, California
and
A. DUKHOVNYScientific Research Institute on Automation
Odessa, U.S.S.R.
ABSTRACT
This paper isolates and studies a class of Markov chains with aspecial quasi-triangular form of the transition matrix [so-called Am,n(A’m,n)-matrix]. Many discrete stochastic processes encountered in
applications (queues, inventories and dams) have transition matriceswhich are special cases of a Am,n (A’m,n)-matrix. Necessary andsufficient conditions for the ergodicity of a Markov chain withtransition Am,n (A’m,n)-matrix are determined in the article in two
equivalent versions. According to the first version, these conditions areexpressed in terms of certain restrictions imposed on the generatingfunctions A (x) of the elements of the i-th row of the transition matrix,
0, 1, 2 in the other version they are connected with thecharacterization of the roots of a certain associated function in the unitcircle of the complex plane. Results obtained in the article generalize,complement, and refine similar results existing in the literature.
Key words: Markov chains, queues, inventories, dams, ergodicity,necessary condition, sufficient condition
AMS subject classification: 60J, 60K25
1. INTRODUCTION
A class of Markov stochastic processes with transition (or infinitesimal) matrix of the
14 JOURNAL OF APPLIED MATHEMATICS AND SIMULATION Vol. 1, Issue 1, 1987
(1.1) G
go,o go,1 go,2
gl,0 al,1 al,2
g2,0 g2,1 g2,2
gin,0 gin,1 gm,2
0 gm+l,1 gm+l,2
0 0 gm/2,2
(a "Am-matrix" or of the form G’ (a A’m-matrix*) was first investigated for the case offinite set of states by Abolnikov in 1]. A method was proposed for reducing the problemof finding the state probabilities for such a process to the simpler problem of finding thestate probabilities for an auxiliary process whose transition matrix is related to the Am(A’m)-matrix of the original process. (This was called, "the continuation method"). Inparticular, this method provides the solution for the finite dimensional case if a solution isavailable for the infinite-dimensional case. Markov chains with an infinite transition matrixwhich is a slightly modified analog of (1.1) was investigated by Neuts 11 ]. In this paper anew effective approach for finding the steady-state probabilities of the process wasproposed.
However, the questions related to the conditions for the ergodicity of suchprocesses were not considered in [1,11]. The present paper is devoted to this problem. Asthe main object of consideration we study a countable Markov chain whose transitionmatrix has form (1.1). Keeping in mind the specific features of processes which arefrequently encountered in application, we do not impose any restrictions on the transientprobabilities gij if < m j < n ), but suppose at the same time that the gij for > n > ndepend only on the difference of the indices and j. That is, the transition matrix ofprocesses considered in the paper has the following form:
*G’ is the transpose of G.
Necessary and Sufficient Conditions for the Ergodicity of Markov Chains:Abolnikov and Dukhovny 15
(1.2)
a00
al0
A= ano0
0
alo ao n-m ao n-m+l ao n-m+2
all al n-m al n-m+2 al n-m+2
anl an n-m an n-m+l an n-m+2
0 0 ko k
0 0 0 k0
or its transpose. Many stochastic processes encountered in applications have transitionmatrices which are special cases of A (or A’). As examples there are embedded Markovchains describing the behavior of a queue in the queueing systems M/G/l, G/M/1, G/M/nwith bulk arrivals, bulk service, with feedback, with a threshold, with "warmup", withswitchings, with hysteresis service, and with queue buildup; the state of a storage in thetheory of inventory control; the state of a dam in the Moran problem. In this papernecessary and sufficient conditions for the existence of the ergodic distribution of theprobabilities of states of a Markov chain with infinite transition Am,n (A’m,n-matrix) aredetermined. These conditions then are connected with the existence and the location of theroots of the function
EZm
k’ (z), k (z) k z.i=0
Some special cases (as a rule, connected with queueing systems) of this problem yieldingmostly sufficient conditions were discussed in Abolnikov/Dshalalow [2], Abolnikov [3],Bahary/Kolesar [4], Bailey [5], Delbrouck [6], Harris/Marlin [7], Kopocinska [8], andLoris-Teghem [9].
However, the arguments given suffer either from insufficient analytical justificationor from considerable restriction of the generality. The same can be said about thecharacterization of the roots of the function zrn- k (z)in the circle z < 1. InBahary/Kolesar [4] the question of determination of the roots of zm k (z) is left open. InHarris/Marlin [7] the proof of the statement about the number of the roots of the functionzrn k (z) is given only under the suggestion of the possibility of the analytical extension ofthe k (z) outside of the unit circle. The solution of the problem was obtained inAbolnikov/Dshalalow [2] and Abolnikov [3] only for the sufficient conditions and m 1.
2. DEFINITIONS AND TERMINOLOGY
Definition 1. We shall say that a finite (not necessarily square) or an infinite stochasticmatrix A aij is a Am,n-matrix (resp. A’m,n-matrix), n > m >_ 1, if aij 0 for > n and
> m (resp. aij 0 for > m and > m). If in addition aij kj_i+rn for > n, > m(resp. aij ki_j/m for > n, > m) then the matrix A is called homogeneous.
16 JOURNAL OF APPLIED MATHEMATICS AND SIMULATION Vol. 1, Issue 1, 1987
Another definition of a homogeneous Am,n-matrix (that is more convenient in thecase of infinite matrices) can be given in the following way. Let
A (z) aij zi,j=o
be a generating function of elements aij of the i-th row of a stochastic matrix A aij ),0, 1, 2 The set of functions A (z), 0, 1, 2, obviously, completely
determines the matrix A.
Definition 2. An infinite stochastic matrix A aij ), i, j 0, 1, 2 is called ahomogenous Am,n-matrix if
A (z) zi-m
r=OkrZ, i=n+ 1, n+ 2
Let r, r 0, 1, 2 be a homogeneous Markov chains with transition Am,n-matrix A aij ), i, j 0, 1, 2, We denote
(0)Pi-(r) p { r } ;/i Pi 0, 1, 2,
P(r) _(r) (r) _(r)=(P0 ’Pl ’P2 ),r=0, 1,2,...; zi(Z) gi
i=O
(r) rU (x) Pi X; W(X,Z)= Ui(X) Zi" Ixl<l Izl<l
i-0
U (X) U0 (X), U (x), U2 (x),(r)
p x;r=0
(r) Elim Pi Pi’ P (P0’ Pl’ P2 ) ;P(z) Pi z.r-, i=O
Let us point out some useful relationships among the vectors
p, P(r), u (x) and the matrix A, which will be used later.
First, since the chain r is homogeneous(2.1) pr+l pr AIf the chain r is ergodic then(2.2) P=PA,and, finally, it is easy to see that(2.3) u (x)= xu (x) A + p0.
Necessary and Sufficient Conditions for the Ergodicity of Markov Chains:Abolnikov and Dukhovny 17
The problem consists in finding necessary and sufficient conditions for the ergodicity of thechain r.
3. NECESSARY AND SUFFICIENT CONDITIONS FOR THE ERGODICITY OFMARKOV CHAINS WITH TRANSITION Am,n and A’m,n-MATRICES.
Theorem 1. A Markov chain r with transition irreducible and nonperiodic homogeneous
Am,n-matrix A =( aij ), i, 0, 1, 2,... is ergodic if and only if
(3 1) A’ (1)<oo i=0,1 2, n,
and(3.2) k’ (1) < m.Proof"Sufficiency. Let us put xj j, j 0, 1, 2,...; then it follows from (1.2) that
f A’ (1)-i, ifi= 0, 1, n
Z alj xj xj=0
k’ (1) m, if n+ 1,n+2,...
By the corollary of Moustafa [10] to Foster’s theorem we conclude that (3.1) and (3.2) are
sufficient for the ergodicity of the chain r.Necessity. Let us first prove the following lemma.Lemma. Consider a function
H (z) h z,i=0
which converges for z 1. Suppose that h < 0 if 0, 1, 2,..., k, and h > 0 if k+ 1,k+2,Assume further that H 0. If there exists z0 e (0,1), such that H (z0) > 0, the H (z) > 0
for any z e (z0, 1].Proof. Under the conditions of the lemma the following chain of inequalities for anyz e (z0, 1] is valid:
_z-(k+l) Z hizi<-z;(k+l) Z hizi0 _< -z;(k+l) Z hizi0 < z-(k+l) Z hi zii=0 i=0 i=k+l i=k+l
Since H (z) 0, at least one of the hi, 0, 1, 2,... is not equal to 0. Therefore, at leastone of the extreme inequalities in the chain is strict. The lemma then follows.
We now return to the proof of the theorem. If the chain {r is ergodic, then there exists astrictly positive vector P P0, Pl, P2, such that P PA.Multiplying this equation by the vector (0, 0, 0, zm, zre+l, with O’s in the first n + 1
places, where z [0,1] and taking into account (1.1) we obtain
18 JOURNAL OF APPLIED MATHEMATICS AND SIMULATION Vol. 1, Issue 1, 1987
ZH (z)
Pn+i+l mi=0 Z k (z)
where
n m-1 m-i-1
H(z) hi Zi zm Pi n+j+l- Pn+i+l hj.i=0 i=0 =0 i=0 j=0
It is easy to see that
(3.3) 0 < limH (z)
Pn+i+l < 1,zl- z
TMk (z) i-0
and since zm k (z)] z=l 0 then
(3.4) h (1)= lim H (z) 0.
It follows that H (z) < 0 for any z (0,1). Indeed, suppose that there exists z0 (0,1)such that H (z0) > 0. Then, since
m-1
H (0) h0 -k0
_Pn+i+l < 0
i-0
the Lemma implies h (1) > 0, which contradicts (3.4).
We shall need an additional property ofH (z), namely
(3.5) 0 < H’(1) <To prove this let us notice that k’ (1) < m, because otherwise there exists 5 > 0 such that
zm k (z) > 0 for any z e (1-5,1). Therefore,
H (z)< 0 in the same region, which is impossible because of (3.3).
mz -k(z)
Now to prove the right inequality in (3.5) it is enough to pass to the limits in (3.3) as z1-. Let us first consider the left inequality in (3.5) for m 1. It is obvious that ih > 0 for
any 0, 1, 2 On the other hand, since the chain r is irreducible, among thenumbers a n+j+l, 0, 1, 2, h, 0, 1, 2 so there exists at least one which isstrictly positive. It follows that there exists 0, such that
0 hi0 > 0 and, therefore, H’ (1) > 0 hi0 > 0.
If m > 1 thenm-2
H’(0) h -k1L Pn+i+l <0"i=0
In addition, since H (0) < 0 and H (1) 0, there exists z0 (0,1) such that H’ (z0) > 0. Bythe lemma it follows that H’ (1) > 0 and (3.5) is proved.
Necessary and Sufficient Conditions for the Ergodicity of Markov Chains:Abolnikov and Dukhovny 19
Now both of the statements (3.1) and (3.2) of the theorem 1 easily follow from(3.3) and (3.5). In fact, if we suppose that there exists i, such that A i’ (1) then, as canbe seen from the form of H (z), lim H’ (z) oo, which contradicts (3.5). Finally, thestatement (3.2) of the theorem can be obtained by passing to the limits in (3.3) as z -- 1-and taking into account (3.5). The theorem is proven.
Next we establish an analogous statement for a Markov chain r, r 0, 1, 2,with transition matrix of A’m,n-type.
Theorem 2: A Markov chain r, r 0, 1, 2, with transition irreducible aperiodic
homogeneous A’m,n-matrix A aij ), i, j 0, 1, 2, is ergodic if and only if
(3.6) k’ (1) > m
Proof"Sufficiency.Let us put x max {j-n,0} 0, oo. Then taking into account (1.2) it is easy to show that
0 ,ifi <n- m
a x x Di-n/m k (z) zif > n m
-0 z )z(1-z
where Dix is an operator defined by the relation
(3.7)1 )iF (x,y)Dix F (x,y)
Ox x=0
Note that DK (z)- z
TM j0 kj _> 0,
z 1-z ’j=0
and, therefore, if < m,m
(3.8) D k(z)-zm, iz k()-z _>0,z )2 z1-z j=0
and if > m 1m
Dk (z)- z
m
(3.9) Di+1 k(z)’- z_<
z )2 z )2(1-z (1-zHence, the sequence
ifi<m
k.- 1 <O, ifi> m
20 JOURNAL OF APPLIED MATHEMATICS AND SIMULATION Vol. 1, Issue 1, 1987
ri k (z) zTM
(l-z)2
i=1
either has a finite limit as --) ,,, or tends to +,. If k’(1) > m then using a Tauberiantheorem in the case of the finite limit we obtain:
rn Ill
lim Dk (z) z
limk (z) z
-k’ (1) + m < 0z 2 1-zi---), 1- z ) z-l-0
so that in this case there exist s and M < 0 such that
Z x.-x.<M, ifi>s.aijj=0
Again appealing to the criterion of Moustafa 10] we conclude that (3.6) is sufficient for theergodicity of the chain {r.Necessity. If the chain r is ergodic then there exists a strictly positive probability vector P
(P0, Pl, P2, such that P PA. Multiplying this equation by the vector (x 0, xl, x2,), where xj max {j-n,0}, j 0, 1, 2 after some simple transformations, we
obtain
rn
Dk(z)- z
(3.10) Pn+l-m+l zi=0 (l-z)
2
Suppose that k’(1) < m. Then
rn rn
limDk(z)- z
lim *’-’__ --Dzk(z)-z
k(l)+m>0.z )2 1 zioo 1 z i---)oo j=0
Together with (3.8) and (3.9), this implies thatrn
D .k(z)-z >0_z (1_z)2
for all 0, 1, 2,..., which contradicts (3.10). The theorem is proved.
4. ERGODICITY OF THE MARKOV CHAIN WITH TRANSITION Am,n(A’m,n)-MATRIX
AND CHARACTERIZATION OF THE ROOTS OF THE FUNCTION zm k (z).
The conditions of ergodicity proved in section 3 are closely connected with thenumber and location of the roots of the function z rn k (z). More precisely, the followingtheorem is valid.
Theorem 3
Necessary and Sufficient Conditions for the Ergodicity of Markov Chains:Abolnikov and Dukhovny 21
A. If k’(1) < m, then the function zm k (z) has exactly m roots (countingmultiplicities) in the disc z < 1. The roots lying on the boundary are simple and,for some integer r, they are all r-th roots of 1.
g. If k’(1) > m, then the function zm k (z) has exactly m roots (countingmultiplicities) in the open disc z I< 1; on the boundary there can be r additionalsimple roots which are all the r-th roots of 1, where r is an integer, 1 < r _< m.
Proof: We will need the following result.Lemma. The function zm k (z) has roots on the unit circle z 1 if and only if there
exists r such that m r and for any such that k 0, r. If these conditions are satisfied,
then all roots of this kind coincide with the roots of the equation zr 1 0, where r is themaximal number having this property.
Proof: Suppose that z0 is a root of the equation zm k (z) 0 such that z01 1. Then
k (z0)I- zl- 1.
On the other hand,
Ik(z0) I= IE kizi01 < kilz01i=li=0 i=0
and the equality is attained if and only if
mz0 Iz01 lforanyi, suchthatkl 0andz0 1.
The realization of both of these conditions simultaneously is possible only if the conditionsof the Lemma are satisfied. In this case, obviously,
z0 1.
Otherwise, if the conditions of the Lemma are satisfied then for any root z0 of the
equation zr 1 0 we obtainrn
k(z0) 1 z0,
which implies that the root z0 is a root of the equation k (z) zm 0. The Lemma isproved.
Now we return to the proof of the theorem.
A. First we suppose that the number r, appearing in the Lemma, is 1. Then thefunction zm k(z) has only 1 root on the unit circle. This root is equal to 1 and is simple,since k’(1) m’. We will prove that in this case zm k (z) has exactly m- 1 roots inside F:Iz 1. Consider an auxiliary function
1- z-mk (z)f (z)
-11-z
22 JOURNAL OF APPLIED MATHEMATICS AND SIMULATION Vol. 1, Issue 1, 1987
Clearly, f (z) 0 for all z, zl 1 since the numerator of f (z) may be zero only ifz 1, but f (1) m -k’(1) > 0. Suppose now that Indr f (z) denotes the difference between
the number of the roots and the number of the poles of the function f (z) inside F.
By the principle of the argument:
[ ] (),(4.1) Indrf(Z -.A Argf(z) Ind 1-z-ink(z) -Ind 1-z-1
where Ar Arg f (z) is the increment of the argument of f (z) when the argument of z ei
increases from 0 to 2ft.
Let us consider the right side of (4.1). Since Arg f (1) 0, it is easy to notice that
[ (ei)] lim [1 e-i] =+lim Arg 1 e-imk(I)-)_+O --->+0 "’"At the same time if
-imk ( )(0,2), e eia’
< 1 and, hence,
It follows that
z-1Indr[1-z-mk(z)] =-= Indr[ 1- ]and, therefore,
From the fact that
Indrf(Z) 0
f (z)rnz k (z)rrrl
z (z-l)
has exactly m 1 poles inside F, we conclude that the number of roots of f (z) inside F isalso m- 1.
Thus, the total number of roots (together with z 1) of the function zm k (z) is theunit disc z < is equal to m.
(4.2)
Now let us consider the general case r > 1. We introduce the function
m
F(z) z k.zi=0
By the definition of r it follows that all exponents of z in (4.2) are integers.Applying the previous reasonings to (4.2) we obtain that f (z) has exactly
Necessary and Sufficient Conditions for the Ergodicity of Markov Chains:Abolnikov and Dukhovny 23
m1 roots in the region zl < i and one root z 1 on the boundary.
The set of all r-th roots of the roots of F (z) gives us all roots of zm k (z). On the otherhand, it is obvious, that any root of zm k (z) raised to the r-th power is a root of F (z).Therefore, the set of all roots of zm k (z) is described completely. The number of them is
m; m r are inside the region z <- 1, and r roots are on the boundary.
Suppose now that z0 is a root of zm k (z) such that z0l 1. Then:
ik,(z0) l=li ikizi-ll= iliki ikiZo, -< i=l
k’(1)_Z =lk’(1) l<rn
It follows that k’(z0) e mzm-l, which implies that all roots on the boundary are simple. PartA of the theorem is proved.
Next we shall prove the part B. Since k’(1) > m, there exists > 0 such that
k(p) < pm for any p e [1 8,1]. Therefore, the inequality k (z) <1 zrn is correct for
any z, such that zl p. Using the theorem of Rouche, we obtain that zTM k (z) has
exactly m roots in the region zl < 1. The number of roots on the boundary, as before,depends on the value of r. If r > 1 then, by the lemma, there are r roots on the boundarywhich represent a group of r-th roots of 1. The simplicity of these roots is proved asbefore.
REFERENCES
[1]
[2]
[3]
[41
[5]
[61
[7]
Abolnikov, L.M. Investigation of a Class of Discrete Markov Processes. Izv.ANSSR, Tekhnicheskaya Kibernetika, N2:69-82 (1977). (English translation:Engineering Cybernetics, 15(1977)N2:51-63).
Abolnikov, L.M., Dshalalow, E. A. Feedback Queueing Systems: Duality Principleand Optimization. Automation and Remote Control 39:11-20 (1978).
Abolnikov, L. M. A Single Level Control in Moran’s Problem with Feedback.Operations Research Letters 2, N1:16-19 (1983).
Bahary, E., Kolesar, P. Multilevel Bulk Service Queues. Oper. Res. 20:407-420(1972).
Bailey, N. T. On Queueing Process with Bulk Service. J. Roy. Statist. Soc. B.16:80-87 (1954).
Delbrouek, L. A Feedback Queueing System with Batch Arrivals, Bulk Service,and Queue-Dependent Service Time. J. Assoc. Comput. Mach. 17:314-323 (1970).
Harris, C. M., Marlin, P. E. A note on Feedback Queues with Bulk Service. J.Assoc. Comput. Mach. 19:727-733 (1972).
24 JOURNAL OF APPLIED MATHEMATICS AND SIMULATION Vol. 1, Issue 1, 1987
[8]
[9]
[10]
[11]
Kopocinska, I. GI/M/I Queueing Systems with Service Rates Depending on theLength of Queue. Zastosow. Matem. 11:265-279 (1970).
Loris-Teghem, J. Condition Necessaire d’Ergodisme pour un ProcessusStochastique lie a un Systeme d’Attente a Arrivees et Services en Groupesd’Effectif Aleatore. Bull. C1. Sci. Acad. Roy. Belg. 52:382-389 (1966).
Moustafa, M.D. Input-Output Markov Processes. Proc. Koninkijke NederlandeAkad. Wetenschappen 60:112-118 (1957).
Neuts, M. F. Queues Solvable without Rouche’s Theorem. Operat. Res.27(4):767-781 (1979).
Submit your manuscripts athttp://www.hindawi.com
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttp://www.hindawi.com
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
CombinatoricsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
International Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
The Scientific World JournalHindawi Publishing Corporation http://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com
Volume 2014 Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Stochastic AnalysisInternational Journal of