Emhhhhhhommmls - DTIC · 2014. 9. 27. · Ccnsider a Markcv chain LX:t = 0,1,2,...) with finite...
Transcript of Emhhhhhhommmls - DTIC · 2014. 9. 27. · Ccnsider a Markcv chain LX:t = 0,1,2,...) with finite...
CHRINS(U) NAVAL POSTGRADUATE SCHOOL MONTEREY CA
M T SUH SEP 34
UNCLASSIFIED FG 2/1 M
mmmommmil
Emhhhhhhommmls
:.0.
1132 2-2'liiiE '-2 IllllN
M!CROCOPY RESOLUTION TEST CHART
140-
ilia- . "" -i" . " i .. .
1111IL2
.. " ~ ~ ~ MCR C P REO UTO TES CH . .i . . . - , - - ART . ,. , , ,, ' , ' '" w - " '
Lfl NAVAL POSTGRADUATE SCHOOLMonterey, California
DTICSELECTEAPR 4 S98
THESIS A
85 03 18 121
r • - . -- - ' ".--
SECURITY CLASSIFICATION OF THIS PAGE 0l17mm Data Entered)REPORT DOCUMENTATION PAGE READ INSTRUCTIONS
BEFORE COMPLETING FORMIREPORT NUMBER 2.GO ACCESSION NO. 3. RECIPIENT'S CATALOG NUMER
4. TITLE (and Subtitle) S. TYPE OF REPORT & PERIOD COVERED
€ -..
'~ ~'~1S. PERFORMING ORG. REPORT NUMBER
7. AUTHOR(#) S. CONTRACT OR GRANT NUMBER(#)
* .on T*+', u
S. PERFORMING ORGANIZATION NAME AND ADDRESS I0. PROGRAM ELEMENT, PROJECT. TASKAREA & WORK UNIT NUMBERS
< O_, iN_ ' e 'z,cool
I I. CONTROLLING OFFICE NAME AND ADDRESS 12. REPORT DATE
.W 1 J ! -t,-raJuat ,':hool e~temr.er 1' i 413. NUMBER OF PAGES
'~~o'a ~ ll rn ia 23 31LL 3
MONITORING AGENCY NAME & ADDRESS(t/ different from Controlling Office) IS. SECURITY CLASS. (of this report)
T'nclassif s; e.ISa. DECLASSIFICATION, DOWNGRADING
SCHEDULE
16. DISTRIBUTION STATEMENT (of thl Report)
"............. ..... 7u_!ic releas-; istribution unlimited.
17. DISTRIBUTION STATEMENT (of the ebetract mitered in Block 20, If dillf -it from Report)
I0. SUPPLEMENTARY NOTES
19. K EY WORDS (Continue on roerse side If necessary and Identify by block number)
"~r' ov - - 'D " "
20 ABSTRACT Contjnue on reveree side If neceeeary aid Identify by block number)
w h-:, flst-n -ro. . .
7 t 1, 1 7n0 1 - M- 1 7
DO ,.AN 73 1413 EDITION OF 1NOV 55, SSOLEiE! *'T"L: - r - ' 4, '.-T"T
S N 0102- LF- 014- b601 1 SECURITY CL.ASIFICATION OF THIS PAGE (Wl~en Date Entered)
:: : - :: :: : -. • . 7.
SECURITY CLASSIFICATION Of THIS PAGE (Vtn Daa EatenCO
nu
,as- ore 7e n n,- r t ev in
75e r s se, r ii r'-asons c7, : onhih y lt
"h nher t:- "riown mr Srllne7d SteOns p C
I model (th 'rlov hovrr- ) 7eoon. hpre v
1l m< 7, b v s mancreasedr o nj,! m 791 estimates oe n a .n5 fon
I i- " h " t 1_ro.cess j1n. d invest - :,n.4 ;.''. a, rtm ' i-'. t ri.'ses ln oonn'iction :.ih -v .cc .a
1 Sa P-. i,.--~c v c- n .n ,-a ! " V c a : modc els is t o f ter- ine w e h
'I'm I z e~cq=!' v
t tr~i n o tn rokl a 1 e s ISs t "a i- i". -: n is case i- is desirable no ns
4~ V1-r-2t -rror - 2 . e st I.-a*'na
- th m' nv1.-sf ih e7 na -v 11 of the m lr-'S :tn2 base-? on . -t'ma o ri ceneral, i-
-~~~~~~~~~~~~~~~~~~~~~~~ r- e .. , -_te,. iat - OS%] C1
[L ]n ,@ e x t[ r rep 1 v _
_
hhl~lh I 1 w< .an *e Te. - .... . .. .... . . - . I . . . . . , " . .nd it ions - v
!02 .',1 -.- '
-.. .. T. ASFC TO OF TI 1.GE1h21 Dta E-ntered)
* SECURITY CL.ASIFICATION OF THIS PAOE4lfle Data Entetad)
Approved for public release; distribution unlimited.
Sensitivity of Conditionsfor Luupin
Finite Markov Chains
by
Noon Taek. SUHMajor, Republic of Korea Army i
B.S., Korea Military Academy, 1975 -
Submitted in partial fulfillment of the
requirements for tae degree of
AASTER OF SCIENCE IN OPERATIONS RESEARCH
from the
NAVAl POSTGRADUATE SCHOOLSeptember 1984 ~M t--- - -- '.-o '- -a k.-
Author: I"_ _
Approved by: ___ __ .
partment of Operations Research
Dean of Information and PolicyNiences
3
ABST EACT
:iarkcv chains with large transition probability matrices
occur in many applications such as manpower models. Under
certain conditions the state space of a stationary discrete
parameter finite Markov chain may be partitioned into
subsets, each of which may be treated as a single state of a
smaller chain that retains the Markov pro4 erty. Such a chain
is said to he "lumpatlew and the resulting lumped chain is a
special case of more general functions of Markov chains.
There are several reasons why one might wish to lump.
First, there may be analytical benefits, including relative
simplicity of the reduced model and development of a new
model which inherits known or assumed strong properties of
the original model (the larkov property). Second, there may
he statistical benefits, such as increased robustness of the
smaller chain as well as improved estimates of transition
prohabilities. Finally, the identification of lumps may
irovide new insights about the process under investigation.
However, a problem that arises in connection with Frac-
tical acplications cf Markov chain models is to determine
whether the chain is lumpable. This is especially difficult
when the matrix P = [p..] of transition probabilities is
estimated from transition data. In this case, it is desir-
able to find bounds Cr E , the largest error, F.- ,
estimating p, , for all i and j.
This thesis examines the sensitivity of the lumping
conditicns~tased on E, the estimate of P. In general, it is
found that the classical lumping conditions are extremely
sensitive to the estimation error which can be expectei to
cccur even with large data sets. Thus, these conditions may
he of limited value ir many actual applications.
4
- - -.. - " <XV' . -: ." . ~ ":t "'-"
•.0
(7ABLE OF CONTENTS
I. 15TSCDUCTION . . . . . . . . . . . . . . . . . . . 8
II. TEORY OF LUMPABILITY . . . . 11
A. CONDITIONS FOR LUMPING .... ............ 11
B. PARTITICNS OF A SET OF STATES ........ 14
- C. AN EIGENVECTOR CONDITION FOR MARKOV CHAIN
LUIPABILIIY.. . .... ................. 16
III. BCUNDS ON THE lARGEST ERROR, 4, IN P . . . . . . 23
A. APPROACH ESING CENTRAL LIMIT THEOREM ..... 23
B. APPROACH USING ORDER STATISTICS . ....... 27
C. COMPARISON OF THE THREE EXPRESSIONS ..... 30
D. SENSITIVITY OF LUMPING CONDITIONS . . . . . . 34
IV. SUMMARY AND CCNCLUSIONS .... ............. 36
iIST CE E-EFEINCES ................... 37
INITIAl DISTRIBUTION lIST ...... ................ 39
5
0o
4 LIST OF TIBLES
1. Partitions of a Set of N States..........16
6
LIST OF FIGURES
3.1 Variation of A for Varying K.. and m ...... 31
3.2 Variation of Z by Changing Alpha ()...... 33
7
a
I. INTRODUCTION
Markcv chains with large transition probability matrices
occur iz many applications such as manpower models. Under
certain conditions the state space of a stationary discrete
parameter finite narkov chain may be partitiored into
subsets, each of which may be treated as a single state or a
smafler chain that retains the .larkov property. Such a chain
is said to be "lumpahle" and the resulting lumped chain is a
special case of more general functions of Narkov chains.
Consider a larkov chain (X:t = 0,1,2,...) with finite
state space S = (1,2,...,n), stationary transiticn Frob-
ability matrix P = p }, and a priori distribution of
"initial states", po (po, o,.,3) . Let § denote a
nontrivial partitior of S into m < n "lumps", say
S= {I(1) , (2) ,...,L(m). If [Xt} is lumpable with respect
to W, denote by £ ) the lumped chain with state space and
transiticr probability matrix P.
A well-known characterization [Ref. 2] is that [Xt) is
lumpaLle to [.) if ard only if there exist matrices A and B
such that
BAPB = PB (1.1)
where B consists of m nonzero ortnogonal n-dimensional
column vectors whose components are zeros or ones, and A isE' with rows normalized to probability vectors (i.e,
A = (E'B)-1B'). The icsitions cf the 1's in each column of 3
corresiond to states in S that together form a lump in .
It follows that if BAEB = PB is satisfied, then P APB as
is shown in Chapter 2.
Many of the mathematical juantities associated with (Xt)
can Le transformed directly to corresponding quantities for
I*., using the lumping matrix B. In Chapter 2, for example,we show that if an original .arkov chain [X.j is lumpable to
[X } an i {Xt) is fcrther lumable to c?], then {Xt} is
directly lumpable to [X,, and we give the lumping matrix
for jXtj in terms of the underlying two lumpings.
.herE are several reasons why one might wish to lump
[Ref. 1]. First, there may be analytical benefits,
including relative simplicity of the reduced model and
development of a neu molel which inherits known or assumed
strong properties of the original model (the Markcv prcp-
erty). Second, there may be statistical benefits, such as
increased robustness of the smaller chain as well as
improved estimates of transition probabilities. Firally,
tLe identification of lumps may provide new insights about
the ;rccess under investigation.
iiowever, a problem that arises in connection with Frac-
tical a plications of Markov chain modeis is to determine
whether the chain is lumpable. For chains with large state
spaces S, it is practically impossible to use an exhaustive
search to determine whether limpaDility conditions such as
those given in equation (1. 1) arE met for some matrices B,
lecause of the large rumber of ways partitioning S, i.e, the
larje n,iner of candidate B matrices. For example, if S has
1) elemerts, there are 115,975 partitions of S.
Another problem is to estimate the matrix = [pJ of
transiticn probabilities and to find bounds on A, the
largest error of .- for all i and j. We shall investi-
c~te the sensitivity of the lumping conditions in equation
(1.1) i cr vaLying A. If Ixt] is lumpable with lumping
matrix 3, is ccnditicn (1. 1) satisfied with P replaced by
the EstilatE P?
9
0]
This thesis will attempt to examine the sensitivity of
the lumping conditions based on reasonable estimation err7crs
Z\ when P is not kncwn and must use estimated by e. e
iescribE these facts about lumpability using eigenvalues and
cigenvectors, including the theorem mentioned by D.R.Barr
and M.0.homas [Pef. 3]. We do not review elementary
concepts of Markov chains here; the reader may wish to
consult [Ref. 2] and [Ref. 4] for review of basic facts and
specific terminologies such as lumpability, regular Markov
chai n, etc.
01
a
0 _.u .u - h m d~~-- -mm'--u " " m=~ ,m 1 , --. ,, .
0 U4 t
II. 1aEORY OF LUMPABILITY
'his chapter will ccver general facts about lumping such
as, corditions for lumpinj, tr.e number of partiticrs
possible for any givEn size of state space S, and theorems
associated with eigEnvector conditions for Markov chain
lumpatility.
A. CCNDIIIONS FOR LUMPING
Ccnsider a Markcv chain LX:t = 0,1,2,...) with finite
state space 3 = 1 stationary transiticn prcb-
ability matrix 2 = jp.j), and a priori distributior of
"initial states", F0 = (po, o,...,:pO). Let S denote a
nontrivial iartition of S into m < n "lumps", that is
3 = (L1(), 1 (2), ... , L(m)). If (Kt is lumpable with
ztspect to S, denote by [X) the lumped chain with state
space S and transition probability matrix P.
:;E now show that if the condition (1.1) for lumpability
with respect to the lumping matrix B,
BAP5 = PE (2.1)
ij 6atislicd, then the lumped transition matrix P is ziven
P = AP3 (2.2)
rcca. . is the sum '• were L(j) is the partition
CCntir and i is any element of L (i). By
11
S
0/
the lumpabibility condition, this value is the same for any
i e I (i). But the product PB sums the columns of P in accor-
dance with the partition subsets indicated by the columns of
E. Hence, PB is an n x m matrix with rows repeated in accor-
dancE with the partition sets 1(1), L(2), ... , L(m); the
effect of ere-multiplying by A = (B'B)-1 3' is to "average"
these coamcn rows yielding an m x m matrix 'P without the
repeated rows. But such "averages" are just the common rows
heiny averaged. hence, = APB is the m x m transition
matrix of the lumped chain with state space
{L(1) ,i (2) ....,L (m)).
ExamFle 1. Consider a transition probability matrix P
with 4 states uhich can be partitioned into
S= {(l,2,3), [4)] = (L(1),L(2) ,L(3)}. Let
1/4 1/16 3/16 1/21P 0 1/12 1/12 5/6
0 1/12 1/12 5/6
.7/8 1/32 3/32 ad
* = 0 1 and A 0 0.5 0.5 0
0 10! 0 0 1J
S0 0
Ve know eua tion (2. 1) is satisfied with partitioning
)= (IT),L(2),L(3)}. Thus, the lumped t insilion matrix is
0.25 0.25 0.5
APB = = 0.167 0.833
0.875 0.125 1.Many of the mathematical cuantities associated with (Xt)
can he transformed directly to corresponding quantities for
S
12
0A
{Xt), using A and B cf equation (2. 1). For example, since
AB is the m-dimensional identity matrix, it follows that fors a positive integer,
M (APB? =A (PE),A (PB). .A (PB) =AiPS 3 (2.3)j
We now show that = APSB
P = (AB) (APB) (APB) (APB)
= AP (BAPB) (APB) . . . (APB)
= AP (PB) (APE) . . . (APB)
= AP2 (BAPB) . . . (APB)
= AP2 PB. . . (APB)
= APS B
-ut A sB = P, since
BAPB = PE
PBAPB = p2B
BAFBAPB = P2B
EAP 2B = p2p
EAP;B = B
so PS is lumpable hith the same matrix B and P = AP B.
This imilies in turn that if [Xt) has steady state distribu-tion IT, then [X,} has steady state distribution IT TrB.
Theorem I. The steady state distribution 1T of the
lumped chain (i) is WB where T = -tP.
Proof. iTB = (WPE) B
= (APB)
TherEfore, WB.
13
Similarly,the a priori distribution of the initial
state of the lumped chain corresponding to that of the orig-
inal chain pO, is given by = pOB, since by eguaticns
(2.1) and (2.3),
pOpSB pOP ... PB
= pOP ... PBAPB
= pOP ... PBP
= pOB f.
Note that pOP B is the distribution of lumped states cccu-
pied ty the lumped chain after s transitions. Since this
equals pOBP= P-OP, it follows that pO = pOB.
B. PARTITIONS OF A SET OF STATES
The matrix B consists of m nonzero orthogonal
n-dimensional column vectors whose components are zeros and
cnes which determine a specific partition of
S = [1,2,...,n). Example 1 illustrates this, where the
state space S = (1,2,3,4) is partitioned into
S =If(, (2,31, [4)) = [L(1),L(2) ,L(3)) , and
101 0 0-B[ 10
0Permutations of these columns give a matrix which also
lumps [Xt]. In order to see this, let B be B with cclumns
permuted in some order. Then B- = BI9, where I is the iden-
tity matrix with its columns permuted in the same order. Now
if BAPB = PB, then
14
?B*' P B*(B*E*)rl BPB*
B I * (B, BI*)-II*B , PBi
ai = *r, (BI B)-I (i *')- t I*Bl PBI*
= B (B1 B4) B, PBI*
= BAPBI*
= PBI*
= P B *
so it follows that (Xe) is also lumpable with respect to the
matrix B*.
Now, how many candidate lumping matrices are there? This
would be the number of partiticns of S. [Ref. 5] gives a
recursion relation for the number Am of ways of partitioning
a set S =(,2, N)
N-1
AN = Ay (N Z 1 , AO = 1) (2.4)
1Mo K
From this relation we find A,= 1, A2= 2, As= 5, Ar= 15, etc.
The sizes of the entries in Table 1 show that it would be
impossible to use a trial and error approach to finding
lumping ratrices B for lumping a chain with larger state
spaces, say with 10 or more elements. Values of AN for
larger N are shown in Table 1.
15
I
TABLE 1
Partiticns of a Set of N States
N Partiticns N Partitions
52 20 5.172415 x 1013
6 203 30 8.467490145 x 1023
7 877 40 1.574505884 x 10 3
8 4140 50 1.857242688 x 1C 4 7
9 21147 60 9.769393075 x 1059
10 115975 70 1.80750039 x 13 7 3
it is of interest to be able to systematically prescribe
alternative lumjings by generating matrices B for a given
transiticn matrix P, using some method other than trial and
error. In the next section, we describe an approach to
finding E matrices using the eigenvalues and eigenvectcrs of
C. AN EIGENVECTOR CCIDITION FOR HARKOV CHAIN LUMPABILITY
Many problems in science and mathematics deal with a
linear operator T : V--)V, and it is of imFortance to deter-
mine thcse scalars for which the equation Tx = Ax has
nonzetc solutions x. In this section we discuss this
problem and its relationship with finding matrices B.
Theorem 2. The value 1 is always an eigenvalue fcr any
Markov chain transiticn probability matrix.
Proof. Let P be any n x n transition probability matrix
cf (Xp , x 'e a left eigenvector in Rf, and A be the ccrre-
sponding eijenvalue c P. Then xP = xA which is equivalent
to
16
x(P - AI) = 0 (2.5)
For A to be an eigervalue, there must !e a nonzero solution
x of equation (2.5). Equation (2.5) will have a nonzero
solution if and only if
dEt(P -AI) = 0 . (2.6)
-1his is callel the characteristic equation. To show that
A = 1 always satisfies eguaticn (2.6), we need only show
that the columns of the matrix in euation (2.6) are
linedrly dependent.Ncte that
*~~~P (P-)=[P p. 1 0 . 0]
P : 1 :01LO 1
P E . - . 1]
= p-1 p... p . . p
P P-. . p ... p
(2.7)
p p.... .. . . ...n
Since L P, 1 for Markov chains, it follows that the rowsJ: I IJ
in equation (2.7) sum to zero, so the determinant in elua-
tion (2.E) is zero with A = I. It follows that A = 1 is an
Eigenvalue of the Markov chain [X,}. We'd next like to see
properties of eigenvectors corresponding to the eigenvalue
* A-I.
Theorem 3. For any regular Markov chain, components of
the eigenvector corresponding to / = 1 are proportional to
the steady state distribution of (Xt).
17
Froot. Let x be a left eigenvector of P, and ?k be the
correspondin eigenvalue of P, such tnat xP = xA, anm
assume xi= 1. For given A = 1, xP = x. The steady state
distribution of [Xt) is unique [Ref. 4]. Therefore, x aust
te the steady state distribution IT since 1. xi= 1.
The following example demonstrates Theorem 3.
Example 2. Let
1/4 1/16 3/16 1/21
P = 0 1/12 1/12 5/6
0 1/12 1/12 5/6
7/8 1, 2 3/32 0
The eigenvectors corresponding to the eigenvalues of P are
displayed as column vectors below:
Ei~envalues : 1 0 -0.25 -0.333--
'C.7367 0 10.5 0.8247
Eicenvectors : 0.09209 -0.7201 -0.375 -0.03436
C.2236 0.7201 -4.125 -3.2405
0.6315 0 -6 -0.5498
Note that T = -T r, -T. i, -I+)
= (0.4375, 0.0547, 0.1328, 0.375),
v ere - 7 etc.
Theozem 4. Eigervectors corresponding to eigenvalues
other than 1 are orthogonal to e (,,...,I).
Proof. xe' x(Pe') = (xP)e' : xAe'. Therefcre, xe'
must le zerc for A # 1.
WE are also interested in finding the relationship
!etwEn eigenvalues cf P and those of lumijed transition
probahility matrix P, where ? = APB as described above.
18
•I
I.
Theorem 5. SuPIcse [Xtj with transition matrix F is
lumpable to (X with transition matrix P. !he eigenvalues
of - are eigenvalues cf P.
Proof. Let Q(%J = 0 be the (n*e degree) characteristic
equaticn of P. By the Cayley - Hamilton theorem [Ref. 6],
OAP) = 01PP O-"... + QP Q, = 0
which together with eguation (2.3) implies
Ao, (P) B = h,," a . . . + .1
= (P)=0
Since P satisfies 2's characteristic equation and since
eigenvalues of A(P) are of the form 02A), it follows that
( ) = 0. Thus all eigenvalues A of P are also eigenva-
lues cf P.
We next examine the eigenvectors of P and P, with the
aim cf identifying lumpings of [X,-} directly in terms of the
eigenvectcrs of P. We have seen that 'pV is obtained directly
as pOr; a similar relationship holds with eigenvectcrs cf P.
Theorez 6. Sup~cse x is a left eigenvector of P ccrre-
sponing to eigenvalue >A, and suppose (Xy is lumpable to a
chain with transition matrix P = APB. Then xB satisfies the
e-uaticn (x_)P = (xB)A
Prroof. By equation (2.1) , (xB)? = xBAPB = xFE. Eut
xP = xX , and the result follows.
We ncte that xB is not necessarill an eigenvector of P
lecause it may he zero. In fact, it easily follcws that
xB = 0 if A is not an eigenvalue of '. But xB may be null
even if A is an eigenvalue of I, in cases of where A is a
repeatej eijenvalue cf P more times than o ?.
19
[ef. 7] pointed out some other useful properties asso-
ciated with eigenvalues and eigenvectors such as : 1) if the
matrix P is symmetric , then eigenvalues are real and eigen-
vectcrs are different for repeated eigenvalues, 2) if the
matrix is not symmetric, then the eigenvectors are the same
for reFEated ei envalues.
Theorem 7. If [Xp) with transition matrix P is lumpakle
to [Xt1 with transition matrix P, and [X} is lumiable to
jX-t with transiticL matrix , then fX t} is directly
lumiyahle to w } where CXI is the lumped chain of [(T].
Proof. Let (X be lumpable to [X+1, and { t be
lumpat1e tc (+) by matrices B, and 3 ,, where B, and B, are
lumping matrices in which the dimension n x m of B, is
0 greater than that of Bz . By equation (2.1), P = APB, and
P = AP2,. Thus,
F =A PB, = A {(A PPI)B= (A.Ai)P(BIB,)
ao see tLat B,-B. is a lumping matrix and AiAI is cf the
required form, we need to show that (AiAI)(BI*B.) is the
identity matrix as mentioned in Stction A. But
(AA%).(BBa) = Aa-(AIB9I)-BA = A2-I.2, = A'B, = I
Also, note that B,.B, is 31 lumped by B,, so E,-B has columns
cf the required form. Therefore, (Xt} is directly lulalle
to JX't1, hy the lumping matrix B = B,'B.
?0
0 •
• 0
Example 3. Consider a Markov chain with 5 states, and
transiticn robability matrix
0.3 0.1 0.2 0.1 0.3
0.1 0.3 0.1 0.3 0.2
0.5 0.1 0 0.1 0.3
0.1 0.5 0.2 0.1 0.1
0.5 0 0.1 0.2 0.2,
First, consider S = (1,2,3,4,5) which can be partitioned to
S ill ,[2,41, [3,5) = L(1),L(2), L(3)}. The correspcnding
lumping matrices are
'1 0 01
B= 0 0 1 and A1= 0.5 0 0.5 00 1 0 1O 0 01
01 3 0 0.5 0 0.5,
.0 0 1
and the lumped transition protability matrix is
0.3 0.2 0.51
1 AI BI 0.1 0.6 0.5 0.2 0.3
Secondly, consider S with 3 states which can be partitioned
to S = [[1,3), (2)) = [L' (1),L' (2)), with matrices
3 : 0 1] and A1 = [0.5 0 0.5JI 0L 0J 1 0
The corresponding lumped transition matrix is
A = P ~B~ 0. [ 0. 20.4 0.6 .
inaily, consider lumping the transition probability matrix
directly. For partitioning,
r2 21
-L7- --. -_
S =[(1, 3,5], [2,4)1 [L" (1), L"(2)), and
001 1 0 1R1 = .. AZA = [1/3 0 1/3 9 1/3
0 1 0 1 0 10 0 1/2 0 1/2 0 ,
Ll 01 LO 0o1
ani the directly lumped transition probability matrix is
Theorem 7 shows that lumping is "-:ansitive", in tne
follcwing sense. Define two transition matrices P and Q to
te eguivalent, (P Q), if Q = ,P for a lumping matrix P
4 whose columns are thcse of the identity matrix, in scme
-ermuted order. (Thus the chain iXtj and [Y.] differ only in
the labels associated with their states). Define a relation
< between transition matrices as follows: Q 5 P if and
cnly if Q = for scme lumping matrix 3. Then theorem 7
saows that Q _< P, RQ 4 E < P. This relation "<"is
reflexive, since Q 5 Q using the lumping matrix I (iden-
tity). Finally, " _< " is antisymmetric since Q :_ P and
A P 5 P ". Thus, the set if all transition probability
matrices is partially ordered by the "lumping" partial
crder, _ S
22
L . - ,,-- - - - -... - "" "
III. BOUNDS CY THE LARGEST ERROR , , IN P
in this chapter we consider three procedures to find
hounds on A. First, we use the central limit thecrem for
given i and j. Secondly, we use a binomial approximation or,
the hasis of the first procedure. Finally, we get the
largest error /%, using the asymptotic extreme value distri-
hution. These three approximations are only designed to give
a couch idea of the relationships between 6 and the number
M of elements in the state space, the total number of
cbserved transitions K.., and the probability oC.
A. APPECACH USING CENTRAL LIMIT THEOREM
t'e are interested in the sizes of the errors between the
estimate p and the unknown P, where P is the transition
probability matrix cf [X.}. We assume tae transiticn -rob-
ability matrix P is of size I x M.
Let Kij be the number of observed transitions of [Xt)
from state i to state j, and let K-. be the number of
observed transitions from state i. Similarly K.j is the
number of observed transitions into state j.
Let -. Le an unknown transition probability from state i
to state 3 and b be an estimate of p6 based on K.. observed
transiticns. Then the usual estimate *pi' of pjj is the ratio
of K<" to K-.. Now, as a rough approximation, imagine that
K-,. is fixed, and the number cf transitions from state i to
state j, K1 , is Pinomial (.i.,p ). Then by the cen-tral
limit theorem,
23
,.i approximate by Normai~p,
sincc
ELPJ " - -= E , and
Varf[ F'j -1=Va[
4 ~ Va ri(K4.)
0(3.1)
We want to find a bound /, on the estimation error - .l
which occurs with picbability at least &(; that is, the
largest Ax for which
Now
- 'L.- ; , t-? &j = ( .s)(3.2)
CK.P
0 ,
let z = j , then
24
,*
Z i3 ap-roxcilaite by standard Normal. Rewrite equation (3.2)ds
E[IZl - ] _ . ; 3 < (< 1
E.uation (3.2) is approximately
>r >. -.K" P
since the Normal distribution is symmetric. Solving for A,
we have
%herr N-1 (1 -is the (1 - - t- quantile of the stan-
dard Normal listribution. Suppose tne steady state prcb-
ability TrL of state i is - based on the eually likely
case, and su-pose the worst case in which
J - p.. = -
T.en an approximate value for A is given by
2
[ 25
N-, (1- ) (0.5)2.< IL
= N-' (1 - ) (0.5);L K.-
= N-I 1 ) (0.5)M
0() (0.5) (3.3)
Equation (3.3) ccncerns the error I"'- p,,l for fixed iPL
and j. We'd now like to find an error bound A overall i
and j. That is, we wish to find the largest A for which
P Pz - p j1 a for some i and jj _oL,
which is rcughly the same as
P[%- P _ for some i and j] (3.4)
We aply the binomial approximation in equation (3.4), so
that
P[- 6 ij > A for soue i and j]
= 1 - [o- p. < a for all i and J
= I - (1 - )M (3.5)2-
let 1-(1 - - -)M= p for some 0 < < 1. Solve for p(, which
gives
-2-2 I- . (3.6)
26
I'
*i
Substitute the value cf o( in equation (3.6) into eiuation
(3.3). Finally we get the approximate bound &l for all i and( j:
N( 1 ) (0.5) . (3.7)
E-uation (3.7) gives an approximate expression for z,, using
binomial approximation.
B. APPECACH USING ORLER STATISTICS
Assume ZI , Z , ... , Z4 are independent ccntinucus
random variables, each with density function f,(z) and
distribution F3 (z). Now let Z(h , Z(,) , ... , Z,, ) denote
their ordered values, from ssallest Z , , to largest Z(M)
these are called the order statistics of Z,, Z,, ... ,
We now consider the Erobability law for Z(,) (Ref. 8], the
largest or maximum value.
The event (Z(M)_< z) occurs if and only if the event
(Z1 1 z, Z< z, .. . , ZM _ z] occurs, since if the largest
Z is smaller than z, all M of the random variables must be
* smaller than z, where z is any fixed real number. The
distribution function for Z(M) is
F (z) = PrZ 5 z]
= PCZ 1 _ z, Z Z :_ z, ... , Z1 <_ z]
= P-Z, _ z] P"Z < z] ... P ZM :_ z]
since Z ,Z,,...,ZM are assumed independent. But each of Z1 ,ZL
,.... 2M has the same distribution F (z). SoMM
F (z) = [F3 (z)]J
27
-I
The density function for Z then isd
f ) (Z) = dz)
d=-i [F (z)
= I [FZ(z) ] Mf(z)
where f (z) = az
Consider the limiting distribution function of the
maximum Z(M) as n tends to infinity. [Ref. 9, 10] show this
distribution is
imF(z) M = e (Z) (3.8)
* ~M- .
if Z1 , Zz, ... , Z. is a random sample from standard Ncrmal
Fopulaticn. We want find a bound 4 on the largest of M2
errors between estimates in P and the unknown components of
P. The random variables p- p.- are very roughly Normal with
mean 0 and variance -- which is derived from e~uation+, X..
(3. 1) fcr i,j = 1,2,...,M. Recall that K.. is the total
number o. transitions observed.
:..et = pq- p. where £ 1,2,...,12. Then we know the
random variable X is approximately equai to V 1-TO Z, where
z Las a standard Normal distribution. Let the random vari-
able X(M) be e ual to maxi i - p. Then
lir PEXM4) _ , J = lir PI max | j: < A ]
28
--
t.av 9et f
Now XO) and X(M, are asymptotically independent, so for large
Ic M,
lira PCX 5M) <
re= { 2X 1 -- A] ... P - .F(X) ] M
X= F,(A)] • (3.9)
Frow equation (3.9) we derive an expression for A asfollows. Let A be the largest value for which
PC A 1 <5 1 - O<
'his is the complemertary probability because we wish tohave PCif' pj.j _ 4 for some i and j) > , as in the
previous section. The limiting distribution function of the
maximum X( )is the sane as
0 2±~lmfF, (4) = imf F ( 6 .
7hen, aproximately,
los(o-,) e
and
29
A - - •
log 9-lcg(1 - )~ log2MI2 (2 A -- -2log2M2).
finally,
0 (05) - (V'2 o -g2 M) (3. 10)
Equation (3.10) is an approximate expression for A tasel oru
the asymftctic distritution of the extreme order statistic.
We will ccmpare the central limit theorem 6,'s with those
cbtained with the extreme value distribution, in the next
section.
C. COMPAISOI OF THE THREE EXPRESSIONS
The three expressions for . obtained using the central
limit theorem and order statistics have been developed under
aeproximations such as: 1) the steady state distribution
of {xt) is -- (equally likely), 2) the variances of
Ij- pLI have -F.. as a maximum value (worst case), and 3)
all transitions are irdependent. Information about [X..} is
from the estimate I because we don't have informaticr about
the unkncwn P. In a view of the above approximations and
comutations, our expressions for & are very rough. Hcwever
they dc Frcvide some insight into the occurrences and sizes
of estimation errors in .
Figure 3.1 contains 3 graphs showing A as a function cf
K.. and M for fixed o. = 0.90 based on the three expressions
(3.3), (3.7) and (3.10).* 'he first graph shows error bounds using the central liiit
theorem cnr - p for fixed i and j. The second Sra'h is
given by the same approach as the first yraph, except
overall estimation errors are considered, for all i and j.
30
ALPHA =0.90
PON?4 E~11MATION fff USINGC C.0. -OvTRALL ESTwATiON By USING C.LY.oGRAP#I CRAP" 2
-K. S(.Q
0 20 40 60 10 g0o 0 2 '0 60 so 100S'Zr(&d) OF TRAkiSION MARI SiZ(&a1 Of TRAt4SIlICN MATRIX
OVE RAL L ESTIMATION BY USING OPO(R 4STATtil CS
o CRAPC
0
0 0
330
'The third graph is Lased on the asymptotic distritution of
the largest value of I~q- p'I over all i and ~
C From Figure 3.1 we see that the ldrgest estimaticn Error
depends very much on the number of transition observaticns
and matrix size, but not so ffuch on the c0( value as seen
from Figure 3.2
32
VARIATION CF A BY CHANG.NC ALPHA
OW(VALJ. C''11YM BY US" C.LT. OVRAL MIUATION OY USINC ORDER STATISTCSo .. aOo o K..SOOO
- °0 LECLO .0
I ALPHA .050 t2~ ?A4Y.A . 05 0
S:ALf' kA . 0 9C
4 :ALPHA . 095
13 0
22- 1.
o4
m so so Ic 2 20 40 10 So
g ZE(m) TRA rTK ATMRIX SIZE(m) OF TRANS1tON MATRIX
0
Figure 3.2 variation of a by Changing lpba Coo.
3
. . ..- .. . . .- . . " .- .
Graphs 2 and 3 in Figure 3.1 are very similar even though
they use different approaches. They give an idea of howlarge likely values cf A are for given K.. and M, in the
"worst case".
If we consider a Markov chain (X+} with M = 20 or 30
states, and we have observed K = 5000 transitions then,
roughly, it is likely (prob = 0.90) that at least one
element of ' is in error by at least 0.1. In general,
expressions (3.7) and (3.10) may be useful for Markov chains
[Xt) with M = 20 or 30 states and large numbers of observed
transiticns.
D. SENSITIVITY OF LUBPING CONDITIONS
We have developed expressions for A, using the central
limit theorem and order statistics. We want to examine the
sensitivity of the lumping conditions applied to the
estimate of P. If equation (2. 1), which is a necessary
condition for lumping a Markov chain with transition matrix
, is satisfied, then the lumped transition matrix is
given by eguation (2.2). However, even though P satisfies
the lumping conditions, it is extremely unlikely that its
estimate will also satisfy these conditions, as we shall
now demonstrate.
In order to simulate the difference between and P,
consider a matrix of errors RHA, where R is a random matrixwith dimension the same as P, whose components are 1's,
-1's, and O's where the sum of each row is zero. Now
consider the lumpability of the simulated estimate P, which
is constructed by taking P plus the random matrix R times
* AI, that is, P* = P + R-6.
7c show the sensitivity of the lumping conditions, we
assume the unknown P is lumpable with lumping matrix B, and
consider the difference (BAP*B - P*B). If eqgation (2. 1) is
satisfied by P then all if these components must be zero.
34
.0
Theorem 8. The difference (BAP*B - P*B) is a linear
. function of A.
Proof. Let R be the random matrix as defined atove and
let P* = P + R.&. Then (BAP*B - pV*B) is given by
[BA (F+iB.A) B - (P+R-)E] = (BAPE + BAR. .3 - PB - R B)= (BAPB - PB + (BARB- RB).)
= (BARB - RB).a
= C4.
Therefore the difference of EAP*B - P*B is linearly depen-Udent cn 4 and P* is not lumpable unless BARB = RB (i.e., R
is "lumable,'), which is not likely to occur.
Since is likely to have elements differing appreciatl"
from the ccrresponding elements in P (errors of size a ), it
can be seen that tic lumpability conditions will not be
satisfied (not even nearly so) by , even though [Xt) is
lumpable. We conclude that attempting to check the lumpa-
bility of tie estimate when P is not known is not useful.
0
35
0 ..
! "' "- - " "" '" " "" ". "" ' " -- " " " '. - ,_.,. :_ L . .m . , -
,, ,,,*,,,.--.,,, ,, .'
IV. SUMMARY AND CONCLUSIONS
Ve have given several theorems associated with eigenva-
lues and eigenvectors for lumpable Markov chains (X. with
finite state spaces. We have derived rough, approximate
mathematical expressicns for the largest error made iL esti-
mating P by P based cn transition data.
Both expressions (3.7) and (3.10) are very similar even
though the estimated A 's for the first expression are
slightly less than those in the second expression. These
expressions show that the largest estimation errors depEn!
very much on the number of transition observations ard on
the matrix size, but rot so much on the o< value.
Since P is likely to have elements differing appreciably
from the ccrresponding elements in P, it is of interest to
examine whether the equation BAPB = PB is likely tc be
nearly satisfied with N, i.e., will (BAPB - PB) be neariv
zero? This is examined by simulation of "estimates" L of
F, using random perturbations of elements oz P of sizes 6
which are likely to cccur as errors in .
This shows that the classical lumping conditions are
Extremely sensitive to estimation errors which can be
expected to occur even when a large number of transitions
have been observed. Thus, the classical lumpinj conditicns
may be of limited value in many actual applications.
As further research, it is recommended that scme
constructive approach to finding matrices 3 for lumping a
lumpatle Markov chair [Xt) be developed , perhaps along the
lines of the theorers mentioned in Chapter 2. It is hoped
that the present study will be useful to those who might
ctherwise have endeavored to check the classical condition
for lumpability of a larkov chain [Xt} when the transition
matrix P has been estimated.
36
.6 -o
LIST OF REFERENCES
1. D.E. Barr and M.U. Thomas, "An Approximation Test ofMarkov Chain Lumpability", J. Am. StatisticalAssociation, vcl.72, pp. 175-179, 1977 -----------
2. J.G. Kemeny and J.L. Snell, Finite larkov Chains, VanNostrand, Prinston, N.J., 1967-.--------
3. D.R. Barr and M.U. Thomas, "An Eigenvector Conditionfor Markov Chain Lumpability", O~erations Research,vol.25, no.6, FE.1028-1031, Rovembr;-!377.
4. S.P. Ross, Introduction to Probability sAcademic Press-2 -IT[on,-p. 12-T.
Altert Nijenhuis and Herbert S. Wilf, CombinatcrialAlSorithms for Computers and Calculators, ZeK
c97-2- "i e~Iioin, fp- 9 3 -9 [-
6. Marvin Marcus and Henryk Minc, Introduction to LinearAlcetra, Macmillan Company, pp. 163---T 6 5.
7. W.E. Boyce and B.C. Diprima, Elementary DifferentialEuations and Boundary Value PoITEm, JoEU-WrI7-"o73.- ticn, pp. 287=297.- ...
8. H.J. Larson, Introduction to Probailit Theory andInterference - -i ey - -ons; 3rd-- i-:.-:13320";- 1982.
9. A.E. Sarhan and B.G. Greenberg, Contributions to CrderStatistics, John Wiley & Sons, ppT------
10. M.G. Kendall and Alan Stuart, The Advanced Thecry ofStatistics, Griffith, pF.333-37, 13 S.-
11. Howard Anton, Elementary Linear Aliebra, John Wiley &Sons, 3rd editicn.
12. C.J. Burke and M. Rosenblatt, "A larkovian Function ofa Markov Chain", Ann.rMath.Statist., vol.29,pp. 112 - 1 2 2 , 1958.
13. Volker Abel "Conditions for Lumping the Grades of aHiErarchical Manpower System", Oerations Fesearch,vol.17, no.4, 1E.379-383, NovembeE,-1gr3.---
37
0I
14. Volker A.el, "Test on !nrna. I lit,', for ~a--'ian "- P,
Mces:, Ole:lnspser:h , ncc
/
K . -. 7. .. t:: > : .. - -x ~ ~ . .
INIIAL DISTRIBUTION LIST
No. Copies
1. Defense Technical Informaticn Center 2Cameron StationAlexandria, Virginia 22314
2. Iitrary, Code 0142 2Naval Postgraduate School!Montery, Califorria 93943
3. Department Chairnan,.Code 55 2Department of OperafionsNaval Postgraduate SchoolMontery, Ca 93943
4. Professor D.R.Barr, Code 55Bn 2Department of OperationsNaval Postgraduate School1ontery, Ca 93943
5. Professor F.R.Richards Code 55RhDeFartment of OerationsNaval Postgraduate SchoolMontery, Ca 93943
6. ! ajcr Moon Taek. SUH 10Ecstal Code 100-ElFyungkido Koyangkun1Jidcmyun Hangjunaeri 151Seoul, Korea
7. Majcr Tae Nam. AHNCcmputer CenterF.O.Eox 77Gongneunq Dong fLobong GuSeoul 130-09, Kcrea
8. libraryP.O.Eox 77Gcngneund Dong. Dobong GuSeoul 13 -09, Kcrea
9. Ma or Hee Sun. SCNGSMC 1359 NPGSMcntery, Ca 93943
39
4 . .' ," . " ' " - " .. " " " " - : ,,..,.. , ,. ',, . " - - - " - . -
. . . . . . .
0
6
0
0
S
0
0
S
. . . . .
._°
FILMED
5-85
-1
DTICg
[_. ,.- .% ., . . - .