SEVERAL NEW PROBLE1'1S OF THE THEORY OF MARKOFF CF.AINS by V. Romanovsky Institute of Statistics Mi.meo o Series No. 102 M9.y, 19.54 Paris 1938 QUELQUES PROBLEr.1ES NCUVEAUX DE LA THEORIE DES CHAINES DE IJrARKOFF Par v 0 Romanovsky ACTUALITIES SCIENTIFIQUES ET INDUSTRIELLES No. 737, PPo 43-66, 1938, Paris, France Translated by Dan Teichrow under the supervis ion of Imogene Riddick, Instructor of Modern La~O'Uages at North Carolina State College of the University of North Carolina, Raleigh, N. C. at the request of the Department of Experimental Statistics, North Carolina state College. (Edited Jointly by the Translation Service and the Deplrtment of EXJ;Brimental Statistics.) May Paris 17, 1954 1938 SEVERAL NEW PROBLEMS OF THE THEORY CF MARKOFF CHAINS v. ltezna.r,oTsky The object of IT\Y article is to indicate several new directions which have been pursued in the theory of Harkoff chains and to state several results I have obtained in my research in these directions. ~hich I am going to consider cor- related chains, cyclic chains, and several statistical problems which are related to probabilities in a chilin. I. 1. Correlated Chains So far as I know, no one has yet considered sequences of random variablE;s which are correlated and subject, at the same time, to a law of probability in a chain. But the problems which arise in the case of such sequences are not less important than the other problems of probabi1it;ies in chains and they must be dealt with in order to complete the theory of the former. To specify and simplify my article, I shall consider only the following case. Let be an infinite sequence of random variables connected in a simple 11arkoff chain and associated ""lith experiments number 0, 1, 2, ••• respectively. in the experiment number h, the corresponding variable n h tve shall suppose that, can take only one of the different values xl' x 2' • u, x n which remains the sam8 for all the experiments and .., have the probabilities POI' P02' ... , Pon' ~ Po..<. = 1) in the initial experiment, number 0, and the transition probabilities I)~ are defined by the equations I , I; ..l/i...LA ,., (.,(,p = P( 11.n+1 = x ~ I 11.n = X.,( = 1, 2, ••• , u; h ) = 1, 2, ••• ), 'lr1here the symbols on the right hand side designate the probability that the v-ariable ~ + 1 takes the value of x B when it is known that ~ has taken the value x.,(. -2- il ~ .. II¢.,(~ Thus the matrix represents the law of the chain being considered, more briefly of the chain Cn which connects the variable ~. Desigilate (I) further by Pk/~ the probability P{uk • xp) calculated under o' the hypothesis that the values of all the other variables u ~+l' ~+2' up ••• , uk_I' ••• , are indeterminate and consider a second infinite sequence of chance variables v o' vI' v 2 ' ••• , associated to the same experiments and connected to the variables uh by a correlation which is defined by the equations rk'~"'" (1) where (~ = 1, ! P{u k .. x~, 2, ••• ,n; v k '" Yr ) c(l\:/f3)(-i/'f3r) ~ = 1, 2, ••• , m; k .. 0, 1, 2, ••• ), r ~ ~( are the conditional probabilities, the same for all the experimentE of the equations v k .. Y~, the condition being that uk ... xf3' and where Y1' Y2' ••• , Ymare the values which can be taken by each of the variables vo' vI' v 2 ' ••• , in the experiments considered. Let 'f = II yrp?f It • of the sequences (~) The matrice s q> and Ifdefine the law of correlation and ( vIi) or of the chain Cuv • We shall study several of the most important properties of this correlated chain. 2. First of all we can propose the following problem: of the chain C which connects the variables vh ? v For greater clarity we shall consider on the simplest case where ~ is a pDimitive and indecomposable matrix. what is the nature ~jd most important In this case its characteristic equation ~(A) (2) (1) ... I>.E-~I·o, I will use the notation of my article in Acta Mathematica, Vol. 66, Which will be cited below as A.M. -j- has aJfAa == I as a simple root, the absolute values of all the other roots of (2) are < I and the probabilities Pk/~ have for k ...,. 00 the determinate limits p~ which represent the limiting probabilities of the equation ~ == x~(l). Let then obviously (3) Likewise, if we designate by ~+k == ~ (4) ¢~} the probability of the equation when it is known only that t %) ; p( vh k + Further, let the equation == yy llh ... f ~ ... x.,(, we shall have xc(} ... ~ ¢!;)t~~ ~ · nJ~} be the transition probability of the equation vk_1 v k == y~ to I 0< They satisfy the relation == y~. (5) and can be obtained by the following procedure. We have equations a qk-l qk-l P2 = P(u k Pj : p( ~ ... X(1' u k _1 ... Y1r'1 == X) =¢ rJ,(1' uk'" X(1} ... r(1¥' ' of which the first is obtained by Bayes Theorem. I~ 'f~ -4It is clear that (k) _ n~~ : whence (6) n(k) ~ 1'" ~Pk-l\~Y-:13¢ 1~ 0: qk - 1 f ~ .<, (J The probabilities n~~ define the law of chain geneous and multiple, because • .J..(J '(J If Cv. This chain is not homo- nJ~ depends on k and the probability PI changes its value according to what is known about the values of'k '-2' u quentlyof the sequence, k-3' ••• and, conse- v k _2 , v k _ , ••• It becomes homogeneous and remains 3 multiple if the chain Cu is stabilized because then (1) for all k and the probabilities (7) do not depend on k, but the information on the values of the vk_2 ' v k• ' ••• 3 can be such that certain values of the uk_l become impossible and the sum on the right hand side changes. 3. For k 4'00 the correlation of the sequences limit a well determined limiting correlation. indecomposable and primitive matrix, ...j: lim Pk I ~ "" p~, lim qk' yo "" q (' lim r k ,~"( II r~ r 0: • and (vh ) has for a In fact we have in our case of an r p~ t~ ~ (~) 'P~'!~ Y" , -5and the various moments of chain C have for their limits the corresponding moments uv of this limiting correlation, which remains a correlated chain. 4. Let us pass to the important problem of the limiting law of the proba- bility of the sums We have for k 5-1 ~ u. h=O n 8-1 and ~ vh • h=rQ -.,.00 = Eu k = and we shall consider the sums s-l ~ , ~uh = .~ ~ (uh - xO) h=O We can calculate the moments of the sums ~ u~ by the method which is explained in detail in A.M., and having the moments of the sum ~ v ~ .we notice that (7) where -6- and](~)are the coefficients of the development of ¢ ~) according to powers of ~ = 1, ~, ••• , A~_l of the roots of the equation !(A) = 0, supposed for the greatest simplicity, to be all distinct. moments of the quantities v We see by these relations that the I are obtainable by beginning with the moments h corresponding to the u' and replacing p ~r(k) and lk(g) by q X(k)and x(g) h ~'~~~ T4 'It' ~~ ~'It respectively. We see further that ••• ••• ••• +kr ) ••• where t = Y - YO. For the effective calculation of these moments we can apply the procedures given in A.M. (~k2 ••• of the moments N m ffi l 2 that the structure ••• k ) r ••• m r ••• m r In particular their asymptotic properties are identical. We find, for example (9) where (10) is the studied in A.M• -7This value is entirely equivalent to that of the dispersion of ~u·~ I (11) (z-x-x). o The calculation of the moments of the products of integral and positive powers of the sums .~ I ~ ~ and 'X_ ~v difficulties in principal. I h is naturally more complicated, but does not present any We cite only the moment for the product (12) where (A ) ... g ~p V(g) " (A ) ... g ~ PI3' f~~~2lhl (f z132 tlJ' ~p 1J!13 I ~ll (13) ~ (Ag'~) .. I 1fI13 2 ~ll ~ (Ag'~) ... ~l ~1~2 z ll' yr(gLy(h) 0-< . ~l 13 113 2 ~P OJ. reg) yr(h) -«31 ~1132 t PI ~ , , z t , 2 cr ~l ~ lfi13, ~ z t 13 2 ~ , ~u~ ~v~: -8sA. vl (A.g ) (13) continued v2 ~ (1 - (A.g'~) ... = r .gx ' g A. S) erg. A ) • g (1 - ~) ~ (~s _ A S) (1 - ~) (~g- Ag ) , the sums on the right hand sides being taken over all the values of the greek letters under the corresponding ~ sign. With the aid of the formulae written above we can easily calculate the correlation coefficient of the sums ~ u~ and ~v ~ according to the general relation r .. - - - - - - - - - • Its final value, for k-t 00, is also obtained immediately. 5. It is not difficult to write the generator oZ all the moments of the sums ~u~ and ~v~ it is (14) rI. 'P .~ r. s ... ~ PO-< -<)t" u ¢ .J 1IfR , 't ~R1 iP l ¢A ••• ~ RAt '"'s-2'"'8-1 ~s-l 8-1 e iQ(z. + ~ ZRh)+i't"(t!+ ~ t J! oJ.. .., and satisfies the recurrence relation (15) where (16) >eiQ(ZR • "'s-h +.,,+ZR "'s-l )+i't"(t~ s-h + ... +t~ ) 6-1 -9- (16) cont1ed • The relation (15) gives us and we can show (1) that lim s -¢s S-"'<D [ ¢s-l ... lim log ¢ s'" - ~1 (92 2 + 2rQ1: + 1: ), S~CD i f we set in (14) Therefore, (17) lim ¢ s "'e and we see that the correlation of the sums .z % and :::-, I .~ I ~ vh has for its limit the normal correlation; r is the limiting value of the correlation coefficient defined above. (1) My proof is not yet rigorous and it. is a litt,le long, tharp-fore, I have omitted it here. -10- 6. We have considered only the simplest case of correlated chains characterized, in particular, by the fact that chain C C'v. by But we can form chains where P(u k+1 .. x~) u ~ich is is independent of chain depends not only on the value taken lok-' but also on the value of v k • A still more general case is obtained when we suppose that the probabilities of the simultaneous equations u k+l .. x.,(J v k+l ' \ and 11 simultaneously. k to formulae (1) - (17) ~ich IS y~ depend on the values taken by I have calculated the fundamental formulae equivalent are shown above for the most general case of corre- lated chains, but I omit them here because of their complexity. They do not present any essentially new particulars. II. Cyclic Chains 1. We call a cyclic chain of indElt k a chain in which the law by an identical permutation of lines and columns can be put in the cyclic form -T" (1) j . 0 L12 0 ••• 0 0 0 ~3 ••• 0 • ••••••••••••• t 0 •••••••••••••••••••••• 0 0 ••• 0 0 ••• where L , L , ••• , L are the non zero sub-matrices, all other sub-matrices 12 kl 23 The matrix is L~ being zero (111' ~2' ••• , L kk should be square matrices). said in this case to be a cyclic matrix of index k. Let there be given a stochastic indecomposable matrix. Then in order to say that it is cyclic. of index k, it is necessary and sufficient that we be .able to put -11it in the form (1) (A.M., p.165). The simplest means of demonstrating this theorem consists in considering the cycles of the indecomposable matrix in question, that is, the sequence of non-.ero elements, ¢ 4 1 ' ¢ ~1~2' ... It is evident that, for the matrix have their orders divisible by k. which form the cycles. i J ¢ ~s-l-< presented in the form (1), all the cycles The order of a cycle is the number of elements Inversely, if all the cycles of an indecomposable matrix have their orders divisible by k, we can state that it is cyclic of index k. This last assertion is demonstrated in A.M. by the aid of the characteristic equation of the matrix considered. But we can give a direct and simple procedure for reducing a matrix to cyclic form. We shall describe this procedure because it is new and useful. It is simplest to consider an example of the procedure in question, which will not diminish the generality. Let us take for example, the following matrix; A .. 0 1 0 0 0 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0 0 0 1· 0 0 1 0 1 1 0 0 0 0 0 1 0 1 0 where the non-zero elements have, for simplicity, been replaced by unity. To simplify the writing we shall write the cycles and merely indicate -the indices -12vf their ,;:lements. For example we shall write the cycle 12, 23, 31. We can see without difficulty that the cycles of A are the follol<.ring: 12, 12, 12, 12, 12, 12, 23, 23, 23, 25, 25, 25, 31 34, 36, 51 54, 56, 17, 17, 17, 17, 17, 17, 47, 75, 51 67, 75, 51 47, 73, 31 67, 73, 31 All these cycles have orders divisible by k a 73, 73, 73, 75, 75, 75, 31 34, 36, 51 54, 56, 42, 25, 51 62, 25, 51 42, 23, 31 62, 23, 31 3 and we shall separate their first indices into three groups I: 1, 4, 6 II& 2, 7 IIH 3, 5. In group I, one finds the indices 1, 4, 6 - the first indices of the first and fourth pairs of indices in the table above; in group II one finds the first indices of the second and fifth pairs, and so on in the sequence. It is now evi- dent how to proceed in the general case. Notice that to exhaust all the indices 1, 2, 7, in the groups I, II, III, it is sufficient to apply the procedure described in the three first oyc1es. There- fore, we can diminish very considerably the labor of searching for the cyclic forrE of a matrix if we write at the same time as the eyc1e, the partition of the first indices in groups, the number of groups being defined by any means whatever. -13- Take now the permutation of lines and columns of the matrix defined by the substitution C s .. 2 3 6 4 4 5 6 2 7 3 ~ Then the matrix A will be put in th'e desir..e.C1 cyclic form; A'" 2. Let AI' ~, 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 1 1 1 1 1 0 0 0 0 1 1 1 0 0 0 0 ... , An be independent events which we can observe in an infinite sequence of experiments connected in a simple cyclic chain of and let the matrix 0 of the equation (1) be their law. be the number of lines of the L12 , ~3' ... , L K1 of~. fulfilled, one can divide the events A .t.. (2) ,~. Let also t , t 2J 1 + ••• + # ~k-l ~ , These conditions being . ...... . ~ , . \ l' ••• , h-n ) +, of which the fundamental property is the following. The initial experiment can lead us to any of these groups, say B.t..' but the succeeding experiments must necessarily and consecutively lead us into the groups, (3) k into k groups, . . . . . . . . . . . . .. Bk = (~#~1 + ~2 # ind:ex~ -14the succession Bl , B2 , ••• , Bk being repeated without end. One can say that the succession (3) of groups (2) is entirely determined by the results of one initial experiment. Now we can construct the cyclic law such that the events .1-'0 can still be divided into groups, the succession of these groups not being entirely determined by the initial experiment_ ~, . 1 ... Such is for example, the law ~ 0 ~ 0 0 0 ~3 0 0 0 L31 0 0 0 0 0 0 0 0 0 L 45 0 0 0 0 0 0 r,6 L61 0 0 0 0 0 Let t l , t 2 , t3' t , t , and t 6, be the numbers of lines of the~~Ia23~L.jllI,45,L 4 5 and L61. We ahall now divide events A.,l, into 6 groups Bl' B2 , ... , B6 including t l , t 2 , ••• , t 6J events ~.,l, taken in the order of their indices as above. Consider now the experiments subjected to the law ¢l. Suppose that the initial experiment leads us into group Bl • It is clear that the following experi- ment can lead us either into group B or group B - In the first case the experi2 4 ments number 2 and 3 will necessarily lead us to groups B and B and in the second l 3 case the three experiments numbered 2, 3, 4, have us pass through the succession B , B , Ble 6 6 Thus we shall have either the cycle Bl' B2 J By Bl or the cycle -15In an infinite number of experiments these cycles will be repeated without end, but in an order regulated by chance. We notice immediately that group B has a particular property; from this l group, we can pass either to group B or to group B , whereas all the other groups 2 4 B always lead us into a single determined group. Therefore, group B is a critical l group, a group of indetermination, a group of branching of cycles. It is the existence of cyclic laws similar to Jl' the polycyclic laws, which I should like to draw to the attention of those who are concerned with the theory of Markoff chains. These laws can be very complicated and show us how complex the phenomena of the real world can be, :: e.riori. We can expect that their study will aid 3. US in the understanding of these phenomena. Of the numerous problems which are raised by the notion of a polycyclic law of probabilities in a chain, we shall stop only for the following two' (a) To recognize whether a given law is polycyclic; (b) To study the behavior of the probabilities Pk/~ and ¢ %) for k -+ 00. To simplify and shorten the article we shall consider only a few examples of these problems which will nevertheless be sufficient to show the general methods for their solution. (a) Let us take an example of the first problem. Alii 0 1 1 0 1 0 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 1 0 1 0 1 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 Take as given the matrix -16where to simplify the writing, the unit element is substituted for all non-zero elements. To determine whether it is cyclic, we shall look for its cycles. These cycles are 1,3 3,2 1,3 3,10 1,4 4,2 5,3 3,2 2,9 10,9 2,9 2,9 1,6 1,8 5,6 5,8 9,1 9,1 9,1 9,5 6,7 8,7 6,7 8,7 7,1 7,1 1,1 7,1 etc.; as above we indicate only the indices of the elements of these cycles. see two groups of cycles: of three and of four elements. We Therefore, we can suppose that A is bicyclic and take as the first indices of the cycles the six following groups: I 1,5 II III 2,10 3,4 IV V VI 9 6,8 7 Now, by making in A. the simultaneous permutation of lines and columns defined by the transformation 5 s "" 2 3 3 4 4 7 6 8 9 2 10 9 5 6 8 we obtain for ! the form o o 0 1 1 o o o 1 1 0 1 1 o o o 1 1 o 0 o 0 1 1 0 o 0 o 0 1 1 o o 0 0 o o 0 o o 0 1 1 o 0 o o 0 o 0 o o o o o o .0 0 o o 1 1 o o 0 o o o o o o 1 o I o o o o o o o o o which places in evidence its hicylic nature. 0 0 0 0 0 0 0 -17- ..- The events AI' ~, ... , ~o' subject to a law ~ which has the form of A*, are divided into groups Bl • (AI' ~), B2 • (A3' ~), B3 • B4 ... 5 "" (Ae' (~), B (As' A6 ), ~), B6 • (AIO ) which will succeed each other in two cycles B , B , B , B , B and B , B ' B , l 2 3 4 l l S 6 Bl (1). (b) Let us pass to the second problem and consider only bicyclic laws. such a law events ~, ~, For ••• , An are divided in general into groups ••• , B m which, in the experiments succeed each other in two cycles (1) We can still utilize the examination of the characteristic equation of the chain considered for the solution of the first problem. If, for example, the chain is bycyclic and its cycles in B have orders k and t its characteristic equation is necessarily of the form d.·(1) "" >.,r ~ C il.gk+ht "" 0, g,h where r ... Constant and g, h are integral, positive numbers such that gk + h.e~n - r. Therefore, if the exponents of the characteristic equation of the chain considered have the form r + gk + ht, we can conclude that we have a chain which is bicyclic and whose cycles in B have the orders k and -e. That does away with the necessity of looking for all the cycles formed by the elements ¢ -43' for us. -18- and (5) for example, the orders k and t • m - k + 1. If k and t, or, more generally, the orders of the most numerous cycles (when a law is polycyclic) and similar to cycles (4) and (5), do not have a common divisor and i f the matrix corresponding to ~ is not decomposable, we easily see that §< does not have roots wi. th absolute value 1 other than ~ p~ = lim Pk/~' babilities 1. c Therefore, in this case, the final probabilities of.~, will exist and be determinate, their limits for the transition pro- ¢~) depending only on ~. We can prove this proposition by using the well known formulae and representing the probabilities of the roots of Pk/~ and ¢ ~~) as functions i. On the other hand, if the orders of the cycles (4) and (5) have a common divisor, say v, the matrix p will necessarily have among its roots the roots of the equation >..V - 1 = 0 and the probabilities k -+ 00, Pk/~ and ¢ 5;) will not have, for a unique limit, but will have different limits depending on the initial PO~ probabilities and the values taken by k. For example, for the bicyclic law j we have~· (>..) = = >..2(>.. _ 1) 0 a 0 0 b 0 0 0 1 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 (J + >..2 + >.. + a) and , a +b co 1, o < a < 1, -19¢(k) .,. ~ 1.<{3 (1) II (1) where AI' }.~@ (A) + D[ ] (X - 1) (}.) + }.2 + }. + a) ~, ~, }. =0 are the roots of the equations }.3 + >..2 + }. + a ... 0 By calculating the minor~(}.), we shall verify at once that the limiting values of p ~) are ¢ (00) ... ¢ (00) = ¢( 00) ... .,(,2 .4 .,(.J ¢ (00) .. ~5 ¢ Pk/~ • 3a + 2b ' b • - 1 + 3a + 2b ' .,(,6 = 1, (~ The lim (00) _ a 1 + 2, ••• , 6) has the same values. Again let 1= 0 a b 0 0 1 0 0 0 0 0 0 0 1 0 0 0 O' 0 1 1 0 0 0 0 J a +b =1, 0< a < 1, We now have 4 --;t I (}.) .,. }.(A - and ¢(k) ..q3 .. 2 a). - ~ f (1) b) ... }.(A2 - 1) (A2 + b) + _.~-l)ki='43(~l) _f ( - 1) }.k~~~ (}.) + D (}.2 _ 1) (}.2 + b) -20- wh~re Al and ~ are the roots of the equation A2 Therefore, (k) ¢ r-_ + b :s O. .:t.,(~ (1) ~ (1) 043 and the n(k) have the values given in the table below where 043 1 + (_l)k .. ' 1 _ C-l)k = 1 + a + 36 1 + a + 3$ and Qk Qk Table of Values of ~ 1 2 1 Q k aQ 2 Q k 3 Q k aQ 4 Q k aQ Q k aQ 5 t t , aQ 3 t k k k I k k bQ bQ bQ bQ bQ t k k k , k k n~) 4 5 bQ bQ bQ bQ bQ bQ k t k t k k t k bQ bQ bQ bQ t k k k , k k It is evident that, in the present case, the values of the limits of the ¢ (~)for k ...,. 00 are different according to whether k passes through even or odd .,(1" • or indeterminate values when k takes all the integral values. Besides, the two systems of limits of ¢ ~)for k increasing by even or odd values depends on therefore the asymptotic values of the Pk/~ which are equal to are oscillating and depend on the initial probabilities. .J.., -21- III. 1. Statistical Problems The statistical problems connected with the Markoff chain have not been considered up to the present though they appear quite naturally when dependent experiments are studied. In fact, the simplest hypothesis about the nature of the dependence which regulates the experiments considered, consists of supposing that it represents a simple Markoff chain. The first of these problems is to establish the law of the assumed chain, that is, finding the unlmown probabilities ¢.,(~ of the chain frequencies of the events AI' ~, ... , ~ I by observing the in the experiments examined. This can be done in two ways. a) We can observe N series of which each one consists of s experiments. $ny one series, number h, will be divided into partial series ~(h) (h) SAl' of observations "after u~' .•• , A:1.", "after S(h) A r A:2" ••• , II after ~" •. be the number of observations which constitute the repetition of the event ~ in this series. is an empirical value of the probability ¢ 4' The most probable value of ¢ co<l3 is defined by the approximate equality ~ s(h) do 'fI .,(~ h 4 =::::--- ~ (h) ~ s.,( h We well know how the precision in this equality is estimateq and we shall not insist on this point. We only note that we can judge the quality of the adjustment obtained by comparing the numbers s~), ~ = 1, 2, ".J n, -22with the numbers s,.(, ¢043' ~ '" 1, 2, ••• , n, s.,(, '" of the test of K. Pearson i . (h) ~ ~ ( _ ne' 8 ~ s~h) , with the aid h )2 c43 ...__,.(,..:.,,:-'P_c43 ......_ S.,( ¢.(~ -.;8 b) The first procedure gives"tB series S(h) with different lengths, which is A,.(, To avoid this inconvenience, we can observe N ~n~nient in practice. series such that in each of them, every partial series sih)iS of the same length, which we can obtain, for example, by keeping 5, with only s observations. e~ch h series si ) Go( The series having been obtained, we proceed as above. Let us say several words on the determination of the limiting probabilities p~ a lim Suppose N = KL series, consisting of K partial Pk/~. series with L sequences of observations each have been observed. , these partial series by Sk' k a 1, 2, ••• , K. , Designate In Sk' we find L sequences each consisting of s observations; let n(k) be the number of cells of those ~ sequences which terminate by the observation of the event .z ~ ~J we have n(k)= L. ~ n(k) 7- is evidently an empirical value approaching the probability Pk/;3': the most probable value of this latter is .~ (k) fn~ • K1 2. kn important problem which can be set for a series of experiments whioh are supposed to be in a chain is to recognize whether the chain is simple or not. We can solve it as follows. -23Let us designate by p~~.J the probability of events 11",/ 11 ll" ~ in the experi- ments number k, k + 1, k + 2 respectively; evidently (1) P ( ~) = p. k/~ kj"" rA. d.. lfJ"" l' lfJ 113 • Let us now consider table I in which Table I ~ ~ • • 0 ~ p(l{ ) k , .p ~ (t) Pk I 11 p( <r) ~ K (t) P k I nl 0 0 0 •• p( ~) kin. (~) Pk I nn 0.. p( ~ ) k , .1 • o. P ( 't) k I on p( ~) k I.p =~ P =~ "" (~ ) Pk I""p p( ( ) k = Pk I .. ~ I "" .,.( ~ , ("IS) .p Pk I ~ k+lf~ p( ~) _ ~ ("If) k , •• Pk I c43 ""'P 1.,(,. p( ~) k 11• • ('() P k I In ................. p( (l') k I aI,,. (2) o•• = is given. g~' Pk+ll ~ • We lmow that, in a simple chain, events A"" and A~ ¢ ~ should be independent when Therefore, the coefficient of contingency of table I, (3) where 2 (4) -24 should reduce to zero for a simple chain, which we verify without difficulty for 6~i~; 0 in consequence of (1) and (2). We now see how we would proceed to verify the simplicity of a chain from the observations. We must construct s series each one consisting of three observations in the experiments number k, k + 1 and k + 2; let us call such a series a k-triad. Of these sk-triads, there are S ( 'lJ) which may have A ~ in and among these there may be th~ (k + 1 s!;l) which give the sequence A...<., h (J' ret experiment ~. Therefore, for s large, we can write the approximate equalities s(-a') ..g? '""'"'p ( ~ ) s -.. "k(-43 ' = ~ s ( (f ) ~ ~ '""'-' ,...., sp ( (f ) kf-<..' from which we deduce: (5) Therefore, in the case of a simple chain, we should have approximately from which also -25where K is the empirical coefficient of contingency constructed for our observations. By the well known formulae of the theory of contingency we can estimate the precision of equations (6) and, in this manner, render our conclusions on the nature of the chain considered more exact and more solid. We can proceed further. The observations which we have described just now can be summarized by table II which correspond to table I. Now if the chain considered is simple, the Table II ~ ( If) ( Al sll • • • • • • • • ( An ~ snl (~ s ( is ) ·~ s. I ) ) s ( lS ) A n • • • Al .,t, • ~) s ( l5 ) • • • sln • • 1 • ·• . • • • • • s( • s( ?f ) n • ZS ) nn s( ~) • n • • • s ( CS) rows and columns in table I are proportional as we can verify immediately. There- fore, if our observations relate to such a chain the rows and columns of the table II should be approximately proportional. This proportionality can be estimated as follows. We construct the sequence , of which the sum is ~ = 1, 2, ••• , s(~). In this case the criterion _( () ( 2=~S-43 X.,t, ~ _ (?( ) ) s.p . s ( ?f) ~ 2 n; -26and the table of Elderton of the corresponding probability P (we take n' ... n to enter this table in our case) willallow us to estimate the degree of accordance -((5 ) of the sequences s-<13 and s~ ~), ~ ... 1, 2, ... , n, for any.,(. If we calculate and enter this table of Elderton with n' ... n(n - 1) + 1 we whal1 find the probability p which estimates the simultaneous discrepancy for all the sequences ;~ ), ~ ... 1, 2, ... , n; -< ... 1, 2, ... , n and ( 2)' ) ~ , ~ s. ... 1, 2, ••• , n (See R.A. Fisher, Statistical Methods for Research Workers, 1934, p.104). Notice that for a chain which we can suppose stable, the triads can be formed n~~bers by the results of the experiments 1, 2, 3; 2, 3, 4; and so on. In the case of a non-stable sequence, the number k for the k triads can be any whatsoever, but is fixed by the entirety of the triads considered. Finally we can apply a third method. Among the consecutive triads, for ex- ample formed from results of the experiments 1, 2, 3; 2, 3, those which contain A"6' bility ~ for A~, ~ 4; etc, we will choose being fixed, and will determine the empirical proba- <;:) of the A~, ~ ... 1, 2, ... , n, in the last element of our triad. and -< ... 1, 2, ••• , n, in the first element. Then, if the observed chain is simple, the quantities ¢( (J 1 ~ should be approximately equal. ) , ¢«f) 2 ~ , ••• , ¢('If) n ~ - ( If ) The probabilities ¢co<j3 will be more precise if -27we determine them by the means n«?f) .. 'P 4 ! .e ~ f ;(Cn 'P..<I3 i ' ¢~~) being calculated to divide the partial sequences of triads. In this case we must have tn2 in such partial sequences. The agreement of results obtained the with the hypothesis that the chain in question is simple is verified as follows. The numbers ~(o ) m( t) .. ~ ~ m 'P..<I3 ,,,( .. 1, 2, ••• , n, i 4 ~ where m is the length of the partial sequences supposed equal, should be approximately equal; by estimating their agreement by one of the well known methods, we will judge if our hypothesis is admissable or not. 3. Let us consider, in conclusion, a new notion which can be useful in the theoretical or empirical examination of Markoff chains, the notion of the rigidity of a given chain. We call rigidity or the coefficient of rigidity of the chain 1: =/I ¢ ..(~ 11 the number (7) which can also be written in the form (8 ) This number depends, in general, on the number k, that is, on the place in the infinite chain considered; they are not equally rigid throughout, but they become more and more uniformly rigid for determined limiting probabilities. constant and equal to (9) k~ 00, i f the chain considered possesses If the chain is stabilized, its rig~dity is -28- The name given to the number K is justified by its following properties. k (a) For the independent events ~, ~, ... , An subject to the law we j, have K :: O. k We can verify in this case this equality, either directly or a consequence of the general theory contingence, because we perceive that K is nothing else k than the coefficient of contingence of events A,,( in the experiments number k and ~f k + 1. (b) We always have K ~ 1, and Kk"... 1 is possible only for the A..< connected k by an absolutely rigid chain, that is, one such when the initial experiment, which is regulated only by chanqe, is made and given us for example, A..<; we have neces- $rily in the following experiments the events AI' ~, In fact, K k ~ ••• , An succeeding themselves in an entirely defined order. 1, because in general ... n, therefore, Kk < 1. But if the chain is cyclic of index n (in this case it is T evidently absolutely rigid) we can put the law 5t in the form where ¢ 12 IS ¢ 2) ... ••• "" ¢ n-l, n :: ¢ nl ... 1 and all the other ¢ -<j3 are zero. Then -29- r-, 1 .. 1 \I /, Kk = -m-l 1-·- L.,(~ 1 rPk + 1/2 = m-l LPKlI n 1 n-.1. n...1. l/~ I J Pk + III + .. ·PkJn 1 ~.J .. n-l =..,-~=l because Pk + If~ + 1= 1, if in the experiment number k we have if we have ~ in the experiment number k, therefore A~, and Pk + 1: 1= 1, , ... , (c) Conversely, if K = 0, the events A are independent and if K = 1, the chain ~ is necessarily cyclic of indice n, therefore absolutely rigid, as we can easily verify. In this way the value of the number K shows us, the more or less strong k dependence of events AI' ~, ••• , An in the experiments connected in a chain. At the same time K shows also a more or Ie ss apparent, more or less temporary succesk sion of these events and represents in this wayan essenti21 characteristic of the processes described by the chain considered.
© Copyright 2024 Paperzz