influenced by a construction that assumes as initial objects of study left without formal definition theset U = {u} of elementary events, <strong>and</strong> the function P(A) called probability having as its domain somesystem F of subsets of the main set U. In 1933 Kolmogorov [24] offered the appropriate axiomatics 4in a complete form although the French school (Borel), <strong>and</strong>, from its very beginning, the <strong>Soviet</strong>Moscow school had begun to develop a related range of ideas much earlier.From the viewpoint of logic <strong>and</strong> philosophy, this system of constructing the theory of probability isnot either the only possible, or preferable to other systems, see our last section. Its great success isapparently due to the following circumstances:1) It is the simplest system of full axiomatization of the theory from among those offered until now5 .2) It enabled to cover, by a single simple pattern, not only the classical branches of the theory, butalso those new chapters that were put forth by the requirements of natural sciences <strong>and</strong> are connectedwith distributions of probabilities for r<strong>and</strong>om functions.3) It connected the theory with the theory of measure <strong>and</strong> the metric theory of functions whichboast a rich collection of subtle methods of research.We shall indeed concentrate on the two last points. According to the concept which we are nowdiscussing, the probability P(A) is nothing else but an abstract measure obeying the normingcondition P(U) = 1; the r<strong>and</strong>om variable is a function (u) measurable with respect to this measure;the expectation E is the Lebesgue integralE = u(u)dP, etc.3.1. The Joint Distribution of Probabilities of an Infinite System of R<strong>and</strong>omVariables. A number of findings of the Moscow school discussed in §2 (<strong>and</strong> especially thoserelated to the convergence of series with r<strong>and</strong>om terms; to the strong law of large numbers; <strong>and</strong> to thelaw of the iterated logarithm) is based on considering a sequence of r<strong>and</strong>om variables 1 , 2 , …, n , … (3.1.1)as a sequence of functions n (u) of one <strong>and</strong> the same argument u with probability P(A) being definedas measure on its domain U. We still have to indicate some more general results.Problems of estimating the probabilities of events depending on the values of a finite number ofvariables (3.1.1) are solved by means of appropriate finite-dimensional distributionsFξ 1 , ξ2,..., ξ n(x 1 ; x 2 ; …; x n ). (3.1.2)If, however, the occurrence of some events depends on values taken by an infinite number ofvariables from (3.1.1), then the law of distribution is naturally considered in the space of numbersequences x 1 , x 2 , …, x n , … of the possible values of variables (3.1.1). This law is uniquely determinedby the totality of distributions (3.1.2).Kolmogorov [24] established a theorem which states that, for the existence of r<strong>and</strong>om variables(<strong>and</strong>, consequently, of their infinite-dimensional law) with given laws of distribution (3.1.2), thecompatibility of these laws, in its usual elementary sense, is not only necessary but also sufficient.The solution of many separate problems repeatedly led to the result that, under certain generalconditions, the probability of some limiting relations concerning sequences of r<strong>and</strong>om variables canonly be equal to 0 or 1. Analogies with the theory of measure prompted Kolmogorov [24] to establishthe following general theorem: If f( 1 , 2 , …, n , …) is a Baire function of independent r<strong>and</strong>omvariables n whose value persists when the values of a finite number of its arguments are changed,then the probability of the equality f( 1 , 2 , …, n , …) = a can only be equal to 0 or 1.3.2. R<strong>and</strong>om Functions. Suppose that a function u (t) belonging to some functional space Ecorresponds to each elementary event, <strong>and</strong> that for any Borel measurable subset A of space E the setof those n for which u (t) ∈ A, belongs to system F. Then we say that a r<strong>and</strong>om function (t) of typeE is given. Following Kolmogorov’s first findings [24], a number of American authors (especially
lim P(|(t+ h) – (t)| > ) = 0 as h ¤ ¥Doob) have developed this logical pattern. Independently of polishing its formal logical side, Slutskysystematically studied the most interesting problems naturally emerging when considering r<strong>and</strong>omfunctions.An n-dimensional law of distribution Ft 1 , t 2 ,..., t n(x 1 ; x 2 ; …; x n ) of the r<strong>and</strong>om variables (t 1 ), (t 2 ),…, (t n ) corresponds to each finite group of values t 1 , t 2 , …, t n of the argument t. One of the mainproblems of the theory of r<strong>and</strong>om functions is, to determine the conditions to be imposed on thefunctions F so that they can correspond to a r<strong>and</strong>om function of some type E. In any case, thefunctions should be consistent in the elementary sense; additional conditions differ, however, fordiffering functional spaces. Slutsky <strong>and</strong> Kolmogorov (especially Slutsky [28]) offered sufficientlygeneral conditions of this type for the most important instances. These conditions are expressedthrough the laws of distribution of the differences (t) – (t), that is, through the two-dimensionaldistributions F t, t (x; x). According to Slutsky, stochastic continuity of (t), i.e., the condition that, forany > 0,for all (or almost for all) values of t, is sufficient for this space of measurable functions. For the samecase, Kolmogorov established a somewhat more involved necessary <strong>and</strong> sufficient condition(Ambrose 1940). For the space of continuous functions Kolmogorov’s sufficient condition is thatthere exist such m > 0 <strong>and</strong> > 1 thatE|(t+ h) – (t)| m = O|h| .The most essential from among the more special results concerning r<strong>and</strong>om functions of a realvariable were obtained in connection with the concepts of the theory of r<strong>and</strong>om processes (where theargument is treated as time), see the next section. For statistical mechanics of continuous mediumsboth scalar <strong>and</strong> vectorial r<strong>and</strong>om functions of several variables (points in space) are, however,essential. Findings in this direction are as yet scarce (Obukhov, Kolmogorov).4. The Theory of R<strong>and</strong>om ProcessesThe direction of research now united under the heading General theory of r<strong>and</strong>om processesoriginated from two sources. One of these is Markov’s work on trials connected into a chain; theother one is Bachelier’s investigations of continuous probabilities which he began in accordance withPoincaré’s thoughts. The latter source only acquired a solid logical base after the set-theoretic systemof constructing the foundations of probability theory (§3) had been created.In its further development the theory of r<strong>and</strong>om processes is closely interwoven with the theory ofdynamic systems. Both conform to the ideas of the classical, pre-quantum physics. The fascination ofboth of them consists in that, issuing from the general notions of determinate process <strong>and</strong> r<strong>and</strong>omprocess, they are able to arrive at sufficiently rich findings by delimiting, absolutely naturally freefrom the logical viewpoint, the various types of phase spaces (the sets of possible states) <strong>and</strong> ofregularities in the changes of the states (the absence, or the presence of aftereffect, stationarity, etc).A similar logical-mathematical treatment of the concepts of the modern quantum physics remains to aconsiderable extent a problem for the future.4.1. Markov Chains. Suppose that the system under study can be in one of the finite or countablestates E 1 , E 2 , …, E n , … <strong>and</strong> that its development is thought to occur in steps numbered by integers t(discrete time). Suppose in addition that the conditional probability of the transition P(E i ¤ E j ) = p ij(t)during step t does not depend on the earlier history of the system (absence of aftereffect). Suchr<strong>and</strong>om processes are called Markov chains.It is easy to see that the probability of the transition E i ¤ E j during t steps having numbers 1, 2, …,t which we shall denote by P ij (t) can be calculated by means of recurrent formulas
- Page 4 and 5:
[3] I bear in mind the well-known p
- Page 6 and 7:
successes of physical statistics. B
- Page 8 and 9:
classes of independent facts whose
- Page 10 and 11: distribution is a corollary of the
- Page 12 and 13: examine in the first place the curv
- Page 14 and 15: 12. According to Bortkiewicz’ ter
- Page 16 and 17: generality, the similarities taking
- Page 18 and 19: one on another, as well as the corr
- Page 20 and 21: is inapplicable because the right s
- Page 22 and 23: Instead, Slutsky introduced new not
- Page 24 and 25: abandoned in August 1936, but it is
- Page 26 and 27: last decades, mathematicians more o
- Page 28 and 29: charged with making the leading ple
- Page 30 and 31: motion and a number of others) are
- Page 32 and 33: phenomena. It is self-evident that
- Page 34 and 35: Such new demands were formulated in
- Page 36 and 37: The addition of independent random
- Page 38 and 39: automatic lathes, etc. Here, the ma
- Page 40 and 41: 11. Kolmogorov, A.N. Grundbegriffe
- Page 42 and 43: period 1 and remained, until the ap
- Page 44 and 45: of the analytical tool rather than
- Page 46 and 47: with probability approaching unity,
- Page 48 and 49: logic. The ensuing vagueness in his
- Page 50 and 51: 2. Gnedenko, B.V. (1949), On Lobach
- Page 52 and 53: will be sufficient, although not ne
- Page 54 and 55: nlimk = 1P(| k (n) - m k (n) | > H
- Page 56 and 57: favorite classical issue as the gam
- Page 58 and 59: and some quite definite (not depend
- Page 62 and 63: P ij (1) = p ij (1) , P ij (t) =kP
- Page 64 and 65: 4.2d. Bebutov [1; 2] as well as Kry
- Page 66 and 67: are yet no limit theorems correspon
- Page 68 and 69: described, from the viewpoint that
- Page 70 and 71: conditional variance and determined
- Page 72 and 73: Romanovsky [45] and Kolmogorov [46]
- Page 74 and 75: Let S be the general population wit
- Page 76 and 77: Part 1. Russian/Soviet AuthorsAmbar
- Page 78 and 79: 2. On necessary and sufficient cond
- Page 80 and 81: Gnedenko, B.V., Groshev, A.V. 1. On
- Page 82 and 83: 52. ( (Mathematical Principl
- Page 84 and 85: Kozuliaev, P.A. 1. Sur la répartit
- Page 86 and 87: Obukhov, A.M. 1. Normal correlation
- Page 88 and 89: 30. Généralisations d’un théor
- Page 90 and 91: 22. Alcune applicazioni dei coeffic
- Page 92 and 93: 10. A.N. Kolmogorov. The Theory of
- Page 94 and 95: Kuznetsov, Stratonovich & Tikhonov
- Page 96 and 97: In the homogeneous case H s t = H t
- Page 98 and 99: to such a generalization. He only s
- Page 100 and 101: In the particular case of a charact
- Page 102 and 103: as it is usual for the modern theor
- Page 104 and 105: 1. {The second reference to Pugache
- Page 106 and 107: Smirnov, Romanovsky and others made
- Page 108 and 109: determined the precise asymptotic c
- Page 110 and 111:
for finite values of N, M and R 2 .
- Page 112 and 113:
Mikhalevich’s findings by far exc
- Page 114 and 115:
Uch. Zap. = Uchenye ZapiskiUkr = Uk
- Page 116 and 117:
Khinchin, A.Ya. 43. Math. Ann. 101,
- Page 118 and 119:
7. DAN 115, 1957, 49 - 52.Pinsker,
- Page 120 and 121:
Anderson, T.W., Darling, D.A. (1952
- Page 122 and 123:
Statistical problems in radio engin
- Page 124 and 125:
observations for its power with reg
- Page 126 and 127:
securing against mistakes (A.N. Kry
- Page 128 and 129:
of the others, then its distributio
- Page 130 and 131:
In Kiev, in the 1930s, N.M. Krylov