i. e., if many observational series can be obta<strong>in</strong>ed under similarstatistically homogeneous conditions. More <strong>of</strong>ten, however, we haveonly one such series, distributions <strong>of</strong> probabilities certa<strong>in</strong>ly can not bereconstructed <strong>and</strong> <strong>the</strong> model <strong>of</strong> an n-dimensional distribution isabsolutely useless. However, if we assume that <strong>the</strong> jo<strong>in</strong>t distribution <strong>of</strong><strong>the</strong> magnitude ξ 1 , ξ 2 , is <strong>the</strong> same as that <strong>of</strong> ξ 2 , ξ 3 , <strong>of</strong> ξ 3 , ξ 4 , etc, <strong>the</strong>n <strong>the</strong>pairs (ξ 1 , ξ 2 ), (ξ 2 , ξ 3 ),..., (ξ n−1 , ξ n ) provide many realizations, althoughperhaps not mutually <strong>in</strong>dependent, <strong>of</strong> that bivariate distribution. Such adistribution is <strong>the</strong>refore determ<strong>in</strong>able <strong>in</strong> pr<strong>in</strong>ciple.It is convenient to generalize somewhat <strong>the</strong> ma<strong>the</strong>matical model. Letus consider a sequence <strong>of</strong> r<strong>and</strong>om variables <strong>in</strong>f<strong>in</strong>ite <strong>in</strong> both directions... ξ −1 , ξ 0 , ξ 1 , ξ 2 , ..., ξ n , ξ n+1 , ... (1.10)called a stochastic process. We assume that <strong>the</strong>oretically <strong>the</strong>re existdistributions <strong>of</strong> probabilities <strong>of</strong> any f<strong>in</strong>ite set{ξ α , ξ β , ξ γ } (1.11)<strong>of</strong> r<strong>and</strong>om variables. Our observational series (1.9) is a part <strong>of</strong> <strong>the</strong><strong>in</strong>f<strong>in</strong>ite sequence (1.10) <strong>and</strong> only allows us to reach some conclusionsabout that whole process if <strong>the</strong> model <strong>in</strong>cludes a rule represent<strong>in</strong>gdistributions <strong>of</strong> magnitudes (1.11) with negative <strong>and</strong> large positivesubscripts through <strong>the</strong> distribution <strong>of</strong> <strong>the</strong> observed variables (1.9).Without such a rule <strong>the</strong> model <strong>of</strong> a stochastic process is useless.In <strong>the</strong> most simple <strong>and</strong> most natural case <strong>the</strong> condition <strong>of</strong>stationarity is imposed: for any τ <strong>the</strong> distribution <strong>of</strong> <strong>the</strong> variables (ξ α+τ ,..., ξ γ+ τ ) co<strong>in</strong>cides with that for τ = 0. The model <strong>of</strong> a stochasticprocess consists <strong>in</strong> that [now] we consider our observations (1.9) as apart <strong>of</strong> <strong>the</strong> realization (1.10) <strong>of</strong> a stationary stochastic process.When assum<strong>in</strong>g a model <strong>of</strong> a stochastic process, only bivariatedistributions are usually applied <strong>and</strong> <strong>in</strong> addition only <strong>the</strong> correlationbetween <strong>the</strong> different values <strong>of</strong> that process are studied. It ought to besaid that <strong>in</strong> spite <strong>of</strong> <strong>the</strong> popularity <strong>of</strong> <strong>the</strong> concept <strong>of</strong> stochastic process,only quite a few examples can be cited <strong>in</strong> which it allowed to describeadequately <strong>the</strong> statistical properties <strong>of</strong> observational series. Mostpublications beg<strong>in</strong> by stat<strong>in</strong>g that a pert<strong>in</strong>ent stochastic processspecified <strong>in</strong> such <strong>and</strong> such a way is given, but <strong>the</strong>re really are only afew works where <strong>the</strong>se specifications are <strong>in</strong>deed determ<strong>in</strong>ed<strong>the</strong>oretically or experimentally.The <strong>the</strong>ory <strong>of</strong> stochastic processes is here suitable for solv<strong>in</strong>gabstract problems: what will happen if a white noise <strong>of</strong> a given<strong>in</strong>tensity <strong>in</strong>fluences some system. Such problems, however, only<strong>in</strong>directly bear on <strong>the</strong> real behaviour <strong>of</strong> a system because under realconditions it is not likely <strong>the</strong> white noise that <strong>in</strong>fluences <strong>the</strong> system, –it does not even concern a stochastic process (lack <strong>of</strong> statisticalhomogeneity). But meanwhile <strong>of</strong>ten no one studies what is reallyact<strong>in</strong>g on <strong>the</strong> system because such <strong>in</strong>vestigations are complicated,difficult <strong>and</strong> expensive so that it is much easier to restrict <strong>the</strong> attentionto arbitrary prior assumptions.54
It is <strong>in</strong>terest<strong>in</strong>g <strong>the</strong>refore to see what occurred when <strong>the</strong> mostem<strong>in</strong>ent statisticians attempted to study actual data by models <strong>of</strong> astochastic process. Ra<strong>the</strong>r <strong>of</strong>ten <strong>the</strong>y experienced failure, see Chapter3. We will also briefly mention <strong>the</strong> statistical <strong>the</strong>ory <strong>of</strong> turbulence <strong>in</strong>which <strong>the</strong> notion <strong>of</strong> stochastic process has been applied with brilliantsuccess.2. The Method <strong>of</strong> Least SquaresGauss discovered <strong>and</strong> <strong>in</strong>troduced it <strong>in</strong>to general usage. The classicalcase which he considered consisted <strong>in</strong> that some known relationsshould be ma<strong>in</strong>ta<strong>in</strong>ed between <strong>the</strong> terms <strong>of</strong> <strong>the</strong> observational seriesx 1 , x 2 , ..., x nhad not <strong>the</strong> observations been corrupted by errors. For example, <strong>in</strong> <strong>the</strong>case <strong>of</strong> <strong>the</strong> path <strong>of</strong> an object <strong>in</strong> space 3 it would have been possible toexpress all terms <strong>of</strong> <strong>the</strong> series through a few <strong>of</strong> its first terms had <strong>the</strong>sebeen known absolutely precisely. This classical case can becomparatively easily studied with<strong>in</strong> <strong>the</strong> boundaries <strong>of</strong> ma<strong>the</strong>maticalstatistics. Practical applications <strong>of</strong> <strong>the</strong> method <strong>of</strong> least squares canencounter more or less essential calculational difficulties which weleave aside. O<strong>the</strong>r difficulties are connected with <strong>the</strong> possible nonfulfilment<strong>of</strong> <strong>the</strong> assumption <strong>of</strong> <strong>the</strong> model <strong>of</strong> trend with error. Thus,errors <strong>of</strong> successive measurements <strong>of</strong> distances by radar apparentlycan not be assumed <strong>in</strong>dependent r<strong>and</strong>om variables. It is <strong>in</strong> generalunclear whe<strong>the</strong>r <strong>the</strong>y possess a statistical character so that statisticalmethods are here unreliable <strong>and</strong> moreover helpless.The observations <strong>the</strong>mselves, however, are highly precise <strong>and</strong> canbe made many times, so that statistical methods are not needed <strong>the</strong>re.In spite <strong>of</strong> all <strong>the</strong> merits <strong>of</strong> <strong>the</strong> classical case, its shortcom<strong>in</strong>g is that itoccurs comparatively rarely. Much more <strong>of</strong>ten we are conv<strong>in</strong>ced thatour observations can be approximated by a smooth dependencex i ≈ f(t i )where t i is a variable describ<strong>in</strong>g <strong>the</strong> conditions <strong>of</strong> <strong>the</strong> i-th experiment.The exact form <strong>of</strong> <strong>the</strong> function f(t) is, however, unknown.Methods strongly resembl<strong>in</strong>g those <strong>of</strong> <strong>the</strong> classical case are appliedhere, but <strong>the</strong>ir study <strong>in</strong>dicates that <strong>the</strong>y are not ma<strong>the</strong>maticallyjustified. Ma<strong>the</strong>matical statistics widely applies ma<strong>the</strong>matics but is notreduced to that comparatively very transparent science. <strong>Statistics</strong> isra<strong>the</strong>r an art <strong>and</strong> as such it has its own secrets <strong>and</strong> we will <strong>in</strong>deedbeg<strong>in</strong> by study<strong>in</strong>g <strong>the</strong>m.2.1. The secrets <strong>of</strong> <strong>the</strong> statistical art. When wish<strong>in</strong>g to apply <strong>the</strong>method <strong>of</strong> least squares we can <strong>in</strong> most cases use a computerprogramme compiled once <strong>and</strong> for all. It is just necessary to enter <strong>the</strong>data, wait for <strong>the</strong> calculations to be made <strong>and</strong> <strong>the</strong> pr<strong>in</strong>ter will provide aformula for a curve fitt<strong>in</strong>g <strong>the</strong> observations. However, he who passesall <strong>the</strong>se procedures to a mach<strong>in</strong>e will be wrong. It is absolutelynecessary to represent <strong>the</strong> available data <strong>in</strong> a visible way <strong>and</strong> at least toglance at <strong>the</strong> figure.55
- Page 1 and 2:
Studies in the History of Statistic
- Page 3 and 4: Introduction by CompilerI am presen
- Page 5 and 6: (Lect. Notes Math., No. 1021, 1983,
- Page 7 and 8: sufficiently securely that a carefu
- Page 9 and 10: is energy?) from chapter 4 of Feynm
- Page 11 and 12: demand to apply transfinite numbers
- Page 13 and 14: for stating that Ω consists of ele
- Page 15 and 16: chances to draw a more suitable apa
- Page 17 and 18: Let the space of elementary events
- Page 19 and 20: 2.3. Independence. When desiring to
- Page 21 and 22: Eξ = ∑ aipi.Our form of definiti
- Page 23 and 24: absolutely precisely if the pertine
- Page 25 and 26: where x is any real number. If dens
- Page 27 and 28: probability can be coupled with an
- Page 29 and 30: Nowadays we are sure that no indepe
- Page 31 and 32: λ = λ(T)with λ(T) being actually
- Page 33 and 34: (1/B n )(m − A n )instead of the
- Page 35 and 36: along with ξ. For example, if ξ i
- Page 37 and 38: µ( − p0) ÷np0 (1 − p0)nhas an
- Page 39 and 40: distribution of the maximal term |s
- Page 41 and 42: ξ (ω) + ... + ξ (ω)n1n{ω :|
- Page 43 and 44: P{max ξ(t) ≥ x} = 0.01, 0 ≤ t
- Page 45 and 46: 1. This example and considerations
- Page 47 and 48: IIV. N. TutubalinTreatment of Obser
- Page 49 and 50: structure of statistical methods, d
- Page 51 and 52: Suppose that we have adopted the pa
- Page 53: and the variances are inversely pro
- Page 57 and 58: is applied with P(t) being a polyno
- Page 59 and 60: ut some mathematical tricks describ
- Page 61 and 62: It is clear therefore that no speci
- Page 63 and 64: of various groups of machines, and
- Page 65 and 66: nnA(λ) x sin λ t, B(λ) = x cosλ
- Page 67 and 68: of the mathematical model of the Br
- Page 69 and 70: dF(λ) = f (λ) dλ, so that B( t
- Page 71 and 72: usually very little of them. Indeed
- Page 73 and 74: This is the celebrated model of aut
- Page 75 and 76: applications of the theory of stoch
- Page 77 and 78: achieved by differentiating because
- Page 79 and 80: u(x 1 , x 2 , t 1 , t 2 ) = v(x 1 ,
- Page 81 and 82: Reasoning based on common sense and
- Page 83 and 84: answering that question is extremel
- Page 85 and 86: IIIV. N. TutubalinThe Boundaries of
- Page 87 and 88: periodograms. It occurred that work
- Page 89 and 90: at point x = 1. However, preceding
- Page 91 and 92: He concludes that since the action
- Page 93 and 94: The verification of the truth of a
- Page 95 and 96: In the purely scientific sense this
- Page 97 and 98: ought to learn at once the simple t
- Page 99 and 100: the material world science had inde
- Page 101 and 102: values of (2.1) realized in the n e
- Page 103 and 104: *several dozen. The totality µ ica
- Page 105 and 106:
Mendelian laws. It is not sufficien
- Page 107 and 108:
example, the problem of the objecti
- Page 109 and 110:
a linear function is not restricted
- Page 111 and 112:
258 - 82 - 176 cases or 68.5% of al
- Page 113 and 114:
The Framingham investigation indeed
- Page 115 and 116:
or, for discrete observations,IT(ω
- Page 117 and 118:
What objections can be made? First,
- Page 119 and 120:
eliability and queuing are known to
- Page 121 and 122:
Kolman E. (1939 Russian), Perversio
- Page 123 and 124:
measurement is provided. Recently,
- Page 125 and 126:
which means that sooner or later th
- Page 127 and 128:
The foundations of the Mises approa
- Page 129 and 130:
A rather subtle arsenal is develope
- Page 131 and 132:
4.3. General remarks on §§ 4.1 an
- Page 133 and 134:
BibliographyAlimov Yu. I. (1976, 19
- Page 135 and 136:
processes are now going on in the s
- Page 137 and 138:
obtaining a deviation from the theo
- Page 139 and 140:
VIOscar SheyninOn the Bernoulli Law