h * = S χ *2,21Γ(n / 2)χ( n−2) / 2 *0<strong>and</strong> h1* determined by equalities n-1 exp(– 2 /2)d = 1/2, (99)h1* = S 1 χ1* 2,21Γ(n / 2)χ 1( n−3) / 2 *0 n-2 exp(– 2 /2)d = 1/2 (100)respectively. I adduce the values of ( χ *) 2 for n ≤ 20. {The table, actually for n = 1(1)10, isomitted here.} For n > 10 we may consider that, with an error less than 0.01,( χ *) 2 ~ n – 2/3. (101)For n observations χ 1* is equal to χ * corresponding to (n – 1) observations 17 . If the choiceof the approximate values of a <strong>and</strong> h is considered in itself, then it is nevertheless natural torestrict it to the values satisfying the following two conditions 18 .1. In each problem, the approximate values depend, in addition to the parameters supposedto be known, only on the appropriate sufficient statistics, or sufficient systems of statistics.2. The approximate values are invariant with respect to the transformations of the Ox axisof the type (64).It is possible to conclude from these dem<strong>and</strong>s that only x may be taken as an approximatevalue of a; <strong>and</strong>, as an approximate value of h, we ought to assume, in Problem 2, h = B /S;<strong>and</strong> in Problem 3, h1= B1/S where B <strong>and</strong> B 1only depend on n.We may apply various additional conditions for determining the most sensible values ofthe factors B <strong>and</strong> B 1. For example, it is possible to dem<strong>and</strong> that, for any values of a <strong>and</strong> h,the conditions E( h |a; h) = h, E( h 1|a; h) = h be satisfied. These dem<strong>and</strong>s can only be met ifB = B 1, B =Γ [( n + 2) / 2]Γ [( n + 1)/ 2], B1=Γ [( n + 1) / 2],Γ(n / 2)i.e., if, assuming these values for B <strong>and</strong> B 1, we will haveh = B /S <strong>and</strong> h1= B1/S. (102)We shall soon see that the finality of this last result should not be overestimated.The dem<strong>and</strong> made use of above that the systematic error be absent, can be formulated withrespect to the approximate value θ of any parameter under consideration. In a generalform, this dem<strong>and</strong> is expressed thus: The equalityE(θ |; 1 ; 2 ; …; n ) = (103)should hold for all possible values of the parameters , 1 , 2 , …, s of a given problem. Theapproximation x for the center of scattering a satisfies this dem<strong>and</strong> both in Problem 1 <strong>and</strong>Problem 3. Now we will determine approximations devoid of systematic error for the meansquare deviation (9). It is natural to restrict our attention here to approximations of the formσ = kS (Problem 2) <strong>and</strong> σ1= k 1 S 1 (Problem 3) 19 . The dem<strong>and</strong> that the systematic error beabsent, E(σ |a; h) = , E( σ1|a; h) = , leads then to the necessity of assuming
Γ(n / 2)Γ [( n −1) / 2]k = (1/2), k 1 = (1/2), (104)Γ [( n + 1) / 2]Γ(n / 2)i.e., of choosing σ <strong>and</strong> σ1(see above) in accord with these magnitudes. However, after thatit is natural to take as an approximation to hΓ [( n + 1) / 2]h = (1/S)Γ(n / 2), h1= (1/S 1 )Γ(n / 2)Γ [( n −1) / 2](102bis)in Problems 2 <strong>and</strong> 3 respectively.If the lack of systematic error is dem<strong>and</strong>ed in the determination of 2 (of the variance), <strong>and</strong>if the approximation to 2 is being determined in the form σ 2 = qS 2 (Problem 2) <strong>and</strong> σ 2 1=q 1 S 2 1 (Problem 3), then we have to assumeq = 1 – n, q 1 = 1/(n – 1), σ 2 =S 2 /n, σ 2 1= S 2 1 /(n – 1). (105)To achieve conformity with formula (103), it is then natural to assume the approximationto h ash =n / 2S, h1= ( n − 1) / 2S1(98ter)in Problems 2 <strong>and</strong> 3 respectively. The last approximations are the most generally accepted inmodern mathematical statistics. We have applied them in §4, where, however, the choicebetween (98), (98bis) <strong>and</strong> (98ter) was not essential since the limit theorems of §4 persistedanyway.From the practitioner’s viewpoint, the differences between the three formulas are notessential when only once determining the measure of precision h by means of (1). Indeed, thedifferences between these approximations have order h/n, <strong>and</strong> the order of their deviationsfrom the true value 20 , h, is higher, h/n. Therefore, since we are only able to determine h towithin deviations of order h/n, we may almost equally well apply any of theseapproximations which differ one from another to within magnitudes of the order h/n.The matter is quite different if a large number of various measures of precision (forexample, corresponding to various conditions of gunfire) has to be determined, each timeonly by a small number of observations. In this case, the absence of a systematic error insome magnitude, calculated by issuing from the approximate value of the measure ofprecision, can become very essential. Depending on whether this magnitude is, for example,h, or 2 , the approximate values of h should be determined by formulas (98), (98bis) or(98ter) respectively.In particular, from the point of view of an artillery man, according to the opinion of Prof.Gelvikh [10; 11] 21 , it is most essential to determine without systematic error the expectedexpenditure of shells required for hitting a target. In the most typical cases (two-dimensionalscattering <strong>and</strong> a small target as compared with the scattering) this expected expenditure,according to him, is proportional to the product (1) (2) of the mean square deviations in thetwo directions. Suppose that we estimate (1) <strong>and</strong> (2) by their approximations σ (1) <strong>and</strong> σ (2)derived from observations (x 1 (1) , x 2 (1) , …, x n (1) ) <strong>and</strong> (x 1 (2) , x 2 (2) , …, x m (2) ) respectively. If thex i (1) are independent of x j (2) , then, for any (1) , (2) , a (1) <strong>and</strong> a (2) (where the last two magnitudesare the centers of scattering for x i (1) <strong>and</strong> x j (2) respectively), we haveE(σ (1) σ (2) |a (1) ; a (2) ; (1) ; (2) ) = E(σ (1) |a (1) ; (1) ) E(σ (2) |a (2) ; (2) )
- Page 3:
of All Countries and to the Entire
- Page 6 and 7:
(Coll. Works), vol. 4. N.p., 1964,
- Page 8 and 9:
individuals of the third class, the
- Page 10:
From the theoretical point of view
- Page 13 and 14:
Second case: Each crossing can repr
- Page 15 and 16:
On the other hand, for four classes
- Page 17 and 18:
f i = i S + i , i = 1, 2, 3, 4, (
- Page 19 and 20:
f 1 = C 1 P(f 1 ; …; f n+1 ), C 1
- Page 21 and 22:
ut in this case f = 2 , f 1 = 2 ,
- Page 23 and 24:
I also note the essential differenc
- Page 25 and 26:
A 1 23n1 + 1 A 1 A 1 … A 11A 2 A
- Page 27 and 28:
coefficient of 2 in the right side
- Page 29 and 30:
h(A r h - c h A r 0 ) = - A r0we tr
- Page 31 and 32:
Notes1. Our formulas obviously pres
- Page 33 and 34:
Bernstein’s standpoint regarding
- Page 35 and 36:
Corollary 1.8. A true proposition c
- Page 37 and 38:
It is important to indicate that al
- Page 39 and 40:
ut for the simultaneous realization
- Page 41 and 42:
devoid of quadratic divisors and re
- Page 43 and 44:
propositions (B i and C j ) can be
- Page 45 and 46:
A ~ A 1 and B = B 1 , we will have
- Page 47 and 48:
included in a given totality as equ
- Page 49 and 50:
For unconnected totalities we would
- Page 51 and 52:
proposition given that a second one
- Page 53 and 54:
On the other hand, let x be a parti
- Page 55 and 56:
totality is perfect, but that the j
- Page 57 and 58:
In this case, all the finite or inf
- Page 59 and 60:
probabilities p 1 , p 2 , … respe
- Page 61 and 62:
where x is determined by the inequa
- Page 63 and 64:
totality of the second type (§3.1.
- Page 65 and 66:
x = /2 + /(23) + … + /(23… p n
- Page 67 and 68:
that the fall of a given die on any
- Page 69 and 70:
infinitely many digits only dependi
- Page 71 and 72:
10. (§2.1.5). Such two proposition
- Page 73 and 74:
F(x + h) - F(x) = Mh, therefore F(x
- Page 75 and 76:
“confidence” probability is bas
- Page 77 and 78:
x1+ Lp n (x) x1− Lx1+ Lf(t)dt < x
- Page 79 and 80:
|(x 1 ; t 0 ; t 1 ) - 1 t0tf(t)dt|
- Page 81 and 82:
5. The distribution ofξ , the arit
- Page 83 and 84:
P(x 1i < x) = F(x; a i ) = C(a i )
- Page 85 and 86:
egards his promises. Markov shows t
- Page 87 and 88:
other solely and equally possible i
- Page 89 and 90:
notion of probability and of its re
- Page 91 and 92:
However, already in the beginning o
- Page 93 and 94:
the revolution. My main findings we
- Page 95 and 96:
Nevertheless, Slutsky is not suffic
- Page 97 and 98:
path that would completely answer h
- Page 99 and 100:
on political economy as well as wit
- Page 101 and 102:
scientific merit. Borel was indeed
- Page 103 and 104:
[3] Already in Kiev Slutsky had bee
- Page 105 and 106:
different foundation. The difficult
- Page 107 and 108:
5. On the criterion of goodness of
- Page 109 and 110:
--- (1999, in Russian), Slutsky: co
- Page 111 and 112:
Here also, the author considers the
- Page 113 and 114:
second, it is not based on assumpti
- Page 115 and 116:
experimentation and connected with
- Page 117 and 118:
Russian, and especially of the Sovi
- Page 119 and 120:
station in England. This book, as h
- Page 121 and 122:
Uspekhi Matematich. Nauk, vol. 10,
- Page 123 and 124:
variety and detachment of those lat
- Page 125 and 126:
46. On the distribution of the regr
- Page 127 and 128: 119. On the Markov method of establ
- Page 129 and 130: No lesser difficulties than those e
- Page 131 and 132: Separate spheres of work considerab
- Page 133 and 134: 10. Anderson, O. Letters to Karl Pe
- Page 135 and 136: Hier sind, im Allgemeinen, ganz ana
- Page 137 and 138: Jedenfalls, glaube ich erwiesen zu
- Page 139 and 140: werde ich das ganze Material in kur
- Page 141 and 142: considered as the limiting case of
- Page 143 and 144: and, inversely,] = m ...1 2 N[ ch h
- Page 145 and 146: µ 2 2 = m 2 2 - 2m 2 m 1 2 + m 1 4
- Page 147 and 148: (x k - x k+1 ) … (x k - x +) = E(
- Page 149 and 150: the thus obtained relations as pert
- Page 151 and 152: [1/S(S - 1)(S - 2)][(Si = 1Sx i ) 3
- Page 153 and 154: ( N −1)((S − N )(2NS− 3S− 3
- Page 155 and 156: µ 5 + 2µ 2 µ 3 = U [S/S] 5 + 2U
- Page 157 and 158: case, the same property is true wit
- Page 159 and 160: It follows that the question about
- Page 161 and 162: then expressed my doubts). And Gned
- Page 163 and 164: For Problem 1, formula (7) shows th
- Page 165 and 166: Let us calculate now, by means of f
- Page 167 and 168: ϕ′1(x)1E(a|x 1 ; x 2 ; …; x n
- Page 169 and 170: Theorem 3. If the prior density 3
- Page 171 and 172: P( ≤ ≤ |, 1 , 2 , …, s )
- Page 173 and 174: 6. A Sensible Choice of Confidence
- Page 175 and 176: 0 = A 0 n, = B2, = B2, 0 = C 0 n
- Page 177: Note also that (95),(96), (83),(85)
- Page 181 and 182: f (x 1 , x 2 , …, x n ) = 1 if x
- Page 183 and 184: and the probability of achieving no
- Page 185 and 186: E = kEµ. (14)In many particular ca
- Page 187 and 188: a = np, b = np 2 = a 2 /n, = a/nand
- Page 189 and 190: with number (2k - 2), we commit an
- Page 191 and 192: (67)which is suitable even without
- Page 193 and 194: " = 1/[1 - e - ], = - ln [1 - (1/
- Page 195 and 196: Such structures are entirely approp
- Page 197 and 198: 11. As a result of its historical d
- Page 199 and 200: exaggeration towards a total denial