Scribd
Upload
32 views7 pages

Aeroian - The Probabability Function of The Product of Two Normally Distributed Variables

The Probabability Function of the Product of Two Normally Distributed Variables

Uploaded by

David C Houser
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
32 views7 pages

Aeroian - The Probabability Function of The Product of Two Normally Distributed Variables

The Probabability Function of the Product of Two Normally Distributed Variables

Uploaded by

David C Houser
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 7
THE PROBABILITY FUNCTION OF THE PRODUCT OF TWO NORMALLY DISTRIBUTED VARIABLES' By Lo A. ARorAN Hunter College 1, Introduction and summary. Let «x and y follow a normal bivariate prob- ability function with means X, Y, standard deviations 01, 02, respectively, r the coefficient of correlation, and p. = X/o1, p: = Y/oz. Professor C. C. Craig [1] has found the probability function of z = «y/eio2 in closed form as the difference of two integrals. For purposes of numerical computation he has expanded this result in an infinite series involving powers of z, p: , pz, and Bessel functions of a certain type; in addition, he has determined the moments, semin- variants, and the moment generating function of z, However for p: and p: large, as Craig points out, the series expansion converges very slowly. Even for p: and p; as small as 2, the expansion is unwieldy. We shall show that as pi and p2 > ©, the probability function of z approaches a normal curve and in case r = 0 the Type III function and the Gram-Charlier Type A series are excel- lent approximations to the z distribution in the proper region. Numerical in- tegration provides a substitute for the infinite series wherever the exact values of the probability funetion of z are needed. Some extensions of the main theorem are given in section 5 and a practical problem involving the probability function of zis solved. 2, Theorems on approach to normality. ‘The moment gencrating function of z, M.(0), is [1] (oi +p — 2rpspa)6* + 2pr 020 exp (94 98 — 2s fl psf (+ ne + 0 = 6) (2.1) M.() = vi Let 2, and o, be the mean and the standard deviation of z, and é, = (¢ — 2)/ar+ Now 2) Z=om+r, of = Vol +03t Qoimti +t Using (2.2) we find in the usual way the moment generating function of t, —2rw + (pi + 92 + 2rp pa)w* + 4r*w* — Qw*(r* — 1)(o102 + 7) ss Vil OF ult + (1 = rel exp (23) Mi, = where w = 6/o:. * Presented to the American Mathematical Society, Oct. 28, 1944, New York City. 265 266 LEO A, AROIAN Consider r 20. ‘Then in the limit as p, and p, —> in any manner whatever, (2.4) lim M,,@) =e”, paren and by the theorem of Curtiss [2] on moment generating functions we see in the limit as p: , p, > «© the probability function of z approaches a normal curve with mean, Z, and variance o? , r = 0. Incase —1-+ € 0, some care is required wherever occurs. If one uses pi + pi = 2pim, the proof goes forward quite readily. Hence we have proved the theorem: Traore (2.5). The distribution of 2 approaches normality with mean 2, and variance 0; as p: and p, > «© in any manner whatever, —1 +¢ <1 S 1, «>0. It is evident in Theorem (2.5) we may allow p:, p: > — © without any other changes. Theorems (2.6) and (2.7) are proved in essentially the same way as (2.5). Tanorem (2.6). The distribution of z approaches normality with mean 2, and variance 03, if > ©,m—> —%, -1Sr<1l—ee>0. Taworem (2.7). The distribution of 2 approaches normality if p remains constant pp—> ©, —1 +¢€ 0; or tf p; remains constant p, > — ~, -1Sr0. Naturally in any of the theorems p; and p; may be interchanged. In practice px and p are usually positive. The approach to normality is more rapid if both p and p: have the same sign as r. 3. Numerical values. In order to show how closely the Type IIT and the Gram-Charlier Type A series approximate the probability function of z, f(2), or more precisely f(z, px , pr , 7), we use numerical integration where SG, pr, p57) = Le) — Le), 1 1 aoe een LO = eal ov - gts {c ~ oot = arte = w({~ m) @.1) . ( - y} as and J,(z) is the integral of the same function over (— 2, 0), [1]. Now f(z) may be written as @2) LO= Graal ewewooey®, where PROBABILITY FUNCTION OF A PRODUCT 267 We readily obtain L:(z) \/1 — 72 by forming the product of g(t), ¢(4), B(ts), and 1/z using numerical integration applying Weddle’s formula, the Gregory- Newton formula, or the simple reetangular formula depending on circumstances ‘The rectangular formula [3] is remarkably accurate when the function 7 (t))e(t) (ts) /x in the interval 0 to » or 0 to —& is somewhat symmetrical. Appropriate tables for g(i:), ot) (see [4)), 6(s) (see [5]) and 1/x (see [6]) are readily available. In the important case of the independence of x and y, r = 0 and (3.2) becomes ° dx 6s) nee [ ote 2, a-e-m, ami 4, Approximations to fiz). When r = 0, the standard seminvariants &, and & of z are pipe _ 6{2(ot + 2) + 1 = = oy Se Gydspe 8G ean remembering Z=pm,0.= Vatat i. In the Pearson system (see (7)) 6, the criterion, is — 2 — 38 (4.2) a= STE and for the probability function of z (43) 5 = Roi + ot + [2G + pt) + 1) = 18 pts , (ot + e+ I(t + pb + 1)? + oi + v2) + 1) and if pr: = pp = p _ __2(4p* + 1)(2p" + 1) — 189! a Qe + DIG + 1 + Ge + DI Now 6 = 0, & = 0, for the Type III function, and clearly lim 8 = 0. pueve By use of (3.3) the accurate values of f(z) have been calculated for various com- binations of p; and p, and compared with the Type III approximation using 2, oz, bs. (4.5) Investigations so far completed show that for p: 2 4.and p, = 4 simul- taneously, and || < .008, the Type III approximation will provide values of t, correct to three significant figures at least where i fe (4.6) f fle) = a jy fe) =a and 05 Sa S005. = he ‘These are the values of f, which would be needed in testing hypotheses. ‘The exact values of ¢{? and for #2 for various values of p; and p less thian 4 will be 268 LEO A. AROIAN determined it is hoped in the future and will be published along with the com: parisons of the Type III values of ¢, with the accurate values of ¢, in the im- portant borderline cases of ps = p2 = 2, and p, = pm = 3. The values of f(z) for 1 = px = 2and p: = p2 = 4 have been calculated but these are being with- held for a more complete table. The table of values of Z, o:, é&, &, and 3 (Table IT) shows then that the Type III function is excellent along a band about pi = pe, since & # 0, and 3 is very small. We use the Gram-Charlier Type A series of three terms to approximate the probability function of z in ¢, units. 47) fe) ~ 00 = 8 0% + 9, in the usual notation. TABLE I 4 f(s) Correct value | Normal Curve Grann-Charlier «9950372 2406367 - 2431716 + 2408235 1.4925558 «1275209 - 130970 «127484 1.9900744 .0538243 .0550708 -053704 2.4875930 -0184606 -0180791 0184500 2.9851116 0052477 0046338 -0052944 34826302 -0012609 0009272 -0012804 3.9801488 -0002611 -0001449 -000260 4.4776674 .0000467 -0000177 0000425, 4.9751860 .00000745, -00000168, 00000555 (4.8) For | &| < .5 and & < 4 simultaneously the Gram-Charlier Type A series is quite adequate for finding probability levels such as those of (4.6). ‘These will in general give 3 significant figures for ¢” or ¢{°. In the special case 1 = 0, p = 10, the Gram-Charlier Type A series differs from (¢,) very slightly in the range 1 S$ |t,| < « (see Table 1). Naturally the Gram-Charlier will be used wherever Type III is not indicated, although there exist some over- lapping regions where either one may be used. It should be noticed that the approach of f(z) to normality is more rapid along a row than down a diagonal. In case either p; or p: is negative, we may make use of the equation (4.9) F@, —pr, m7) = f(—2, mr, m, —7). We note that when r = 0, f(z, p1 , px) always possesses a discontinuity at z = 0, (see [1]). A table of 2, 02, &, &, and 6 is provided for values of p; and p: from 0 to 10 inclusive, PROBABILITY FUNCTION OF A TABLE II* PRODUCT 269 N ma oN 2 4 6 8 10 0 0 0 0 0 2.236068 4.123106 6.082762 8.062258 10.049876 0 0 0 0 0 0 2.160 -685121 +319942 + 183195, = 118224 529 205 101 +059 039 4 8 12, 16. 20. . 3 4.582576 6.403124 8.306624 10. 246951 2 8 498784 «274256, - 167493 111531 1, 259259 -557823 289114 -172653 .113742 020 056 056 042 031 16. 24, 32. 40 5.744563 | 7.280110 | 9. 10. 816654 4 -506408 -373206 -263874 |. 189641 -358127 + 224279 +147234 | .102126 — 0084 -0049 014 016 36. 48. 60 8.544004 | 10.049876 11.704700 6 346314 + 28373 224503 = 163258 «118224 .087272 — .0054 — .00083 -0038 64, 80 11.357817 | 12.845233 8 = 262088 226472 192663 072507 — .0034 —.0015 100. 14.177447 10 -210551 059553 — .0023 * The first value in a cell is Z, the second oz, the third &, the fourth &, the fifth 5. 270 LEO A. AROIAN 5. Some extensions. We may generalize our results to any case where x and y are distributed approximately in a normal distribution such as the dis- tribution of the product of two means, when the sizes of the samples N; and N, are large and consequently p; and p, will be large. Another example oceurs if x and y each follows a Bernoulloi probability function with parameters p; and ps respectively where the number of trials in each case is large. We must warn the reader that the condition p;—> ©, py -> © alone does not mean that the dis- tribution of z approaches normality. Both x and y must be distributed normally. ‘The actual problem which gave rise to this investigation was the question of determining the sum of a great many variates [8]. Let 7’ variates 1, v, +++, p be given whose sum A = )074 0; is desired. Clearly r A=TV,,V,= wt. Now let us estimate A by A = 7,V, where 7, is an estimate of T and J, is an estimate of 7. If og, is very small, p, = T'/og, will be large and » = V,/oy, = VNV,/oy will be very large. Assuming 7’, is distributed normally and obviously V, is distributed normally for N large, we see by the theorems of this paper that A will be distributed normally. Confidence limits for A may be calculated in the usual fashion as A +: yoz, where 7 is determined by L, olddt = a, with a generally chosen as .025 or less and 3 = VT, + Wiad + 7, Stratification is also possible. It is interesting to note that many functions which oceur in life insurance are products. Such applications will be treated fully elsewhere. Naturally the critical region whether both tails or one tail of the distribution should be used depends on the alternatives to the hypothesis being tested. Generalizations of the main theorem are possible for the probability function of 2 = [Ji1a; where 1, a, «++, 2, follow a multivariate normal probability function. These will be investigated in a later paper. It may be noted that J.B. S. Haldane has investigated the distribution of a product along different lines [9]. REFERENCES {1] Cxctt C. Crara, “On the frequency funetion of zy,” Annals of Math. Stat., Vol. 7 (1936), pp. 1-15. 12] J. H. Corriss, “A note on the theory of moment generating functions,” Annals of Math. Siat., Vol. 13 (1942), pp. 430-434. [3] A. L. O’Tooue, “On the degree of approximation of certain quadrature formulas,” Annals of Math. Stat., Vol. 4 (1938), pp. 143-153. PROBABILITY FUNCTION OF A PRODUCT 271 [4] Arnoup N. Lowan, Technical Director, Tables of Probability Functions, Vol. II. Na- tional Bureau of Standards, Washington, D. C. [5] Annoy N. Lowan, Technical Director, Tables of the Exponential Function e*. Na- tional Bureau of Standards, Washington, D. C. [6] Arnot N, Lowan, Technical Director, Tables of Reciprocals of Integers from 100,000 through 200,009. Columbia Univ. Press, New York. [7] Cee C. Crare, “A new exposition and chart for the Pearson system of frequency curves,” Annals of Math. Stat., Vol. 7 (1936), pp. 16-28. [8] Leo A. Arotan, “Some methods for the evaluation of a sum,” Amer, Stat, Ass. Jour., Vol. 39 (1944), pp. 511-515. 19] J. B.S. Haupane, “Moments of the distribution of powers and produets of normal variates,” Biometrika, Vol. 32 (1942), pp. 226-242.

You might also like