Continuous bootstrapping Naoto Niki1 and Yoko Ono2 1 2 Tokyo University of Science, niki@ms.kagu.tus.ac.jp Niigata University of International and Information Studies, onoyk@nuis.ac.jp Summary. Bayesian bootstrap method is proposed by Rubin where prior distribution is Diriclet distribution. However, Efron’s bootstrap method has Multinomial prior from the Bayesian’s point of view. According to the difference in these prior, we have proposed another type of prior distribution and continuous version of bootstrapping. Key words: Bayesian Bootstrapping, Prior distribution 1 Introduction 1.1 Bootstrapping as parameter sampling Let x = (x1 , . . . , xn ) be a random vector of observations independently drawn from an unknown population F , for estimating a population parameter θ = T (F ) by using θx = T (Fx ), where Fx is the empirical distribution based on x. In “nonparametric bootstrapping” due to Efron [Efr79], a large number of samples of size n from Fx are drawn for numerical evaluation of properties of the distribution of θx , or more essentially, for simulating the distribution of the “random distribution” Fx . Hereafter, for the sake of simplicity, it is assumed that i 6= j ⇒ xi 6= xj . Let b = (b1 , . . . , bn ) be a random vector drawn from Fx in place of F . Then, the empirical distribution Fb is a multinomial distribution Mul(n; p1 , . . . , pn−1 ) on x of which parameter vector p = (p1 , . . . , pn ), providing pn = 1 − p1 − · · · − pn−1 , is distributed as np ∼ Mul(n; 1/n, . . . , 1/n). The common marginal distribution of npi (i = 1, . . . , n) is a binomial distribution Bin(n, 1/n). 1.2 Continuous analogue to bootstrapping It is well known that, if n is large enough, the distribution of θb = T (Fb ) furnishes as good estimates for the low order moments of the sampling distribution of θx , as statisticians need in practical applications. But, for smaller n, discreteness involved in the values of p brings fatal influence on the accuracy of the estimates. See, e.g., Bickel and Freedman [?] for more details. 1076 Naoto Niki and Yoko Ono The purpose of this article is to propose a vector r = (r1 , . . . , rn ) of continuous random variables approximately distributed with Mul(n; 1/n, . . . , 1/n), in the sense that r1 ≥ 0, . . . , rn ≥ 0, r1 + · · · + rn = 1; 1 1 Pr k − ≤ ri < k + ≈ n Ck 2 2 k 1 n 1 ≈ 2 Pr ri < 1 1− n 1− n−k 1 n n , i, k ∈ {1, . . . , n} . 1.3 Bayesian bootstrapping One possible solution may be the use of a Dirichlet variate d = (d1 , . . . , dn−1 ) ∼ Dir(n; 1, . . . , 1) employed in the “Bayesian Bootstrapping” due to Rubin [?]. However, this distribution has clearly heavier tails than desired, besides the fact that time-consuming sorting operation is involved in bootstrapping. For example, for large n, the common marginal distribution Beta(1, n − 1) of di has the right tail deceasing geometrically with fixed ratio e, whereas the tail of Bin(n, 1/n) reduces in factorial descend, as demonstrated in Table 1, where Beta∗ (1, ∞) is the limiting distribution of n di as n tends to infinity. Table 1. Difference in Pr{X ∈ [k − 0.5, k + 0.5)}. Distribution of X k = 0 k = 1 k = 2 1 1 1 X ∼ Poison(1) e e 2!e √ e−1 e−1 e−1 √ √ X ∼ Beta∗ (1, ∞) √ e e e e2 e k=3 1 3! e e−1 √ e3 e k=4 1 4! e e−1 √ e4 e k=5 1 5! e e−1 √ e5 e 2 Continuous Bootstrapping Let u = (u1 , . . . , un ) be a vector of random variables i.i.d. with the uniform distribution on [0, 1]. For some constant α > 1, we consider a simple rational transformation gα (x) = x/(α − x), and write gα (u) = (gα (u1 ), . . . , gα (un )). Then, the common marginal distribution of xi = gα (ui ) (i = 1, . . . , n) has pdf, cdf, mean and variance as given below: 8 > > < 1 α 0≤x≤ 2 α − (1 + x) 1 fα (x) = 1 > > x> , :0 α−1 Continuous bootstrapping 1077 Table 2. Marginal probabilities for the intervals [k − 0.5, k + 0.5). Distribution k=1 k=2 Bin(n, 1/n) 10 20 50 100 ∞ 0.349 0.358 0.364 0.366 0.368 0.387 0.377 0.372 0.370 0.368 0.194 0.189 0.186 0.185 0.184 CB(n) 10 20 50 100 500 0.351 0.362 0.367 0.369 0.370 0.395 0.381 0.374 0.371 0.370 Beta∗ (1, n − 1) 10 20 50 100 ∞ 0.370 0.382 0.389 0.391 0.393 0.399 0.391 0.386 0.385 0.383 8 > > < Fα (x) = n k=0 0.011 0.013 0.015 0.015 0.015 0.001 0.002 0.003 0.003 0.003 0.000 0.000 0.000 0.001 0.001 0.199 0.195 0.190 0.187 0.186 0.049 0.057 0.067 0.072 0.074 0.006 0.005 0.002 0.001 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.157 0.148 0.144 0.142 0.141 0.054 0.053 0.052 0.052 0.052 0.016 0.018 0.019 0.019 0.019 0.004 0.006 0.007 0.007 0.007 0.001 0.002 0.003 0.004 0.004 0≤x≤ > > :1 0.057 0.060 0.061 0.061 0.061 1 α − 1 1 0≤x≤ α−1 αx 1+x µα = α log k=3 k=4 k=5 k≥6 α α−1 − 1, σα2 = α − α2 log α−1 α α−1 2 , respectively. Discussion in this article is focused mainly upon the distribution of r = gα (u) X n gα (ui ) . i=1 and its use in place of the bootstrapping and the Bayesian bootstrapping. 2.1 Numerical Comparisons The constant α employed here is determined through numerical experiments and for mnemonic’s sake as 4 α= ≈ 1.4715. e 4 Setting α = yields e µ= 4 log e 4 4−e − 1 ≈ 0.6747, σ 2 = 4 16 − 2 log 4−e e 4 4−e 2 ≈ 0.3161. For α = 4/e, the densities fα (x) and fα∗10 (x) for gα (u1 ) and gα (u1 ) + · · · + gα (u10 ), respectively, and the normal approximation to the latter one (thin line) are shown in Fig. 1. Fig. 2 shows densities fα∗2 (x) to fα∗6 (x). Table 2 illustrates numerical comparisons of the three bootstrapping concerned. 1078 Naoto Niki and Yoko Ono 1.4 0.2 1.2 1 0.15 0.8 0.1 0.6 0.4 0.05 0.2 0.5 1 1.5 5 2 15 10 fα (x) 20 fα∗10 (x) Fig. 1. Densities fα (x) and fα∗10 (x) 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 1 2 3 1 4 2 fα∗2 (x) 3 4 5 6 8 10 12 fα∗3 (x) 0.35 0.25 0.3 0.25 0.2 0.2 0.15 0.15 0.1 0.1 0.05 0.05 2 4 6 fα∗4 (x) 8 2 4 6 fα∗6 (x) Fig. 2. Densities fα∗2 (x) to fα∗6 (x) References [BF81] Bickel, P.J. and Freedman, D.A.: Some asymptotic theory for the bootstrap. Ann. Statist., 7, 1–26 (1979) [Efr79] Efron, B.: Bootstrap methods: another look at the jackknife. Ann. Statist., 9, 1196–1217 (1981) [Rub81] Rubin, D. B.: The Bayesian bootstrap. Ann. Statist., 9, 130–134 (1981)