Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
An Intuitive Interpretation of Ratios of Moment Generating Functions Via One-Step Conditioning Donghui Chen AND Kenneth S. Berenhaut, Department of Mathematics, Wake Forest University Motivation Motivation This paper studies a measure of the quality of local approximations to distributions for sums of independent and identically distributed (i.i.d.) random variables. In particular, suppose X is a random variable satisfying E(X) = µ and V ar(X) = σ 2 < ∞ and suppose that X1, X2, . . . is a sequence of i.i.d. random variables each with the same distribution as X. In addition, suppose that X has probability function p, where p is either a density or a probability mass function, and denote the support set of X by SX = {x : p(x) > 0}. Let p(i) be the i-fold convolution of p ( i.e. the probability function for the partial sum Si = X1 + X2 + · · · + Xi). ⋆ One step conditioning and independence lead to p(n)(x) = EX (p(n−1)(x − X)). 1 φnµ,nσ2 (x) = √ √ e 2πσ n , −∞ < x < ∞ (1) Our present consideration is motivated by the following two well-known facts: ⋆ There are many well-known local limit theorems for random variables where the associated attracting distribution is the normal. Two such results are the following two theorems. Theorem 1. Suppose X takes values of the form b + N h, N = 0, ±1, ±2, ..., with probability 1, where b is a constant and h > 0 is the maximal span of the distribution. Then √ (n) sup σ n p (nb + N h) − hφnµ,nσ2 (nb + N h) → 0. N Theorem 2. Suppose p is a density, and that there exists an N such that pN (x) is bounded. Then √ (n) sup σ n p (x) − φnµ,nσ2 (x) → 0. X Definition 1. For any two functions f and g, define the ”convolution ratio”, Hf,g of f and g with respect to X via for all x ∈ R such that g(x) 6= 0. LX,N (t) = mX mN = e (t2 −p2 ) 2p(p−1) = RX,N − pe (t2 −p2 ) 2p(p−1) t−p p(1 − p) + pe (t−p)(t+p−2) 2p(p−1) For simplicity, we will write L(t, p) to represent LX,N (t). We have the following theorem: Theorem 5. Suppose X is a Bernoulli random variable with parameter p. Then, for all 0 < t < 1, 1 1 − 32 2 LX,N (t) = L(t, p) ≥ L(t, 1/2) ≥ L(0, 1/2) = (e + e ) ≈ 0.9359. 2 5 0.5 4 0.4 3 0.3 2 0.2 1 0.1 0 0.75 0.5 p Figure 2: 0.25 0.5 0.25 0.0 t 1.0 0.75 For n > 1, set where p̃ = (p̃i)i≥1 is a sequence of real valued functions such that p̃i(x) approximates p(i)(x) for each x. For any random variable Y , let mY be the moment generating function of Y defined via mY (t) = E(etY ) for −∞ < t < ∞. As well, for random variables Y1 and Y2, define RY1,Y2 via P∞ mY1 (t) ti k (i) RY1,Y2 (t) = mY (t) . Also, let KY = log(mY (t)) = i=1 Y i! be the cumulant 2 generating function for Y . Note that for a normal random variable, N , with density as in (1), 0.75 0.5 p 0.25 1.0 0.75 The function LX,N (left) and | log(L(X, N ))| (right) for X Bernoulli(p) with 0 < p < 1. Theorem 3. Suppose p̃i = φnµ,nσ2 , if X is continuous (or p̃i = hφnµ,nσ2 if X is defined on a lattice with maximal spread h). Then 2 2 tµ+ t 2σ mN (t) = e . Table 2: Optimal bounds for LX,N (t) for t ∈ UX for some common distributions Lower bound 0.9359257154 . . . 0.9359257154 . . . 0.7450966791 . . . 0.7450966791 . . . 0.8243606355 . . . 0.3032653298 . . . e−λ(0.1321205588... ) Upper bound 1 Unbounded 1 1 Unbounded Unbounded Unbounded • Discrete Uniformly Distributed Random Variables Suppose X takes value on B = {0, 1, . . . T −1} with constant probability P (X = i) = T1 for i ∈ B (where we assume T > 3). The mean and the variance of X are T 2−1 2 and σ = µ = T −1 2 12 , respectively and TX = [0, T − 1]. In light of Theorem 3, we have mX mN t−µ σ2 t−µ σ2 −1 −6T −2t+T 2 T −1 − 1 e e = −1 −6 −2t+T 2 e T −1 − 1 2 2 3 (T −1) −4t 2 T 2 −1 T Theorem 6. Suppose X is uniformly distributed on B. Then LX,N (t) is maximized at t = T −1 2 and minimized at t = 0. In addition, for all t ∈ [0, T − 1], 1 > L(t, T ) ≥ L(0, T ) = − T6T +1 e 6 − T +1 e −1 −1 ! e lim C (tn, n) = n→∞ X,p̃ mN (3) provided that the numerator on the right hand side exists. We will refer to the right hand side of (3) as LX,N (t). t−µ Proof. Suppose that t is fixed and mX σ2 < ∞. Then, E(p̃n−1(tn − X)) =E p̃n(tn) CX,p̃(tn, n) = E √ p̃n−1(tn − X) . p̃n(tn) 3 T −1 2 T +1 T . Note that limT →∞ L(0, T ) = (−1/6 e−6 + 1/6)e3/2 ≈ 0.7451. • Continuous Uniformly Distributed Random Variables Suppose that X possesses a uniform distribution on the interval [a, b]. The mean (b−a)2 b−a 2 and the variance of X are µ = 2 and σ = 12 , respectively and TX = [a, b]. tb −eta The moment generating function of X is given by mX (t) = et(b−a) . Note that Theorem 4 reduces our consideration through scaling to (a, b) = (0, 1). In light 2 2 (t−X)2 − t −µ2 − 2(n−1)σ2 2σ n (t−µ)X √ e σ2 e n−1 (4) e ! 2 √ (t−µ)X t2 −µ2 − (t−X) − √ n e σ2 e 2σ2 e 2(n−1)σ2 . n−1 for n > 1. Now, set Yn = Since for any n > 1, we have 0 < Yn < 2e (t−µ)X σ2 2 2 − t −µ2 e P Yn → e and 2σ (t−µ)X σ2 2 2 − t −µ2 e 2σ , Lebesgue’s Dominated Convergence Theorem, implies that lim E e (t−µ)X σ2 2 2 − t −µ2 2σ e where we have used the fact that mN X Bernoulli(1/2) Bernoulli(p) Uniform on {0, 1, . . . , T − 1} Uniform on [a, b] Exponential(λ) Geometric(p) Poisson(λ) t−µ σ2 t−µ , σ2 mX n→∞ For simplicity, here we will write L(t, T ) to represent LX,N (t). 0.5 0.25 0.0 t Corollary 1. Suppose that µ = 0 and σ 2 = 1 and p̃i is as in the statement of Theorem 3. Then Simplifying (4) gives E(p̃n−1(x − X)) CX,p̃(x, n) = Hp̃ ,p̃ (x) = n−1 n p̃n(x) X LX,N (t) = 0.0 Among our observations is the following theorem for when p̃ is taken as suggested by Theorems 1 and 2. CX,p̃(tn, n) = Example of Uniformly Distributed Random Variables Suppose X is a Bernoulli random variable with parameter p, i.e. P (X = 1) = p and P (X = 0) = 1 − p with 0 < p < 1. Then µ = p, σ 2 = p(1 − p) and TX = [0, 1]. 2 2 Thus, we are interested in bounding the ratio R(v) for v ∈ [(0−µ)/σ , (1−µ)/σ ]= i h −1 1 1−p , p . In light of Theorem 3, we have t−µ σ2 t−µ σ2 E(f (x − X)) g(x) X (x) = Hf,g x Example for Bernoulli Random Variables Corollary of Theorem 3 and Conclusion Essentially, of interest here is the nature of the relationship in (2) when p(n) and p(n−1) are replaced by approximating functions. Let φnµ,nσ2 be the density for a normal random variable, N , with mean nµ and variance nσ 2, i.e. 2 x−nµ 1 √ −2 nσ (2) Main Theorem and Its Proof t−µ σ2 = =e mX mN t−µ σ2 t−µ , σ2 (t−µ)2 t−µ µ+ σ2 2σ 2 =e t2 −µ2 2σ 2 . Examples of Distributions with Unbounded Support • Exponential random variables Suppose X is an exponential random variable with parameter λ > 0, i.e. P (X = x) = λe−λx for xT> 0. Then µ = 1/λ, σ 2 = 1/λ2, mX (t) = λ/(λ − t) for t < λ and TX = [0, ∞] {t : (t − λ−1)/λ−2} = {t : 0 ≤ t < 2/λ}. In light of Theorem 3, we have LX,N (t) = mX mN = − t−1/λ 1/λ2 t−1/λ 1/λ2 − 12 (tλ−1)(tλ+1) e tλ − 2 , Theorem 8. Suppose X is an exponential random variable with parameter λ. Then, for all 0 < t < 2/λ, 1 1 LX,N (t) = L(t, λ) ≥ L(0, λ) = e 2 ≈ 0.8244. 2 • Poisson Random Variables Suppose X is a Poisson random variable with parameter λ > 0, i.e. P (X = t i) = e−λλi/i! for i = 0, 1, 2, . . . . Then µ = λ, σ 2 = λ, mX (t) = eλ(e −1) and TX = [0, ∞). In light of Theorem 3, we have mX t−λ λ t−λ mN λ −t+λ − λ −t2 λ−1 1/2 −λ2 +2λ2 e = e , Theorem 9. Suppose X is an Poisson random variable with parameter λ. Then, for all 0 < t < ∞, LX,N (t) = L(t, λ) ≥ L(0, λ) = e t2 2 where Z possesses a standard normal distribution, and hence mZ (t) = e . The next theorem regarding changes of scale and origin follows directly from (3). Theorem 4. Suppose a, b ∈ R and a 6= 0, then t−b LaX+b,N (t) = LX,N . a Remark. Note that ratios as on the right side of (5) occur when considering Cornish-Fisher expansions. Remark. The value limn→∞ CX,p̃(tn, n) in (5) and (3) addresses the behavior of the approximation to the probability function for X̄n = Sn/n, in the limit, near the value t. Employing (3), we have that RX,N (t) gives information on the local behavior of the approximation near the value µ + tσ 2. In light of Theorem 3, and since the probability function of X̄n is zero outside the range of X, we will be primarily concerned with values of LX,N (t) for t ∈ TX , where \ t−µ TX = [m, M ] <∞ , t : mX σ2 where m = inf(SX ) and M = sup(SX ). As per the title for this paper, via Theorem 3, we can interpret the ratio of moment mX (t) as giving information on the quality of the generating functions RX,N (t) = m N (t) normal approximation for X n locally near the value µ + σ 2t. We will be interested def in values of RX,N (t) for t ∈ Q = ((m − µ)/σ 2 , (M − µ)/σ 2). Note that Q is an interval containing zero. References 1. Chow, Y. S. AND Teicher, H. Probability Theory, Second Edition, SpringerVerlag, New York, 1988. 2. Billingsley, P. Probability and Measure, Third edition, John Wiley Inc., New York, 1995. 3. Petrov, V. V. Sums of Independent Random Variables, Springer, Berlin, 1975. 5. Petrov, V. V. A Local Theorem for Densities of Sums of Independent Random Variables, Theory of Probability and Its Application, Vol 1 Issue 3, pp.316-322, 1956. 6. Sirazhdinov, S. KH. AND Mamatov, M. Short Communication on Convergence in the Mean for Densities, Theory of Probability and Its Application, Vol 7, Issue 4, pp.424-428, 1962. 7. Gnedenko, B. V. A Local Limit Theorem for Densities, Dokl. Akad. Nauk SSSR.,95 pp.5-7, 1954. 8. Gnedenko, B. V. AND Kolmogorov A. N. Limit Distributions for Sums of Independent Variables, Addison-Wesley, Boston, 1954. 9. Prokhorov, Y. V. A Local Theorem for Densities, Dokl. Akad. Nauk SSSR., 83 pp.797-800,1952. 10. Feller, W. Probability Theory and Its Applications, Vol. 2, Wiley, New York, 1966. 11. Wallace, D. L. Asymptotic Approximatio to Distributions, The Annals of Mathematics Statistics, Vol.29, No.3, pp.635-654, 1958. for t > 0. For simplicity, we will write L(t, λ) to represent LX,N (t). −λ e−2 2e (5) 4. Petrov, V. V. On Local Limit Theorems for Sums of Independent Random Variables, Theory of Probability and Its Application, Vol. 9, Issue 2, pp.312-320, 1962. for t < 2/λ. For simplicity, we will write L(t, λ) to represent LX,N (t). LX,N (t) = mX (t) lim C (tn, n) = n→∞ X,p̃ mZ (t) ≈ e−.1321λ. 12. Cornish, E. A. AND Fisher, R. A. Moments and Cumulants in the Specification of Distributions, Review of the International Statistical Institute, Vol.5, No.4, pp.307-320, 1938. 13. Mitrinovic, D. S. Analytic Inequalities, Springer-Verlag, New York, 1970.