Seminar 2012 - Counterexamples in Probability Seminar | 19.11.2012 | Presenter : Joung In Kim Seite 2 Seminar – Counterexamples in Probability Ch8. Characteristic and Generating Functions Ch9. Infinitely divisible and stable distributions Seite 3 Seminar – Counterexamples in Probability Ch8. Characteristic and Generating Functions Ch9. Infinitely divisible and stable distributions Seite 4 Notation and Abbreviations • r.v. : random variable • ch. f. : characteristic function (ฯ(t)) • d.f. • i.i.d. : independent and identically distributed • d = : distribution function (F) : equality in distribution Seite 5 Definition (Characteristic function) Let X be a r. v. defined on Ω, โฑ, Ρ with distribution function F โถ โ → [0,1] The ๐๐ก๐๐ซ๐๐๐ญ๐๐ซ๐ข๐ฌ๐ญ๐ข๐ ๐๐ฎ๐ง๐๐ญ๐ข๐จ๐ง ฯ: โ → โ of X is defined by ฯ t = Ε eitX = ∞ itx e −∞ dF x ∞ itx e −∞ f x dx, if X is ๐๐๐๐๐๐๐๐๐๐ ๐๐๐๐๐๐๐๐๐๐ with density f = n eit x n pn , if X is ๐ ๐๐๐๐๐๐๐ with P X = xn = pn , pn > 0, n pn = 1 Seite 6 Properties of a characteristic function i) ฯ 0 = 1, ฯ −t = ฯ t , ฯ t ii) If Ε X ฯ n iii) If ฯ n < ∞, then ฯ n 0 exists and Ε X n = i−n ฯ n 0 t = n dn dn t ฯ t 0 exists and n is even, then Ε X If n is odd , then Ε X iv) If Ε X ≤ 1, t ∈ โ n n−1 ฯ t = k=0 <∞ <∞ < ∞ and hence Ε X n n k < ∞ for k < ๐ then (it)k E X k + ο(t n ) k! t→0 Seite 7 Properties of a characteristic function v) If X1 and X 2 are independent random variables with ch. f. s ฯ1 and ฯ2 . ⇒ the ch. f. of X1 + X 2 is given by ฯ t = ฯ1 t ฯ2 t , t ∈ โ vi) If we know the ch.f. ฯ of a r.v. X then the d.f. F of X is given by the inversion formula: 1 P a < ๐ < ๐ = lim c→∞ 2π c e−ita − e−itb ฯ t dt it −c If ฯ is absolutely integrable over โ, it follows that X is absolutely continuous with density: f x = ∞ −itx e ฯ(t)dt 2π −∞ 1 (inverse Fourier-transformation) Seite 8 Fourier expansion of a periodic function g(t+T)=g(t) ∞ => ๐ t = (an cos(nw0 t) + bn sin(nw0 t)) n=0 where, ๐๐ = 1 T T 2 T − 2 2π w0 = , T: Period, an , bn โถ Fourier coefficient T g t dt, ๐๐ง = 2 T T 2 T − 2 g t cos nw0 t dt , ๐๐ง = 2 T T 2 T − 2 g t sin nw0 t dt Seite 9 Example 1. Discrete and absolutely continuous distributions with the same characteristic functions on [-1, 1] continuous ฯ t = 1 1 ๐ ๐ฅ = 2π discrete ΡY=0 = 1− t , 0 , ∞ −it๐ฅ e −∞ 1 2 if t ≤ 1 otherwise. 1 − cos ๐ฅ ฯ1 t dt = , ๐ฅ∈โ 2 π๐ฅ , P Y = 2๐ − 1 π = 1 ฯ2 (t) = + 4π−2 2 ∞ k=1 2 2k−1 2 π 2 cos( 2k − 1 πt) (2k − 1)2 , k = 0, ±1, ±2,โโโ Seite 10 Example 1. Discrete and absolutely continuous distributions with the same chacteristic functions on [-1, 1] ∞ n=1 an h t = t = a0 + t ≤ 1, 1 = + 2 ∞ k=1 1 a0 = , 2 2(cos nπ − 1) an = n2 π2 −4cos( 2k − 1 πt) (2k − 1)2 π2 1 ฯ1 t = 1 − t = + 2 ⇒ ฯ1 t = ฯ2 t , cos nπt (Fourier series) ∞ k=1 4cos( 2k − 1 πt) (2k − 1)2 π2 t ∈ [−1,1] Seite 11 Example 2. The absolute value of a characteristic function is not necessarily a characteristic function. ฯ t = 1 8 1 + 7eit , t ∈ โ , ฯ(t) = 1 8 50 + 7e−it + 7eit 1/2 If ฯ(t) is a ch. function, then ψ(t) โถ= ฯ(t) must be of the form ψ t = p โ eit x 1 + (1 − p) โ eit x 2 Where 0 < ๐ < 1 ๐๐๐ x1 , x2 are different real numbers. Comparing ψ t 2 and ฯ(t) => p2 = (1 − p)2 = 7 64 2 , 2p(1 − p) = 50 64 ๐๐จ๐ง๐๐ฅ๐ฎ๐ฌ๐ข๐จ๐ง โถ ฯ is a ch. function, but not ฯ . => contradiction! Seite 12 Decomposable and Indecomposable We say that a ch.f. ฯ is decomposable if it can be represented as a product of two non-trivial ch.f.s. ฯ1 and ฯ2, i.e. ฯ(t) = ฯ1(t) ฯ2(t) and neither ฯ1 nor ฯ2 is the ch.f. of a probability measure which is concentrated at one point. Otherwise ฯ is called indecomposable. Seite 13 Example 3. The factorization of a characteristic function into indecomposable factors may not be unique. (i) discrete case X : discrete uniform distribution on the set {0, 1, 2, 3, 4, 5}. Characteristic function of X : 5 itX ฯ t =E e itk = P(X = k)e k=0 1 = 6 5 eitk k=0 We can factorize the ch. f. in the following way : ฯ1 t = 1 ψ1 t = 1 3 3 1 + e2it + e4it , ฯ2 t = 1 1 + eit + e2it , ψ2 t = 1 2 2 1 + eit 1 + e3it ๏ฐ ฯ t = ฯ1 t โ ฯ2 t = ψ1 t โ ψ2 t , t ∈ โ Seite 14 Example 3. The factorization of a characteristic function into indecomposable factors may not be unique. Need to check : − Are ฯ1 , ฯ2 , ψ1 and ψ2 ch. f. s? − Are ฯ1 , ฯ2 , ψ1 and ψ2 indecomposable? ⋅ ฯ2 and ψ2 correspond to two − point distribution. => ๐๐๐๐๐๐๐๐๐๐ ๐๐๐๐ ⋅ Suppose that ψ1 t = ψ11 t โ ψ12 t , ψ11 t , ψ12 t : non − trivial Let ψ11 t = peitx 1 + 1 − p eitx 2 , ψ12 t = qeity 1 + 1 − q eity 2 0 < p < 1, 0 < ๐ < 1 then pq = 1 − p 1 − q = p 1 − q + q 1 − p = 1 => contradiction 3 ⇒ ψ1 is indecomposable. ⋅ ฯ1 t = ψ1 2t ⇒ ฯ1 is also indecomposable. โ Seite 15 Example 3. The factorization of a characteristic function into indecomposable factors may not be unique. (ii) Continuous case โ Let X be a r.v. which is uniformly distributed on -1,1). โ Ch. f. of X : ฯ t = t −1 sin t , t ∈ โ Factorization 1. n ฯ t = t −1 sin t = k=1 ∞ => Is cos cos lim ฯ t = n→ ∞ t 2k t cos k 2 cos k=1 t 2n −1 t 2k an indecomposable ch.f.? t 1 it/2k −it /2k = (e + e ) 2 2k sin t 2n Seite 16 Example 3. The factorization of a characteristic function into indecomposable factors may not be unique. Factorization 2. −1 t t t −1 sin t = sin 3 3 ฯ t = t −1 sin t = 2 cos 2t + 1 /3 3 1 2t [2 cos + 1] 3 3 ∞ cos k=1 ⇒ We have two different factorizations. t 3 โ 2k Seite 17 Seminar – Counterexamples in Probability Ch8. Characteristic and Generating Functions Ch9. Infinitely divisible and stable distributions Seite 18 Definition (infinitely divisible distribution) โ X : a r.v. with d.f. F โ ฯ : ch.f. of X โ X is called infinitely divisible if for each n≥1 there exist i.i.d. r.v.s Xn1, ..., Xnn such that d X = Xn1 + โโโ + Xnn Equivalent : โ ฦ d.f. Fn with F=(Fn)*n โ ฦ ch.f. ฯn with ฯ =(ฯ n)n Seite 19 Definition (stable distribution) โ X : a r.v. with d.f. F โ ฯ : ch.f. of X โ X is called stable if for X1 and X2 independent copies of X and any positive numbers b1 and b2, there is a positive number b and a real number γ s.t. : d b1X1+b2X2 = bX + γ Equivalent : ฯ b1 t ฯ b2 t = ฯ bt eiγt Seite 20 Properties of infinitely divisible and stable distributions • The ch.f. of an infinitely divisible r.v. does not vanish. • If a r.v. X is stable, then it is infinitely divisible. Seite 21 Example 4. A non-vanishing characteristic function which is not infinitely divisible random variable X X -1 0 1 P(X=x) 1/8 3/4 1/8 1 −it 3 it0 1 it 1 ฯ t = e + e + e = (3 + cos t) 8 4 8 4 => ฯ t > 0, ∀๐ก ∈ โ => ฯ does not vanish. Seite 22 Example 4. A non-vanishing characteristic function which is not infinitely divisible Is X infinitely divisible? d Assume X = X1+X2, (X1, X2 are iid r.v.s) Since X has three possible values, each of X1 and X2 can take only two values, say a and b, a<b. Let P[Xi=a]=p, P[Xi=b]=1-p for some p, 0<p<1, i=1,2 X1+X2 2a a+b 2b P(X1+X2=x) p2 2p(1-p) (1-p)2 ๏ 2a= -1, a+b=0, 2b=1, p2=1/8, 2p(1-p)=3/4, (1-p)2=1/8 => contradiction! d ๏ X= X1+X2 is not possible. => X is not infinitely divisible. Seite 23 Example 5. Infinitely divisible distribution, but not stable i) X ~ Poi(λ) ๏[ X ๏ฝ n] ๏ฝ ๏ฌn n! e ๏ญ๏ฌ , n=0, 1, 2, โโโ, λ>0 Characteristic funtion of X : ฯ t = exp ๐ ๐ ๐๐ก − 1 , ๐ก ∈ โ Characteristic funtion of Xn ~Poi(λ/n) : ฯn t = exp ๐ ๐๐ก ๐ −1 ๐ => ฯ t = [ฯn (t)]n => X is infinitely divisible Seite 24 Example 5. Infinitely divisible distribution, but not stable Is X a stable distribution? If yes, for any b1 and b2 >0, there exist b>0 and γ∈โ s.t. ฯ b1 t ฯ b2 t = ฯ(bt)eiγt โ Ch. f. of X โถ ฯ t = exp ๐ ๐ ๐๐ก − 1 โ ฯ b1 t ฯ b2 t = exp λ eib 1 t − 1 exp λ eib 2 t − 1 = expโก [λ eib 1 t + eib 2 t − 2 ] โ ฯ bt eiγt = exp λ eibt − 1 exp ๐๐พ๐ก = exp λeibt + iγt − λ => ฯ b1 t ฯ b2 t ≠ ฯ bt eiγt X is not stable. , i. e. Seite 25 Example 5. Infinitely divisible distribution, but not stable ii) Let see the gamma distribution with parameter θ=1, k=1/2 g x = x k−1 −θ x e Γ k θk ฯ t = (1 − itθ) ฯn t = 1 − it 1 = −k π 1 − −x x 2e = (1 − 1 − 2n , x>0 1 − it) 2 โถ ch. f. of gamma distr. with θ = 1, k = ฯ t = [ฯn (t)]n , i. e. X is infinitely divisible 1 2n Seite 26 Example 5. Infinitely divisible distribution, but not stable Is X a stable distribution? If yes, for any b1 and b2 >0, there exist b>0 and γ ∈ โ s.t. ฯ b1 t ฯ b2 t = ฯ(bt)eiγt โ Ch. f. of X โถ ฯ t = (1 − โ ฯ b1 t ฯ b2 t = (1 − iγt โ ฯ bt e = (1 − 1 − it) 2 1 − ib1 t) 2 (1 − 1 − ib2 t) 2 1 − iγt ibt) 2 e => ๐ b1 t ฯ b2 t ≠ ฯ bt eiγt , i. e. X is not stable. = (1 − i b1 + b2 t − 1 2 −2 b1 b2 t ) Seite 27 REFERENCES [1] J. Stoyanov. Counterexamples in probability (2nd edition). Wiley 1997 [2] G. Samorodnitsky, M. S. Taqqu. Stable Non-Gaussian Random Processes. Chapman&Hall, 1994 [3] K. L. Chung. A course in probability theory. Academic Press, 1974 [4] E. Lukacs. Characteristic functions. Griffin, 1970 Seite 28 Thank you very much !!!