advertisement

Continuous Distributions by R.J. Reed These notes are based on handouts for lecture modules given jointly by myself and the late Jeff Harrison. The following continuous distributions are covered: gamma, χ2 ,normal, lognormal, beta, arcsine, t, Cauchy, F, power, Laplace, Rayleigh, Weibull, Pareto, bivariate normal, multivariate normal, bivariate t, and multivariate t distributions. Important. There are many theoretical results in the exercises—full solutions are provided at the end. The colour red indicates a hyperlink. Comments are welcome—even comments such as “not useful because it omits xxx”. Please send comments and details of errors to my Wordpress account: bcgts.wordpress.com. Contents 1 Univariate Continuous Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1 Revision of some basic results: 3. 2 Order statistics: 3. 3 Exercises: 7. 4 The uniform distribution: 8. 5 Exercises: 14. 6 The exponential distribution: 15. 7 Exercises: 18. 8 The Gamma and χ2 distributions: 19. 9 Exercises: 22. 10 The normal distribution: 23. 11 Exercises: 26. 12 The lognormal distribution: 27. 13 Exercises: 30. 14 The beta and arcsine distributions: 31. 15 Exercises: 33. 16 The t, Cauchy and F distributions: 34. 17 Exercises: 37. 18 Non-central distributions: 39. 19 Exercises: 41. 20 The power and Pareto distributions: 42. 21 Exercises: 45. 22 Size, shape and related characterization theorems: 47. 23 Exercises: 50. 24 Laplace, Rayleigh and Weibull distributions: 51. 25 Exercises: 52. 2 Multivariate Continuous Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 1 General results: 55. 2 Exercises: 59. 3 The bivariate normal: 60. 4 Exercises: 64. 5 The multivariate normal: 65. 6 Exercises: 72. 7 The bivariate t distribution: 74. 8 The multivariate t distribution: 75. 9 Exercises: 78. Appendix: Answers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Chapter 1. Exercises 3 (on page 7): 79. Exercises 5 (on page 14): 82. Exercises 7 (on page 18): 85. Exercises 9 (on page 22): 87. Exercises 11 (on page 26): 89. Exercises 13 (on page 29): 92. Exercises 15 (on page 33): 93. Exercises 17 (on page 37): 94. Exercises 19 (on page 41): 97. Exercises 21 (on page 45): 97. Exercises 23 (on page 50): 102. Exercises 25 (on page 52): 103. Chapter 2. Exercises 2 (on page 58): 105. Exercises 4 (on page 64): 106. Exercises 6 (on page 72): 109. Exercises 9 (on page 77): 110. Appendix: References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 Bayesian Time Series Analysis by R.J. Reed Aug 1, 2018(17:54) Page 1 Page 2 Aug 1, 2018(17:54) Bayesian Time Series Analysis CHAPTER 1 Univariate Continuous Distributions 1 Revision of some basic results 1.1 Conditional variance and expectation. For any random vector (X, Y ) such that E[Y ] is finite, the conditional variance of Y given X is defined to be 2 var(Y | X) = E[Y 2 | X] − E[Y | X] = E (Y − E[Y | X])2 | X (1.1a) This is a function of X. It follows that E var(Y | X) + var E[Y | X] = E[Y 2 ] − (E[Y ])2 = var(Y ) (1.1b) Equation(1.1b) is often called the Law of Total Variance and is probably best remembered in the following form: var(Y ) = E[conditional variance] + var(conditional mean) which is similar to the decomposition in the analysis of variance. Definition(1.1a). For any random vector (X, Y, Z) such that E[XY ], E[X] and E[Y ] are all finite, the conditional covariance between X and Y given Z is defined to be cov(X, Y | Z) = E[XY | Z] − E[X | Z] E[Y | Z] An alternative definition is cov(X, Y | Z) = E X − E[X | Z] Y − E[Y | Z] Z (1.1c) Note that cov(X, Y | Z) is a function of Z. Using the results cov(X, Y ) = E[XY ] − E[X]E[Y ] and cov E[X|Z], E[Y |Z] = E E[X|Z] E[Y |Z] − E[X]E[Y ] gives the Law of Total Covariance cov(X, Y ) = E cov(X, Y | Z) + cov E[X|Z], E[Y |Z] (1.1d) This can be remembered as cov(X, Y ) = E[conditional covariance] + cov(conditional means) Note that setting X = Y in the Law of Total Covariance gives the Law of Total Variance. 1.2 Conditional independence. Recall that X and Y are conditionally independent given Z iff Pr[X ≤ x, Y ≤ y | Z] = Pr[X ≤ x | Z] Pr[Y ≤ y | Z] a.e. for all x ∈ R and all y ∈ R. Example(1.2a). Conditional independence does not imply independence. Here is a simple demonstration: suppose box 1 contains two fair coins and box 2 contains two coins which have heads on both sides. A box is chosen at random—denote the result by Z. A coin is selected from the chosen box and tossed— denote the result by X; then the other coin from the chosen box is tossed independently of the first coin—denote the result by Y . Clearly X and Y are conditionally independent given Z. However 5 3 but Pr[X = H] = Pr[Y = H] = Pr[X = H, Y = H] = 8 4 2 Order statistics 2.1 Basics. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with a continuous distribution which has distribution function F and density f . Then X1:n , X2:n , . . . , Xn:n denote the order statistics of X1 , X2 , . . . , Xn . This means X1:n = min{X1 , . . . , Xn } Xn:n = max{X1 , . . . , Xn } and the random variables X1:n , X2:n , . . . , Xn:n consist of X1 , X2 , . . . , Xn arranged in increasing order; hence X1:n ≤ X2:n ≤ · · · ≤ Xn:n Bayesian Time Series Analysis by R.J. Reed Aug 1, 2018(17:54) §2 Page 3 Page 4 §2 Aug 1, 2018(17:54) Bayesian Time Series Analysis 2.2 Finding the density of (X1:n , . . . , Xn:n ). Let g(y1 , . . . , yn ) denote the density of (X1:n , . . . , Xn:n ). Note that (X1:n , . . . , Xn:n ) can be regarded as a transformation T of the vector (X1 , . . . , Xn ). • Suppose n = 2. Let A1 = {(x1 , x2 ) ∈ R : x1 < x2 } and let T1 denote the restriction of T to A1 . Similarly let A2 = {(x1 , x2 ) ∈ R : x1 > x2 } and let T2 denote the restriction of T to A2 . Clearly T1 : A1 → A1 is 1 − 1 and T2 : A2 → A1 is 1 − 1. Hence for all (y1 , y2 ) ∈ A1 (i.e. for all y1 < y2 ), the density g(y1 , y2 ) of (X1:2 , X2:2 ) is fX1 ,X2 (T1−1 (y1 , y2 )) fX1 ,X2 (T2−1 (y1 , y2 )) g(y1 , y2 ) = + ∂(y1 ,y2 ) ∂(y1 ,y2 ) ∂(x1 ,x2 ) ∂(x1 ,x2 ) fX1 ,X2 (y1 , y2 ) fX1 ,X2 (y2 , y1 ) + |1| |−1| = 2f (y1 )fx (y2 ) • Suppose n = 3. For this case, we need A1 , A2 , A3 , A4 , A5 and A6 where A1 = {(x1 , x2 , x3 ) ∈ R : x1 < x2 < x3 } A2 = {(x1 , x2 , x3 ) ∈ R : x1 < x3 < x2 } etc. There are 3! = 6 orderings of (x1 , x2 , x3 ). So this leads to g(y1 , y2 , y3 ) = 3!f (y1 )f (y2 )f (y3 ) • For the general case of n ≥ 2, we have g(y1 , . . . , yn ) = n!f (y1 ) · · · f (yn ) for y1 < · · · < yn . = (2.2a) 2.3 Finding the distribution of Xr:n by using distribution functions. Dealing with the maximum is easy: n Y Fn:n (x) = P[Xn:n ] ≤ x] = P[X1 ≤ x, . . . , Xn ≤ x] = P[Xi ≤ x] = {F (x)}n i=1 n−1 fn:n (x) = nf (x) {F (x)} and provided the random variables are positive, using the result of exercise 5 on page 7 gives Z ∞ E[Xn:n ] = 1 − {F (x)}n dx 0 Now for the minimum: X1:n : P[X1:n > x] = P[X1 > x, . . . , Xn > x] = n Y i=1 P[Xi > x] = {1 − F (x)}n F1:n (x) = 1 − P[X1:n > x] = 1 − {1 − F (x)}n f1:n (x) = nf (x) {1 − F (x)}n−1 and provided the random variables are positive, using the result of exercise 5 on page 7 gives Z ∞ E[X1:n ] = {1 − F (x)}n dx 0 Now for the general case, Xr:n where 2 ≤ r ≤ n − 1. The event {Xr:n ≤ x} occurs iff at least r random variables from X1 , . . . , Xn are less than or equal to x. Hence n X n P[Xr:n ≤ x] = {F (x)}j {1 − F (x)}n−j (2.3a) j j=r = n−1 X j=r n {F (x)}j {1 − F (x)}n−j + {F (x)}n j Differentiating gives fr:n (x) = n−1 X n j=r j jf (x) {F (x)}j−1 {1 − F (x)}n−j − n−1 X n j=r j (n − j)f (x) {F (x)}j {1 − F (x)}n−j−1 + nf (x) {F (x)}n−1 1 Univariate Continuous Distributions = n X j=r Aug 1, 2018(17:54) §2 Page 5 n! f (x) {F (x)}j−1 {1 − F (x)}n−j − (j − 1)!(n − j)! n−1 X j=r n! f (x) {F (x)}j {1 − F (x)}n−j−1 j!(n − j − 1)! n! f (x) {F (x)}r−1 {1 − F (x)}n−r (r − 1)!(n − r)! Note that equation(2.3b) is true for all r = 1, 2, . . . , n. = (2.3b) 2.4 Finding the distribution of Xr:n by using the density of (X1:n , . . . , Xn:n ). Recall that the density of (X1:n , . . . , Xn:n ) is g(y1 , . . . , yn ) = n!f (y1 ) · · · f (yn ) for y1 < · · · < yn . Integrating out yn gives Z ∞ g(y1 , . . . , yn−1 ) = n!f (y1 ) · · · f (yn−1 ) f (yn )dyn = n!f (y1 ) · · · f (yn−1 ) 1 − F (yn−1 ) yn−1 and hence g(y1 , . . . , yn−2 ) = n!f (y1 ) · · · f (yn−2 ) Z ∞ yn−2 f (yn−1 ) 1 − F (yn−1 ) dyn−1 2 1 − F (yn−2 ) = n!f (y1 ) · · · f (yn−2 ) 2! 3 1 − F (yn−3 ) g(y1 , . . . , yn−3 ) = n!f (y1 ) · · · f (yn−3 ) 3! and by induction for r = 1, 2, . . . , n − 1 [1 − F (yr )]n−r g(y1 , . . . , yr ) = n!f (y1 ) · · · f (yr ) for y1 < y2 < · · · < yr . (n − r)! Assuming r ≥ 3 and integrating over y1 gives Z y2 [1 − F (yr )]n−r [1 − F (yr )]n−r dy1 = n!F (y2 )f (y2 ) · · · f (yr ) g(y2 , . . . , yr ) = n! f (y1 ) · · · f (yr ) (n − r)! (n − r)! y1 =−∞ Integrating over y2 gives [F (y3 )]2 [1 − F (yr )]n−r g(y3 , . . . , yr ) = n! f (y3 ) · · · f (yr ) for y3 < · · · < yr . 2! (n − r)! And so on, leading to equation(2.3b). 2.5 Joint distribution of Xj:n and Xr:n by using the density of (X1:n , . . . , Xn:n ). Suppose X1:n , . . . , Xn:n denote the order statistics from the n random variables X1 , . . . , Xn which have density f (x) and distribution function F (x). Suppose 1 ≤ j < r ≤ n; then the joint density of (Xj:n , Xr:n ) is j−1 r−1−j n−r f(j:n,r:n) (u, v) = cf (u)f (v) F (u) F (v) − F (u) 1 − F (v) (2.5a) where n! c= (j − 1)!(r − 1 − j)!(n − r)! The method used to derive this result is the same as that used to derive the distribution of Xr:n in the previous paragraph. Example(2.5a). Suppose X1 , . . . , Xn are i.i.d. random variables with density f (x) and distribution function F (x). Find expressions for the density and distribution function of Rn = Xn:n − X1:n , the range of X1 , . . . , Xn . Solution. The density of (X1:n , Xn:n ) is n−2 f(1:n,n:n) (u, v) = n(n − 1)f (u)f (v) F (v) − F (u) for u < v. Now use the transformation R = Xn:n − X1:n and T = X1:n . The absolute value of the Jacobian is one. Hence n−2 f(R,T ) (r, t) = n(n − 1)f (t)f (r + t) F (r + t) − F (t) for r > 0 and t ∈ R. Integrating out T gives Z ∞ n−2 fR (r) = n(n − 1) f (t)f (r + t) F (r + t) − F (t) dt t=−∞ Page 6 §2 Aug 1, 2018(17:54) Bayesian Time Series Analysis The distribution function is, for Z 0, Z vv > ∞ n−2 dt dr f (t)f (r + t) F (r + t) − F (t) FR (v) = n(n − 1) r=0 t=−∞ Z ∞ Z v n−2 = n(n − 1) f (t) f (r + t) F (r + t) − F (t) dr dt t=−∞ r=0 Z Z ∞ ∞ h n−1 n−1 v dt f (t) F (v + t) − F (t) dt = n f (t) F (r + t) − F (t) =n r=0 t=−∞ t=−∞ 2.6 Joint distribution of Xj:n and Xr:n by using distribution functions. Suppose u < v and then define the counts N1 , N2 and N3 as follows: n n n X X X N1 = I(Xi ≤ u) N2 = I(u < Xi ≤ v) and N3 = n − N1 − N2 = I(Xi > v) i=1 i=1 i=1 Now P X1 ≤ u = F (u); P u < X1 ≤ v = F (v) − F (u) and P X > v = 1 − F (v). It follows that the vector (N1 , N2 , N3 ) has the multinomial distribution with probabilities F (u), F (v) − F (u), 1 − F (v) . The joint distribution function of Xj:n .Xr:n is: n X ` X P Xj:n ≤ u and Xr:n < v = P N1 ≥ j and (N1 + N2 ) ≥ r = P N1 = k and N1 + N2 = ` `=r k=j = n X ` X `=r k=j k `−k n−` n! F (u) F (v) − F (u) 1 − F (v) k!(` − k)!(n − `)! The joint density of Xj:n .Xr:n is: ∂2 P Xj:n ≤ u and Xr:n < v ∂u∂v Using the abbreviations a = F (u), b = F (v) − F (u) and c = 1 − F (v) gives n ` X X n! ∂ P Xj:n ≤ u and Xr:n < v = f (u) ak−1 b`−k cn−` ∂u (k − 1)!(` − k)!(n − `)! f(j:n,r:n) (u, v) = `=r k=j − = f (u) n X `=r `−1 X k=j n! ak b`−k−1 cn−` k!(` − k − 1)!(n − `)! n! aj−1 b`−j cn−` (j − 1)!(` − j)!(n − `)! and hence ∂2 n! P Xj:n ≤ u and Xr:n < v = f (u)f (v) aj−1 br−j−1 cn−r ∂u∂v (j − 1)!(r − j − 1)!(n − r)! as required—see equation(2.5a) on page 5. 2.7 Asymptotic distributions. Here is the result about the asymptotic distribution of the median. For other results, see for example chapter 8 in [A RNOLD et al.(2008)]. Proposition(2.7a). Suppose the random variable X has an absolutely continuous distribution with density f which is positive and continuous at the median, µ̃. Suppose in = n/2 + 1. Then D √ 2 nf (µ̃) Xin :n − µ̃ =⇒ N (0, 1) as n → ∞. This means that Xin :n is asymptotically normal with mean µ̃ and variance 4n(f1(µ̃) )2 . Proof. See page 223 in [A RNOLD et al.(2008)]. Example(2.7b). Suppose U1 , . . . , U2n−1 are i.i.d. random variables with the U (0, 1) distribution. Then the median is √ D √ √ Un:(2n−1) and by the proposition 8n − 4 Un:(2n−1) − 1/2 =⇒ N (0, 1) as n → ∞. Of course an = 8n/ 8n − 4 → 1 √ D as n → ∞. Hence by Lemma 23 on page 263 of [F RISTEDT & G RAY(1997)] we have 8n Un:(2n−1) − 1/2 =⇒ N (0, 1) as n → ∞ and 1 t lim P Un:(2n−1) − < √ = Φ(t) for t ∈ R. n→∞ 2 8n 1 Univariate Continuous Distributions §3 Page 7 Aug 1, 2018(17:54) 3 Exercises (exs-basic.tex) Revision exercises. 1. The following assumptions are made about the interest rates for the next three years. Suppose the interest rate for year 1 is 4% p.a. effective. Let V1 and V2 denote the interest rates in years 2 and 3 respectively. Suppose V1 = 0.04 + U1 and V2 = 0.04 + 2U2 where U1 and U2 are independent random variables with a uniform distribution on [−0.01, 0.01]. Hence V1 has a uniform distribution on [0.03, 0.05] and V2 has a uniform distribution on [0.02, 0.06]. (a) Find the expectation of the accumulated amount at the end of 3 years of £1,000 invested now. (b) Find the expectation of the present value of £1,000 in three years’ time. 2. Uniform to triangular. Suppose X and Y are i.i.d random variables with the uniform distribution U (−a, a), where a > 0. Find the density of W = X + Y and sketch the shape of the density. 3. Suppose the random variable X has the density fX (x) = 1 2 for −1 < x < 1. Find the density of Y = X 4 . 4. Suppose X is a random variable with X > 0 a.e. and such that both E[X] and E[ 1/X ] both exist. Prove that E[X] + E[ 1/X ] ≥ 2. 5. Suppose X is a random variable with X ≥ 0 and density function f . Let F denote the distribution function of X. Show that Z Z ∞ (a) E[X] = 0 ∞ [1 − F (x)] dx E[X r ] = (b) 0 rxr−1 [1 − F (x)] dx for r = 1, 2, . . . . 6. Suppose X1 , X2 , . . . , Xn are independent and identically distributed positive random variables and Sn = X1 + · · · + Xn . (a) Show that E 1 1 ≥ Sn nµ (b) Show that Z ∞ n 1 E = E[e−tX ] dt Sn 0 7. Suppose X1 , X2 , . . . , Xn are independent and identically distributed positive random variables. (a) Suppose E[1/Xi ] is finite for all i. Show that E[1/Sj ] is finite for all j = 2, 3, . . . , n where Sj = X1 + · · · + Xj . (b) Suppose E[Xi ] and E[1/Xi ] both exist and are finite for all i. Show that Sj j E for j = 1, 2,. . . , n. = Sn n 8. Suppose X and Y are positive random variables with E[Y ] > 0. Suppose further that X/Y is independent of X and X/Y is independent of Y . 2 (a) Suppose E[X 2 ], E[Y 2 ] and E[ X /Y 2 ] are all finite. Show that E[X] = E X/Y ] E[Y ]. Hence deduce that there exists b ∈ R with X/Y = b almost everywhere. (b) Use characteristic functions to prove there exists b ∈ R with X/Y = b almost everywhere. Conditional expectation. 9. The best predictor of the random variable Y . Given the random vector (X, Y ) with E[X 2 ] < ∞ and E[Y 2 ] < ∞, find that random variable Yb = g(X) which is a function of X and provides the best predictor of Y . Precisely, show that Yb = E[Y |X], which is a function of X, minimizes 2 E Y − Yb 10. Suppose the random vector (X, Y ) satisfies 0 < E[X 2 ] < ∞ and 0 < E[Y 2 ] < ∞. Suppose further that E[Y |X = x] = a + bx a.e.. (a) Show that µY = a + bµX and E[XY ] = aµX + bE[X 2 ]. Hence show that cov[X, Y ] = b var[X] and E[Y |X] = Y µY + ρ σσX (X − µX ) a.e.. 2 (b) Show that var E(Y |X) = ρ2 σY2 and E Y − E(Y |X) = (1 − ρ2 )σY2 . (Hence if ρ ≈ 1 then Y is near E(Y |X) with high probability; if ρ = 0 then the variation of Y about E(Y |X) is the same as the variation about the mean µY .) 2 (c) Suppose E(X|Y ) = c + dY a.e. where bd < 1 and d 6= 0. Find expressions for E[X], E[Y ], ρ2 and σY2 /σX in terms of a, b, c and d. Page 8 §4 Aug 1, 2018(17:54) Bayesian Time Series Analysis 11. Best linear predictor of the random variable Y . Suppose the random vector (X, Y ) satisfies 0 < E[X 2 ] < ∞ and 0 < E[Y 2 ] < ∞. Find a and b such that therandom variable Yb = a + bX provides the best linear predictor of Y . Precisely, find a ∈ R and b ∈ R which minimize E ( Y − a − bX )2 . Note. Suppose E[Y |X] = a0 + b0 X. By exercise 9, we know that E[Y |X] = a0 + b0 X is the best predictor of Y . Hence a0 + b0 X is also the best linear predictor of Y 12. Suppose the random vector (X, Y ) has the density 6 2 f(X,Y ) (x, y) = 7 (x + y) for x ∈ [0, 1] and y ∈ [0, 1]; 0 otherwise. (a) Find the best predictor of Y . (b) Find the best linear predictor of Y . (c) Compare the plots of the answers to parts (a) and (b) as functions of x ∈ [0, 1]. Order statistics. 13. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the uniform U (0, 1) distribution. (a) Find the distribution of Xj:n . (b) Find E[Xj:n ]. 14. Suppose k > r. It is known1 that if the random variable X has an absolutely continuous distribution with distribution function F then the conditional distribution function P[Xk:n < y|Xr:n = x] is the same as the distribution function of the (k − r)th order statistic in a sample of size (n − r) from the distribution function F (y)−F (x) if y > x; 1−F (y) F1 (y) = 0 otherwise. Suppose X1 and X2 are i.i.d. absolutely continuous non-negative random variables with density function f (x) and distribution function F (x). By using the above result, show that X2:2 − X1:2 is independent of X1:2 if and only if X ∼ exponential (λ). 15. Suppose X1 , X2 , . . . , Xn are i.i.d. absolutely continuous non-negative random variables with density function f (x) and distribution function F (x). Define the vector (Y1 , Y2 , . . . , Yn ) by X2:n Xn:n Y1 = X1:n , Y2 = , . . . , Yn = X1:n X1:n (a) Find an expression for the density of the vector (Y1 , Y2 , . . . , Yn ) in terms of f and F . (b) Hence derive expressions for the density of the vector (Y1 , Y2 ) = (X1:n , X2:n/X1:n ) and the density of the random variable Y1 = X1:n . 4 The uniform distribution 4.1 Definition of the uniform distribution. Definition(4.1a). Suppose a ∈ R, b ∈ R and a < b. Then the random variable X has the uniform U (a, b) distribution iff X has density 1 for x ∈ (a, b). f (x) = b − a 0 otherwise. The distribution function is if x < a; 0 x−a F (x) = if x ∈ (a, b); b−a 1 if x > b. If X ∼ U (0, 1) and Y = a + (b − a)X then Y ∼ U (a, b). The uniform distribution is also called the rectangular distribution. Moments. All the are finite. Z moments Z b b a2 + ab + b2 (b − a)2 a+b 2 E[X] = xf (x), dx = E[X ] = x2 f (x) dx = var[X] = 2 3 12 a a Z b n+1 n+1 b −a E[X n ] = xn f (x) dx = for n 6= −1, n ∈ R. (n + 1)(b − a) a 1 For example, page 38 of [G ALAMBOS & KOTZ(1978)]. 1 Univariate Continuous Distributions fX (x) ................ .. ... .. ... ............................................................................................. ... .. .. ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... . ... ... . .. ... ... . .. ... ... . .. ... ... . .. ... ... . .. ... ... . .. ... ... . ... .. .... . ........................................................................................................................................................................................................ . ... . 1 b−a 0 a §4 Page 9 Aug 1, 2018(17:54) b x FX (x) .......... .. ... .. .......... ......................................... .... ..... . .. ..... .. ..... ... . ..... . . . ... . . .. . ... ..... . ..... ... . ..... . . . ... . . ... . . . . ... . ... . . . . ... . . ... . . . ... . . ... . . . . ... . ... . . . . ... . . ... . . . ... . . ... . . . . ... . ... . . . . ... . . . .. ......................................................................................................................................................................................................... . ... . 1 0 a b x Figure(4.1a). Left: plot of density of U (a, b). Right: plot of distribution function of U (a, b). (PICTEX) The moment generating function and characteristic function. tb (b−a)t Z b tx e − eta 2 sin( 2 )ei(a+b)t/2 e for t 6= 0; itX tX dx = and E[e ] = E[e ] = t(b − a) t(b − a) a b−a 1 for t = 0. 1 for t 6= 0; for t = 0. 4.2 Sums of i.i.d. uniforms. First, the sum of two i.i.d. uniforms. Example(4.2a). Suppose X ∼ U (0, a), Y ∼ U (0, a) and X and Y are independent. Find the distribution of Z = X + Y . Solution. Clearly Z ∈ (0, 2a). The usual convolution integral gives Z Z 1 a fZ (z) = fX (x)fY (z − x) dx = fY (z − x) dx a 0 x where fY (z − x) = 1/a when 0 < z − x < a; i.e. when z − a < x < z. Hence z Z min{a,z} min{a, z} − max{0, z − a} a2 1 dx = = fZ (z) = 2 2a − z a max{0,z−a} a2 a2 if 0 < z < a; if a < z < 2a. and the distribution function is 2 z if 0 < z < a; 2 2a FZ (z) = z2 (2a − z)2 2z − 2 −1=1− if a < z < 2a. a 2a 2a2 A graph of the density and distribution function is shown in figure(4.2a). For obvious reasons, this is called the triangular distribution. fX (x) ............... .... ... .. ... .......... ... ..... . ...... ..... . ..... ... ..... .. ........ . ... . . ..... ... . ... ..... .... . ..... ... ..... . ..... ..... . ... ..... .... . . . . ... ..... . . .... ..... . . ... . . .. ..... . . . . ... . ..... .... . . . ... ..... . . ... ..... . . ... . . .. ..... . . . . ... . ..... .. . ... ....... ..... . ..... ... ....... . .... ........ . . . ........................................................................................................................................................................................................ . .... 1/a 0 a 2a x FX (x) .......... .... ... .. .......... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ............................................................. . ........ .... . ....... ... . ....... ...... . ... ..... . . . . . ... . ... . . . . ... . ... . . . . . ... ... . . . . . ... . .. . . . . . . . . . . . . . . . . . .......... . .. . . . . . .... ... . . . . . ... . . . ... . . . ... . . . .... . . . . . ... . ... . . . . . . . ... . . .... . . . . . . ... . . .... . . . . .. . . . . . . . . . ............................................................................................................................................................................................................... . .... 1 1/2 0 a 2a x Figure(4.2a). Plot of density (left) and distribution function (right) of triangular distribution. (PICTEX) Now for the general result on the sum of n independent and identically distributed uniforms. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the U (0, 1) distribution. Let Sn = X1 + · · · + Xn . Then the density and distribution function of Sn are given by n n 1 X k n Fn (t) = (−1) for all t ∈ R and all n = 1, 2, . . . . (4.2a) (t − k)+ n! k k=0 n X n−1 1 k n fn (t) = (−1) (t − k)+ for all t ∈ R and all n = 2, 3, . . . . (n − 1)! k Proposition(4.2b). k=0 Page 10 §4 Aug 1, 2018(17:54) Bayesian Time Series Analysis Proof. We prove the result for Fn (t) by induction on n. Now if n = 1, then the right hand side of equation(4.2a) gives [t+ − (t − 1)+ ] which equals ( 0 if t < 0; t if 0 < t < 1; 1 if t > 1. as required. Also, for t ∈ R and n = 2, 3, . . . , we have Z t Z 1 Z 1 fn−1 (y) dy = Fn−1 (t) − Fn−1 (t − 1) fn−1 (t − x) dx = fn−1 (t − x)f1 (x) dx = fn (t) = y=t−1 0 0 Assume that equation(4.2a) is true for n; to prove it true for n + 1: fn+1 (t) = [Fn (t) − Fn (t − 1)] " n # n X 1 X k n + n k n + n = − (−1) (t − k) (−1) (t − k − 1) n! k k k=0 k=0 " n # n+1 X 1 X n n n n = (−1)k (t − k)+ + (−1)` (t − `)+ n! k `−1 k=0 `=1 n+1 X n n+1 1 (−1)k (t − k)+ = n! k k=0 n by using the combinatorial identity nk + k−1 = n+1 k . Integrating gives Z t n+1 X n+1 1 k n+1 Fn+1 (t) = fn+1 (x) dx = (−1) (t − k)+ (n + 1)! k 0 k=0 This establishes the proposition. The proposition easily extends to other uniforms. For example. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the U (0, a) distribution and Sn = X1 + · · · + Xn . Then the proposition can be applied to the sum Sn0 = Y1 + · · · + Yn , where Yj = Xj/a. Hence Fn (t) = P[X1 + · · · + Xn ≤ t] = P[Sn0 ≤ t/a] n n 1 X k n = n (−1) (t − ka)+ for all t ∈ R and all n = 1, 2, . . . . a n! k k=0 n X n−1 1 k n (−1) (t − ka)+ fn (t) = n for all t ∈ R and all n = 2, 3, . . . . k a (n − 1)! k=0 Similarly if the random variables come from the U (a, b) distribution. An alternative proof of proposition(4.2b) is based on taking the weak limit of the equivalent result for discrete uniforms; the proof below is essentially that given on pages 284–285 in [F ELLER(1968)]. Proof. (An alternative proof of proposition(4.2b).) For m = 1, 2, . . . and n = 2, 3, . . . , suppose Xm1 , Xm2 , . . . , Xmn are i.i.d. random variables with the discrete uniform distribution on the points 0, 1/m, 2/m, . . . , m/m = 1. 0 0 0 0 Let Xmj = mXmj + 1. Then Xm1 , Xm2 , . . . , Xmn are i.i.d. random variables with the discrete uniform distribution on the points 1, 2, . . . , m + 1. Now the discrete uniform distribution on the points 1, 2, . . . , m + 1 has probability generating function s(1 − sm+1 ) for |s| < 1. (m + 1)(1 − s) 0 0 + · · · + Xmn is Hence the probability generating function of Xm1 n m+1 n s (1 − s ) for |s| < 1. (m + 1)n (1 − s)n 0 0 + · · · + Xmn ≤ j] is Hence the probability generating function of the sequence P[Xm1 n m+1 n s (1 − s ) for |s| < 1. (4.2b) (m + 1)n (1 − s)n+1 Also 0 0 P[Xm1 + · · · + Xmn ≤ j] = P[Xm1 + · · · + Xmn ≤ mj + n] mj+n and this is the coefficient of s in the expansion of equation(4.2b), which in turn is n 1 (1 − sm+1 )n X n sm`+` mj × coefficient of s in the expansion of = (−1)` n n+1 (m + 1) (1 − s) ` (1 − s)n+1 `=0 1 Univariate Continuous Distributions §4 Page 11 Aug 1, 2018(17:54) which is n m`−mj+` X 1 n `s 0 × coefficient of s in the expansion of (−1) (m + 1)n ` (1 − s)n+1 `=0 which is clearly 0 if l > j. Otherwise it is n n X 1 n 1 X (mj − m` − ` + n)! ` mj − m` − ` + n ` n (−1) = (−1) (4.2c) (m + 1)n ` n n! ` (m + 1)n (mj − m` − `)! `=0 `=0 ` P∞ by using the binomial series 1/(1−z)n+1 = `=0 `+n n z for |z| < 1. Taking the limit as m → ∞ of the expression in (4.2c) gives n n n 1 X (−1)` (j − `)+ n! ` `=0 This proves the result when j is an integer. If j is any rational, take a series m tending to ∞ with mj an integer. Hence the result in equation(4.2a) for any t by right continuity. A note on the combinatorialPidentity implied by equation(4.2a) on page 9. Now Fn (t) = P[Sn ≤ t] = 1 for t ≥ n. By n equation(4.2a), this implies k=0 (−1)k nk (t − k)n = n! for t ≥ n. How do we prove this identity without probability? For all t ∈ R we have the identity n X n (−1)k etk = (1 − et )n k k=0 Pn n Setting t = 0 gives k=0 k (−1)k = 0. Differentiating the identity once and setting t = 0 gives n X n (−1)k k = 0 k k=0 Similarly, differentiating r times and setting t = 0 gives n X n 0 if r = 0, 1, 2. . . . , n − 1; (−1)k k r = (−1)n n! if r = n. k k=0 and hence n X n k=0 For all t ∈ R k n X n k=0 k (−1)n−k k r = n 0 if r = 1, 2. . . . , n − 1; n! if r = n. n X n X n n j t (−1)n−j k n−j j k k=0 j=0 n n X X n n j j = (−1) t (−1)n−k k n−j = n! j k (−1)k (t − k)n = (−1)k j=0 k=0 This generalizes the combinatorial result implied by equation(4.2a). See also question 16 on page 65 in [F ELLER(1968)]. 4.3 Representing the uniform distribution as the sum of independent Bernoulli random variables. every y ∈ [0, 1) can be represented as a ‘binary decimal’. This means we can write y = 0.x1 x2 x3 . . . where each xj is either 0 or 1. This representation motivates the following result: Proposition(4.3a). Suppose X1 , X2 , . . . are i.i.d. random variables with the Bernoulli distribution 1 with probability 1/2; 0 with probability 1/2. Then the random variable V = ∞ X Xk k=1 2k has the uniform U (0, 1) distribution. Proof. Let Sn = Pn k=1 Xk/2k for n = 2, 3, . . . . Now the moment generating function of k k 1 1 E[etXk /2 ] = + et/2 2 2 Xk/2k is Now Page 12 §4 Aug 1, 2018(17:54) Bayesian Time Series Analysis Hence E[etSn ] = n i 1 Y h t/2k e + 1 2n k=1 Using the identity e t/2n+1 n+1 n − 1 et/2 + 1 = et/2 − 1, and induction, it is possible to show n h Y i k et/2 + 1 = k=1 et − 1 et/2n − 1 and hence 1 et − 1 et − 1 → as n → ∞. 2n et/2n − 1 t Because (et − 1)/t is the moment generating function of U (0, 1), we see that V ∼ U (0, 1) as required. E[etSn ] = Using the same approach we can also prove the following representations: suppose n ≥ 1, Vn ∼ U (0, 1/2n ), and X1 , X2 , . . . , Xn , Vn are all independent. Then n X Xk Vn + ∼ U (0, 1) 2k k=1 4.4 Order statistics for the uniform distribution. Suppose X1 , . . . , Xn are i.i.d. with the U (0, 1) distribution. Maximum and minimum. It is straightforward to check the following results: n P[Xn:n ≤ x] = xn for 0 ≤ x ≤ 1, and E[Xn:n ] = n+1 1 n P[X1:n ≤ x] = 1 − (1 − x) for 0 ≤ x ≤ 1, and E[X1:n ] = n+1 Note that E[X1:n ] = 1 − E[Xn:n ] → 0 as n → ∞. n For 0 ≤ x < 1 we have P[nX1:n > x] = P[X1:n > x/n] = 1 − nx → e−x as n → ∞. This implies D D nX1:n =⇒ exponential (1) as n → ∞ where =⇒ denotes weak convergence (also called convergence in distribution). This is generalised below to Xk:n . n Also, for x < 0, we have P[n(Xn:n − 1) < x] = P[Xn:n < 1 + x/n] = 1 + nx → ex as n → ∞. The distribution of Xk:n . By equations(2.3a) and (2.3b) we have n X n j P[Xk:n ≤ t] = t (1 − t)n−k j j=k n − 1 k−1 t (1 − t)n−k fXk:n (t) = n k−1 The density is clearly the Beta distribution Beta(k, n − k + 1). The limiting distribution of Xk:n as n → ∞. for 0 ≤ t ≤ 1. P[nXk:n > t] = 1 − P[Xk:n ≤ 1/n] n j X n t t n−j =1− 1− j n n j=k k−1 t n n t t n−1 n t t n−k = 1− + 1− + ··· + 1− n 1 n n k−1 n n tk−1 e−t → e−t + te−t + · · · + k! which is equal to P[Y > t] where Y ∼ Gamma(k, 1) distribution—this will be shown in §8.2. D We have shown that nXk:n =⇒ Gamma(k, 1) as n → ∞ for any fixed k ∈ {1, 2, . . . , n}. (4.4a) 1 Univariate Continuous Distributions §4 Page 13 Aug 1, 2018(17:54) 4.5 The probability integral transform. Every probability distribution function F is monotonic increasing; if F is a strictly increasing probability distribution function on the whole of R, then we know from elementary analysis that F has a unique inverse G = F −1 . If U ∼ U (0, 1) and Y = G(U ), then clearly Y has distribution function F . Hence we can simulate variates from the distribution F by simulating variates x1 , x2 , . . . xn from the U (0, 1) distribution and calculating G(x1 ), G(x2 ), . . . , G(xn ). Now for the general case. Suppose F : R → [0, 1] is a distribution function (not necessarily continuous). We first need to define the “inverse” of F . Proposition(4.5a). Suppose F : R → [0, 1] is a distribution function and we let G(u) = min{x: F (x) ≥ u} for u ∈ (0, 1). Then Proof. Fix u ∈ (0, 1); then {x: G(u) ≤ x} = {x: F (x) ≥ u}. x0 ∈ R.H.S. ⇒ F (x0 ) ≥ u ⇒ G(u) ≤ x0 Conversely by definition of G ⇒ x0 ∈ L.H.S. x0 ∈ L.H.S. ⇒ G(u) ≤ x0 ⇒ min{x: F (x) ≥ u} ≤ x0 Let x∗ = min{x: F (x) ≥ u}. Choose a sequence {xn }n≥1 with xn ↓↓ x∗ as n → ∞ (this means that the sequence {xn }n≥1 strictly decreases with limit x∗ ). Hence F (xn ) ≥ u for all n = 1, 2, . . . . Now F is a distribution function; hence F is right continuous; hence F (x∗ ) ≥ u. Also x0 ≥ x∗ and F is monotonic increasing; hence F (x0 ) ≥ F (x∗ ) ≥ u. Hence x0 ∈ R.H.S. If the distribution function F is continuous at α ∈ R, then G(β) = α implies F (α) = β (because G(β) = α implies F (α) ≥ β and for every x < α we have F (x) < β). The above result implies the following. • Suppose F is a distribution function and G is defined in terms of F as explained above. If U has a uniform distribution on (0, 1) and X = G(U ) then Pr[X ≤ x] = Pr [G(U ) ≤ x] = Pr [F (x) ≥ U ] = F (x). Hence If U ∼ U (0, 1), then the distribution function of G(U ) is F . • If the random variable X has the distribution function F and F is continuous, then the random variable F (X) has the U (0, 1) distribution. As explained before the proposition, if the distribution function F is strictly increasing on the whole of R then F −1 , the inverse of F , exists and G = F −1 . If F is the distribution function of a discrete distribution, then F is constant except for countably many jumps and the inverse of F does not exist. However, G(u) is still defined by the proposition and this method of simulating from the distribution F still works. 4.6 Using the probability integral transformation to prove results about order statistics. Suppose X1 , . . . , Xn are i.i.d. with the U (0, 1) distribution and we have established that n! fXk:n (x) = xk−1 (1 − x)n−k (k − 1)!(n − k)! Suppose Y1 , . . . , Yn are i.i.d. with an absolutely continuous distribution with distribution function FY and density function fy and we wish to find the distribution of Yk:n . Then X1 = FY (Y1 ), . . . , Xn = FY (Yn ) are i.i.d. with the U (0, 1) distribution and hence n! fXk:n (x) = xk−1 (1 − x)n−k (k − 1)!(n − k)! Now FY is monotonic increasing and continuous and the transformation (X1 , . . . , Xn ) → (Y1 , . . . , Yn ) is order preserving; hence Z FY (y) P[Yk:n ≤ y] = P[FY (Yk:n ) ≤ FY (y)] = P[Xk:n ≤ FY (y)] = fXk:n (x) dx −∞ and hence Page 14 §5 Aug 1, 2018(17:54) Bayesian Time Series Analysis n! {FY (y)}k−1 {1 − FY (y)}n−k fY (y) (k − 1)!(n − k)! This approach provides a general method for proving results about the order statistics of a sample from a continuous distribution function. fYk:n (y) = 4.7 Random partitions of an interval. Suppose U1 , . . . , Un are i.i.d. random variables with the uniform U (0, 1) distribution and let U1:n , . . . , Un:n denote the order statistics. These variables partition the interval [0, 1] into n + 1 disjoint intervals with the following lengths: D1 = U1:n , D2 = U2:n − U1:n , . . . , Dn = Un:n − U(n−1):n , Dn+1 = 1 − Un:n Clearly D1 + · · · + Dn+1 = 1. The absolute value of the Jacobian of the transformation (U1:n , . . . , Un:n ) → (D1 , . . . , Dn ) is ∂(d1 , . . . , dn ) ∂(u1:n , . . . , un:n ) = 1 The density of (U1:n , . . . , Un:n ) is given by (2.2a). Hence the density of (D1 , . . . , Dn ) is P (4.7a) f(D1 ,...,Dn ) (d1 , . . . , dn ) = n! for d1 ≥ 0,. . . , dn ≥ 0, n`=1 d` ≤ 1. There are many results on random partitions of an interval—see [F ELLER(1971)], [DAVID & BARTON(1962)] and [W HITWORTH(1901)]. 5 Exercises (exs-uniform.tex) 1. Suppose X ∼ U (a, b). Show that n (b − a) n n+1 (b − a) 1 − (−1) n E[(X − µ) ] = = (n + 1)2n (n + 1)2n+1 0 if n is even; if n is odd. 2. Transforming uniform to exponential. Suppose X ∼ U (0, 1). Find the distribution of Y = − ln X. 3. Product of independent uniforms. (a) Suppose X and Y are i.i.d. with the U (0, 1) distribution. Let Z = XY . Find the density and distribution function of Z. (b) (Note: this part makes use of the fact that the sum of independent exponentials is a gamma distribution—see proposition(8.6a) on page 20.) Suppose X1 , . . . , Xn are i.i.d. with the U (0, 1) distribution and let Pn = X1 · · · Xn . Find the density of − ln Pn and hence find the density of Pn . 4. Suppose X1 ∼ U (0, 1), X2 ∼ U (0, X1 ), X3 ∼ U (0, X2 ), . . . , Xn ∼ U (0, Xn−1 ) for some n ≥ 2. (a) Prove by induction that the density of Xn is n−1 ln 1/xn f (x) = for 0 < xn < 1. (n − 1)! (b) By using the result of part (b) exercise 3, find the density of Xn . 5. Suppose X is a random variable with an absolutely continuous distribution with density f . Then entropy of X is defined to be Z H(X) = − f (x) ln f (x) dx Suppose X ∼ U (a, b). Find the entropy of X. (It can be shown that the continuous distribution on the interval (a, b) with the largest entropy is the uniform.) 6. Sum and difference of two independent uniforms. Suppose X ∼ U (0, a) and Y ∼ U (0, b) and X and Y are independent. (a) Find the density of V = X + Y and sketch its shape. (b) Find the density of W = Y − X and sketch its shape. 7. Suppose X ∼ U (0, a) and Y ∼ U (0, b) and X and Y are independent. Find the distribution of V = min{X, Y } and find P[V = X]. 8. A waiting time problem. Suppose you arrive at a bus stop at time t = 0. The stop is served by two bus routes. From past observations, you assess that the time X1 to wait for a bus on route 1 has the U (0, a) distribution and the time X2 to wait for a bus on route 2 has the U (0, b) distribution. Also X1 and X2 are independent. (Clearly this assumption will not hold in practice!!) A bus on route 1 takes the time α to reach your destination whilst a bus on route 2 takes the time α + β. Suppose the first bus arrive at the stop at time t0 and is on route 2. Should you catch it if you wish to minimize your expected arrival time? 1 Univariate Continuous Distributions §6 Page 15 Aug 1, 2018(17:54) 9. Suppose U ∼ U (0, 1) and the random variable V has an absolutely continuous distribution with finite expectation and density f . Also U and V are independent. Let W denote the fractional part of U + V ; this means that W = U + V − bU + V c. Show that W ∼ U (0, 1). (See also Poincaré’s roulette problem; pages 62–63 in [F ELLER(1971)] for example. ) 10. Suppose X and Y are i.i.d. random variables with the U (0, 1) distribution. Let V = min{X, Y } and W = max{X, Y }. (a) Find the distribution functions, densities and expectations of V and W . (b) Find P[V ≤ v, W ≤ w] and hence derive f(V,W ) (v, w), the joint density of (V, W ). (c) Find the density of (W |V ≤ v) and hence derive E[W |V ≤ v]. 11. Suppose two points are chosen independently and at random on a circle with a circumference which has unit length. (a) Find the distribution of the lengths of the intervals (X1 , X2 ) and (X2 , X1 ). (b) Find the distribution of the length of the interval L which contains the fixed point Q. (See page 23 in [F ELLER(1971)].) 12. Suppose n points are distributed independently and uniformly on a disc with radius r. Let D denote the distance from the centre of the disc to the nearest point. Find the density and expectation of D. 13. Suppose the random vector (X1 , X2 ) has a distribution which is uniform over the disc {(x, y) ∈ R2 : x2 + y 2 ≤ a2 }. Find the density of X1 . 14. Suppose a < b and X1 , X2 , . . . , Xn are i.i.d. random variables with the U (a, b) distribution. Let Sn = X1 + · · · + Xn . Find expressions for the distribution function and density function of Sn . 15. Suppose X1 , . . . , Xn are i.i.d. with the U (0, 1) distribution. (a) Find E[Xk:n ] and var[Xk:n ] for k = 1, 2, . . . , n. (b) Find the joint density of (Xj:n , Xk:n ). (c) Find E[Xj:n Xk:n ], cov[Xj:n , Xk:n ] and corr[Xj:n , Xk:n ]. 6 The exponential distribution 6.1 The basics The random variableX has an exponential distribution, exponential (λ), iff it has an absolutely continuous distribution with density λe−λx if x > 0; f (x) = 0 if x < 0. where λ > 0. Definition(6.1a). The distribution function. F (x) = 1 − e−λx 0 Moments. These can be obtained by integrating by parts. 1 2 E[X] = E[X 2 ] = 2 λ λ if x > 0; if x < 0. var[X] = 1 λ2 (6.1a) The moment generating function and characteristic function. λ λ E[etX ] = φ(t) = E[eitX ] = λ−t λ − it Multiple of an exponential distribution. Suppose X ∼ exponential (λ) and Y = αX where α > 0. Then P[Y > t] = P[X > t/α] = e−λt/α and hence Y ∼ exponential ( λ/α). Page 16 §6 Aug 1, 2018(17:54) Bayesian Time Series Analysis 6.2 The exponential as the limit of geometric distributions. Suppose events can only occur at times δ, 2δ, 3δ, . . . , and events at different times are independent. Let P[event occurs at time kδ] = p Let T denote the time to the first event. Then P[T > kδ] = (1 − p)k Hence P[T = kδ] = (1 − p)k−1 p which is the geometric distribution and E[T ] = δ/p. Now suppose δn → 0 as n → ∞ in such a way that E[Tn ] = δn/pn = 1/α is constant. Then lim P[Tn > t] = lim (1 − pn )t/δn = lim (1 − αδn )t/δn = e−αt n→∞ n→∞ n→∞ (6.2a) D and hence Tn =⇒ T as n → ∞ where T ∼ exponential (α). Here is another approach to defining points distributed randomly on [0, ∞). Suppose X1 , . . . , Xn are i.i.d. random variables with the U (0, `n ) distribution. Then P[X(1) > t] = (1 − t/`n )n for t ∈ (0, `n ). Now suppose n → ∞ in D such a way that n/`n = λ is fixed. Then limn→∞ P[X(1) > t] = e−λt , which means that X(1) =⇒ T as n → ∞ where T ∼ exponential (α). Informally, this result says that if points are distributed randomly on the line such that the mean density of points is λ, then the distance to the first point has the exponential (λ) distribution. 6.3 The lack of memory or Markov property of the exponential distribution. Suppose the random variable T models the lifetime of some component. Then the random variable T is said to have the lack of memory property iff the remaining lifetime of an item which has already lasted for a length of time x has the same distribution as T . This means P[T > x + t|T > x] = P[T > t] and hence P[T > x + t] = P[T > t] P[T > x] for all t > 0 and x > 0. Similarly, the distribution with distribution function F has the lack of memory property iff [1 − F (x + t)] = [1 − F (x)][1 − F (t)] for all x > 0 and all t > 0. If X ∼ exponential (λ) then 1 − F (x) = e−λx and hence X has the lack of memory property. Conversely Proposition(6.3a). Suppose X is an absolutely continuous random variable on [0, ∞) with the lack of memory property. Then there exists λ > 0 such that X ∼ exponential (λ). Proof. Let G(x) = P[X > x]. Then G(x + y) = G(x)G(y) for all x ≥ 0 and all y ≥ 0. m n mn Suppose x = m/n is rational. Then G(m/n) = G(1/n) . Raising to the nth power gives G(m/n) = G(1/n) = [G(1)]m . Hence G(m/n) = [G(1)]m/n . Now suppose x is any real number in [0, ∞). Choose sequences qn and rn of rationals such that qn ≤ x ≤ rn and qn → x as n → ∞ and rn → x as n → ∞. Hence G(1)qn = G(qn ) ≥ G(x) ≥ G(rn ) = G(1)rn . Letting n → ∞ gives G(x) = G(1)x . Now let λ = − ln [G(1)]. Then G(x) = G(1)x = e−λx . See also [F ELLER(1968)], page 459. The proof of proposition(6.3a) depends on finding a solution of the functional equation f (x + y) = f (x)f (y). Taking logs of this equation gives the Cauchy functional equation f (x + y) = f (x) + f (y). Both of these equations have been studied very extensively—see [S AATY(1981)], [ACZ ÉL(1966)], [K UCZMA(2009)], etc. 6.4 Distribution of the minimum. Suppose Xj ∼ exponential (λj ) for j = 1, 2, . . . , n. Suppose further that X1 , X2 , . . . , Xn are independent. Let X1:n = min{X1 , . . . , Xn }. Then P[X1:n > t] = P[X1 > t] · · · P[Xn > t] = e−(λ1 +···+λn )t for t > 0. Hence X1:n ∼ exponential (λ1 + · · · + λn ). In particular, we have shown that if X1 , . . . , Xn are i.i.d. random variables with the exponential (λ) distribution, then X1:n ∼ exponential (nλ). Hence X1 X1:n has the same distribution as n It can be shown that this property characterizes the exponential distribution; see, for example, page 39 of [G ALAM BOS & KOTZ (1978)]. 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §6 Page 17 Suppose we think of X1 , . . . , Xn as the times to n events occurring. So we have shown that the time to the first event has the exponential (nλ) distribution. Using the lack of memory property suggests that the extra time to the second event, X2:n − X1:n , should have the exponential ( (n − 1)λ ) distribution. And so on. This result is established in the following proposition. Proposition(6.4a). Suppose X1 , . . . , Xn are i.i.d. random variables with the exponential (λ) distribution. De- fine Z1 , . . . , Zn by Z1 = nX1:n , Z2 = (n − 1)(X2:n − X1:n ), . . . , Zn = Xn:n − X(n−1):n Then Z1 , . . . ,Zn are i.i.d. random variables with the exponential (λ) distribution. Pn −λ xj j=1 for 0 < x1 < · · · < xn . Proof. We know that the density of (X1:n , . . . , Xn:n ) is g(x, . . . , xn ) = n!λn e Also Z1 Z1 Z2 Z1 Z2 Zn−1 X1:n = X2:n = + . . . , Xn:n = + + ··· + + Zn n n n−1 n n−1 2 and hence the Jacobian of the transformation is ∂(x1:n , . . . , xn:n ) 1 = ∂(z1 , . . . , zn ) n! Hence the density of (Z1 , . . . , Zn ) is 1 f(Z1 ,...,Zn ) (z1 , . . . , zn ) = n!λn e−λ(z1 +···+zn ) for z1 > 0, . . . , zn > 0. n! This establishes the proposition. The following result links the order statistics from a uniform distribution and the exponential distribution. Proposition(6.4b). Suppose Y1 , . . . , Yn+1 are i.i.d. random variables with the exponential (λ) distribution, and let S` = Y1 + · · · + Y` Then for ` = 1, . . . , n + 1. Yn Y1 ,..., is independent of Sn+1 . (6.4a) Sn+1 Sn+1 Suppose U1 , . . . , Un are i.i.d. random variables with the U (0, 1) distribution and denote the vector of order statistics by (U1:n , . . . , Un:n ). Let D1 = U1:n , D2 = U2:n − U1:n , . . . , Dn = Un:n − U(n−1):n , Dn+1 = 1 − Un:n Then Y1 Yn+1 ,..., has the same distribution as (D1 , . . . , Dn+1 ) (6.4b) Sn+1 Sn+1 Proof. By equation(2.2a), the density of the vector (U1:n , . . . , Un:n ) is n n! if 0 < x1 < · · · < xn < 1 g(x1 , . . . , xn ) = 0 otherwise Also f(Y1 ,...,Yn+1 ) (y1 , . . . , yn+1 ) = λn+1 e−λ(y1 +···+yn+1 ) for y1 > 0,. . . , yn+1 > 0. Consider the transformation: Y1 Y2 Yn X1 = , X2 = , . . . , Xn = , Xn+1 = Y1 + · · · + Yn+1 Sn+1 Sn+1 Sn+1 Or Y1 = X1 Xn+1 , Y2 = X2 Xn+1 , . . . , Yn = Xn Xn+1 , Yn+1 = (1 − X1 − X2 − · · · − Xn )Xn+1 The absolute value of the Jacobian of the transformation is: ∂(y1 , . . . , yn+1 ) n ∂(x1 , . . . , xn+1 ) = xn+1 The determinant can be easily evaluated by replacing the last row by the sum of all the rows—this gives an upper triangular determinant. Hence the density of the transformed distribution is Pn+1 n+1 −λxn+1 n xn+1 for x1 ≥ 0, . . . , xn+1 ≥ 0, `=1 x` = 1 f(X1 ,...,Xn+1 ) (x1 , . . . , xn+1 ) = λ e 0 otherwise Now Xn+1 = Y1 + · · · + Yn+1 is the sum of (n + 1) i.i.d. exponentials; hence by proposition(8.6a) on page 20, Xn+1 ∼ Gamma(n + 1, λ) and has density λn+1 xnn+1 e−λxn+1 fXn+1 (xn+1 ) = n! Page 18 §7 Aug 1, 2018(17:54) Bayesian Time Series Analysis It follows that (X1 , . . . , Xn ) is independent of Xn+1 and (X1 , . . . , Xn ) has density Pn f(X1 ,...,Xn ) (x1 , . . . , xn ) = n! for x1 ≥ 0, . . . , xn ≥ 0, `=1 x` ≤ 1. This establishes (6.4a). Using equation(4.7a) on page 14 shows that the density of (D1 , . . . , Dn ) is the same as the density of (X1 , . . . , Xn ). Also D1 + · · · + Dn + Dn+1 = 1 and X1 + · · · + Xn + Yn+1/Sn+1 = 1. Hence (6.4b). Corollary(6.4c). With the same notation as the proposition, S1 Sn ,..., Sn+1 Sn+1 has the same distribution as (U1:n , . . . , Un:n ) (6.4c) Also S1 Sn ,..., Sn+1 Sn+1 is independent of Sn+1 . (6.4d) Proof. The proposition shows that (X1 , . . . , Xn ) has the same distribution (D1 , . . . , Dn ). Taking partial sums of both sides gives (6.4c). Result (6.4d) follows directly from (6.4a). 7 Exercises (exs-exponential.tex) 1. Suppose X ∼ U (0, 1) and λ > 0. Prove that Y =− ln(1 − X) ∼ exponential (λ) λ 2. Suppose the random variables X and Y are i.i.d. with the exponential (1) distribution. Let U = min{X, Y } and V = max{X, Y }. Proves that U ∼ exponential (2) and V ∼ X + 21 Y . 3. Suppose X ∼ exponential (µ) and Y ∼ exponential (δ) where 0 < δ ≤ µ. Suppose further that f : (0, ∞) → (0, ∞) with f differentiable and f 0 (x) > 0 for all x > 0. Prove that E[f (X)] ≤ E[f (Y )]. 4. Suppose X and Y are i.i.d. random variables with the exponential (λ) distribution. Find the conditional density of X given X + Y = z. What is E[X|X + Y ]? 5. Suppose X1 and X2 are i.i.d. random variables with the exponential (λ) distribution. Let Y1 = X1 −X2 and Y2 = X1 +X2 . (a) Find the densities of Y1 and Y2 . (b) What is the density of R = |X1 − X2 |? 6. A characterization of the exponential distribution. Suppose X1 and X2 are i.i.d. random variables which are non-negative and absolutely continuous. Let Y = min{X1 , X2 } and R = |X1 − X2 |. Then Y and R are independent iff X1 and X2 have the exponential distribution. 7. Suppose X1 ∼ exponential (λ1 ), X2 ∼ exponential (λ2 ) and X1 and X2 are independent. (a) Find P[min{X1 , X2 } = X1 ]. (b) Show that {min{X1 , X2 } > t} and {min{X1 , X2 } = X1 } are independent. (c) Let R = max{X1 , X2 } − min{X1 , X2 }. Find P[R > t]. (d) Show that min{X1 , X2 } and R are independent. 8. Most elementary analysis texts contain a proof of the result: n 1 lim 1 − =e n→∞ n By using this result, show that if {δn } is a real sequence in (0, ∞) such that δn → 0 as n → ∞ and α ∈ (0, ∞), then lim (1 − αδn )t/δn = e−αt n→∞ This is used in (6.2a) on page 16. 9. (a) Ratio of two independent exponentials. Suppose X ∼ exponential (λ), Y ∼ exponential (µ) and X and Y are independent. Find the distribution of Z = X/Y . (b) Product of two independent exponentials. Suppose X ∼ exponential (λ), Y ∼ exponential (µ) and X and Y are independent. (i) Find the distribution function of Z = XY . Express your answer in terms of the modified Bessel function of the second kind, order 1, which is: Z ∞ K1 (y) = cosh(x)e−y cosh(x) dx for <(x) > 0. x=0 1 Univariate Continuous Distributions §8 Page 19 Aug 1, 2018(17:54) (ii) Find the density of Z = XY . Express your answer in terms of the modified Bessel function of the second kind, order 0, which is: Z ∞ K0 (y) = e−y cosh(x) dx for <(x) > 0. x=0 (iii) Write down what these answers become when λ = µ. 10. Suppose X1 , . . . , Xn are i.i.d. random variables with the exponential (λ) distribution. Define Y1 , . . . , Yn as follows: Y1 = X1 , Y2 = X1 + X2 , . . . , Yn = X1 + · · · + Xn . Find the density of the vector (Y1 , . . . , Yn ). 11. Order statistics. Suppose X1 , . . . , Xn are i.i.d. random variables with the exponential (λ) distribution. Find (a) E[Xk:n ] (b) var[Xk:n ] (c) cov[Xj:n , Xk:n ]. 12. Suppose X1 , . . . , Xn are i.i.d. random variables with the exponential (λ) distribution. Prove that n X X1:n is independent of (X` − X1:n ) `=1 8 The gamma and chi-squared distributions 8.1 Definition of the Gamma distribution. Definition(8.1a). Suppose n > 0 and α > 0. Then the random variable X has the Gamma(n, α) distribution iff X has density f (x) = By definition, Γ(n) = R∞ 0 αn xn−1 e−αx Γ(n) for x > 0. (8.1a) xn−1 e−x dx for all n ∈ (0, ∞). It follows that Z ∞ Γ(n) xn−1 e−αx dx = n provided α > 0 and n > 0. α 0 (8.1b) 8.2 The distribution function. There is a simple expression only when n is a positive integer and then αx (αx)2 (αx)n−1 −αx Gn (x) = P[Sn ≤ x] = 1 − e 1+ + + ··· + (8.2a) 1! 2! (n − 1)! This is easy to check—just differentiate Gn (x) and obtain the density in equation(8.1a). Note that P[Sn ≤ x] = P[Y ≥ n] where Y has a Poisson distribution with expectation αx. In terms of the Poisson process with rate α, this means that the nth event occurs before time x iff there are at least n events in [0, x]. 8.3 Multiple of a gamma distribution. Suppose n > 0 and α > 0 and X ∼ Gamma(n, α) with density fX (x). Suppose further that Y = βX where β > 0. Then the density of Y is given by: fX (x) αn ( y/β )n−1 exp(−αy/β) fY (y) = = dy βΓ(n) dx Hence Y = βX ∼ Gamma(n, α/β ). 8.4 Moments and shape of the gamma distribution. Using the result that E[X k ] = Γ(n + k) αk Γ(n) R f (x) dx = 1 easily gives for n + k > 0. (8.4a) and so n E[X] = α n and var(X) = 2 α 1 α and E = for n > 1. X n−1 Page 20 §8 Aug 1, 2018(17:54) Bayesian Time Series Analysis n = 21 n=1 n=2 0.8 0.6 0.4 0.2 0.0 1 2 Figure(8.4a). Plot of gamma density function for n = 3 1 2, 4 n = 1 and n = 2 (all with α = 1). (wmf/gammadensity-fig001,121mm,73mm) By §8.3, we know that if X ∼ Gamma(n, α) then Y = X/α ∼ Gamma(n, 1). So without loss of generality, we consider the shape of the density of Gamma(n, 1) distribution. Let g(x) = xn−1 e−x . If n ≤ 1, then g(x) = e−x /x1−n is monotonic decreasing and hence the density of the Gamma(n, 1) distribution is monotonic decreasing. If n > 1, then g 0 (x) = e−x xn−2 [n − 1 − x]. Clearly, if x < n − 1 then g 0 (x) > 0; if x = n − 1 then g 0 (x) = 0 and if x > n − 1 then g 0 (x) < 0. Hence the density first increases to the maximum at x = n − 1 and then decreases. By using §8.3, it follows that the maximum of the density of a Gamma(n, α) density occurs at x = (n − 1)/α. 8.5 The moment generating function of a gamma distribution. Suppose X ∼ Gamma(n, α). Then Z ∞ Z ∞ n n−1 e−αx αn tX tx α x dx = xn−1 e−(α−t)x dx MX (t) = E[e ] = e Γ(n) Γ(n) 0 0 αn Γ(n) 1 = = for t < α. (8.5a) n Γ(n) (α − t) (1 − t/α)n Hence the characteristic function is 1/(1 − it/α)n ; in particular, if n = 1, the characteristic function of the exponential(α) distribution is α/(α − it). Equation(8.5a) shows that for integral n, the Gamma distribution is the sum of n independent exponentials. The next paragraph gives the long proof of this. 8.6 Representing the gamma distribution as a sum of independent exponentials. The following proposition shows that the distribution of the waiting time for the nth event in a Poisson process with rate α has the Gamma(n, α) distribution. Suppose X1 , X2 . . . . , Xn are i.i.d. random variables with the exponential density αe−αx for x ≥ 0. Then Sn = X1 + · · · + Xn has the Gamma distribution Γ(n, α). Proposition(8.6a). Proof. By induction: let gn denote the density of Sn . Then for all t > 0 we have Z t Z t n α (t − x)n−1 e−α(t−x) −αx −αx αe dx gn+1 (t) = gn (t − x)αe dx = Γ(n) 0 0 Z x=t αn+1 e−αt t αn+1 e−αt (t − x)n = (t − x)n−1 dx = − Γ(n) Γ(n) n 0 x=0 = αn+1 tn e−αt Γ(n + 1) as required. The result that the sum of n independent exponentials has the Gamma distribution is the continuous analogue of the result that the sum of n independent geometrics has a negative binomial distribution. Link with the Poisson distribution. Suppose Sn = X1 + · · · + Xn as in the proposition. Let Nt denote the number of indices k ≥ 1 with Sk ≤ t. Then P[Nt = n] = P[Sn ≤ t and Sn+1 > t] = Gn (t) − Gn+1 (t) e−αt (αt)n = n! by using equation(8.2a) on page 19. 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §8 Page 21 8.7 Normal limit and approximation. Suppose Gn ∼ Gamma(n, α). It follows from proposition(8.6a) and the Central Limit Theorem that for n large, Gn − n/α √ is approximately N (0, 1) n/α and hence for large n αx − n √ P[Gn ≤ x] ≈ Φ n 2 The local central limit theorem√ showsthat √ 1 2 n n+z n 1 fGn = n(z) where n(x) = √ e− 2 x (8.7a) lim n→∞ α α 2π See exercise 10 on page 23 below. 8.8 Lukacs’ characterization of the gamma distribution. Suppose X and Y are both positive, non-degenerate 3 and independent random variables. Then X/(X+Y ) is independent of X+Y iff there exist k1 > 0, k2 > 0 and α > 0 such that X ∼ Gamma(k1 , α) and Y ∼ Gamma(k2 , α). Proposition(8.8a). Proof. ⇐ This is exercise 5 on page 23. ⇒ This is proved in [L UKACS(1955)] and [M ARSAGLIA(1989)]. We can easily extend this result to n variables: Proposition(8.8b). Suppose X1 , X2 , . . . , Xn are positive, non-degenerate and independent random variables. Then Xj /(X1 + · · · + Xn ) is independent of X1 + · · · + Xn for j = 1, 2, . . . , n iff there exist α > 0, k1 > 0, . . . , kn > 0 such that Xj ∼ Gamma(kj , α) for j = 1, 2, . . . , n. Proof. ⇐ Now W = X2 + · · · + Xn ∼ Gamma(k2 + · · · + kn , β) and X1 ∼ Gamma(k1 , β). Also W and X1 are independent positive random variables. Hence X1 /(X1 + · · · + Xn ) is independent of X1 + · · · + Xn by proposition(8.8a).. Similarly Xj /(X1 + · · · + Xn ) is independent of X1 + · · · + Xn for j = 2, . . . , n. ⇒ Let Wj = X1 + · · · + Xn − Xj . Then Wj and Xj are independent positive random variables. Also Xj /(Wj + Xj ) is independent of Wj + Xj . By proposition(8.8a), there exist kj > 0, kj∗ > 0 and αj > 0 such that Xj ∼ Gamma(kj , αj ) and Zj ∼ Gamma(kj∗ , αj ). Hence X1 + · · · + Xn = Zj + Xj ∼ Gamma(kj + kj∗ , αj ). Applying the same argument to W1 , . . . , Wn gives α1 = · · · = αn . The result follows. 8.9 The χ2 distribution. For n ∈ (0, ∞) the Gamma( n/2, 1/2) distribution has density: xn/2−1 e−x/2 f (x) = n/2 n for x > 0. 2 Γ( /2) This is the density of the χ2n distribution. If n is a positive integer, then n is called the degrees of freedom. In particular, if n ∈ (0, ∞) and X ∼ Gamma( n/2, α) then 2αX ∼ Gamma( n/2, 1/2) = χ2n . If Y ∼ χ2n = Gamma( n/2, 1/2), then equation(8.4a) shows that the k th moment of Y is given by ( n k Γ(k+ /2) 2 if n > −2k; k n Γ( /2) E[Y ] = (8.9a) ∞ if n ≤ −2k. √ √ In particular E[Y ] = n, E[Y 2 ] = n(n + 2), var[Y ] = 2n, E[ Y ] = 2Γ( (n+1)/2)/Γ( n/2) and E[1/Y ] = 1/(n − 2) provided n > 2. By equation(8.5a), the c.f. of the χ2n distribution is 1/(1 − 2it)n/2 . It immediately follows that if X ∼ χ2m , Y ∼ χ2n and X and Y are independent, then X + Y ∼ χ2m+n . 2 The local central limit theorem. Suppose Y1 , Y2 , . . . are i.i.d. random variables with mean 0 and variance 1 and characteristic k function φY . Suppose further that √ |φY | is integrable for some positive k and sup{|φY (t)| : |t| ≥ δ} < 1 for all δ > 0. Let Sn = Y1 + · · · + Yn ; then Sn / n has a bounded continuous density fn for all n ≥ k and supx∈R |fn (x) − n(x)| → 0 as n → ∞. This formulation is due to Michael Wichura: galton.uchicago.edu/~wichura/Stat304/Handouts/L16.limits.pdf. See also page 516 in [F ELLER(1971)]. 3 To exclude the trivial case that both X and Y are constant. In fact if one of X and Y is constant and X/(X +Y ) is independent of X + Y , then the other must be constant also. Page 22 §9 Aug 1, 2018(17:54) Bayesian Time Series Analysis 8.10 The generalized gamma distribution. Definition(8.10a). Suppose n > 0, λ > 0 and b > 0. Then the random variable X has the generalized gamma distribution GGamma(n, λ, b) iff X has density bλn bn−1 −λxb x e for x > 0. (8.10a) f (x) = Γ(n) Note that if n = b = 1, then the generalized gamma is the exponential distribution, if b = 1, the generalized gamma is the gamma distribution, if n = 1, the generalized gamma is the Weibull distribution—introduced below in §24.3 on page 52, if n = 1, b = 2 and λ = 1/2σ2 , the generalized gamma is the Rayleigh distribution—introduced below in §24.2 on page 52, and finally if n = 1/2, b = 2 and λ = 1/2σ2 then the generalized gamma is the half-normal distribution—introduced in exercise 1.11(7) on page 26. It is left to an exercise (see exercise12 on page 23) to check: • The function f in equation(8.10a) integrates to 1 and so is a density. • If X ∼ GGamma(n, λ, b) then Y = X b ∼ Gamma(n, λ). • The central moments are given by the expression: Γ( k/b + n) E[X k ] = b/k λ Γ(n) The generalized gamma distribution is used in survival analysis and reliability theory to model lifetimes. 8.11 Summary. The gamma distribution. • Density. X has the Gamma(n, α) density for n > 0 and α > 0 iff αn xn−1 e−αx for x > 0. Γ(n) • Moments. E[X] = n/α; var[X] = n/α2 and E[X k ] = Γ(n+k)/αk Γ(n) for n + k > 0. • M.g.f. and c.f. fX (x) = MX (t) = E[etX ] = 1 (1 − t/α)n for t < α. φX (t) = E[eitX ] = 1 (1 − it/α)n • Properties. Gamma(1, α) is the exponential (α) distribution. If X ∼ Gamma(n, α) and β > 0 then βX ∼ Gamma(n, α/β ). The Gamma(n, α) distribution is the sum of n independent exponential (α) distributions. If X ∼ Gamma(m, α), Y ∼ Gamma(n, α) and X and Y are independent, then X + Y ∼ Gamma(m + n, α). The χ2n distribution. • This is the Gamma( n/2, 1/2) distribution. • If X ∼ χ2n , then E[X] = n, var[X] = 2n and the c.f. is φ(t) = 1/(1 − 2it)n/2 . • If X ∼ χ2m , Y ∼ χ2n and X and Y are independent, then X + Y ∼ χ2m+n . • The χ22 distribution is the exponential ( 1/2) distribution. 9 Exercises (exs-gamma.tex) R∞ x−1 −u 1. The Gamma function. This is defined to be Γ(x) = 0 u e du for x > 0. Show that (a) Γ(x + 1) = x Γ(x) for all x > 0; (b) Γ(1) = 1;√ (c) Γ(n) = (n − 1)! for all integral n ≥ 2; (d) Γ( 1/2) = π 1.3.5 . . . (2n − 1) √ (2n)! √ (e) Γ n + 1/2 = π = 2n π for integral n ≥ 1 2n 2 n! 2. Suppose X ∼ Gamma(m, α) and Y ∼ Gamma(n, α) and X and Y are independent. Find E[ Y /X ]. 3. By §8.4 on page 19, we know that if n > 1, the maximum of the Gamma(n, 1) density occurs at x = n − 1. Show that the maximum value of the density when n > 1 is approximately 1 √ 2π(n − 1) √ Hint: Stirling’s formula is n! ∼ nn e−n 2πn as n → ∞. 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §10 Page 23 4. Gamma densities are closed under convolution. Suppose X ∼ Gamma(n1 , α), Y ∼ Gamma(n2 , α) and X and Y are independent. Prove that X + Y has the Gamma(n1 + n2 , α) distribution. 5. Suppose X ∼ Gamma(m, α) and Y ∼ Gamma(n, α) and X and Y are independent. (a) Show that U = X + Y and V = X/(X + Y ) are independent.. (b) Show that U = X + Y and V = Y /X are independent. In both cases, find the densities of U and V . 6. Suppose X ∼ Gamma(n, α). Show that n 2n 2 P X≥ ≤ α e 7. Suppose X ∼ Gamma(m, α) and Y ∼ Gamma(n, α) and X and Y are independent. Show that n mv if v > 0; E[X|X + Y = v] = m+n 0 otherwise. 8. Suppose the random variable X ∼ exponential (Y ) where Y ∼ Gamma(n, α). (a) Find the density of X and E[X]. (b) Find the conditional density of Y given X = x. 9. Suppose X ∼ exponential(λ), and given X = x, the n random variables Y1 , . . . , Yn are i.i.d. exponential(x).4 Find the distribution of (X|Y1 , . . . , Yn ) and E[X|Y1 , . . . , Yn ]. 10. Suppose Gn ∼ Gamma(n, Pn Pn α) and Sn = α(Gn − n/α) where n > 1 is an integer. Hence Sn = α(Gn − n/α) = α(X − 1/α) = i i=1 i=1 Yi where each Xi ∼ exponential (α) and each Yi has mean 0 and variance 1. Check that the conditions of the local central limit theorem (§8.7 on page 21) are satisfied and hence verify the limiting result (8.7a) on page 21. 11. Length biased sampling in the Poisson process. Suppose {Xj }j≥1 is a sequence of i.i.d. random variables with the exponential (α) distribution. For n ≥ 1, let Sn = X1 + · · · + Xn and suppose t ∈ (0, ∞). Define the random variable K to be the unique integer with SK−1 < t ≤ SK ; equivalently K = min{j : Sj ≥ t}. (a) Find the density of XK . Find E[XK ] and compare with 1/α, the expectation of an exponential (α) distribution. Note that a longer interval has a higher chance of containing t! (b) Let Wt denote the waiting time to the next event after time t; hence Wt = SK − t. Find the distribution of Wt . 12. The generalized gamma distribution. (a) Show that the function f defined in equation(8.10a) is a density. (b) Suppose X ∼ GGamma(n, λ, b). Show that Y = X b ∼ Gamma(n, λ). (c) Suppose X ∼ GGamma(n, λ, b). Find the central moments E[X k ] for k = 1, 2, . . . . 10 The normal distribution 10.1 The density function. Suppose µ ∈ (−∞, ∞) and σ ∈ (0, ∞). Then the random variable X has the normal N (µ, σ 2 ) distribution if it has density (x − µ)2 1 for x ∈ R. (10.1a) fX (x) = √ exp − 2σ 2 σ 2π The normal density has the familiar “bell” shape. There are points of inflection at x = µ − σ and x = µ + σ—this means the f 00 (x) = 0 at these points and the curve changes from convex, when x < µ − σ, to concave and then to convex again when x > µ + σ. Definition(10.1a). 4 This means that f(Y1 ,...,Yn )|X (y1 , . . . , yn |x) = Πni=1 fYi |X (yi |x) = xn e−x(y1 +···+yn ) . Page 24 §10 Aug 1, 2018(17:54) A µ−σ Bayesian Time Series Analysis B µ µ+σ Figure(10.1a). The graph of the normal density. Points A and B are points of inflection. (wmf/normaldensity,72mm,54mm) To check that the function fX defined in equation(10.1a) is a density function: Clearly fX (x) ≥ 0 for all x ∈ R. Using the substitution t = (x − µ)/σ gives Z ∞ Z ∞ 1 (x − µ)2 √ exp − dx I= fX (x) dx = 2σ 2 ∞ ∞ σ 2π r Z ∞ Z ∞ 2 2 1 2 2 =√ exp − t /2 dt = √ exp − t /2 dt = J π 2π −∞ 2π 0 where Z ∞Z ∞ Z π/2 Z ∞ π 2 2 1 2 J = exp[− 2 (x + y )]dy dx = r exp[− 21 r2 ]dr dθ = 2 0 0 0 0 and hence r π J= 2 This shows that fX integrates to 1 and hence is a density function. 10.2 The distribution function, mean and variance. The standard normal distribution is the normal distribution N (0, 1); its distribution function is Z x 1 √ exp[− 12 t2 ] dt Φ(x) = 2π −∞ This function is widely tabulated. Note that: • Φ(−x) = 1 − Φ(x). See exercise 1 on page 26. • If X has the N (µ, σ 2 ) distribution, then for −∞ < a < b < ∞ we have Z (b−µ)/σ Z b 1 (x − µ)2 1 √ exp − P[a < X ≤ b] = dx = √ exp − t2/2 dt 2 2σ 2π (a−µ)/σ a σ 2π b−µ a−µ =Φ −Φ σ σ The mean of the N (µ, σ 2 ) distribution: Z ∞ 1 (x − µ)2 E[X] = [(x − µ) + µ] √ exp − dx = 0 + µ = µ 2σ 2 σ 2π −∞ because the function x 7−→ x exp[− x2/2] is odd. The variance of the N (µ, σ 2 ) distribution: use integration by parts as follows Z ∞ Z ∞ 1 (x − µ)2 σ2 2 var[X] = (x − µ) √ exp − dx = √ t2 exp[ −t2/2] dt 2 2σ σ 2π 2π −∞ −∞ Z ∞ 2 Z ∞ 2σ 2σ 2 =√ t t exp[ −t2/2] dt = √ exp[− t2/2] dt = σ 2 2π 0 2π 0 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §10 Page 25 10.3 The moment generating function and characteristic function. Suppose X ∼ N (µ, σ 2 ); then X = µ+σY where Y ∼ N (0, 1). For s ∈ R, the moment generating function of X is given by Z ∞ 1 2 1 sX sµ sσY sµ MX (s) = E[e ] = e E[e ] = e esσt √ e− 2 t dt 2π −∞ 2 Z ∞ Z ∞ sµ sµ e e t − 2σst (t − σs)2 σ 2 s2 dt = √ + dt =√ exp − exp − 2 2 2 2π −∞ 2π −∞ = exp sµ + 12 σ 2 s2 1 2 2 Similarly the characteristic function of X is E[eitX ] = eiµt− 2 σ t . Moments of a distribution can be obtained by expanding the moment generating function as a power series: E[X r ] is the coefficient of sr /r! in the expansion of the moment generating function. It is easy to find the moments about the mean of a normal distribution in this way: if X ∼ N (µ, σ 2 ) and Y = X − µ then E[esY ] = exp 21 σ 2 s2 which can be expanded in a power series of powers of s. Hence E (X − µ)2n+1 = E Y 2n+1 = 0 for n = 0, 1, . . . and (2n)!σ 2n E (X − µ)2n = E Y 2n = for n = 0, 1, . . . 2n n! For example, E[(X − µ)2 ] = σ 2 and E[(X − µ)4 ] = 3σ 4 . Similarly we can show that (see exercise 9 on page 27): 2n/2 σ n n+1 n for n = 0, 1, . . . . E |X − µ| = √ Γ 2 π There are available complicated expressions for E[X n ] and E[|X|n ]; for example, see [W INKELBAUER(2014)]. 10.4 Linear combination of independent normals. 2 Suppose Pn X1 , X2 , . . . , Xn are independent random variables with Xi ∼ N (µi , σi ) for i = 1, 2, . . . , n. Let T = i=1 di Xi where di ∈ R for i = 1, 2, . . . , n. Then ! n n X X 2 2 T ∼N di µi , di σi Proposition(10.4a). i=1 i=1 Proof. Using moment generating functions gives n n n Y Y Y 1 MT (s) = E[esT ] = E[esdi Xi ] = Mxi (sdi ) = exp sdi µi + s2 d2i σi2 2 i=1 i=1 i=1 ! n n X 1 X 2 2 = exp s di µi , + s2 di σi 2 i=1 i=1 Pn Pn 2 2 which is the mgf of N i=1 di µi , i=1 di σi . Corollary(10.4b). If X1 , . . . , Xn are i.i.d. N (µ, σ 2 ), then Xn has the normal distribution N (µ, σ 2 /n). 10.5 Sum of squares of independent N (0, 1) variables. Proposition(10.5a). Suppose X1 ,. . . , Xn are i.i.d. random variables with the N (0, 1) distribution. Let Z = X12 + · · · + Xn2 . Then Z ∼ χ2n . Proof. Consider n = 1. Now X1 has density 2 1 fX1 (x) = √ e−x /2 2π for x ∈ R. Then Z = X12 has density √ dx 1 z −1/2 e−z/2 1 for z > 0. fZ (z) = 2fX1 ( z) = 2 √ e−z/2 √ = 1/2 1 dz 2 z 2 Γ( /2) 2π Thus Z ∼ χ21 . We know that if X ∼ χ2m , Y ∼ χ2n and X and Y are independent, then X + Y ∼ χ2n+m . Hence Z ∼ χ2n in the general case. Page 26 §11 Aug 1, 2018(17:54) Bayesian Time Series Analysis 10.6 Characterizations of the normal distribution. There are many characterizations of the normal distribution5 —here are two of the most useful and interesting. Cramér’s theorem. Suppose X and Y are independent random variables such that Z = X + Y has a normal distribution. Then both X and Y have normal distributions—although one may have a degenerate distribution. Proposition(10.6a). Proof. See, for example, page 298 in [M ORAN(2003)]. Proposition(10.6b). The Skitovich-Darmois theorem. Suppose n ≥ 2 and X1 , . . . , Xn are independent random variables. Suppose a1 , . . . , an , b1 , . . . , bn are all in R and L1 = a1 X1 + · · · + an Xn L2 = b2 X1 + · · · + bn Xn If L1 and L2 are independent, then all random variables Xj with aj bj 6= 0 are normal. Proof. See, for example, page 89 in [K AGAN et al.(1973)]. 10.7 Summary. The normal distribution. • Density. X has the N (µ, σ 2 ) distribution iff it has the density 1 (x − µ)2 fX (x) = √ exp − for x ∈ R. 2σ 2 σ 2π • Moments: E[X] = µ and var[X] = σ 2 • The distribution function: P[X ≤ x] = Φ(x) which is tabulated. • The moment generating function: MX (t) = E[etX ] = exp[tµ + 12 t2 σ 2 ] • The characteristic function: φX (t) = E[eitX ] = exp[iµt − 12 σ 2 t2 ] • A linear combination of independent normals has a normal distribution. • The sum of squares of n independent N (0, 1) variables has the χ2n distribution. 11 Exercises (exs-normal.tex) 1. Show that Φ(−x) = 1 − Φ(x). 2. Suppose X ∼ N (µ, σ 2 ). Suppose further that P[X ≤ 140] = 0.3 and P[X ≤ 200] = 0.6. Find µ and σ 2 . 3. Suppose Y has the distribution function FY (y) with 1 Φ(y) if y < 0; FY (y) = 21 1 + Φ(y) if y ≥ 0. 2 2 Find E[Y n ] for n = 0, 1, . . . . 4. Suppose X is a random variable with density fX (x) = ce−Q(x) for all x ∈ R where Q(x) = ax2 − bx and a 6= 0. (a) Find any relations that must exist between a, b and c and show that X must have a normal density. (b) Find the mean and variance of X in terms of a and b. 5. (a) Suppose X and Y are i.i.d. random variables with the N (0, σ 2 ) distribution. Find the density of Z = X 2 + Y 2 . (b) Suppose X1 ,. . . , Xn are i.i.d. random variables with the N (0, σ 2 ) distribution. Let Z = X12 + · · · + Xn2 . Find the distribution of Z. 6. Suppose X ∼ N (µ, σ 2 ). Suppose further that, given X = x, the n random variables Y1 , . . . , Yn are i.i.d. N (x, σ12 ).6 Find the distribution of (X|Y1 , . . . , Yn ). 7. The half-normal distribution. Suppose X ∼ N (0, σ 2 ). (a) Find the density of |X|. (b) Find E[|X|]. 5 For example, see [M ATHAI & P EDERZOLI(1977)] and [PATEL & R EAD(1996)] 6 This means that f(Y1 ,...,Yn )|X (y1 , . . . , yn |x) = Πni=1 fYi |X (yi |x). 1 Univariate Continuous Distributions §12 Page 27 Aug 1, 2018(17:54) 8. The folded normal distribution. Suppose X ∼ N (µ, σ 2 ). Then |X| has the folded normal distribution, folded (µ, σ 2 ). Clearly the half-normal is the folded (0, σ 2 ) distribution. Suppose Y ∼ folded (µ, σ 2 ). (a) Find the density of Y . b) Find E[Y ] and var[Y ]. (c) Find the c.f. of Y . 9. Suppose X ∼ N (µ, σ 2 ). Show that 2n/2 σ n n+1 E |X − µ|n = √ Γ 2 π This also gives E[|X|n ] for the half-normal distribution. for n = 0, 1, . . . . 10. Suppose X and Y are i.i.d. N (0, 1). (a) Let Z1 = X + Y and Z2 = X − Y . Show that Z1 and Z2 are independent. 2 2 (b) By using the relation XY = X+Y − X−Y , find the characteristic function of Z = XY . 2 2 R ∞ (c) By using the relation E[eitXY ] = −∞ E[eityX ]fY (y) dy, find the characteristic function of Z = XY . (d) Now suppose X and Y are i.i.d. N (0, σ 2 ). Find the c.f. of Z = XY . (e) Now suppose X and Y are i.i.d. N (µ, σ 2 ). Find the c.f. of Z = XY . 11. Suppose X1 , X2 , X3 and X4 are i.i.d. N (0, 1). Find the c.f. of X1 X2 + X3 X4 and the c.f. of X1 X2 − X3 X4 . See also exercise 1.25(16) on page 53. 12. (a) Suppose b ∈ (0, ∞). Show that ∞ r b2 π −b 1 u2 + 2 e du = exp − 2 u 2 0 (b) Suppose a ∈ R with a 6= 0 and b ∈ R. Show that Z ∞ π 1/2 1 b2 2 2 exp − au + 2 du = e−|ab| 2 u 2a2 0 Z (11.12a) (11.12b) This result is used in exercise 1.25(22) on page 53. 12 The lognormal distribution 12.1 The definition. Definition(12.1a). The random variable X has the lognormal (µ, σ 2 ) distribution iff ln(X) ∼ N (µ, σ 2 ). We shall denote this distribution by logN (µ, σ 2 ). Hence: • if X ∼ logN (µ, σ 2 ) then ln(X) ∼ N (µ, σ 2 ); • if Z ∼ N (µ, σ 2 ) then eZ ∼ logN (µ, σ 2 ). 12.2 The density and distribution function. Suppose X ∼ logN (µ, σ 2 ) and let Z = ln(X). Then ln(x) − µ FX (x) = P[X ≤ x] = P[Z ≤ ln(x)] = Φ σ hence the distribution function of the logN (µ, σ 2 ) distribution is ln x − µ FX (x) = Φ for x > 0. σ Differentiating the distribution function gives the density: ln x − µ 1 1 (ln x − µ)2 fX (x) = φ =√ exp − for x > 0. σx σ 2σ 2 2πσx Z The density can also be obtained by transforming the Z ∼ N (µ, σ 2 ). Hence normal density. Now X = e where dx dz z 2 | dz | = e = x; hence fX (x) = fZ (z)| dx | = fZ (ln x) x where fZ is the density of N (µ, σ ). Page 28 §12 Aug 1, 2018(17:54) Bayesian Time Series Analysis µ = 0, σ = 0.25 µ = 0, σ = 0.5 µ = 0, σ = 1.0 1.5 1.0 0.5 0.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Figure(12.2a). The graph of the lognormal density for (µ = 0, σ = 0.25), (µ = 0, σ = 0.5) and (µ = 0, σ = 1). In all 3 cases, we have median = 1, mode < 1 and mean > 1—see exercise 4 on page 30. (wmf/lognormaldensity,72mm,54mm) 12.3 Moments. particular Suppose X ∼ logN (µ, σ 2 ). Then E[X n ] = E[enZ ] = exp nµ + 21 n2 σ 2 for any n ∈ R. In 1 E[X] = exp µ + σ 2 2 (12.3a) var[X] = E[X 2 ] − {E[X]}2 = e2µ+σ 2 2 eσ − 1 12.4 Other properties. • Suppose X1 , . . . , Xn are independent random variables with Xi ∼ logN (µi , σi2 ) for i = 1, . . . , n. Then ! n n n Y X X Xi = X1 · · · Xn ∼ logN µi , σi2 i=1 i=1 • Suppose X1 , . . . , Xn are i.i.d. with the i=1 logN (µ, σ 2 ) distribution. Then σ2 1/n (X1 · · · Xn ) ∼ logN µ, n • If X ∼ logN (µ, σ 2 ) , b ∈ R and c > 0 then cX b ∼ logN ln(c) + bµ, b2 σ 2 (12.4a) See exercises 5 and 2 below for the derivations of these results. 12.5 The multiplicative central limit theorem. Proposition(12.5a). Suppose X1 , . . . , Xn are i.i.d. positive random variables such that E[ ln(X) ] = µ and var[ ln(X) ] = σ 2 both exist and are finite. Then √ X1 · · · Xn 1/ n D −→ logN (0, σ 2 ) as n → ∞. enµ Proof. Let Yi = ln(Xi ) for i = 1, 2, . . . , n. Then " 1/√n # Pn (Yi − µ) D X1 · · · Xn ln = i=1√ −→ N (0, σ 2 ) enµ n D as n → ∞. 7 D Now if Xn −→X as n → ∞ then g(Xn ) −→g(X) as n → ∞ for any continuous function, g. Taking g(x) = ex proves the proposition. 7 The classical central limit theorem asserts that if X1 , X2 , . . . is a sequence of i.i.d. random variables with finite expectation µ and finite variance σ 2 and Sn = (X1 + · · · + Xn )/n, then √ D n (Sn − µ) −→ N (0, σ 2 ) as n → ∞. See page 357 in [B ILLINGSLEY(1995)]. 1 Univariate Continuous Distributions §12 Page 29 Aug 1, 2018(17:54) Using equation(12.4a) shows that if X ∼ logN (0, σ 2 ) then X 1/σ ∼ logN (0, 1). It follows that " # √1 X1 · · · Xn σ n lim P ≤ x = Φ(x) for all x > 0. n→∞ enµ Also, if we let √ √ √ √ X1 · · · Xn 1/ n 1/ n µ n 1/n µ 1/ n then (X · · · X ) = e W and (X · · · X ) = e W W = n n 1 1 enµ and hence by equation(12.4a), (X1 · · · Xn )1/n is asymptotically logN (µ, σ 2 /n). We can generalise proposition (12.5a) as follows: Proposition(12.5b). Suppose X1 , X2 , . . . is a sequence of independent positive random variables such that for all i = 1, 2, . . . E[ ln(Xi ) ] = µi , var[ ln(Xi ) ] = σi2 and E |ln(Xi ) − µi |3 = ωi3 all exist and are finite. For n = 1, 2, . . . , let n n n X X X 2 2 3 µ(n) = µi s(n) = σi ω(n) = ωi3 i=1 i=1 Suppose further that ω(n) /s(n) → 0 as n → ∞. Then X1 · · · Xn 1/s(n) D −→ logN (0, 1) eµ(n) i=1 as n → ∞. Proof. Let Yi = ln(Xi ) for i = 1, 2, . . . , n. Then 1/s(n) Pn (Yi − µi ) D X1 · · · Xn ln = i=1 −→N (0, 1) as n → ∞. eµ(n) s(n) Using the transformation g(x) = ex proves the proposition. 8 Also, if we let W = X1 · · · Xn eµ(n) 1/s(n) then X1 · · · Xn = eµ(n) W s(n) and hence by equation(12.4a), the random variable (X1 · · · Xn ) is asymptotically logN µ(n) , s2(n) . 12.6 Usage. The multiplicative central limit theorem suggests the following applications of the lognormal which can be verified by checking available data. • Grinding, where a whole is divided into a multiplicity of particles and the particle size is measured by volume, mass, surface area or length. • Distribution of farm size (which corresponds to a division of land)—where a 3-parameter lognormal can be used. The third parameter would be the smallest size entertained. • The size of many natural phenomena is due to the accumulation of many small percentage changes—leading to a lognormal distribution. 12.7 Summary. The lognormal distribution. • X ∼ logN (µ, σ 2 ) iff ln(X) ∼ N (µ, σ 2 ). 2 2 • Moments: if X ∼ logN (µ, σ 2 ) then E[X] = exp µ + 12 σ 2 and var[X] = e2µ+σ ( eσ − 1 ) • The product of independent lognormals is lognormal. • If X ∼ logN (µ, σ 2 ) , b ∈ R and c > 0 then cX b ∼ logN ln(c) + bµ, b2 σ 2 • The multiplicative central limit theorem. 8 Lyapunov central limit theorem with δ = 1. Suppose X1 , X2 , . . . is a sequence of independent random variables such that E[Xi ] = µi and var[Xi ] = σi2 are both finite. Let sn = σ12 + · · · + σn2 and suppose Pn n 1 X 3 i=1 (Xi − µi ) D lim E |Xi − µi | = 0, then −→N (0, 1) as n → ∞. n→∞ s3n sn i=1 See page 362 in [B ILLINGSLEY(1995)]. Page 30 §13 Aug 1, 2018(17:54) Bayesian Time Series Analysis 13 Exercises (exs-logN.tex) 1. An investor forecasts that the returns on an investment over the next 4 years will be as follows: for each of the first 2 years he estimates that £1 will grow to £(1 + I) where I is a random variable with E[I] = 0.08 and var[I] = 0.001; for each of the last 2 years he estimates that £1 will grow to £(1 + I) where I is a random variable with E[I] = 0.06 and var[I] = 0.002. Suppose he further assumes that the return Ij in year j is independent of the returns in all other years and that 1 + Ij has a lognormal distribution, for j = 1, 2, . . . , 4. Calculate the amount of money which must be invested at time t = 0 in order to ensure that there is a 95% chance that the accumulated value at time t = 4 is at least £5,000. 2. Suppose X ∼ logN (µ, σ 2 ). (a) Find the distribution of 1/X. (b) Suppose b ∈ R and c > 0. Find the distribution of cX b . 3. The geometric mean and geometric variance of a distribution. Suppose each xi P in the data set {x1 ., . . . , P xn } satisfies n xi > 0. Then the geometric mean of the data set is g = (x1 · · · xn )1/n or ln(g) = n1 i=1 ln(xi ) or ln(g) = n1 j fj ln(xj ) where fj is the frequency of the observation xj . This definition motivates the following. Suppose X is a random variable with X > 0. Then GMX , the geometric mean of X is defined by ln(GMX ) = R∞ ln(x)f X (x) dx = E[ln(X)]. 0 Similarly, we define the geometric variance, GVX , by ln(GVX ) = E (ln X − ln GMX )2 = var[ln(X)] √ and the geometric standard deviation by GSDX = GVX . Suppose X ∼ logN (µ, σ 2 ). Find GMX and GSDX . 4. Suppose X ∼ logN (µ, σ 2 ). (a) Find the median and mode and show that: mode < median < mean. (b) Find expressions for the lower and upper quartiles of X in terms of µ and σ. (c) Suppose αp denotes the p-quartile of X; this means that P[X ≤ αp ] = p. Prove that αp = eµ+σβp where βp is the p-quartile of the N (0, 1) distribution. 5. (a) Suppose X1 , .Q . . , Xn are independent random variables with Xi ∼ logN (µi , σi2 ) for i = 1, . . . , n. Find the n distribution of i=1 Xi = X1 · · · Xn . (b) Suppose X1 , . . . , Xn are i.i.d. with the logN (µ, σ 2 ) distribution. Find the distribution of (X1 · · · Xn )1/n . (c) Suppose X1 , . . . , Xn be independent random variables with Xi ∼ logN (µi , σi2 ) for i = 1, . . . , n. Suppose further that a1 , . . . , an are real constants. Show that n Y Xiai ∼ logN (mn , s2n ) i=1 for some mn and sn and find explicit expressions for mn and sn . 6. Suppose X1 and X2 are independent random variables with Xi ∼ logN (µi , σi2 ) for i = 1 and i = 2. Find the distribution of X1 /X2 . 7. Suppose X ∼ logN (µ, σ 2 ). Suppose further that E[X] = α and var[X] = β. Express µ and σ 2 in terms of α and β. 8. Suppose X ∼ logN (µ, σ 2 ) and k > 0. Show that 2 Φ ln(k)−µ−σ σ 1 2 E[X|X < k] = eµ+ 2 σ Φ ln(k)−µ σ and 1 E[X|X ≥ k] = eµ+ 2 σ 2 Φ µ+σ 2 −ln(k) σ 1−Φ ln(k)−µ σ 9. Suppose X ∼ logN (µ, σ 2 ). Then the j th moment distribution function of X is defined to be the function G : [0, ∞) → [0, 1] with Z x 1 G(x) = uj fX (u) du E[X j ] 0 (a) Show that G is the distribution function of the logN (µ + jσ 2 , σ 2 ) distribution. (b) Suppose γX denotes the Gini coefficient of X (also called the coefficient of mean difference of X). By definition Z ∞Z ∞ 1 γX = |u − v|fX (u)fX (v) dudv 2E[X] 0 0 | Hence γX = E|X−Y 2E[X] where X and Y are independent with the same distribution. Prove that γX = 2Φ( σ/√2) − 1 1 Univariate Continuous Distributions §14 Page 31 Aug 1, 2018(17:54) 14 The beta and arcsine distributions 14.1 The density and distribution function. Definition(14.1a). Suppose α > 0 and β > 0. Then the random variable X has a beta distribution, Beta(α, β), iff it has density f (x; α, β) = Γ(α + β) α−1 x (1 − x)β−1 Γ(α)Γ(β) for 0 < x < 1. (14.1a) Note: R∞ • Checking equation(14.1a) is a density function. Now Γ(α) = 0 uα−1 e−u du by definition. Hence Z ∞Z ∞ Γ(α)Γ(β) = uα−1 v β−1 e−u−v du dv 0 0 Now use the transformation x = u/(u + v) and y = u + v; hence u = xy and v = y(1 − x). Clearly 0 < x < 1 and 0 < y < ∞. Finally ∂(u,v) ∂(x,y) = y. Hence Z 1 Z ∞ Γ(α)Γ(β) = y α+β−1 xα−1 (1 − x)β−1 e−y dx dy x=0 y=0 1 Z = Γ(α + β) x=0 xα−1 (1 − x)β−1 dx • The beta function is defined by Z 1 Γ(α)Γ(β) for all α > 0 and β > 0. B(α, β) = tα−1 (1 − t)β−1 dt = Γ(α + β) 0 Properties of the beta and gamma functions can be found in most advanced calculus books. Recall that Γ(n) = (n − 1)! if n is a positive integer. • The distribution function of the beta distribution, Beta(α, β) is Z x Z x Ix (α, β) 1 tα−1 (1 − t)β−1 dt = for x ∈ (0, 1). F (x; α, β) = f (t; α, β) dt = B(α, β) B(α, β) 0 0 The integral, Ix (α, β), is called the incomplete beta function. R1 14.2 Moments. Using the fact that 0 xα−1 (1 − x)β−1 dx = B(α, β), it is easy to check that α (α + 1)α αβ E[X] = E[X 2 ] = and hence var[X] = α+β (α + β + 1)(α + β) (α + β)2 (α + β + 1) (14.2a) By differentiation, we get f 0 (x; α, β) = 0 implies x(2 − α − β) = α − 1. This has a root for x in [0, 1] if either (a) α + β > 2, α ≥ 1 and β ≥ 1 or (b) α + β < 2, α ≤ 1 and β ≤ 1. By checking the second derivative, we see α−1 mode[X] = if α + β > 2, α ≥ 1 and β ≥ 1. α+β−2 14.3 Shape of the density. The beta density can take many different shapes. α = 1/2 , β = 1/2 α = 5, β = 1 α = 1, β = 3 3.0 2.5 2.5 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0.0 0.0 0.0 0.2 0.4 0.6 α = 2, β = 2 α = 2, β = 5 α = 5, β = 2 3.0 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Figure(14.3a). Shape of the beta density for various values of the parameters. (wmf/betadensity1,wmf/betadensity2,72mm,54mm) Page 32 §14 Aug 1, 2018(17:54) Bayesian Time Series Analysis 14.4 Some distribution properties. • Suppose X ∼ Beta(α, β), then 1 − X ∼ Beta(β, α). • The Beta(1, 1) distribution is the same as the uniform distribution on (0, 1). • Suppose X ∼ Gamma(n1 , α) and Y ∼ Gamma(n2 , α). Suppose further that X and Y are independent. Then X/(X + Y ) ∼ Beta(n1 , n2 ). See exercise 5 on page 23. In particular, if X ∼ χ22k = Gamma(k, 1/2), Y ∼ χ22m = Gamma(m, 1/2) and X and Y are independent, then X/(X + Y ) ∼ Beta(k, m). • If X ∼ Beta(α, 1) then − ln(X) ∼ Exponential(α). See exercise 2 on page 33. 14.5 The beta prime distribution. Suppose α > 0 and β > 0. Then the random variable X is said to have the beta prime distribution, Beta (α, β), iff it has density xα−1 (1 + x)−α−β for x > 0. (14.5a) f (x) = B(α, β) Properties of the beta prime distribution are left to the exercises. Its relation to the beta distribution is given by the next two observations. Definition(14.5a). 0 • If X ∼ Beta(α, β), then X 1−X ∼ Beta0 (α, β). See exercise 3 on page 33. • If X ∼ Beta(α, β), then X1 − 1 ∼ Beta0 (β, α). This follows from the previous result: just use Y = 1 − X ∼ Beta(β, α) and 1/X − 1 = (1 − X)/X = Y /(1 − Y ). We shall see later (see §16.8 on page 37) that the beta prime distribution is just a multiple of the F -distribution. 14.6 The arcsine distribution on (0, 1). The arcsine distribution is the distribution Beta 1/2, 1/2 . Definition(14.6a). The random variable has the arcsine distribution iff X has density 1 for x ∈ (0, 1). fX (x) = √ π x(1 − x) The distribution function. Suppose X has the arcsine distribution; then √ 2 arcsin(2x − 1) 1 FX (x) = P[X ≤ x] = arcsin( x) = + for x ∈ [0, 1]. (14.6a) π π 2 Moments of the arcsine distribution. Using the results in equation(14.2a) on page 31 above and figure (14.6a) below we get 1 3 1 1 E[X] = E[X 2 ] = var[X] = mode(X) = {0, 1} median(X) = 2 8 8 2 Shape of the distribution. 3.0 2.5 2.0 1.5 1.0 0.5 0.0 0.2 0.4 0.6 0.8 1.0 Figure(14.6a). Plot of the arcsine density (wmf/arcsineDensity,72mm,54mm) 1 Univariate Continuous Distributions §15 Page 33 Aug 1, 2018(17:54) 14.7 The arcsine distribution on (a, b). Definition(14.7a). Suppose −∞ < a < b < ∞. Then the random variable X has the arcsine distribution on (a, b), denoted arcsin(a, b), iff X has density 1 fX (x) = √ for x ∈ (a, b). π (x − a)(b − x)) This means that the distribution defined in definition(14.6a) can also be described as the arcsin(0, 1) distribution. The distribution function is 2 F (x) = arcsin π r x−a b−a for a ≤ x ≤ b. If X ∼ arcsin(a, b) then kX + m ∼ arcsin(ka + m, bk + m). In particular, if X ∼ arcsin(0, 1) then (b − a)X + a ∼ arcsin(a, b); if X ∼ arcsin(a, b) then (X − a)/(b − a) ∼ arcsin(0, 1). The proof of this and further properties of the arcsine distribution can be found in exercises 6 and 7 on page 34. 14.8 Summary. The beta distribution. Suppose α > 0 and β > 0; thenX ∼ Beta(α, β) iff X has density f (x; α, β) = • Moments: Γ(α + β) α−1 x (1 − x)β−1 Γ(α)Γ(β) for 0 < x < 1. α αβ var[X] = 2 α+β (α + β) (α + β + 1) • Suppose X ∼ Beta(α, β), then 1 − X ∼ Beta(β, α). • The Beta(1, 1) distribution is the same as the uniform distribution on (0, 1). • Suppose X ∼ Gamma(n1 , α) and Y ∼ Gamma(n2 , α). Suppose further that X and Y are independent. Then X/(X + Y ) ∼ Beta(n1 , n2 ). nX • If X ∼ Beta(α, 1) then − ln(X) ∼ Exponential(α). If X ∼ Beta( m/2, n/2) then m(1−X) ∼ Fm,n . E[X] = The arcsine distribution. If X ∼ arcsin(0, 1) then X has density 1 fX (x) = √ for x ∈ (0, 1). π x(1 − x) • Moments: E[X] = 1/2 and var[X] = 1/8. The beta prime distribution. Suppose α > 0 and β > 0; then X ∼ Beta0 (α, β) iff the density is xα−1 (1 + x)−α−β B(α, β) ∼ Beta0 (α, β). f (x) = • If X ∼ Beta(α, β), then X 1−X for x > 0. 15 Exercises (exs-betaarcsine.tex.tex) The beta and beta prime distributions. 1. Suppose X ∼ Beta(α, β). Find an expression for E[X m ] for m = 1, 2, . . . . 2. Suppose X ∼ Beta(α, 1). Show that − ln(X) ∼ Exponential(α). 3. Suppose X ∼ Beta(α, β). Show that X 1−X ∼ Beta0 (α, β). Page 34 §16 Aug 1, 2018(17:54) Bayesian Time Series Analysis 4. The beta prime distribution. Suppose X has the beta prime distribution, Beta0 (α, β). (a) Show that E[X] = α/(β − 1) provided β > 1. (b) Show that var[X] = α(α + β − 1)/(β − 2)(β − 1)2 provided β > 2. (c) Show that the mode occurs at (α − 1)/(β + 1) if α ≥ 1 and at 0 otherwise. (d) 1/X ∼ Beta0 (β, α). (e) Suppose X ∼ Gamma(n1 , 1) and Y ∼ Gamma(n2 , 1). Suppose further that X and Y are independent. Show that X/Y ∼ Beta0 (n1 , n2 ). (f) Suppose X ∼ χ2n1 , Y ∼ χ2n2 and X and Y are independent. Show that X/Y ∼ Beta0 (n1 /2, n2 /2). The arcsine distribution. 5. Prove the equality in equation(14.6a) on page 32: √ 2 arcsin(2x − 1) 1 arcsin( x) = + π π 2 for x ∈ [0, 1] 6. (a) Suppose X ∼ arcsin(a, b). Prove that kX + m ∼ arcsin(ka + m, bk + m). (b) Suppose X ∼ arcsin(−1, 1). Prove that X 2 ∼ arcsin(0, 1). (c) Suppose X ∼ Uniform(−π, π). Prove that sin(X), sin(2X) and − cos(2X) all have the arcsin(−1, 1) distribution. 7. Suppose X ∼ Uniform(−π, π), Y ∼ Uniform(−π, π) and X and Y are independent. (a) Prove that sin(X + Y ) ∼ arcsin(−1, 1). (b) Prove that sin(X − Y ) ∼ arcsin(−1, 1). 16 The t, Cauchy and F distributions 16.1 Definition of the tn distribution. Definition(16.1a). freedom iff Suppose n ∈ (0, ∞). Then the random variable T has a t-distribution with n degrees of X T =p Y /n (16.1a) where X ∼ N(0, 1), Y ∼ χ2n , and X and Y are independent. Density: Finding the density is a routine calculation and is left to exercise 2 on page 37; that exercise shows that the density of the tn distribution is −(n+1)/2 −(n+1)/2 Γ n+1/2 1 t2 t2 √ √ fT (t) = 1+ = 1+ for t ∈ R. (16.1b) n n B( 1/2, n/2) n Γ n/2 πn We can check that the function fT defined in equation(16.1b) is a density for any n ∈ (0, ∞) as follows. Clearly fT (t) > 0; also, by using the transformation θ = 1/(1 + t2 /n), it follows that −(n+1)/2 −(n+1)/2 Z ∞ Z ∞ Z 1 √ t2 t2 1+ dt = 2 1+ dt = n θ(n−2)/2 (1 − θ)−1/2 dθ n n −∞ 0 0 √ = n B(1/2, n/2) Hence fT is a density. Now Y in equation(16.1a) can be replaced by Z12 + · · · + Zn2 where Z1 , Z2 , . . . , Zn are i.i.d. with the N (0, 1) distribution. Hence Y /n has variance 1 when n = 1, but its distribution becomes more clustered about the constant 1 as n becomes larger. Hence T has a larger variance then the normal when n = 1 but tends to the normal as n → ∞. Figure(16.1a) graphically demonstrates the density of the t-distribution is similar to the shape of the normal density but has heavier tails. See exercise 4 on page 37 for a mathematical proof that the distribution of T tends to the normal as n → ∞. 1 Univariate Continuous Distributions §16 Page 35 Aug 1, 2018(17:54) t2 t10 normal 0.4 0.3 0.2 0.1 0.0 −4 −2 0 2 4 Figure(16.1a). Plot of the t2 , t10 and standard normal densities. (wmf/tdensity,72mm,54mm) R∞ 16.2 Moments of the tn distribution. Suppose T ∼ tn . Now it is well-known that the integral 1 x1j dx converges if j > 1 and diverges if j ≤ 1. It follows that Z ∞ tr dt converges if r < n. (n + t2 )(n+1)/2 1 Hence the function tr fT (t) is integrable iff r < n. √ √ Provided n > 1, E[T ] exists and equals nE[X]E[1/ Y ] = 0. Provided n > 2, var(T ) = E[T 2 ] = nE[X 2 ] E[1/Y ] = n/(n − 2) by using equation(8.9a) on page 21 which gives E[1/Y ] = 1/(n − 2). 16.3 Linear transformation of the tn distribution. Suppose m ∈ R, s > 0 and V = m + sT . Then !−(n+1)/2 1 1 v−m 2 fT (t) √ 1+ fV (v) = dv = n s B( 1/2, n/2)s n dt Also E[V ] = m for n > 1 and var(V ) = s2 This is called a tn (m, s2 ) distribution. n n−2 for n > 2 (16.3a) (16.3b) 16.4 The Cauchy distribution. The standard Cauchy distribution, denoted Cauchy(1), is the same as the t1 distribution. Hence its density is: 1 for t ∈ R. γ1 (t) = π(1 + t2 ) More generally, suppose s > 0. Then the Cauchy distribution, Cauchy(s), is the same as the t1 (0, s2 ) distribution. Its density is s γs (t) = for t ∈ R. 2 π(s + t2 ) Clearly if X ∼ Cauchy(1) and s > 0 then sX ∼ Cauchy(s). 0.4 0.3 normal γ1 = t1 γ2 = t1 (0, 4) 0.2 0.1 0.0 −4 −2 0 2 4 Figure(16.4a). Plot of the normal, standard Cauchy and the Cauchy(2) = t(0, 4) densities. (wmf/cauchydensity,72mm,54mm) Page 36 §16 Aug 1, 2018(17:54) Bayesian Time Series Analysis 16.5 Elementary properties of the Cauchy distribution. • Moments. The expectation, variance and higher moments of the Cauchy distribution are not defined. • The distribution function. This is 1 Fs (t) = tan−1 π This is probably better written as 1 1 t −1 Fs (t) = + tan 2 π s t s where now tan−1 t/s ∈ (− π/2, π/2). • The characteristic function. Suppose the random variable T has the standard Cauchy distribution Cauchy(1). Then φT (t) = E[eitT ] = e−|t| and hence if W ∼ Cauchy(s), then W = sT and E[eitW ] = e−s|t| . Note. The characteristic function can be derived by using the calculus of residues, or by the following trick. Using integration by parts gives Z ∞ Z ∞ Z ∞ Z ∞ e−y cos(ty) dy = 1 − t e−y sin(ty) dy and e−y sin(ty) dy = t e−y cos(ty) dy 0 0 and hence 0 Z ∞ e−y cos(ty) dy = 0 0 1 1 + t2 Now the characteristic function of the bivariate exponential density f (x) = 12 e−|x| for x ∈ R is Z Z ∞ 1 ∞ 1 −|y| φ(t) = (cos(ty) + i sin(ty))e dy = e−y cos(ty) dy = 2 −∞ 1 + t2 0 Because this function is absolutely inegrable, we can use the inversion theorem to get Z ∞ −ity Z ∞ ity 1 −|t| 1 e 1 e e = dy = dy 2 2 2π −∞ 1 + y 2π −∞ 1 + y 2 as required. Further properties of the Cauchy distribution can be found in exercises 5 –12 starting on page 37. 16.6 Definition of the F distribution. Definition(16.6a). independent. Then Suppose m > 0 and n > 0. Suppose further that X ∼ χ2m , Y ∼ χ2n and X and Y are X/m has an Fm,n distribution. Y /n Finding the density of the Fm,n distribution is a routine calculation and is left to exercise 13 on page 38; that exercise shows that the density of the Fm,n distribution is Γ( m+n ) mm/2 nn/2 xm/2−1 for x ∈ (0, ∞). (16.6a) fF (x) = m 2 n Γ( 2 )Γ( 2 ) [mx + n](m+n)/2 F = F10,4 density F10,50 density 0.8 0.6 0.4 0.2 0.0 0 1 2 3 4 5 Figure(16.6a). Plot of the F10,4 and F10,50 densities. (wmf/fdensity,72mm,54mm) 1 Univariate Continuous Distributions §17 Page 37 Aug 1, 2018(17:54) 16.7 The connection between the t and F distributions. Recall the definition of a tn distribution: X Tn = p Y /n where X and Y are independent, X ∼ N (0, 1) and Y ∼ χ2n . Now X 2 ∼ χ21 ; hence X2 ∼ F1,n Tn2 = 2 Y /n (16.7a) Example(16.7a). Using knowledge of the F density and equation(16.7a), find the density of Tn . Solution. Let W = Tn2 ; hence W ∼ F1,n . Then √ √ 1 fW (w) = √ fTn (− w) + fTn ( w) 2 w But equation(16.7a) clearly implies the distribution of Tn is symmetric about 0; hence for w > 0 −(n+1)/2 Γ( n+1 Γ( n+1 nn/2 w−1 w2 2 ) 2 ) and fTn (w) = wfW (w ) = w 1 =√ 1+ n nπΓ( n2 ) Γ( 2 )Γ( n2 ) [w2 + n](n+1)/2 Finally, by symmetry, fTn (−w) = fTn (w). √ 1 fW (w) = √ fTn ( w) w 2 16.8 Properties of the F distribution. The following properties of the F -distribution are considered in exercises 16–20 on pages 38–39. • If X ∼ Fm,n then 1/X ∼ Fn,m . • If X1 ∼ Gamma(n1 , α1 ), X2 ∼ Gamma(n2 , α2 ) and X1 and X2 are independent then n2 α1 X1 ∼ F2n1 ,2n2 n1 α2 X2 • If X ∼ Beta( m/2, n/2) then • If X ∼ Fm,n then mX n+mX nX m(1−X) (16.8a) ∼ Fm,n . See exercise 17 on page 38. ∼ Beta( m/2, n/2). • If X ∼ Fm,n then mX/n ∼ Beta0 ( m/2, n/2). D • Suppose X ∼ Fm,n . Then mX −→χ2m as n → ∞. (16.8b) 16.9 Fisher’s z distribution. Definition(16.9a). If X ∼ Fm,n , then ln(X) ∼ FisherZ (m, n) 2 It follows that if X ∼ FisherZ (n, m) then e2X ∼ Fn,m . 17 Exercises (exs-tCauchyF.tex.tex) The t distribution. 1. Suppose X, Y1 , . . . , Yn are i.i.d. random variables with the N (0, σ 2 ) distribution. Find the distribution of X Z=q 2 (Y1 + · · · + Yn2 )/n 2. Using the definition of the tn distribution given in definition(16.1a) on page 34, show that the density of the tn distribution is given by equation(16.1b). 3. Suppose n > 0, s > 0 and α ∈ R. Show that Z ∞ −∞ 1 1 + (t − α)2 /s2 n/2 dt = sB 1 n−1 , 2 2 4. Prove that the limit as n → ∞ of the tn density given in equation(16.1b) is the standard normal density. The Cauchy distribution. 5. Suppose X and Y are i.i.d. with the N (0, σ 2 ) distribution. Find the distribution of: (a) W = X/Y ; (b) W = X/|Y |; (c) W = |X|/|Y |. Page 38 §17 Aug 1, 2018(17:54) Bayesian Time Series Analysis 6. (a) Suppose U has the uniform distribution U (− π/2, π/2). Show that tan(U ) ∼ Cauchy(1). (b) Suppose U has the uniform distribution U (−π, π). Show that tan(U ) ∼ Cauchy(1). 7. Suppose X1 , . . . , Xn are i.i.d. with density γs . (a) Show that Y = (X1 +···+Xn ) n also has the Cauchy(s) distribution. (b) Let Mn = median(X1 , . . . , Xn ). Show that Mn is asymptotically normal with mean 0 and a variance which tends to 0. 8. (a) Suppose X ∼ Cauchy(s). Find the density of 2X. (This shows that 2X has the same distribution as X1 + X2 where X1 and X2 are i.i.d. with the same distribution as X.) (b) Supppose U and V are i.i.d. with the Cauchy(s) distribution. Let X = aU + bV and Y = cU + dV . Find the distribution of X + Y . 9. Suppose X and Y are i.i.d. with the N (0, 1) distribution. Define R and Θ by R2 = X 2 + Y 2 and tan(Θ) = Y /X where R > 0 and Θ ∈ (−π, π). Show that R2 has the χ22 distribution, tan(Θ) has the Cauchy(1) distribution, and R and Θ are 2 independent. Show also that the density of R is re−r /2 for r > 0. 10. Suppose X has the Cauchy(1) distribution. Find the density of 2X 1 − X2 2 Hint: tan(2θ) = 2 tan(θ) 1 − tan (θ) . 11. From a point O, radioactive particles are directed at an absorbing line which is at a distance b from O. Suppose OP denotes the perpendicular from the point O to the absorbing line—and hence the length of OP is b. The direction of emission is measured by the angle Θ from the straight line OP . Suppose Θ is equally likely to be any direction in (− π/2, π/2). Formally, Θ ∼ U (− π/2, π/2). (a) Determine the density of X, the distance from P where the particle hits the absorbing line. (b) What is the density of 1/X ? 12. The symmetric Cauchy distribution in R2 . Define the function f : R2 → (0, ∞) by 1 f (x, y) = 2π(1 + x2 + y 2 )3/2 (a) Show that f is a density function. (b) Find the marginal densities. (c) Suppose (X, Y ) has the density f and we transform to polar coordinates: X = R cos Θ and Y = R sin Θ. Show that R and Θ are independent and find the distributions of R and Θ. The last question can be generalized to produce this density—in this case, the direction must be uniform over the surface of a hemisphere. The F distribution. 13. Using definition(18.6a) on page 41, show that the density of the Fm,n distribution is given by equation(16.6a) on page 36. 14. Suppose X and Y are i.i.d. N (0, σ 2 ). Find the density of Z where 2 2 Z = Y /X if X 6= 0; 0 if X = 0. 15. Suppose F has the Fm,n distribution where n > 2. Find E[F ]. 16. (a) Suppose X ∼ Fm,n . Show that 1/X ∼ Fn,m . (b) Suppose X1 ∼ Γ(n1 , α1 ), X2 ∼ Gamma(n2 , α2 ) and X1 and X2 are independent. Show that n2 α1 X1 ∼ F2n1 ,2n2 n1 α2 X2 17. Suppose X ∼ Beta( m/2, n/2). Show that 18. (a) Suppose X ∼ Fm,n . Show that nX m(1−X) mX n+mX ∼ Fm,n . ∼ Beta( m/2, n/2). (b) Suppose X ∼ F2α,2β where α > 0 and β > 0. Show that αX/β ∼ Beta0 (α, β). 19. Suppose W ∼ Fm,n . Show that mW/n ∼ Beta0 ( m/2, n/2). 1 Univariate Continuous Distributions Aug 1, 2018(17:54) D 20. Suppose W ∼ Fm,n . Show that mW −→χ2m as n → ∞. §18 Page 39 18 Non-central distributions 18.1 The non-central χ2 -distribution with 1 degree of freedom. We know that if Z ∼ N (0, 1), then Z 2 ∼ χ21 . Now suppose W = (Z + a)2 where Z ∼ N (0, 1) and a ∈ R. Then W is said to have a non-central χ21 distribution with non-centrality parameter a2 . Hence W ∼ Y 2 where Y ∼ N (a, 1). The moment generating function of W . Z ∞ 1 2 1 2 t(Z+a)2 E[e ]= √ et(z+a) e− 2 z dz 2π −∞ But t(z + a)2 − 1/2 z 2 = z 2 t + 2azt + a2 t − 1/2 z 2 = z 2 (t − 1/2) + 2azt + a2 t # " 2 2t 2t 2 t2 2a 1 − 2t 2a 4a 4azt 2at − =− − − = (t − 1/2) z 2 − z− 1 − 2t 1 − 2t 2 1 − 2t 1 − 2t (1 − 2t)2 # " 1 − 2t 2a2 t 2at 2 =− − z− 2 1 − 2t (1 − 2t)2 and hence, if α = 2at/(1 − 2t) and t < 1/2, 2 Z ∞ a t 1 (1 − 2t)(z − α)2 t(Z+a)2 √ E[e ] = exp exp − dz 1 − 2t 2 2π −∞ 2 a t −1/2 = (1 − 2t) exp 1 − 2t (18.1a) The density of W . For w > 0 we have √ √ √ √ φ( w − a) + φ(− w − a) φ( w − a) + φ( w + a) √ √ fW (w) = = 2 w 2 w √ √ 1 = √ exp − 1/2(w + a2 ) exp(a w) + exp(−a w) (18.1b) 2 2πw √ 1 =√ exp − 1/2(w + a2 ) cosh(a w) because cosh(x) = (ex + e−x )/2 for all x ∈ R. 2πw Using the standard expansion for cosh and the standard property of the Gamma function that Γ(n + 1/2) = √ (2n)! π/(4n n!) for all n = 0, 1, 2, . . . , gives ∞ ∞ X X 1 (a2 w)j 1 (a2 w/4)j √ fW (w) = √ exp − 1/2(w + a2 ) =√ exp − 1/2(w + a2 ) j!Γ(j + 1/2) π(2j)! 2w 2w j=0 j=0 √ 1/2 √ 1 a w 2 1 exp − /2(w + a ) I−1/2 (a w) =√ 2 2w 2 1/4 a √ 1 = exp − 1/2(w + a2 ) I−1/2 (a w) (18.1c) 2 w where, for all x > 0, ∞ x 2j− 1/2 X 1 I−1/2 (x) = j!Γ(j + 1/2) 2 j=0 is a modified Bessel function of the first kind. Note. The general definition of a modified Bessel function of the first kind is ∞ x ν X x2j Iν (x) = for all ν ∈ R and x ∈ C. 2 4j j!Γ(ν + j + 1) j=0 (18.1d) Page 40 §18 Aug 1, 2018(17:54) Bayesian Time Series Analysis 18.2 The non-central χ2 -distribution with n degrees of freedom. Suppose X1 ∼ N (µ1 , 1), X2 ∼ N (µ2 , 1), . . . , Xn ∼ N (µn , 1) are independent. Then n X W = Xj2 j=1 P has a non-central distribution with non-centrality parameter λ = nj=1 µ2j . Caution!!! Some authors define the non-centrality parameter to be λ/2. χ2n The moment generating function of W . By equation(18.1a) the moment generating function of X12 is 2 µ1 t 1 tX12 exp E[e ] = 1/2 1 − 2t (1 − 2t) Hence 1 λt tW E[e ] = exp for t < 1/2. 1 − 2t (1 − 2t)n/2 (18.2a) 18.3 The non-central χ2 -distribution with n degrees of freedom—the basic decomposition theorem. Proposition(18.3a). Suppose W has a non-central χ2n distribution with non-centrality parameter λ > 0. Then W has the same distribution as U + V where: U has a non-central χ21 distribution with non-centrality parameter λ; V has a χ2n−1 distribution; U and V are independent. p Pn Proof. Let µj = λ/n for j = 1, . . . , n; hence j=1 µ2j = λ. Pn We are given that W has a non-central χ2n distribution with non-centrality parameter λ > 0. Hence W ∼ j=1 Xj2 where X1 , . . . , Xn are independent with Xj ∼ N (µj , 1) for j = 1, . . . , n. √ √ Let e1 = (1, 0, . . . , 0), . . . , en = (0, . . . , 0, 1) denote the standard basis of Rn . Set b1 = (µ1 / λ, . . . , µn / λ). Then {b1 , e2 , . . . , en } form a basis of Rn . Use the Gram-Schmidt orthogonalization procedure to create the basis {b1 , . . . , bn }. Define B to be the n × n matrix with rows {b1 , . . . , bn }; then B is orthogonal. √ Suppose X = (X1 , . . . , Xn ) and set Y = BX. Then Y ∼ N (Bµ, BIBT = I) where µ = (µ1 , . . . , µn ) = λb1 . Hence n×1 n×1 √ Y1 ∼ N (bT1 µ = λ, 1) and Yj ∼ N (bTj µ = 0, 1) for j = 2, . . . , n and Y1 , . . . , Yn are independent. Also YT Y = XT X. Pn Finally, let U = Y12 and V = j=2 Yj2 . 18.4 The non-central χ2 -distribution with n degrees of freedom—the density function. We use proposition(18.3a). Now U has a non-central χ21 distribution with non-centrality parameter λ. Using equation(18.1b) gives h √ √ i 1 1 fU (u) = 3/2 1 √ e− 2 (u+λ) e λu + e− λu for u > 0. 2 Γ( /2) u Also, V ∼ χ2n−1 has density e−v/2 v (n−3)/2 for v > 0. 2(n−1)/2 Γ( (n−1)/2) Using independence of U and V gives √ i u−1/2 v (n−3)/2 e−(u+v)/2 e−λ/2 h √λu − λu f(U,V ) (u, v) = e + e 2(n+2)/2 Γ( 1/2)Γ( (n−1)/2) Now use the transformation X = U + V and Y = V . The Jacobian equals 1. Hence for y > 0 and x > y √ # 1/2 " √λ(x−y) e−x/2 e−λ/2 x(n−4)/2 y (n−3)/2 x e − e− λ(x−y) f(X,Y ) (x, y) = n/2 1 x−y 2 2 Γ( /2)Γ( (n−1)/2) x fV (v) = Now x x−y √ # 1/2 " √λ(x−y) 1/2 X ∞ e − e− λ(x−y) x λj (x − y)j = 2 x−y (2j)! j=0 = ∞ X (λx)j j=0 (2j)! 1− y j−1/2 x 1 Univariate Continuous Distributions §19 Page 41 Aug 1, 2018(17:54) and so we have ∞ y j−1/2 e−x/2 e−λ/2 x(n−4)/2 X (λx)j y (n−3)/2 1 − f(X,Y ) (x, y) = n/2 1 x 2 Γ( /2)Γ( (n−1)/2) j=0 (2j)! x for y > 0 and x > y. We need to integrate out y. By setting w = y/x we get Z x (n−3)/2 Z 1 y y j−1/2 1− dy = x w(n−3)/2 (1 − w)j−1/2 dw x x y=0 w=0 = xB( (n−1)/2, j + 1/2)) Γ( (n−1)/2)Γ(j + 1/2)) =x Γ( n/2 + j) and hence for x > 0 ∞ e−x/2 e−λ/2 x(n−2)/2 X (λx)j Γ(j + 1/2)) Γ( n/2) fX (x) = (2j)! Γ( 1/2) Γ( n/2 + j) 2n/2 Γ( n/2) = j=0 ∞ −x/2 −λ/2 (n−2)/2 X e e x 2n/2 j=0 (λx)j 4j j!Γ( n/2 + j) (18.4a) The expression for the modified Bessel function of the first kind in equation(18.1d) on page 39 gives ∞ √ (λx)(n−2)/4 X (λx)j n I 2 −1 ( λx) = 2n/2−1 j=0 4j j!Γ( n/2 + j) Hence an alternative expression for the density is x (n−2)/4 √ 1 fX (x) = e−x/2 e−λ/2 I n2 −1 ( λx) 2 λ This is the same as equation(18.1c) if we set n = 1 and λ = a2 . (18.4b) 18.5 The non-central t distribution. Suppose n ∈ (0, ∞) and µ ∈ R. Then the random variable T has a non-central tdistribution with n degrees of freedom and non-centrality parameter µ iff X +µ T =p (18.5a) Y /n where X ∼ N(0, 1), Y ∼ χ2n , and X and Y are independent. Definition(18.5a). 18.6 The non-central F distribution. Definition(18.6a). Suppose m > 0 and n > 0. Suppose further that X has a non-central χ2m distribution with non-centrality parameter λ, Y ∼ χ2n and X and Y are independent. Then X/m has a non-central Fm,n distribution with non-centrality parameter λ. F = Y /n 19 Exercises 1. Suppose X1 , . . . , Xn are independent random variables such that Xj has a non-central centrality parameter λj for j = 1, . . . , n. Find the distribution of Z = X1 + · · · + Xn . (exs-noncentral.tex.tex) χ2kj distribution with non- 2. Suppose W has a non-central χ2n distribution with non-centrality parameter λ. (a) Find the expectation of W . (b) Find the variance of W . 3. Suppose the random variable V has the Poisson distribution with mean λ/2. Suppose further that the distribution of W given V = j is the χ2k+2j distribution. Show that the distribution of W is the non-central χ2k with non-centrality parameter λ. Page 42 §20 Aug 1, 2018(17:54) Bayesian Time Series Analysis 20 The power and Pareto distributions There are many results about these distributions in the exercises. 20.1 The power distribution. Suppose a0 ∈ R, h > 0 and α > 0. Then the random variable X is said to have the power distribution Power(α, a0 , h) iff X has density α(x − a0 )α−1 for a0 < x < a0 + h. f (x) = hα The distribution function is (x − a0 )α F (x) = for a0 < x < a0 + h. hα The standard power distribution is Power(α, 0, 1); this has density f (x) = αxα−1 for 0 < x < 1 and distribution function F (x) = xα for 0 < x < 1. If X ∼ Power(α, 0, 1), the standard power distribution, then E[X] = α/(α+1), E[X 2 ] = α/(α + 2) and var[X] = α/(α + 1)2 (α + 2); see exercise 1 on page 45. Clearly, X ∼ Power(α, a0 , h) iff (X − a0 )/h ∼ Power(α, 0, 1). 4 α = 1/2 α=2 α=4 3 2 1 0 0.0 0.2 0.4 0.6 0.8 1.0 Figure(20.1a). The density of the standard power distribution for α = 1/2, α = 2 and α = 4. (wmf/powerdensity,72mm,54mm) 20.2 A characterization of the power distribution. Suppose X ∼ Power(α, 0, h); then αxα−1 xα f (x) = and F (x) = for x ∈ (0, h). hα hα Also, for all c ∈ (0, h) we have Z c αxα−1 α c E[X|X ≤ c] = x α dx = c = E[X] c α+1 h 0 The next proposition shows this result characterizes the power distribution (see [DALLAS(1976)]). Proposition(20.2a). Suppose X is a non-negative absolutely continuous random variable such that these exists h > 0 with P[X ≤ h] = 1. Suppose further that for all c ∈ (0, h) we have c E[X|X ≤ c] = E[X] h Then there exists α > 0 such that X ∼ Power(α, 0, h). Proof. Let f denote the density and F denote the distribution function of X. Then equation(20.2a) becomes Z c Z xf (x) c h dx = xf (x) dx F (c) h 0 0 Rh Let δ = h1 0 xf (x) dx. Then δ ∈ (0, 1) and Z c xf (x) dx = cF (c) δ for all c ∈ (0, h). (20.2a) (20.2b) 0 Differentiating with respect to c gives cf (c) = [F (c) + cf (c)] δ and hence F 0 (c) α δ = where α = > 0. F (c) c 1−δ Integrating gives ln F (c) = A + α ln(c) or F (c) = kcα . Using F (h) = 1 gives F (c) = cα /hα for c ∈ (0, h), as required. 1 Univariate Continuous Distributions §20 Page 43 Aug 1, 2018(17:54) The above result leads on to another characterization of the power distribution: Proposition(20.2b). Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with a non-negative absolutely continuous distribution function F and such that there exists h > 0 with F (h) = 1. Then Sn E X(n) = x = c with c independent of x (20.2c) X(n) iff there exists α > 0 such that F (x) = xα /hα for x ∈ (0, h). Proof. ⇒ Writing Sn = X(1) + · · · + X(n) in equation(20.2c) gives (c − 1)x = E X(1) + · · · + X(n−1) |X(n) = x It is easy to see that given X(n) = x, then the vector (X(1) , . . . , X(n−1) ) has the same distribution as the vector of n − 1 order statistics (Y(1) , . . . , Y(n−1) ) from the density f (y)/F (x) for 0 < y < x. Hence Y(1) + . . . + Y(n−1) = Y1 + · · · + Yn−1 and (c − 1)x = (n − 1)E[Y ] where Y has density f (y)/F (x) for y ∈ (0, x). (20.2d) Hence Z x c−1 yf (y) dy = xF (x) n −1 0 Because X(j) < X(n) for all j = 1, 2, . . . , n, equation(20.2c) implies c < nx x = n; also equation(20.2d) implies c > 1. c−1 ∈ (0, 1). So we have equation(20.2b) again and we must have F (x) = xα /hα for x ∈ (0, h). Hence δ = n−1 ⇐ See part (a) of exercise 6 on page 45. The next result is an easy consequence of the last one—it was originally announced in [S RIVASTAVA(1965)] but the proof here is due to [DALLAS(1976)]. Proposition(20.2c). Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with a non-negative absolutely continuous distribution function F and such that there exists h > 0 with F (h) = 1. Then Sn is independent of max{X1 , . . . , Xn } (20.2e) max{X1 , . . . , Xn } iff there exists α > 0 such that F (x) = xα /hα for x ∈ (0, h). Proof. ⇒ Clearly equation(20.2e) implies equation(20.2c). ⇐ See part (b) of exercise 6 on page 45. 20.3 The Pareto distribution. Suppose α > 0 and x0 > 0. Then the random variable X is said to have a Pareto distribution iff X has the distribution function x α 0 FX (x) = 1 − for x ≥ x0 . x It follows that X has density αxα0 fX (x) = α+1 for x ≥ x0 . x More generally, the random variable has the Pareto(α, a, x0 ) distribution iff αxα0 xα0 fX (x) = for x > a + x and F (x) = 1 − for x > a + x0 . 0 X (x − a)α+1 (x − a)α α for x > 1 and distribution The standard Pareto distribution is Pareto(α, 0, 1) which has density f (x) = xα+1 1 function F (x) = 1 − xα for x > 1. Clearly X ∼ Pareto(α, a, x0 ) iff (X − a)/x0 ∼ Pareto(α, 0, 1). It is important to note that if X ∼ Pareto(α, 0, x0 ) then 1/X ∼ Power(α, 0, 1/x) also, if X ∼ Power(α, 0, h) then 1/X ∼ Pareto(α, 0, 1/h). So results about one distribution can often be transformed into an equivalent result about the other. 3.0 x0 = 1, α = 1 x0 = 1, α = 2 x0 = 1, α = 3 2.5 2.0 1.5 1.0 0.5 0.0 0 1 2 3 4 5 Figure(20.3a). The Pareto distribution density for α = 1, α = 2 and α = 3 (all with x0 = 1). (wmf/Paretodensity,72mm,54mm) Page 44 §20 Aug 1, 2018(17:54) Bayesian Time Series Analysis The Pareto distribution has been used to model the distribution of incomes, the distribution of wealth, the sizes of human settlements, etc. 20.4 A characterization of the Pareto distribution. Suppose X ∼ Pareto(α, 0, x0 ). Suppose further that α > 1 so that the expectation is finite. We have αxα0 xα f (x) = α+1 and F (x) = 1 − α0 for x > x0 . x x Also, for all c > x0 we have, because the expectation is finite, Z ∞ Z ∞ αxα0 1 αc c α E[X|X > c] = x α+1 dx = αc dx = = E[X] α x [1 − F (c)] x α − 1 x 0 c c The next proposition shows this result characterizes the power distribution (see [DALLAS(1976)]). Proposition(20.4a). Suppose X is a non-negative absolutely continuous random variable with a finite expec- tation and such that these exists x0 > 0 with P[X > x0 ] = 1. Suppose further that for all c ∈ (0, h) we have c E[X|X > c] = E[X] (20.4a) x0 Then there exists α > 1 such that X ∼ Pareto(α, 0, x0 ). Proof. Let f denote the density and F denote the distribution function of X. Then equation(20.2a) becomes Z ∞ Z ∞ xf (x) c xf (x) dx dx = 1 − F (c) x0 x0 c R∞ Let δ = x10 x0 xf (x) dx. We are assuming E[X] is finite; hence δ ∈ (1, ∞) and Z ∞ xf (x) dx = c[1 − F (c)] δ for all c > x0 . (20.4b) c Differentiating with respect to c gives and hence −cf (c) = [1 − F (c) − cf (c)] δ cf (c)[δ − 1] = [1 − F (c)]δ α δ F 0 (c) = where α = > 1. 1 − F (c) c δ−1 α Integrating gives − ln[1 − F (c)] = A + ln(cα ) or 1 − F (c) = k/cα . Using F (x0 ) = 0 gives 1 − F (c) = xα 0 /c for c > x0 , as required. The above result leads on to another characterization of the Pareto distribution: Proposition(20.4b). Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with a non-negative absolutely con- tinuous distribution function F with a finite expectation and such that there exists x0 > 0 with P[X > x0 ] = 1. Then Sn E X(1) = x = c with c independent of x (20.4c) X(1) iff there exists α > 1 such that F (x) = 1 − xα0 /xα for x > x0 . Proof. ⇒ Writing Sn = X(1) + · · · + X(n) in equation(20.4c) gives (c − 1)x = E X(2) + · · · + X(n) |X(1) = x It is easy to see that given X(1) = x, then the vector (X(2) , . . . , X(n) ) has the same distribution as the vector of n − 1 order statistics (Y(1) , . . . , Y(n−1) ) from the density f (y)/[1 − F (x)] for y > x. Hence Y(1) + . . . + Y(n−1) = Y1 + · · · + Yn−1 and (c − 1)x = (n − 1)E[Y ] where Y has density f (y)/[1 − F (x)] for y > x0 . (20.4d) Hence Z ∞ c−1 yf (y) dy = x[1 − F (x)] n −1 x0 Because X(j) > X(1) for all j = 2, 3, . . . , n, equation(20.4c) implies c > nx x = n. Hence δ = α equation(20.4b) again and we must have F (x) = 1 − xα 0 /x for x ∈ (x0 , ∞). ⇐ See part (b) of exercise 20 on page 46. c−1 n−1 ∈ (1, ∞). So we have The next result is an easy consequence of the last one—it was originally announced in [S RIVASTAVA(1965)] but the proof here is due to [DALLAS(1976)]. 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §21 Page 45 Proposition(20.4c). Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with a non-negative absolutely con- tinuous distribution function F with finite expectation and such that there exists x0 > 0 with P[X > x0 ] = 1. Then Sn is independent of min{X1 , . . . , Xn } (20.4e) min{X1 , . . . , Xn } iff there exists α > 1 such that F (x) = 1 − xα0 /xα for x ∈ (x0 , ∞). Proof. ⇒ Clearly equation(20.4e) implies equation(20.4c). ⇐ See part (c) of exercise 20 on page 46. 21 Exercises (exs-powerPareto.tex) The power distribution. 1. Suppose X has the Power(α, a, h) distribution. Find E[X], E[X 2 ] and var[X]. 2. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Power(α, a, h) distribution. Find the distribution of Mn = max(X1 , . . . , Xn ). 3. Suppose U1 , U2 , . . . , Un are i.i.d. random variables with the Uniform(0, 1) distribution. (a) Find the distribution of Mn = max(U1 , . . . , Un ). 1/n (b) Find the distribution of Y = U1 . (c) Suppose X ∼ Power(α, a, h). Show that X ∼ a + hU 1/α where U ∼ Uniform(0, 1). Hence show that n X α n j n−j n E[X ] = h a for n = 1, 2, . . . . α+j j j=0 4. Transforming the power distribution to the exponential. Suppose X ∼ Power(α, 0, h). Let Y = − ln(X); equivalently Y = ln( 1/X ). Show that Y − ln( 1/h) ∼ exponential (α). 5. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the power distribution Power(α, 0, 1). By using the density of 2 Xk:n , find E[Xk:n ] and E[Xk:n ]. 6. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Power(α, 0, h) distribution. Sn (a) Show that E X(n) = x = c where c is independent of x. X(n) (b) Show that Sn max{X1 , . . . , Xn } is independent of max{X1 , . . . , Xn }. 7. Suppose r > 0 and X1 , X2 , . . . , Xn are i.i.d. random variables with a non-negative absolutely continuous distribution function F such that there exists h > 0 with F (h) = 1. (a) Show that for some i = 1, 2, . . . , n − 1 " # r X(i) E X(i+1) = x = c with c independent of x for x ∈ (0, h) r X(i+1) iff there exists α > 0 such that F (x) = xα /hα for x ∈ (0, h). (b) Assuming the expectation is finite, show that for some i = 1, 2, . . . , n − 1 " # r X(i+1) E r X(i+1) = x = c with c independent of x for x ∈ (0, h) X(i) iff there exists α > 0 such that F (x) = xα /hα for x ∈ (0, h). [DALLAS(1976)] 8. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the power distribution Power(α, 0, 1), which has distribution function F (x) = xα for 0 < x < 1 where α > 0. (a) Let X1:n X2:n X(n−1):n , W2 = , . . . , Wn−1 = , Wn = Xn:n X2:n X3:n Xn:n Prove that W1 , W2 , . . . , Wn are independent and find the distribution of Wk for k = 1, 2, . . . , n. W1 = 2 (b) Hence find E[Xk:n ] and E[Xk:n ]. Page 46 §21 Aug 1, 2018(17:54) Bayesian Time Series Analysis The Pareto distribution. 9. Relationship with the power distribution. Recall that if α > 0, then U ∼ Uniform(0, 1) iff Y = U 1/α ∼ Power(α, 0, 1). (a) Suppose α > 0. Show that U ∼ Uniform(0, 1) iff Y = U −1/α ∼ Pareto(α, 0, 1). (b) Suppose α > 0 and x0 > 0. Show that Y ∼ Pareto(α, a, x0 ) iff Y = a + x0 U −1/α where U ∼ Uniform(0, 1). (c) Show that X ∼ Power(α, 0, 1) iff 1/X ∼ Pareto(α, 0, 1). 10. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Pareto(α, a, x0 ) distribution. Find the distribution of Mn = min(X1 , X2 , . . . , Xn ). 11. Suppose X ∼ Pareto(α, 0, x0 ). (a) Find E[X] and var[X]. (b) Find the median and mode of X. (c) Find E[X n ] for n = 1, 2, . . . . (d) Find MX (t) = E[etX ], the moment generating function of X and φX (t) = E[eitX ], the characteristic function of X. 12. Show that the Pareto( 1/2, 0, 1) distribution provides an example of a distribution with E[1/X] finite but E[X] infinite. 13. Transforming the Pareto to the exponential. Suppose X ∼ Pareto(α, 0, x0 ). Let Y = ln(X). Show that Y has a shifted exponential distribution: Y − ln(x0 ) ∼ exponential (α). 14. Suppose X ∼ Pareto(α, 0, x0 ). Find the geometric mean of X and the Gini coefficient of X. The geometric mean of a distribution is defined in exercise 3 on page 30 and the Gini coefficient is defined in exercise 9 on page 30. 15. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Pareto distribution Pareto(α, 0, 1). By using the density of 2 Xk:n , find E[Xk:n ] and E[Xk:n ]. 16. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Pareto(α, 0, 1) distribution. (a) Let X2:n X(n−1):n Xn:n W1 = X1:n W2 = , . . . , Wn−1 = , Wn = X1:n X(n−2):n X(n−1):n Prove that W1 , W2 , . . . , Wn are independent and find the distribution of Wk for k = 1, 2, . . . , n. 2 (b) Hence find E[Xk:n ] and E[Xk:n ]. See exercise 15 for an alternative derivation. 17. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Power(α, 0, 1) distribution. Suppose also Y1 , Y2 , . . . , Yn are i.i.d. random variables with the Pareto(α, 0, 1) distribution. Show that for k = 1, 2, . . . , n 1 Xk:n and have the same distribution. Y(n−k+1):n 18. Suppose X and Y are i.i.d. random variables with the Pareto(α, 0, x0 ) distribution. Find the distribution function and density of Y /X . 19. Suppose X and Y are i.i.d. random variables with the Pareto(α, 0, x0 ) distribution. Let M = min(X, Y ). Prove that M and Y /X are independent. 20. Suppose X1 , X2 , . . . , Xn are i.i.d. random variables with the Pareto(α, 0, x0 ) distribution. (a) Prove that the random variable X1:n is independent of the random vector ( X2:n/X1:n , . . . , Sn X = x = c where c is independent of x. (b) Show that E (1) X(1) (c) Prove that X1:n is independent of Sn/X1:n = (X1 +···+Xn )/X1:n . Xn:n/X 1:n ). 21. Suppose r > 0 and X1 , X2 , . . . , Xn are i.i.d. random variables with a non-negative absolutely continuous distribution function F with finite expectation and such that there exists x0 > 0 with P[X > x0 ] = 1. (a) Show that for some i = 1,"2, . . . , n − 1 # r X(i+1) E r X(i) = x = c with c independent of x for x > x0 X(i) α iff there exists α > r/(n − i) such that F (x) = 1 − xα 0 /x for x > x0 . (b) Show that for some i = 1,"2, . . . , n − 1 # r X(i) E X(i) = x = c with c independent of x for x > x0 r X(i+1) α iff there exists α > r/(n − i) such that F (x) = 1 − xα 0 /x for x > x0 . [DALLAS(1976)] 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §22 Page 47 22. A characterization of the Pareto distribution. It is known that if X and Y are i.i.d. random variables with an absolutely continuous distribution and min(X, Y ) is independent of X − Y , then X and Y have an exponential distribution—see [C RAWFORD(1966)]. Now suppose X and Y are i.i.d. positive random variables with an absolutely continuous distribution and min(X, Y ) is independent of Y /X . Prove that X and Y have a Pareto distribution. Combining this result with exercise 19 gives the following characterization theorem: suppose X and Y are i.i.d. positive random variables with an absolutely continuous distribution; then min(X, Y ) is independent of Y /X if and only if X and Y have the Pareto distribution. 23. Another characterization of the Pareto distribution. Suppose X1 , X2 , . . . , Xn are i.i.d. absolutely continuous nonnegative random variables with density function f (x) and distribution function F (x). Suppose further that F (1) = 0 and f (x) > 0 for all x > 1 and 1 ≤ i < j ≤ n. Show that Xj:n/Xi:n is independent of Xi:n if and only if there exists β > 0 such that each Xi has the Pareto(β, 0, 1) distribution. Using the fact that X ∼ Pareto(α, 0, x0 ) iff X/x0 ∼ Pareto(α, 0, 1), it follows that if F (x0 ) = 0 and f (x) > 0 for all x > x0 where x0 > 0 then Xj:n/Xi:n is independent of Xi:n if and only if there exists β > 0 such that each Xi has the Pareto(β, 0, x0 ) distribution. 22 Size, shape and related characterization theorems 22.1 Size and shape: the definitions. The results in this section on size and shape are from [M OSSIMAN(1970)] and [JAMES(1979)]. Definition(22.1a). The function g : (0, ∞)n → (0, ∞) is an n-dimensional size variable iff g(ax) = ag(x) for all a > 0 and all x ∈ (0, ∞)n . Suppose g : (0, ∞)n → (0, ∞) is an n-dimensional size variable. Then the function → (0, ∞)n is the shape function associated with g iff x z(x) = for all x ∈ (0, ∞)n . g(x) Definition(22.1b). z: (0, ∞)n 22.2 Size and shape: standard examples. • The standard size function. This is g(x1 , . . . , xn ) = x1 + · · · + xn . The associated shape function is the function z : (0, ∞)n → (0, ∞)n with ! xn x1 , . . . , Pn z(x1 , . . . , xn ) = Pn j=1 xj j=1 xj • Dimension 1 size. This is g(x1 , . . . , xn ) = x1 . The associated shape function is x2 xn z(x1 , . . . , xn ) = 1, , . . . , x1 x1 • Dimension 2 size. This is g(x1 , . . . , xn ) = x2 . The associated shape function is x1 xn z(x1 , . . . , xn ) = , 1, . . . , x2 x2 • Volume. This is g(x1 , . . . , xn ) = (x21 + · · · + x2n )1/2 . The associated shape function is x1 xn z(x1 , . . . , xn ) = , . . . , (x21 + · · · + x2n )1/2 (x21 + · · · + x2n )1/2 • The maximum. This is g(x1 , . . . , xn ) = max{x1 , . . . , xn }. The associated shape function is xn x1 ,..., z(x1 , . . . , xn ) = max{x1 , . . . , xn } max{x1 , . . . , xn } • Root n size. This is g(x1 , . . . , xn ) = (x1 x2 . . . xn )1/n . The associated shape function is x1 xn z(x1 , . . . , xn ) = ,..., (x1 x2 . . . xn )1/n (x1 x2 . . . xn )1/n Page 48 §22 Aug 1, 2018(17:54) Bayesian Time Series Analysis 22.3 Size and shape: the fundamental result. We shall show that: • if any one shape function z(X) is independent of the size variable g(X), then every shape function is independent of g(X); • if two size variables g(X) and g ∗ (X) are both independent of the same shape function z(X), then g(X)/g ∗ (X) is almost surely constant. First a specific example9 of this second result: Example(22.3a). Suppose X = (X1 , X2 , X3 ) ∼ logN (µ, Σ) distribution. By definition, this means that if Y1 = ln(X1 ), Y2 = ln(X2 ) and Y3 = ln(X3 ), then (Y1 , Y2 , Y3 ) ∼ N (µ, Σ). Define the three size functions: √ g3 (x) = (x1 x2 x3 )1/3 g1 (x) = x1 g2 (x) = x2 x3 and let z1 , z2 and z3 denote the corresponding shape functions. Suppose g1 (X) is independent of z1 (X). (a) Show that var[Y1 ] = cov[Y1 , Y2 ] = cov[Y1 , Y3 ]. (b) Show that g1 (X) is independent of z2 (X). (c) Show that g1 (X) is independent of g2 (X)/g1 (X). Now suppose g3 (X) is also independent of z1 (X). (d) Show that cov[Y1 , S] = cov[Y2 , S] = cov[Y3 , S] where S = Y1 + Y2 + Y3 . (e) Show that var[Y2 ] + cov[Y2 , Y3 ] = var[Y3 ] + cov[Y2 , Y3 ] = 2var[Y1 ]. (f) Show that var[2Y1 − Y2 − Y3 ] = 0 and hence g1 (X)/g3 (X) is constant almost everywhere. Solution. We are given X1 is independent of 1, X2 /X1 , X3 /X1 . Taking logs shows that Y1 is independent of (Y2 − Y1 , Y3 − Y1 ) and these are normal. Hence cov[Y1 , Y2 − Y1 ] = cov[Y1 , Y3 − Y1 ] = 0 and hence (a). (b) follows because Y1 is independent of (Y1 − 21 Y2 − 21 Y3 , 12 Y2 − 21 Y3 , 12 Y3 − 21 Y2 ). (c) Now cov[Y1 , 12 (Y2 +Y3 )−Y1 ) = 21 cov[Y1 , Y2 ]+ 12 cov[Y1 , Y3 ]−var[Y1 ] = 0. By normality, ln (g1 (X)) = Y1 is independent of log (g2 (X)) − ln (g1 (X)). Because the exponential function is one-one, we have (c). (d) The assumption g3 (X) is independent of z1 (X) implies, by taking logs, that S is independent of (Y2 − Y1 , Y3 − Y1 ) and these are normal. Hence (d). (e) Expanding cov[Y1 , S] and using part (a) shows that cov[Y1 , S] = 3var[Y1 ]. Similarly, expanding cov[Y2 , S] shows that var[Y2 ] + cov[Y2 , Y3 ] + cov[Y1 , Y2 ] = cov[Y2 , S] = cov[Y1 , S] = 3var[Y1 ]. Hence (e). (f) Now var[2Y1 − Y2 − Y3 ] = 4var[Y1 ] − 4cov[Y 1 , Y2 ] − 4cov[Y1 , Y3 ] + var[Y2 ] + var[Y3 ] + 2cov[Y2 , Y3 ] = 0. Hence var[Y1 − 31 S] = 0; hence var[ln g1 (X)/g3 (X) ] = 0. Hence (f). Now for the general result: Suppose g : (0, ∞)n → (0, ∞) is an n-dimensional size variable and z ∗ : (0, ∞)n → is any shape function. Suppose further that X is a random vector such that z ∗ (X) is non-degenerate and independent of g(X). Then (a) for any other shape function z1 : (0, ∞)n → (0, ∞)n , z1 (X) is independent of g(X); (b) if g2 : (0, ∞)n → (0, ∞) is another size variable such that z ∗ (X) is independent of both g2 (X) and g(X), then g2 (X) is constant almost everywhere. g(X) Proposition(22.3b). (0, ∞)n Proof. Let g ∗ and g1 denote the size variables which lead to the shape functions z ∗ and z1 . Hence x x z ∗ (x) = ∗ and z1 (x) = for all x ∈ (0, ∞)n . g (x) g1 (x) For all x ∈ (0, ∞)n we have x g1 (x) g1 z ∗ (x) = g1 = ∗ by using g1 (ax) = ag1 (x). g ∗ (x) g (x) Hence for all x ∈ (0, ∞)n z ∗ (x) x g ∗ (x) z1 z ∗ (x) = = ∗ × = z1 (x) (22.3a) ∗ g1 ( z (x) ) g (x) g1 (x) Equation(22.3a) shows that z1 (X) is a function of z ∗ (X); also, we are given that z ∗ (X) is independent of g(X). Hence z1 (X) is independent of g(X). This proves (a). (b) Because of part (a), we can assume X X z2 (X) = is independent of g(X) and z(X) = is independent of g2 (X) g2 (X) g(X) 9 Understanding this example is not necessary for the rest of the section. The example makes use of the definition of the multivariate normal and the fact that normals are independent if the covariance is zero. See 2-§3.6 on page 63. 1 Univariate Continuous Distributions Aug 1, 2018(17:54) §22 Page 49 Applying g to the first and g2 to the second gives g(X) g2 (X) g ( z2 (X) ) = is independent of g(X) and g2 ( z(X) ) = is independent of g2 (X) g2 (X) g(X) and hence g2 (X) is independent of both g2 (X) and g(X). g(X) Hence result by part (b) of exercise 8 on page 7. 22.4 A characterization of the gamma distribution. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables. Suppose g ∗ (0, ∞)n → (0, ∞) denotes the size variable n X ∗ g (x) = xj Proposition(22.4a). j=1 Then there exists a shape vector z(X) which is independent of g ∗ (X) iff there exist α > 0, k1 > 0, . . . , kn > 0 such that Xj ∼ Gamma(kj , α) for j = 1, 2, . . . , n. Proof. P ⇐ Now g ∗ (X) = nj=1 Xj ∼ Gamma(k1 + · · · + kn , α). Proposition(8.8b) implies Xj /(X1 + · · · + Xn ) is independent of g ∗ (X) = X1 + · · · + Xn for j = 1, 2, . . . , n. Hence the standard shape vector X1 X2 Xn z(X) = , ,..., X1 + · · · + Xn X1 + · · · + Xn X1 + · · · + Xn is independent of g ∗ (X). Hence all shape vectors are independent of g ∗ (X). ⇒ By proposition(22.3b), if there exists one shape vector which is independent of g ∗ (X), then all shape vectors are independent of g ∗ (X). Hence Xj /(X1 + · · · + Xn ) is independent of g ∗ (X) = X1 + · · · + Xn for j = 1, 2, . . . , n. Hence by proposition(8.8b), there exists α > 0 and kj > 0 such that Xj ∼ Gamma(kj , α) for j = 1, 2, . . . , n. This result implies many others. For example, suppose X1 , X2 , . . . , Xn are independent random variables with Xj ∼ Gamma(kj , α). Then every shape vector is independent of X1 + X2 + · · · + Xn ; in particular, Xj X1 + X2 + · · · + Xn is independent of max{X1 , X2 , . . . , Xn } and X1 + X2 + · · · + Xn X1 + X2 + · · · + Xn is independent of max{X1 , X2 , . . . , Xn } 22.5 A characterization of the Pareto distribution. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables. Suppose g ∗ (0, ∞)n → (0, ∞) denotes the size variable g ∗ (x) = min{x1 , . . . , xn } Then there exists a shape vector z(X) which is independent of g ∗ (X) iff there exists x0 > 0 and αj > 0 for j = 1, 2, . . . , n such that Xj ∼ Pareto(αj , 0, x0 ) for j = 1, 2, . . . , n. Proposition(22.5a). Proof. ⇐ Let Y1 = ln(X1 ) and Y2 = ln(X2 ). Then Y1 −ln(x0 ) ∼ exponential (α1 ) and Y2 −ln(x0 ) ∼ exponential (α2 ) and Y1 and Y2 are independent. By exercise 5 on page 51, we know that if Y1 − a ∼ exponential (λ1 ) and Y2 − a ∼ exponential (λ2 ) and Y1 and Y2 are independent, then min{Y1 , Y2 } is independent of Y2 − Y1 . This establishes U = min{Y1 , Y2 } is independent of V = Y2 − Y1 . But (Y3 , . . . , Yn ) is independent of U and V . Hence min{Y1 , . . . , Yn } is independent of Y2 −Y1 . Similarly min{Y1 , . . . , Yn } is independent of Yj −Y1 for j = 2, . . . , n. Hence min{Y1 , . . . , Yn } is independent of the vector (Y2 − Y1 , Y3 − Y1 , . . . , Yn − Y1 ). And hence g ∗ (X) = min{X1 , . . . , Xn ) is independent of the shape vector (1, X2/X1 , . . . , Xn/X1 ) as required. ⇒ Suppose n = 2. Using the shape vector (1, x2/x1 ) implies that we are given min{X1 , X2 } is independent of X2 /X1 . Taking logs shows that min{Y1 , Y2 } is independent of Y2 − Y1 where Y1 = ln(X1 ) and Y2 = ln(X2 ). It is known (see [C RAWFORD(1966)]) that if Y1 and Y2 are independent random variables with an absolutely continuous distribution and min(Y1 , Y2 ) is independent of Y2 − Y1 , then there exist a ∈ R, α1 > 0 and α2 > 0 such that Y1 − a ∼ exponential (α1 ) and Y2 − a ∼ exponential (α2 ). Hence fY1 (y1 ) = α1 e−α1 (y1 −a) for y1 > a and fY2 (y2 ) = α2 e−α2 (y2 −a) for y2 > a. Hence X1 = eY1 ∼ Pareto(α1 , 0, x0 = ea ) and X2 = eY2 ∼ Pareto(α2 , 0, x0 = ea ) where x0 > 0. Page 50 §23 Aug 1, 2018(17:54) Bayesian Time Series Analysis For n > 2 we are given that Xj is independent of min{X1 , . . . , Xn } for j = 1, 2, . . . , n. min{X1 , . . . , Xn } But min{X1 , . . . , Xn } = min{Xj , Zj } where Zj = min{Xi : i 6= j} Hence for some x0j > 0, λj > 0 and λ∗j > 0, Xj ∼ Pareto(λj , 0, x0j ) and Zj ∼ Pareto(λ∗j , 0, x0j ). Because Zj = min{Xi : i 6= j} we must have x0j ≤ x0i for j 6= i. It follows that all x0j are equal. Hence result. 22.6 A characterization of the power distribution. Because the inverse of a Pareto random variable has the power distribution, the previous proposition can be transformed into a result about the power distribution. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables. Suppose → (0, ∞) denotes the size variable g ∗ (x) = max{x1 , . . . , xn } Then there exists a shape vector z(X) which is independent of g ∗ (X) iff there exists x0 > 0 and αj > 0 for j = 1, 2, . . . , n such that Xj ∼ Power(αj , 0, x0 ) for j = 1, 2, . . . , n. Proposition(22.6a). g ∗ (0, ∞)n Proof. ⇐ Let Y1 = ln( 1/X1 ) and Y2 = ln( 1/X2 ). By exercise 4 on page 45, Y1 − ln( 1/x0 ) ∼ exponential (α1 ) and Y2 − ln( 1/x0 ) ∼ exponential (α2 ) and Y1 and Y2 are independent. By exercise 5 on page 51, we know that if Y1 − a ∼ exponential (λ1 ) and Y2 − a ∼ exponential (λ2 ) and Y1 and Y2 are independent, then min{Y1 , Y2 } is independent of Y1 − Y2 . This establishes U = min{Y1 , Y2 } is independent of V = Y1 − Y2 . But (Y3 , . . . , Yn ) is independent of U and V . Hence min{Y1 , . . . , Yn } is independent of Y1 −Y2 . Similarly min{Y1 , . . . , Yn } is independent of Y1 −Yj for j = 2, . . . , n. Hence min{Y1 , . . . , Yn } is independent of the vector (Y1 − Y2 , Y1 − Y3 , . . . , Y1 − Yn ). And hence g ∗ (X) = max{X1 , . . . , Xn ) is independent of the shape vector (1, X2/X1 , . . . , Xn/X1 ) as required. ⇒ Suppose n = 2. Using the shape vector (1, x2/x1 ) implies that we are given max{X1 , X2 } is independent of X2 /X1 . Set Y1 = ln( 1/X1 ) and Y2 = ln( 1/X2 ). Hence min{Y1 , Y2 } is independent of Y2 − Y1 . It is known (see [C RAWFORD(1966)]) that if Y1 and Y2 are independent random variables with an absolutely continuous distribution and min(Y1 , Y2 ) is independent of Y2 − Y1 , then there exist a ∈ R, α1 > 0 and α2 > 0 such that Y1 − a ∼ exponential (α1 ) and Y2 − a ∼ exponential (α2 ). Hence fY1 (y1 ) = α1 e−α1 (y1 −a) for y1 > a and fY2 (y2 ) = α2 e−α2 (y2 −a) for y2 > a. Hence X1 = e−Y1 ∼ Power(α1 , 0, h = e−a ) and X2 = e−Y2 ∼ Power(α2 , 0, h = e−a ) where h > 0. For n > 2 we are given that Xj is independent of max{X1 , . . . , Xn } for j = 1, 2, . . . , n. max{X1 , . . . , Xn } But max{X1 , . . . , Xn } = max{Xj , Zj } where Zj = max{Xi : i 6= j} Hence for some hj > 0, λj > 0 and λ∗j > 0, Xj ∼ Power(λj , 0, hj ) and Zj ∼ Power(λ∗j , 0, hj ). Because Zj = max{Xi : i 6= j} we must have hj ≥ hi for j 6= i. It follows that all hj are equal. Hence result. 23 Exercises (exs-sizeshape.tex.tex) 1. Suppose X = (X1 , X2 ) is a 2-dimensional random vector with X1 = aX2 where a ∈ R. Show that if z : (0, ∞)2 → (0, ∞)2 is any shape function, then z(X) is constant almost everywhere. 2. Suppose X = (X1 , X2 ) is a 2-dimensional random vector with the distribution given in the following table: X2 1 2 3 6 1 1 0 0 /4 0 1/4 2 0 0 0 X1 1/4 3 0 0 0 1/4 6 0 0 0 √ Define the size variables g1 (x) = x1 x2 and g2 (x) = x1 + x2 . (a) Suppose z is any shape function: (0, ∞)2 → (0, ∞)2 . Show that z(X) cannot be almost surely constant. Also, show that z(X) is independent of both g1 (X) and g2 (X). (b) Find the distribution of g1 (X)/g2 (X). 1 Univariate Continuous Distributions §24 Page 51 Aug 1, 2018(17:54) 3. A characterization of the generalized gamma distribution. Prove the following result. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables. Suppose g ∗ (0, ∞)n → (0, ∞) denotes the size variable 1/b n X g ∗ (x) = xbj where b > 0. j=1 Then there exists a shape vector z(X) which is independent of g ∗ (X) iff there exist α > 0, k1 > 0, . . . , kn > 0 such that Xj ∼ GGamma(kj , α, b) for j = 1, 2, . . . , n. Hint: use the result that X ∼ GGamma(n, λ, b) iff X b ∼ Γ(n, λ) and proposition(22.4a). 4. Suppose X1 ∼ exponential (λ1 ), Y ∼ exponential (λ2 ) and X and Y are independent. (a) Find P[X1 < X2 ]. (b) By using the lack of memory property of the exponential distribution, find the distribution of X1 − X2 . (c) By using the usual convolution formula for densities, find the denisty of X1 − X2 . 5. Suppose Y1 − a ∼ exponential (λ1 ) and Y2 − a ∼ exponential (λ2 ) and Y1 and Y2 are independent. Show that U = min{Y1 , Y2 } is independent of V = Y2 − Y1 . 6. A generalization of proposition(22.5a) on page 49. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables. and θ1 , θ2 , . . . , θn are positive constants. Suppose g ∗ (0, ∞)n → (0, ∞) denotes the size variable x1 xn ∗ , ..., g (x) = min θ1 θn Prove there exists a shape vector z(X) which is independent of g ∗ (X) iff there exists x0 > 0 and αj > 0 for j = 1, 2, . . . , n such that Xj ∼ Pareto(αj , 0, θj x0 ) for j = 1, 2, . . . , n. [JAMES(1979)] 7. A generalization of proposition(22.6a) on page 50. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables and θ1 , θ2 , . . . , θn are positive constants. Suppose g ∗ (0, ∞)n → (0, ∞) denotes the size variable x1 xn ∗ g (x) = max , ..., θ1 θn Prove there exists a shape vector z(X) which is independent of g ∗ (X) iff there exists x0 > 0 and αj > 0 for j = 1, 2, . . . , n such that Xj ∼ Power(αj , 0, θj x0 ) for j = 1, 2, . . . , n. [JAMES(1979)] 24 Laplace, Rayleigh and Weibull distributions There are many results about these distributions in the exercises. 24.1 The Laplace or bilateral exponential distribution. Suppose µ ∈ R and α > 0. Then the random variable X is said to have the Laplace(µ, α) distribution iff X has the density α fX (x) = e−α|x−µ| for x ∈ R. 2 Clearly if X ∼ Laplace(µ, α), then X − µ ∼ Laplace(0, α). As figure(24.1a) shows, the density consists of two equal exponential densities spliced back to back. 3.0 2.5 2.0 1.5 1.0 0.5 0.0 −2 −1 0 1 2 Figure(24.1a). The bilateral exponential density for µ = 0 and α = 6. (wmf/bilateralExponential,72mm,54mm) Page 52 §25 Aug 1, 2018(17:54) Bayesian Time Series Analysis 24.2 The Rayleigh distribution. Suppose σ > 0. Then the random variable X is said to have the Rayleigh (σ) distribution if X has the density r 2 2 fR (r) = 2 e−r /2σ for r > 0. σ The Rayleigh distribution is used to model the lifetime of various items and the magnitude of vectors—see exercise 18 on page 53. 1.2 σ = 0.5 σ = 1.5 σ=4 1.0 0.8 0.6 0.4 0.2 0.0 0 2 4 6 8 10 Figure(24.2a). The Rayleigh distribution density for σ = 0.5, σ = 1.5 and σ = 4. (wmf/Rayleighdensity,72mm,54mm) 24.3 The Weibull distribution. Suppose β > 0 and γ > 0. Then the random variable X is said to have the Weibull (β, γ) distribution iff X has the density βxβ−1 −(x/γ)β fX (x) = e for x > 0. γβ The distribution function is F (x) = 1 − exp(−xβ /γ β ) for x > 0. The Weibull distribution is frequently used to model failure times. The density can take several shapes as figure(24.3a) illustrates. 2.5 β = 1/2, γ = 1 2.0 β = 5, γ = 1 1.5 β = 1.5, γ = 1 1.0 β = 1, γ = 1 0.5 0.0 0.0 0.5 1.0 1.5 2.0 2.5 Figure(24.3a). The Weibull density for β = 1/2, β = 1, β = 1.5 and β = 5; all with γ = 1. (wmf/weibulldensity,72mm,54mm) 25 Exercises (exs-other.tex) The Laplace or bilateral exponential distribution. 8. (a) Suppose α > 0. Suppose further that X has the exponential density αe−αx for x > 0 and Y has the exponential density αeαx for x < 0 and X and Y are independent. Find the density of X + Y . (b) Suppose α > 0 and the random variables X and Y have the exponential density αe−αx for x > 0. Suppose further that X and Y are independent. Find the density of X − Y . 9. Suppose X has the Laplace(µ, α) distribution. (a) Find the expectation, median, mode and variance of X. (b) Find the distribution function of X. (c) Find the moment generating function of X. 1 Univariate Continuous Distributions §25 Page 53 Aug 1, 2018(17:54) 10. Suppose X has the Laplace(0, α) distribution. Find the density of |X|. 11. Suppose X ∼ exponential (λ), Y ∼ exponential (µ) and X and Y are independent. Find the density of Z = X − Y . 12. (a) Suppose X ∼ Laplace(µ, α); suppose further that k > 0 and b ∈ R. Show that kX + b ∼ Laplace(kµ + b, α/k). (b) Suppose X ∼ Laplace(µ, α). Show that α(X − µ) ∼ Laplace(0, 1). Pn (c) Suppose X1 , . . . , Xn are i.i.d. with the Laplace(µ, α) distribution. Show that 2α i=1 |Xi − µ| ∼ χ22n . 13. Suppose X and Y are i.i.d. Laplace(µ, α). Show that |X − µ| ∼ F2,2 |Y − µ| 14. Suppose X and Y are i.i.d. uniform U (0, 1). Show that ln( X/Y ) ∼ Laplace(0, 1). 15. Suppose X and Y are independent random variables with X ∼ exponential (α) and Y ∼ Bernoulli ( 1/2). Show that X(2Y − 1) ∼ Laplace(0, α). 16. Suppose X1 , X2 , X3 and X4 are i.i.d. N (0, 1). (a) Show that X1 X2 − X3 X4 ∼ Laplace(0, 1). (b) Show that X1 X2 + X3 X4 ∼ Laplace(0, 1). (See also exercise 1.11(11) on page 27.) 17. Exponential scale mixture of normals. Suppose X and Y are independent random variables with X ∼ exponential (1) √ √ and Y ∼ N (0, 1). Show that Y 2X ∼ Laplace(0, 1) and µ + Y 2X/α ∼ Laplace(µ, α). Note. Provide two solutions: one using characteristic functions and one using densities. The Rayleigh distribution. √ 18. (a) Suppose X and Y are i.i.d. with the N (0, σ 2 ) distribution. Define R and Θ by R = X 2 + Y 2 , X = R cos Θ and Y = R sin Θ with Θ ∈ [0, 2π). Prove that R and Θ are independent and find the density of R and Θ. (b) Suppose R has the Rayleigh (σ) distribution and Θ has the uniform U (−π, π) distribution. Show that X = R cos Θ and Y = R sin Θ are i.i.d. N (0, σ 2 ). 19. Suppose R has the Rayleigh (σ) distribution: fR (r) = (a) (b) (c) (d) r −r2 /2σ2 e σ2 for r > 0. Find the distribution function of R. Find an expression for E[Rn ] for n = 1, 2, . . . . Find E[R] and var[R]. Find the mode and median of R. √ 20. Suppose U has the uniform distribution U (0, 1) and X = σ −2 ln U . Show that X has the Rayleigh (σ) distribution. 21. (a) Suppose R has the Rayleigh (σ) distribution. Find the distribution of R2 . (b) Suppose R has the Rayleigh (1) distribution. Show that the distribution of R2 is χ22 . Pn (c) Suppose R1 , . . . , Rn are i.i.d. with the Rayleigh (σ) distribution. Show that Y = i=1 Ri2 has the Gamma(n, 1/2σ 2 ) distribution. √ (d) Suppose X has the exponential (λ) = Gamma(1, λ) distribution. Find the distribution of Y = X. 22. Suppose X ∼ Rayleigh (s) where s > 0, and Y |X ∼ N (µ, σ = X). Show that Y ∼ Laplace(µ, 1/s). The Weibull distribution. β 23. Suppose X has the Weibull (β, γ) distribution; hence fX (x) = βxβ−1 e−(x/γ) /γ β for x > 0 where β is the shape and γ is the scale. (a) Show that the Weibull (1, γ) distribution is the same as the exponential (1/γ) distribution. √ (b) Show that the Weibull (2, γ) distribution is the same as the Rayleigh (γ/ 2) distribution. 24. Suppose X has the exponential (1) distribution; hence fX (x) = e−x for x > 0. Suppose further that β > 0 and γ > 0 and W = γX 1/β . Find the density of W . This is the Weibull (β, γ) distribution; β is called the shape and γ is called the scale. β 25. Suppose X has the Weibull (β, γ) distribution; hence fX (x) = βxβ−1 e−(x/γ) /γ β for x > 0. (a) Suppose α > 0; find the distribution of Y = αX. (b) Find an expression for E[X n ] for n = 1, 2, . . . . (c) Find the mean, variance, median and mode of X. (d) Find E[et ln(X) ], the moment generating function of ln(X). Page 54 §25 Aug 1, 2018(17:54) Bayesian Time Series Analysis 26. Suppose X has the Weibull (β, γ) distribution. (a) Find hX (x) = f (x)/[1 − F (x)], the hazard function of X. (b) Check that if β < 1 then hX decreases as x increases; if β = 1 then hX is constant; and if β > 1 then hX increases as x increases. 27. Suppose U has the uniform U (0, 1) distribution. Show that X = γ(− ln U )1/β has the Weibull (β, γ) distribution. CHAPTER 2 Multivariate Continuous Distributions 1 General results 1.1 The mean and variance matrices. If X1 . X = .. n×1 Xn is a random vector, then, provided the univariate expectations E[X1 ], . . . , E[Xn ] exist, we define E[X1 ] . E[X] = .. n×1 E[Xn ] Provided the second moments E[X12 ], . . . , E[Xn2 ] are finite, the variance matrix or covariance matrix of X is the n × n matrix var[X] = E[(X − µ)(X − µ)T ] where µ = E[X]. n×n The (i, j) entry in the variance matrix is cov[Xi , Xj ]. In particular, the ith diagonal entry is var[Xi ]. Clearly: • the variance matrix is symmetric; • if X1 , . . . , Xn are i.i.d. with variance σ 2 , then var[X] = σ 2 I. • var[X] = E[XXT ] − µµT . (1.1a) We shall often omit stating “when the second moments are finite” when it is obviously needed. 1.2 Linear transformations. If Y = X + a then var[Y] = var[X]. More generally, if Y = A + BX where A is m × 1 and B is m × n, then µY = A + BµX and var[Y] = E (Y − µY )(Y − µY )T = E B(X − µX )(X − µX )T )BT = B var[X] BT In particular, if a = (a1 , . . . , an ) is a 1 × n vector, then aX = var[aX] = a var[X] aT = Pn n X n X i=1 ai Xi is a random variable and ai aj cov[Xi , Xj ] i=1 j=1 Example(1.2a). Suppose the random vector X = (X1 , X2 , X3 ) has variance matrix " # 6 2 3 var[X] = 2 4 0 3 0 2 Let Y1 = X1 + X2 and Y2 = X1 + X2 − X3 . Find the variance matrix of Y = (Y1 , Y2 ). Solution. Now Y = AX where 1 1 0 A= 1 1 −1 Hence 14 11 var[Y] = var[AX] = Avar[X]AT = 11 10 Bayesian Time Series Analysis by R.J. Reed Aug 1, 2018(17:54) §1 Page 55 Page 56 §1 Aug 1, 2018(17:54) Bayesian Time Series Analysis 1.3 Positive definiteness of the variance matrix. Suppose X is an n × 1 random vector. Then for any 1 × n vector c we have cT var[X]c = var[cT X] ≥ 0 Hence var[X] is positive semi-definite (also called non-negative definite). Proposition(1.3a). Suppose X is a random vector with finite second moments and such that no element of X is a linear combination of the other elements. Then var[X] is a symmetric positive definite matrix. Proof. No element of X is a linear combination of the other elements; hence. if a is a 1 × n vector with aT X constant then we must have a = 0. Now suppose var[cT X] = 0; this implies cT X is constant and hence c = 0. We have shown that var[cT X] = cT var[X]c = 0 implies c = 0 and hence var[X] must be positive definite. Example(1.3b). Consider the random vector Z = (X, Y, X + Y )T where µX = E[X], µY = E[Y ] and ρ = corr[X, Y ]. Show that var[Z] is not positive definite. Solution. Let a = (1, 1, −1). Then a var[Z] aT = var[aZ] = var[0] = 0. 1.4 The square root of a variance matrix. Suppose C is real, symmetric and positive definite. Because C is real and symmetric, we can write C = LDLT where D = diag(d1 , . . . , dp ) is diagonal and L is orthogonal1 . Because C is also positive definite, we have d1 > 0, . . . , dp > 0. Hence we can write C = (LD1/2 LT )(LD1/2 LT ) = QQ where Q is symmetric and nonsingular. Now suppose X is a random vector with finite second moments and such that no element of X is a linear combination of the other elements; then var[X] is a real symmetric positive definite matrix. Hence var[X] = QQ and if Y = Q−1 X then var(Y) = Q−1 var[X] (Q−1 )T = I. We have shown that if X is a random vector with finite second moments and such that no element of X is a linear combination of the other elements, then there exists a linear transformation of X to independent variables. 1.5 The covariance between two random vectors. Definition(1.5a). If X is an m × 1 random vector with finite second moments and Y is an n × 1 random vector with finite second moments, then ΣX,Y = cov[X, Y] is the m × n matrix with (i, j) entry equal to cov[Xi , Yj ]. Because the (i, j) entry of cov[X, Y] equals cov[Xi , Yj ], it follows that cov[X, Y] = E[(X − µX )(Y − µy )T ] = E[XYT ] − µX µTY Also: • because cov[Xi , Yj ] = cov[Yj , Xi ], it follows that cov[X, Y] = cov[Y, X]T ; • if n = m, the covariance matrix cov[X, Y] is symmetric; • cov[X, X] = var[X]. 1.6 The correlation matrix. Suppose√the n × 1 random vector X has the variance matrix Σ. Let D be the n × n diagonal matrix with diagonal equal to diag(Σ). Then the correlation matrix of X is given by corr[X] = D−1 ΣD−1 Clearly, the (i, j) element of corr[X] is corr(X pi , Xj ). Also, corr[X] is the variance matrix of the random vector Z = (Z1 , . . . , Zn ) where Zj = (Xj − E[Xj ])/ var(Xj ). Conversely, given corr[X] we need the vector of standard deviations in order to determine the variance matrix. In fact, var[X] = D corr[X] D where D is the diagonal matrix with entries stdev(X1 ), . . . , stdev(Xn ) . 1.7 Quadratic forms. Results about quadratic forms are important in regression and the analysis of variance. Definition(1.7a). Suppose A is a real n × n symmetric matrix. Then the quadratic form of A is the function qA : Rn → Rn with qA (x) = n X n X ajk xj xk = xT Ax j=1 k=1 We are interested in situations where x is a random vector. Here are some examples of quadratic forms. Example(1.7b). Suppose A = I, the identity matrix. Then qA (x) = XT AX = Pn k=1 Xk2 . Example(1.7c). Suppose A = 1, the n × n matrix with every entry equal to 1. Then qA (x) = XT AX = 1 Pn k=1 2 Xk . Orthogonal means that LT L = I and hence L−1 = LT . The numbers d1 , . . . , dn are the eigenvalues of C; see page 39 of [R AO(1973)]. 2 Multivariate Continuous Distributions §1 Page 57 Aug 1, 2018(17:54) If A and B are both real n×n symmetric matrices and a, b ∈ R, then aA+bB can be used to create a new quadratic form: qaA+bB (X) = XT (aA + bB)X = aqA (X) + bqB (X) 2 Pn Pn Example(1.7d). Suppose A = I and B = 1. Then qaI+b1 (X) = a k=1 Xk2 + b k=1 Xk . 2 P n Pn Pn = k=1 (Xk − X)2 . In particular qI−1/n (X) = k=1 Xk2 − n1 k=1 Xk Example(1.7e). Suppose 0 1 0 1 0 1 0 1 0 A= ... ... ... 0 0 0 0 0 0 Then qA (X) = XT AX = 2 Pn−2 k=1 Xk Xk+2 . ··· ··· ··· .. . ··· ··· 0 0 0 .. . 0 0 0 .. . 0 1 1 0 1.8 Mean of a quadratic form. Proposition(1.8a). Suppose X is an n × 1 random vector with E[X] = µ and var[X] = Σ. Suppose A is a real n × n symmetric matrix. Then the quadratic form in A has expectation E[XT AX] = trace(AΣ) + µT Aµ (1.8a) Proof. Now and hence XT AX = (X − µ)T A(X − µ) + µT AX + XT Aµ − µT Aµ n n X X T ajk cov[Xj , Xk ] + µT Aµ E[X AX] = E (X − µ) A(X − µ) + µ Aµ = T T j=1 k=1 Pn Now that (j, `) element AΣ is k=1 aP By definition, trace(AΣ) is the sum of the diagonal elements of AΣ. jk σk` .P Pn ofP n n n Hence trace(AΣ) = j=1 k=1 ajk σkj = j=1 k=1 ajk σjk because σkj = σjk . Hence result. Note that the second term in equation(1.8a) is xT Ax evaluated at x = µ; this simplifies some derivations. We now apply this result to some of the examples above: Example(1.8b). Suppose A = I, the identity matrix. Then qA (X) = XT AX = n n n X X X E Xj2 = σj2 + µ2j j=1 j=1 Pn j=1 Xj2 . and equation(1.8a) gives j=1 Example(1.8c). Suppose A = 1, the n × n matrix with every entry equal to 1. Then qA (X) = XT AX = and equation(1.8a) gives 2 2 n n X n n X X X E Xj = σj,k + µj j=1 j=1 k=1 P n j=1 Xj 2 j=1 Example(1.8d). Suppose X1 , . . . , Xn are i.i.d. random variables with mean µ and variance σ 2 . Consider the quadratic Pn form Q = k=1 (Xk − X)2 . Then Q = XT AX where A = I − n1 1. Now var[X] = σ 2 I; hence equation(1.8a) gives n X 1 E[Q] = σ 2 trace(I − 1) + xT Ax = (n − 1)σ 2 + (xk − x)2 = (n − 1)σ 2 n x=µ x=µ k=1 Pn Hence if S 2 = k=1 (Xk − X)2 /(n − 1), then E[S 2 ] = σ 2 . Example(1.8e). with mean µ and variance σ 2 . Pn−1 Suppose X21 , . . . , Xn are2i.i.d. random variables 2 Let Q = k=1 (Xk − Xk+1 ) = (X1 − X2 ) + (X2 − X3 ) + · · · + (Xn−1 − Xn )2 . Then Q = XT AX where 1 −1 0 · · · 0 0 −1 2 −1 · · · 0 0 0 −1 2 · · · 0 0 A= . . . . . .. .. .. .. .. .. . 0 0 0 · · · 2 −1 0 0 0 ··· −1 1 Page 58 §1 Aug 1, 2018(17:54) An alternative expression is Q = 2 Pn Xk2 − X12 − Xn2 − 2 Bayesian Time Series Analysis Pn−1 Xk Xk+1. Now var[X] = Σ = σ I; hence Equation(1.8a) gives E[Q] = σ trace(A) + Q k=1 2 k=1 2 (X1 ,...,Xn )=(µ,...,µ) = σ 2 (2n − 2). 1.9 Variance of a quadratic form. The result is complicated! Proposition(1.9a). Suppose X1 , . . . , Xn are independent random variables with E[Xj ] = 0 for j = 1, . . . , n. Suppose all the random variables have the same finite second and fourth moments; we shall use the following notation: E[Xj2 ] = σ 2 and E[Xj4 ] = µ4 Suppose A is an n × n symmetric matrix with entries ai,j and d is the n × 1 column vector with entries (a1,1 , . . . , an,n ) = diag(A). Then var[XT AX] = (µ4 − 3σ 4 )dT d + 2σ 4 trace(A2 ) 2 Proof. Now var[XT AX] = E (XT AX)2 −PE[XT AX] . n Because E[X] = 0 we have E[XT AX] = σ 2 j=1 ajj = σ 2 trace(A); see also exercise 7 on page 59. P P P P If c = XT A and Z = XXT , then (XT AX)2 = XT AXXT AX = cZcT = j k cj ck Zj,k = j k cj ck Xj Xk . The j th entry P n in c = XT A is i=1 Xi ai,j . Hence XXXX (XT AX)2 = aij ak` Xi Xj Xk X` i Using independence of the X’s gives ( k ` if i = j = k = `; if i = j and k = `, or i = k and j = `, or i = ` and j = k. otherwise. X X X X E[ (XT AX)2 = µ4 a2i,i + σ 4 ai,i aj,j + a2i,j + ai,j aj,i E[Xi Xj Xk X` ] = Hence j µ4 σ4 0 i Now X ai,i aj,j = n X ai,i i=1 i,j i6=j X a2i,j = i,j i6=j X ai,j aj,i = i=1 i,j i6=j i,j i6=j n X aj,j = j=1 j6=i n X i=1 j=1 j6=i n X n X X i,j i6=j a2i,j = n X n X i=1 j=1 i,j i6=j i,j i6=j ai,i trace(A) − ai,i = [trace(A)]2 − dT d a2i,j − dT d = trace(A2 ) − dT d a2i,j = trace(A2 ) − dT d and hence E[ (XT AX)2 = (µ4 − 3σ 4 )dT d + σ 4 trace(A)2 + 2trace(A2 ) and hence the result. Example(1.9b). Suppose X1 , . . . , Xn are i.i.d. random variables with the N (0, σ 2 ) distribution and A is a symmetric n × n matrix. By §10.3 on page 25, we know that E[Xj4 ] = 3σ 4 . Hence var[XT AX] = 2σ 4 trace(A2 ) We can generalize this result to non-zero means as follows: Proposition(1.9c). Suppose X1 , . . . , Xn are independent random variables with E[Xj ] = µj for j = 1, . . . , n. Suppose all the random variables have the same finite second, third and fourth moments about the mean; we shall use the following notation: E[X] = µ and E[(Xj − µj )2 ] = σ 2 E[(Xj − µj )3 ] = µ3 E[(Xj − µj )4 ] = µ4 Suppose A is an n × n symmetric matrix with entries ai,j and d is the n × 1 column vector with entries (a1,1 , . . . , an,n ) = diag(A). Then var[XT AX] = (µ4 − 3σ 4 )dT d + 2σ 4 trace(A2 ) + 4σ 2 µT A2 µ + 4µ3 µT Ad (1.9a) Proof. See exercise 10 on page 60. 2 Multivariate Continuous Distributions §2 Page 59 Aug 1, 2018(17:54) 2 Exercises (exs-multiv.tex) 1. Suppose X is an m × 1 random vector and Y is an n × 1 random vector. Suppose further that all second moments are finite. Finally, suppose a is an m × 1 vector and b is an n × 1 vector. Show that m X n X cov[aT X, bT Y] = aT cov[X, Y]b = ai bj cov[Xi , Yj ] i=1 j=1 2. Further properties of the covariance matrix. Suppose X is an m × 1 random vector and Y is an n × 1 random vector. Suppose further that all second moments are finite. (a) Show that for any m × 1 vector b and any n × 1 vector d we have cov[X + b, Y + d] = cov[X, Y] (b) Show that for any ` × m matrix A and any p × n matrix B we have cov[AX, BY] = Acov[X, Y]BT (c) Suppose a, b, c and d ∈ R; suppose further that V is an m × 1 random vector and W is an n × 1 random vector with finite second moments. cov[aX + bV, cY + dW] = ac cov[X, Y] + ad cov[X, W] + bc cov[V, Y] + bd cov[V, W] Both sides are m × n matrices. (d) Suppose a and b ∈ R and both X and V are m × 1 random vectors. Show that var[aX + bV] = a2 var[X] + ab cov[X, V] + ab cov[V, X] + b2 var[V] 3. Suppose Y1 , Y2 , . . . , Yn are independent random variables each with variance 1. Let X1 = Y1 , X2 = Y1 + Y2 , . . . , Xn = Y1 + · · · + Yn . Find the n × n matrix var[X]. 4. Suppose X is an n × 1 random vector with finite second moments. Show that for any n × 1 vector α ∈ Rn we have E[(X − α)(X − α)T ] = var[X] + (µX − α)(µx − α)T 5. Suppose X is an m-dimensional random vector with finite second order moments and such that such that no element of X is a linear combination of the other elements. Show that for any n-dimensional random vector Y, there exists an n × m matrix A such that cov[Y − AX, X] = 0 6. Suppose X is an n × 1 random vector with E[X] = µ and var[X] = Σ. Prove the following results: (a) E[(AX + a)(BX + b)T ] = AΣBT + (Aµ + a)(Bµ + b)T where A is m × n, a is m × 1, B is r × n and b is r × 1. (b) E[(X + a)(X + a)T ] = Σ + (µ + a)(µ + a)T where a is n × 1. (c) E[XaT X] = (Σ + µµT )a where a is n × 1. 7. Suppose X is an n × 1 random vector with E[X] = µ and var[X] = Σ. Prove the following results: (a) E[(AX + a)T (BX + b)] = trace(AΣBT ) + (Aµ + a)T (Bµ + b) where A and B are m × n, and a and b are m × 1. (b) E[XT X] = trace(Σ) + µT µ (c) E[XT AX] = trace(AΣ) + µT Aµ where A is n × n. (d) E[(AX)T (AX)] = trace(AΣAT ) + (Aµ)T (Aµ) where A is n × n. (e) E[(X + a)T (X + a)] = trace(Σ) + (µ + a)T (µ + a) where a is n × 1. 8. Quadratic forms. Suppose X is an n × 1 random vector with E[X] = µ and var[X] = Σ. Suppose A is a real n × n symmetric matrix and b is an n × 1 real vector. Show that E[(X − b)T A(X − b)] = trace(AΣ) + (µ − b)T A(µ − b) In particular • E[ (X − µ)T A(X − µ) ] = trace(AΣ). √ • If kak denotes the length of the vector a, then kak = aT a and EkX − bk2 = trace(A) + kµ − bk2 . 9. Suppose X1 , . . . , Xn are random variables with E[Xj ] = µj and var[Xj ] = σj2 for j = 1, . . . , n; also cov[Xj , Xk ] = 0 for k > j + 1. If n X Q= (Xk − X)2 k=1 show that (n − 1)α − 2β +γ n Pn where α = σ12 + · · · + σn2 , β = cov[X1 , X2 ] + cov[X2 , X3 ] + · · · + cov[Xn−1 , Xn ] and γ = k=1 µ2k − Note that if all variables have the same mean, then γ = 0. E[Q] = 1 n Pn k=1 2 µk . Page 60 §3 Aug 1, 2018(17:54) Bayesian Time Series Analysis 10. Variance of a quadratic form—proof of proposition (1.9c) on page 58. (a) Show that XT AX = W1 + W2 + c where W1 = (X − µ)T A(X − µ), W2 = 2µT A(X − µ) and c = µT Aµ. (b) Show that var[W2 ] = 4σ 2 µT A2 µ. (c) Show that cov[W1 , W2 ] = E[W1 W2 ] = 2µT A E[YYT AY] = 2µ3 µT Ad where Y = X − µ. (d) Hence show var[XT AX] = (µ4 − 3σ 4 )dT d + 2σ 4 trace(A2 ) + 4σ 2 µT A2 µ + 4µ3 µT Ad 2 11. Suppose X1 , . . . , Xn are random variables Pnwith common2 expectation µ and2 common variance σ . Suppose further that 2 cov[Xj , Xk ] = ρσ for j 6= k. Show that k=1 (Xk − X) has expectation σ (1 − ρ)(n − 1) and hence Pn 2 k=1 (Xk − X) (1 − ρ)(n − 1) is an unbiased estimator of σ 2 . 12. Suppose X1 , . . . , Xn are i.i.d. random variables with the N (µ, σ 2 ) distribution. Let Pn Pn−1 2 (Xk+1 − Xk )2 2 k=1 (Xk − X) S = and Q = k=1 n−1 2(n − 1) (a) Show that var[S 2 ] = 2σ 4 /(n − 1). (b) Show that E[Q] = σ 2 and var[Q] = 2σ 4 (6n − 8)/4(n − 1)2 . 13. Suppose (X1 , X2 , X3 ) has the density 2 2 2 1 1 x1 x2 x3 −(x21 +x22 +x23 ) e− 2 (x1 +x2 +x3 ) + e for x1 , x2 , x3 ∈ R. (2π)3/2 (2π)3/2 2 (Note that the maximum of |x3 e−x3 | occurs at x3 = 1 and has absolute value which is less than 1. Hence f ≥ 0 everywhere.) Show that X1 , X2 and X3 are pairwise independent but not independent. f(X1 ,X2 ,X3 ) (x1 , x2 , x3 ) = 3 The bivariate normal 3.1 The density. Here is the first of several equivalent formulations of the density. Definition(3.1a). The random vector (X1 , X2 ) has a bivariate normal distribution iff it has density (x − µ)T P (x − µ) |P|1/2 exp − fX1 X2 (x1 , x2 ) = 2π 2 x1 µ1 where x = , µ = ∈ R2 and P is a real symmetric positive definite matrix. x µ2 2 2×1 2×1 2×2 Suppose the entries in the 2 × 2 real symmetric matrix P are denoted as follows2 : a1 a2 P= a2 a3 It follows that equation(3.1a) is equivalent to q a1 a3 − a22 a1 (x1 − µ1 )2 + 2a2 (x1 − µ1 )(x2 − µ2 ) + a3 (x2 − µ2 )2 exp − f (x1 , x2 ) = 2π 2 (3.1a) (3.1b) To show that equation(3.1b) defines a density. Clearly f ≥ 0. It remains to check that f integrates to 1. Let y1 = x1 − µ1 and y2 = x2 − µ2 . Then Z Z fX1 X2 (x1 , x2 ) dx1 dx2 x1 x2 a1 y12 + 2a2 y1 y2 + a3 y22 exp − dy1 dy2 2 y1 y2 ( 2 ) 2 Z Z y2 a22 |P|1/2 a1 a2 = exp − y1 + y2 exp − a3 − dy1 dy2 2π y1 y2 2 a1 2 a1 = 2 |P|1/2 2π Z Z It is easy to check that the real symmetric matrix P is positive definite iff a1 > 0 and a1 a3 − a22 > 0. (3.1c) (3.1d) 2 Multivariate Continuous Distributions Aug 1, 2018(17:54) §3 Page 61 √ √ a1 a3 −a22 Now use the transformation z1 = a1 y1 + aa12 y2 and z2 = y2 √ . This transformation has Jacobian a1 q a1 a3 − a22 = |P|1/2 and is a 1 − 1 map R2 → R2 ; it gives 2 Z Z Z Z z + z22 1 exp − 1 dz1 dz2 = 1 fX1 X2 (x1 , x2 ) dx1 dx2 = 2π z1 z2 2 x1 x2 by using the integral of the standard normal density equals one. 3.2 The marginal distributions of X1 and X2 . For the marginal density of X2 we need to find the following integral: Z fX1 X2 (x1 , x2 ) dx1 fX2 (x2 ) = x1 First let Y1 = X1 − µ1 and Y2 = X2 − µ2 and find the density of Y2 : Z Z a1 y12 + 2a2 y1 y2 + a3 y22 |P|1/2 fY1 ,Y2 (y1 , y2 ) dy1 = fY2 (y2 ) = exp − dy1 2π y1 2 y1 Using the decomposition in equation(3.1d) gives ( 2 ) 2 Z y2 a22 |P|1/2 a1 a2 fY2 (y2 ) = dy1 exp − a3 − exp − y1 + y2 2π 2 a1 2 a1 y1 2 r y2 a1 a3 − a22 y22 |P|1/2 2π 1 = exp − =q exp − 2 2π 2 a1 a1 2σ2 2πσ22 where σ22 = a1 /(a1 a3 − a22 ) = a1 /|P|. It follows that the density of X2 = Y2 + µ2 is a1 1 a1 (x2 − µ2 )2 = fX2 (x2 ) = q where σ22 = exp − 2 2 2σ2 (a1 a3 − a2 ) |P| 2πσ22 We have shown that the marginal distributions are normal: X2 has the N (µ2 , σ22 ) distribution. Similarly, X1 has the N (µ1 , σ12 ) distribution. where a3 a1 a1 a3 = and σ22 = = σ12 = a1 a3 − a22 |P| a1 a3 − a22 |P| 3.3 The covariance and correlation between X1 and X2 . Of course, cov[X1 , X2 ] = cov[Y1 , Y2 ] where Y1 = X1 − µ1 and Y2 = X2 − µ2 . So it suffices to find cov[Y1 , Y2 ] = E[Y1 Y2 ]. The density of (Y1 , Y2 ) is: a1 y12 + 2a2 y1 y2 + a3 y22 |P|1/2 fY1 Y2 (y1 , y2 ) = exp − 2π 2 It follows that Z Z a1 y12 + 2a2 y1 y2 + a3 y22 2π dy1 dy2 = q exp − 2 y1 y2 a1 a3 − a22 Differentiating with respect to a2 gives Z Z a1 y12 + 2a2 y1 y2 + a3 y22 2π (−y1 y2 ) exp − dy1 dy2 = a2 2 (a1 a3 − a22 )3/2 y1 y2 and hence a2 cov[X1 , X2 ] = E[Y1 Y2 ] = − a1 a3 − a22 The correlation between X1 and X2 is cov[X1 , X2 ] a2 ρ= = −√ σ1 σ2 a1 a3 These results lead to an alternative expression for the density of a bivariate normal: 1 1 (x1 − µ1 )2 (x1 − µ1 )(x2 − µ2 ) (x2 − µ2 )2 p fX1 X2 (x1 , x2 ) = exp − − 2ρ + 2(1 − ρ2 ) σ 1 σ2 σ12 σ22 2πσ1 σ2 1 − ρ2 (3.3a) Page 62 §3 Aug 1, 2018(17:54) Bayesian Time Series Analysis We have also shown that σ12 cov[X1 , X2 ] var[X] = = P−1 cov[X1 , X2 ] σ22 P is sometimes called the precision matrix—it is the inverse of the variance matrix var[X]. Summarizing some of these results: 1 a3 −a2 a1 a2 2 −1 |P| = a1 a3 − a2 P= P = a2 a3 a1 a3 − a22 −a2 a1 1 σ12 ρσ1 σ2 σ22 2 2 2 −1 −1 |Σ| = (1 − ρ )σ1 σ2 Σ=P = Σ =P= 2 2 2 2 ρσ1 σ2 σ2 (1 − ρ )σ1 σ2 −ρσ1 σ2 −ρσ1 σ2 σ12 (3.3b) Example(3.3a). Suppose (X, Y ) has a bivariate normal distribution with density 1 1 f (x, y) = exp − (x2 + 2y 2 − xy − 3x − 2y + 4) k 2 Find the mean vector and the variance matrix of (X, Y ). What is the value of k? Solution. Let Q(x, y) = a1 (x−µ1 )2 +2a2 (x−µ1 )(y −µ2 )+a3 (y −µ2 )2 . So we want Q(x, y) = x2 +2y 2 −xy −3x−2y +4. Equating coefficients of x2 , xy and y 2 gives a1 = 1, a2 = − 21 and a3 = 2. Hence 4 2 12 1 − 12 −1 P= and Σ = P = − 12 2 7 21 1 √ Also |P| = 74 and hence k = 2π/|P|1/2 = 4π/ 7. = 2a1 (x − µ1 ) + 2a2 (y − µ2 ) and ∂Q(x,y) = 2a2 (x − µ1 ) + 2a3 (y − µ2 ). If ∂Q(x,y) = 0 and ∂Q(x,y) = 0 then Now ∂Q(x,y) ∂x ∂y ∂x ∂y 2 we must have x = µ1 and y = µ2 because |P| = a1 a3 − a2 6= 0. Applying this to Q(x, y) = x2 + 2y 2 − xy − 3x − 2y + 4 gives the equations 2µ1 − µ2 − 3 = 0 and 4µ2 − µ1 − 2 = 0. Hence (µ1 , µ2 ) = (2, 1). 3.4 The characteristic function. Suppose XT = (X1 , X2 ) has the bivariate density defined in equation(3.1a). Then for all t ∈ R2 , the characteristic function of X is 2×1 h φ(t) = E e = itT X i |P|1/2 = 2π Z Z |P|1/2 itT µ e 2π y1 (x − µ)T P(x − µ) e exp − dx dy 2 x1 x2 T 2it y − yT Py dy1 dy2 by setting y = x − µ. exp 2 y2 Z Z itT x But y Py − 2it y = (y − iΣt) P(y − iΣt) + tT Σt where Σ = P−1 = var[X]. Hence Z Z |P|1/2 itT µ− 1 tT Σt (y − iΣt)T P(y − iΣt) 2 φ(t) = e exp − dy1 dy2 2π 2 y1 y2 T T T = eit T µ− 21 tT Σt by using the integral of equation(3.1a) is 1. 3.5 The conditional distributions. We first find the conditional density of Y1 given Y2 where Y1 = X1 − µ1 and Y2 = X2 − µ2 . Now fY Y (y1 , y2 ) fY1 |Y2 (y1 |y2 ) = 1 2 fY2 (y2 ) We use the following forms: n o σ12 2 2ρσ1 2 y1 − σ2 y1 y2 + σ2 y2 1 2 p fY1 Y2 (y1 , y2 ) = exp − 2 2 2 2σ1 (1 − ρ ) 2πσ1 σ2 1 − ρ y22 fY2 (y2 ) = q exp − 2 2σ2 2πσ22 1 and hence 2 Multivariate Continuous Distributions f (y1 |y2 ) = √ §3 Page 63 Aug 1, 2018(17:54) 1 2π q σ12 (1 − ρ2 ) n y12 − exp − 2ρσ1 σ2 y1 y2 + ρ2 σ12 2 y σ22 2 2σ12 (1 − ρ2 ) o n o2 ρσ1 y1 − σ2 y2 1 =√ q exp − 2 2 2σ1 (1 − ρ ) 2π σ 2 (1 − ρ2 ) 1 − ρ2 ) distribution. 2 (1 − ρ2 ) distribution, and hence 1 (x − µ ), σ It follows that the density of X1 given X2 is the N µ1 + ρσ 2 2 1 σ2 which is the density of the N E[X1 |X2 ] = µ1 + ρσ1 σ2 (X2 ρσ1 2 σ2 y2 , σ1 (1 − µ2 ) and var[X1 |X2 ] = σ12 (1 − ρ2 ). In terms of the original notation, σ12 (1 − ρ2 ) = 1/a1 and ρσ1 /σ2 = −a2 /a1 and hence the distribution of X1 given X2 is N µ1 − a2 a1 (x2 − µ2 ), a11 . We have also shown that if the random vector (X1 , X2 ) is bivariate normal, then E[X1 |X2 ] is a linear function of X2 and hence the best predictor and best linear predictor are the same—see exercises 9 and 12 on page 8. 3.6 Independence of X1 and X2 . Proposition(3.6a). Suppose (X1 , X2 ) ∼ N (µ, Σ). Then X1 and X2 are independent iff ρ = 0. Proof. If ρ = 0 then fX1 X2 (x1 , x2 ) = fX1 (x1 )fX2 (x2 ). Conversely, if X1 and X2 are independent then cov[X1 , X2 ] = 0 and hence ρ = 0. In terms of entries in the precision matrix: X1 and X2 are independent iff a2 = 0. 3.7 Linear transformation of a bivariate normal. Proposition(3.7a). Suppose X has the bivariate normal distribution N (µ, Σ) and C is a 2 × 2 non-singular matrix. Then the random vector Y = a + CX has the bivariate normal distribution N (a + Cµ, CΣCT ). Proof. The easiest way is to find the characteristic function of Y. For t ∈ R2 we have 2×1 T φY (t) = E[eit Y T itT CX ] = eit a E[e T T ] = eit a eit Cµ− 21 tT CΣCT t which is the characteristic function of the bivariate normal N (a + Cµ, CΣCT ). We need C to be non-singular in order to ensure the variance matrix of the result is non-singular. 3.8 Summary. The bivariate normal distribution. Suppose X = (X1 , X2 ) ∼ N (µ, Σ) where µ = E[X] and Σ = var[X]. • Density. |P|1/2 (x − µ)T P(x − µ) fX (x) = exp − where P = Σ−1 is the precision matrix. 2π 2 1 1 (x1 − µ1 )2 (x1 − µ1 )(x2 − µ2 ) (x2 − µ2 )2 p = exp − − 2ρ + 2(1 − ρ2 ) σ1 σ2 σ12 σ22 2πσ1 σ2 1 − ρ2 1 a1 a2 a3 −a2 • If P = then Σ = P−1 = . a2 a3 a1 a3 − a22 −a2 a1 1 σ22 −ρσ1 σ2 −1 • P=Σ = 2 2 and |Σ| = (1 − ρ2 )σ12 σ22 . σ12 σ1 σ2 (1 − ρ2 ) −ρσ1 σ2 • The marginal distributions. X1 ∼ N (µ1 , σ12 ) and X2 ∼ N (µ2 , σ22 ). . . . continued Page 64 §4 Aug 1, 2018(17:54) Bayesian Time Series Analysis Summary (continued). • The characteristic function. φ(t) = exp itT µ − 12 tT Σt . • The conditional distributions. 2 2 1 The distribution of X1 given X2 = x2 is N µ1 + ρσ σ2 (x2 − µ2 ), σ1 (1 − ρ ) . 2 (1 − ρ2 ) . 2 (x − µ ), σ The distribution of X2 given X1 = x1 is N µ2 + ρσ 1 1 2 σ1 • X1 and X2 are independent iff ρ = 0. • Linear transformation of a bivariate normal. If C is non-singular, then Y = a + CX has a bivariate normal distribution with mean a + Cµ and variance matrix CΣCT . 4 Exercises (exs-bivnormal.tex) 1. Suppose (X, Y ) has the density 2 fXY (x, y) = ce−(x −xy+y 2 )/3 (a) Find c. (b) Are X and Y independent? 2. Suppose (X, Y ) has a bivariate normal distribution with density f (x, y) = k exp −(x2 + 2xy + 4y 2 ) Find the mean vector and the variance matrix of (X, Y ). What is the value of k? 3. Suppose (X, Y ) has a bivariate normal distribution with density 1 1 f (x, y) = exp − (2x2 + y 2 + 2xy − 22x − 14y + 65) k 2 Find the mean vector and the variance matrix of (X, Y ). What is the value of k? 4. Suppose the random vector Y = (Y1 , Y2 ) has the density 1 1 2 2 f (y1 , y2 ) = exp − (y1 + 2y2 − y1 y2 − 3y1 − 2y2 + 4) k 2 Find E[Y] and var[Y]. for y = (y1 , y2 ) ∈ R2 . 5. Evaluate the integral Z ∞ −∞ exp −(y12 + 2y1 y2 + 4y22 ) dy1 dy2 6. Suppose the random vector Y = (Y1 , Y2 ) has the density 1 f (y1 , y2 ) = k exp − (y12 + 2y1 (y2 − 1) + 4(y2 − 1)2 12 Show that Y ∼ N (µ, Σ) and find the values of µ and Σ. for y = (y1 , y2 ) ∈ R2 . 7. Suppose X = (X1 , X2 ) has a bivariate normal distribution with E[X1 ] = E[X2 ] = 0 and variance matrix Σ. Prove that X2 XT PX − 21 ∼ χ21 σ1 where P is the precision matrix of X. 8. (a) Suppose E[X1 ] = µ1 , E[X2 ] = µ2 and there exists α such that Y = X1 + αX2 is independent of X2 . Prove that E[X1 |X2 ] = µ1 + αµ2 − αX2 . (b) Use part (a) to derive E[X1 |X2 ] for the bivariate normal. 9. An alternative method for constructing the bivariate normal. Suppose X and Y are i.i.d. N (0, 1). Suppose ρ ∈ (−1, 1) p and Z = ρX + 1 − ρ2 Y . (a) Find the density of Z. (b) Find the density of (X, Z). (c) Suppose µ1 ∈ R, µ2 ∈ R, σ1 > 0 and σ2 > 0. Find the density of (U, V ) where U = µ1 + σ1 X and V = µ2 + σ2 Z. 2 Multivariate Continuous Distributions §5 Page 65 Aug 1, 2018(17:54) 10. Suppose (X1 , X2 ) has the bivariate normal distribution with density given by equation(3.3a). Define Q by: e−Q(x1 ,x2 ) p fX1 X2 (x1 , x2 ) = 2πσ1 σ2 1 − ρ2 Hence 1 (x1 − µ1 )2 (x1 − µ1 )(x2 − µ2 ) (x2 − µ2 )2 Q(x1 , x2 ) = − 2ρ + 2(1 − ρ2 ) σ1 σ2 σ12 σ22 Define the random variable Y by Y = Q(X1 , X2 ). Show that Y has the exponential density. 11. (a) Suppose (X1 , X2 ) has a bivariate normal distribution with E[X1 ] = E[X2 ] = 0. Hence it has characteristic function 1 1 φX (t) = exp − tT Σt = exp − σ12 t21 + 2σ12 t1 t2 + σ22 t22 2 2 Explore the situations when Σ is singular. (b) Now suppose (X1 , X2 ) has a bivariate normal distribution without the restriction of zero means. Explore the situations when the variance matrix Σ is singular. 12. Suppose T1 and T2 are i.i.d. N (0, 1). Set X = a1 T1 + a2 T2 and Y = b1 T1 + b2 T2 where a21 + a22 > 0 and a1 b2 6= a2 b1 . (a) Show that E[Y |X] = X(a1 b1 + a2 b2 )/(a21 + a22 ). 2 (b) Show that E Y − E(Y |X) = (a1 b2 − a2 b1 )2 /(a21 + a22 ). 13. Suppose (X, Y ) has a bivariate normal distribution with E[X] = E[Y ] = 0, var[X] = var[Y ] = 1 and cov[X, Y ] = ρ. Show that X 2 and Y 2 are independent iff ρ = 0. 2 and var[Y ] = σY2 then just set X1 = X/σX and Y1 = Y /σY .) (Note. If var[X] = σX 14. Suppose (X1 , X2 ) has a bivariate normal distribution with E[X1 ] = E[X2 ] = 0. Let Z = X1 /X2 . Show that p σ1 σ2 1 − ρ2 fZ (z) = π(σ22 z 2 − 2ρσ1 σ2 z + σ12 ) 15. Suppose (X1 , X2 ) has a bivariate normal distribution with E[X1 ] = E[X2 ] = 0 and var[X1 ] = var[X2 ] = 1. Let ρ = corr[X1 , X2 ] = cov[X1 , X2 ] = E[X1 X2 ]. Show that X12 − 2ρX1 X2 + X22 ∼ χ22 1 − ρ2 16. Suppose (X, Y ) has a bivariate normal distribution with E[X] = E[Y ] = 0. Show that 1 1 sin−1 ρ P[X ≥ 0, Y ≥ 0] = P[X ≤ 0, Y ≤ 0] = + 4 2π 1 1 P[X ≤ 0, Y ≥ 0] = P[X ≥ 0, Y ≤ 0] = − sin−1 ρ 4 2π 5 The multivariate normal 5.1 The multivariate normal distribution. The random vector X has a (non-singular) multivariate normal n×1 distribution iff X has density 1 T f (x) = C exp − (x − µ) P(x − µ) 2 for x ∈ Rn (5.1a) where • C is a constant so that the density integrates to 1; • µ is a vector in Rn ; • P is a real symmetric positive definite n × n matrix called the precision matrix. 5.2 Integrating the density. Because P is a real symmetric positive definite matrix, there exists an orthogonal matrix L with P = LT DL where L is orthogonal and D is diagonal with entries d1 > 0, . . . , dn > 0 which are the eigenvalues of P. This result is explained in §1.4 on page 56. Consider the transformation y = L(x − µ); this is a 1 − 1 transformation: Rn → Rn which has a Jacobian with absolute value: ∂(y1 , . . . , yn ) ∂(x1 , . . . , xn ) = |(det(L)| = 1 Page 66 §5 Aug 1, 2018(17:54) Using x − µ = LT y gives 1 fY (y) = C exp − yT LPLT y 2 Bayesian Time Series Analysis 1 = C exp − yT Dy 2 =C n Y j=1 1 exp − dj yj2 2 It follows that Y1 , . . . , Yn are independent with distributions N (0, 1 ), . . . , N (0, 1/dn ) respectively, and √ √ √ d1 · · · dn det(D) det(P) C= = = n/2 n/2 (2π) (2π) (2π)n/2 Equation(5.1a) becomes √ det(P) 1 T for x ∈ Rn (5.2a) f (x) = exp − (x − µ) P(x − µ) 2 (2π)n/2 Note that the random vector Y satisfies E[Y] = 0 and var[Y] = D−1 . Using X = µ + LT Y gives E[X] = µ T var[X] = var[L Y] = LT var[Y]L = LT D−1 L = P−1 and hence P is the precision matrix—the inverse of the variance matrix. So equation(5.1a) can be written as 1 1 T −1 √ f (x) = exp − (x − µ) Σ (x − µ) for x ∈ Rn (5.2b) 2 (2π)n/2 det(Σ) where µ = E[X] and Σ = P−1 = var[X]. This is defined to be the density of the N (µ, Σ) distribution. Notes. • The matrix M is the variance matrix of a non-singular normal distribution iff it is symmetric and positive definite. • The random vector X is said to have a spherical normal distribution iff X ∼ N (µ, σ 2 I). Hence X1 , . . . , Xn are independent and have the same variance. 1/d 5.3 The characteristic function. Suppose the n-dimensional random vector X has the N (µ, Σ) distribution. We know that using the transformation Y = L(X − µ) leads to Y ∼ N (0, D−1 ). Because L is orthogonal we have X = µ + LT Y and the characteristic function of X is: h T i h T i h T T i T T T φX (t) = E eit X = E eit µ+it L Y = eit µ E eit L Y for all t ∈ Rn . n×1 But Y1 , . . . , Yn are independent with distributions N (0, 1/d1 ), . . . , N (0, 1/dn ), respectively. Hence h T i 1 T −1 2 2 E eit Y = E ei(t1 Y1 +···+tn Yn ) = e−t1 /2d1 · · · e−tn /2dn = e− 2 t D t for all n × 1 vectors t ∈ Rn . Applying this result to the n × 1 vector Lt gives h T T i 1 T T −1 1 T E eit L Y = e− 2 t L D Lt = e− 2 t Σt We have shown that if X ∼ N (µ, Σ) then h T i 1 T T φX (t) = E eit X = eit µ− 2 t Σt for all t ∈ Rn . 5.4 The singular multivariate normal distribution. In the last section we saw that if µ ∈ Rn and Σ is an n × n real symmetric positive definite matrix, then the function φ : Rn → Cn with T 1 T φ(t) = eit µ− 2 t Σt for t ∈ Rn . is a characteristic function. The condition on Σ can be relaxed as follows: Proposition(5.4a). Suppose µ ∈ Rn and V is an n × n real symmetric non-negative definite matrix, then the function φ : Rn → Cn with is a characteristic function. T µ− 1 tT Vt φ(t) = eit 2 for t ∈ Rn . Proof. For n = 1, 2, . . . , set Vn = V + n1 I where I is the n × n identity matrix. Then Vn is symmetric and positive definite and so T 1 T φn (t) = eit µ− 2 t Vn t is a characteristic function. Also φn (t) → φ(t) as n → ∞ for all t ∈ Rn . Finally, φ is continuous at t = 0. It follows that φ is a characteristic function by the multidimensional form of Lévy’s convergence theorem3 . 3 Also called the “Continuity Theorem.” See, for example, page 361 in [F RISTEDT & G RAY(1997)]. 2 Multivariate Continuous Distributions §5 Page 67 Aug 1, 2018(17:54) If V is symmetric and positive-definite, then we know that φ is the characteristic function of the N (µ, Σ) distribution. If V is only symmetric and non-negative definite and not positive definite, then by §1.3 on page 56, we know that some linear combination of the components is zero and the density does not exist. In this case, we say that the distribution with characteristic function φ is a singular multivariate normal distribution. 5.5 Linear combinations of the components of a multivariate normal. Suppose the n-dimensional random vector X has the N (µ, Σ) distribution. Then for any n × 1 vector ` ∈ Rn the random variable Z = `T X has a normal distribution. Proposition(5.5a). Proof. Use characteristic functions. For t ∈ R we have TX φZ (t) = E[eitZ ] = E[eit` and hence Z ∼ N `T µ, `T Σ` . T µ− 1 t2 `T Σ` ] = φX (t`) = expit` 2 Conversely: Proposition(5.5b). Suppose X is an n-dimensional random vector such that for every n × 1 vector ` ∈ Rn the random variable `T X is univariate normal. Then X has the multivariate normal distribution. T Proof. The characteristic function of X is φX (t) = E[expit X ]. Now Z = tT X is univariate normal. Also E[tT X] = tT µ and var[tT X] = tT Σt where µ = E[X] and Σ = var[X]. Hence Z ∼ N (tT µ, tT Σt). Hence the characteristic function of Z is, for all u ∈ R: T µ− 1 u2 tT Σt φZ (u) = expiut 2 Take u = 1; hence T µ− 1 tT Σt φZ (1) = expit itT X But φZ (1) = E[expiZ ] = E[exp 2 ]. So we have shown that TX E[expit T µ− 1 tT Σt ] = expit 2 and so X ∼ N (µ, Σ). Combining these two previous propositions gives a characterization of the multivariate normal distribution: the ndimensional random vector X has a multivariate normal distribution iff every linear combination of the components of X has a univariate normal distribution. 5.6 The marginal distributions. Suppose the n-dimensional random vector X has the N (µ, Σ) distribution. Then the characteristic function of X is 1 T T φX (t) = E[ei(t1 X1 +···+tn Xn ) ] = eit µ− 2 t Σt for t ∈ Rn . and hence the characteristic function of X1 is 1 2 φX1 (t1 ) = φX (t1 , 0, . . . , 0) = expiµ1 t1 − 2 t1 Σ11 and so X1 ∼ N (µ1 , Σ11 ). Similarly, Xj ∼ N (µj , Σjj ) where Σjj is the (j, j) entry in the matrix Σ. Similarly, the random vector (Xi , Xj ) has the bivariate normal distribution with mean vector (µi , µj ) and variance matrix Σii Σij Σij Σjj In general, we see that every marginal distribution of a multivariate normal is normal. The converse is false!! Example(5.6a). Suppose X and Y are independent 2-dimensional random vectors with distrbutions N (µ, ΣX ) and N (µ, ΣY ) respectively where µ1 σ12 ρ1 σ1 σ2 σ12 ρ2 σ1 σ2 µ= ΣX = and ΣY = µ2 ρ1 σ1 σ2 σ22 ρ2 σ1 σ2 σ22 and ρ1 6= ρ2 . Let X with probability 1/2; Z= Y with probability 1/2. Show that Z has normal marginals but is not bivariate normal. Solution. Let Z1 and Z2 denote the components of Z; hence Z = (Z1 , Z2 ). Then Z1 ∼ N (µ1 , σ12 ) and Z2 ∼ N (µ2 , σ22 ). Hence every marginal distribution of Z is normal. Page 68 §5 Aug 1, 2018(17:54) Bayesian Time Series Analysis Now E[Z] = µ and var[Z] = E[(Z − µ)(Z − µ)T ] = 1/2(ΣX + ΣY ). The density of Z is 1 1 fZ (z) = fX (z) + fY (z) 2 2 and this is not the density of N (µ, 1/2(ΣX + ΣY ))—we can see that by comparing the values of these two densities at z = µ. 5.7 Linear transformation of a multivariate normal. Proposition(5.7a). Suppose the n-dimensional random vector X has the non-singular N (µ, Σ) distribution. Suppose further that B is an m × n matrix with m ≤ n and rank(B) = m; hence B has full rank. Let Z = BX. Then Z has the non-singular N (Bµ, BΣBT ) distribution. Proof. We first establish that BΣBT is positive definite. Suppose x ∈ Rm with xT BΣBT x = 0. Then yT Σy = 0 where y = BT x. Because Σ is positive definite, we must have y = 0. Hence BT x = 0; hence x1 αT1 + · · · + xm αTm = 0 where α1 , . . . , αm are the m rows of B. But rank(B) = m; hence x = 0 and hence BΣBT is positive definite. The characteristic function of Z is, for all t ∈ Rm : TZ φZ (t) = E[expit Hence Z ∼ N (Bµ, BΣBT ). T BX ] = E[expit T Bµ− 1 tT BΣBT T ] = φX (BT t) = expit 2 The following proposition shows that we can always transform the components of a non-singular multivariate normal into i.i.d. random variables with the N (0, 1) distribution. Proposition(5.7b). Suppose the random vector X is has the non-singular N (µ, Σ) distribution. Then there exists a non-singular matrix Q such that the QT Q = P, the precision matrix, and Q(X − µ) ∼ N (0, I) Proof. From §5.2 on page 65 we know that the precision matrix P = Σ−1 = LT DL where L√is orthogonal √ and D = diag[d1 , . . . , dn ] with d1 > 0, . . . , dn > 0. Hence P = LT D1/2 D1/2 L where D1/2 = diag[ d1 , . . . , dn ]. Hence P = QT Q where Q = D1/2 L. Because L is non-singular, it follows that Q is also non-singular. If Z = Q(X − µ), then E[Z] = 0 and var[Z] = QΣQT = QLT D−1 LQT = I. Hence Z ∼ N (0, I) and Z1 , . . . , Zn are i.i.d. random variabes with the N (0, 1) distribution. If the random vector has the spherical normal distribution N (µ, σ 2 I) and L is orthogonal, then Y ∼ N (Lµ, σ 2 I) and hence Y also has a spherical normal distribution. 5.8 Partitioning a multivariate normal into two sub-vectors. Suppose X is an n-dimensional random vector with the N (µ, Σ) distribution, and X is partitioned into two sub-vectors: X1 k×1 X = where n = k + `. X2 n×1 Proposition(5.8a). `×1 Now partition µ and Σ conformably as follows: µ1 k×1 µ = and µ2 n×1 `×1 Σ = n×n Σ11 Σ12 k×k k×` Σ21 Σ22 `×k `×` Then the random vectors X1 and X2 are independent iff Σ12 = 0, equivalently iff cov[X1 , X2 ] = 0. Note that Σ21 = ΣT12 . Proof. ⇒ Now cov[X1,i , X2,j ] = 0 for all i = 1, . . . , k and j = 1, . . . , `. Hence Σ12 = 0. ⇐ Because Σ12 = 0 we have Σ11 0 Σ= 0 Σ22 Now the characteristic function of X is h i T T 1 T E eit X = eit µ− 2 t Σt for all t ∈ Rn . Partitioning t conformably into t = (t1 , t2 ) gives h T i T T 1 T 1 T E eit X = eit1 µ1 − 2 t1 Σ11 t1 eit2 µ2 − 2 t2 Σ22 t2 and hence (5.8a) 2 Multivariate Continuous Distributions §5 Page 69 Aug 1, 2018(17:54) h T i h T i h T i T E eit1 X1 +it2 X2 = E eit1 X1 E eit2 X2 and hence X1 and X2 are independent. Repeated application of this result shows that if Y = (Y1 , . . . , Yn ) has a multivariate normal distribution, then Y1 , . . . , Yn are independent iff all covariances equal 0. 5.9 Conditional distributions. To prove the following proposition, we use the following result: suppose W1 is a k-dimensional random vector; W2 is an `-dimensional random vector; W1 and W2 are independent; h is a function : R` → Rk . Let V = W1 + h(W2 ). Then the distribution of V given W2 has density f(V,W2 ) (v, w2 ) f(W1 ,W2 ) (v − h(w2 ), w2 ) = = fW1 ( v − h(w2 ) ) fV|W2 (v|w2 ) = fW2 (w2 ) fW2 (w2 ) In particular, if W1 ∼ N (µ1 , Σ1 ) then the conditional distribution of V = W1 + h(W2 ) given W2 = w2 has the density of the N ( µ1 + h(w2 ), Σ1 ) distribution. We have shown the following. Suppose W1 ∼ N (µ1 , Σ1 ) and W2 is independent of W1 . Then the conditional density of W1 + h(W2 ) given W2 = w2 is the density of N ( µ1 + h(w2 ), Σ1 ). (5.9a) Proposition(5.9a). Suppose X is an n-dimensional random vector with the non-singular N (µ, Σ) distribution, and X is partitioned into two sub-vectors: X = n×1 X1 k×1 where n = k + `. X2 `×1 Partition µ and Σ conformably as in equations(5.8a). Then the conditional distribution of X1 given X2 = x2 has the density of the following distribution: −1 N µ1 + Σ12 Σ−1 (5.9b) 22 (x2 − µ2 ), Σ11 − Σ12 Σ22 Σ21 Proof. We shall give two proofs of this important result; the first proof is shorter but requires knowledge of the answer! Proof 1. Let I −Σ12 Σ−1 22 k×k k×` B = 0 I n×n `×k `×` Note that B is invertible with inverse −1 B = I 0 Σ12 Σ−1 22 I But by proposition(5.7a), we know that BX ∼ N (Bµ, BΣBT ). where X1 − Σ12 Σ−1 µ1 − Σ12 Σ−1 Σ11 − Σ12 Σ−1 22 Σ21 22 X2 22 µ2 BX = Bµ = and BΣBT = X2 µ2 0 0 Σ22 It follows that X1 − Σ12 Σ−1 22 X2 is independent of X2 . Also −1 −1 X1 − Σ12 Σ−1 22 X2 ∼ N µ1 − Σ12 Σ22 µ2 , Σ11 − Σ12 Σ22 Σ21 It follows by equation(5.9a) that the conditional distribution of X1 given X2 = x2 has the density of −1 N µ1 + Σ12 Σ−1 22 (x2 − µ2 ), Σ11 − Σ12 Σ22 Σ21 Proof 2. We want to construct a k × 1 random vector W1 with W1 = C1 X1 + C2 X2 where C1 is k × k and C2 is k × ` and such that W1 is independent of X2 . Now for any C∗ , the random vector C∗ W1 is also independent of X2 ; so the answer is arbitrary up to multiplicative C∗ . So take C1 = I. Now cov[W1 , X2 ] = 0; hence cov(X1 , X2 ) + cov(C2 X2 , X2 ) = 0; hence Σ12 + C2 Σ22 = 0 and hence C2 = −Σ12 Σ−1 22 . −1 We now have W1 = X1 − Σ12 Σ22 X2 independent of X2 . Also X1 = W1 − C2 X2 ; hence var[X1 ] = var[W1 ] + C2 var[X2 ]CT2 ; hence var[W1 ] = Σ11 − Σ12 Σ−1 22 Σ21 . The rest of the proof is as in the first proof. Page 70 §5 Aug 1, 2018(17:54) Bayesian Time Series Analysis 5.10 The matrix of regression coefficients, partial covariance and partial correlation coefficients. Note that Σ12 Σ−1 22 is called the matrix of regression coefficients of X1 on X2 ; it is obtained by multiplying the k × ` matrix cov[X1 , X2 ] = Σ12 by the ` × ` precision matrix of X2 . Similarly, the matrix of regression coefficients of X2 on X1 is Σ21 Σ−1 11 . Let D1 denote the variance matrix of the conditional distribution of X1 given X2 . Hence D1 is a k × k invertible −1 matrix and D1 = Σ11 − Σ12 Σ−1 22 Σ21 . Similarly, let D2 = Σ22 − Σ21 Σ11 Σ12 ; then D2 is an ` × ` invertible matrix. By postmultiplying the following partitioned matrix by the partitioned form of Σ, it is easy to check that D−1 −D−1 Σ12 Σ−1 −1 1 1 22 P=Σ = (5.10a) −1 −D−1 D−1 2 Σ21 Σ11 2 Note that D1 is called the partial covariance of X1 given X2 . The partial correlation between X1,j and X1,k given X2 is defined to be [D1 ]j,k p p [D1 ]j,j [D1 ]k,k 5.11 The special case of the conditional distribution of X1 given (X2 , . . . , Xn ). For this case, k = 1 and ` = n−1. Hence D1 is 1×1; denote the first row of the precision matrix P by [q1,1 , q1,2 , . . . , q1,n ]. By equation(5.10a) −1 −1 [q1,1 , q1,2 , . . . , q1,n ] = [D−1 1 , −D1 Σ12 Σ22 ] Hence 1 Σ12 Σ−1 22 = − 1 [q1,2 , . . . , q1,n ] q1,1 q1,1 By equation(5.9b) on page 69, the conditional distribution of X1 given (X2 , . . . , Xn ) = (x2 , . . . , xn ) is q1,2 (x2 − µ2 ) + · · · + q1,n (xn − µn ) 1 N µ1 − , q1,1 q1,1 From the proof of proposition (5.9a) we know that q1,2 (X2 − µ2 ) + · · · + q1,n (Xn − µn ) X1 − is independent of (X2 , . . . , Xn ) q1,1 The second proof also shows that if X1 − (a2 X2 + · · · + an Xn ) is independent of (X2 , . . . , Xn ), then we must q q1,2 , . . . , an = q1,n . In other words, the unique linear function a2 X2 + · · · + an Xn which makes have a1 = q1,1 1,1 D1 = and X1 − (a2 X2 + · · · + an Xn ) independent of (X2 , . . . , Xn ) is given by a1 = q1,2 q1,1 , . . . , an = q1,n q1,1 . 5.12 The joint distribution of X and S 2 . This is a very important result in statistical inference!! Proposition(5.12a). Suppose X1 , . . . , Xn are i.i.d. random variables with the N (µ, σ 2 ) distribution. Let Pn X= Then X and S2 k=1 Xk n and X∼N σ2 µ, n Pn S = are independent; also 2 and − X)2 n−1 k=1 (Xk (n − 1)S 2 ∼ χ2n−1 σ2 (5.12a) Proof. We shall give three proofs. Pn Method 1. Let Yk = (Xk − µ)/σ for k = 1, . . . , n. Then Y = k=1 Yk /n = (X − µ)/σ and P n n 2 X (n − 1)S 2 k=1 (Xk − X) = = (Yk − Y)2 σ2 σ2 k=1 Because Y1 , . . . , Yn are i.i.d. N (0, 1), the density of the vector Y = (Y1 , . . . , Yn ) is 2 1 y1 + · · · + yn2 fY (y1 , . . . , yn ) = exp − 2 (2π)n/2 Consider the transformation to Z = (Z1 , . . . , Zn ) with Z = AY. Suppose further that Y1 + · · · + Yn √ √ Z1 = = nY n Hence the first row of the matrix A is √1n , . . . , √1n . Construct the other (n − 1) rows of A so that A is orthogonal. For the explicit value of A, see exercise 7 on page 73. This means AAT = I. Hence n n X X Zk2 = ZT Z = YT AT AY = YT Y = Yk2 k=1 k=1 2 Multivariate Continuous Distributions §5 Page 71 Aug 1, 2018(17:54) Pn Pn Also E[Z] = 0 and var[Z] = E[ZZT ] = A var[Y] AT = I. Hence Z1 ,. . . ,Zn are i.i.d. N (0, 1). Now k=1 Zk2 = k=1 Yk2 . Hence n n n n X X X X (n − 1)S 2 2 Zk2 = (Yk − Y)2 = Yk2 − Z12 = Yk2 − nY = σ2 k=2 k=1 k=1 k=1 √ This proves Z1 = n Y is independent of S 2 and hence X is independent of S 2 . The results in (5.12a) are immediate. Method 2. This is based on an algebraic trick applied to moment generating functions. For all t1 ∈ R, . . . , tn ∈ R we have n n X X tk (Xk − X) = tk Xk − nX t and hence for all t0 ∈ R we have t0 X + k=1 k=1 n X n X t0 k=1 tk (Xk − X) = k=1 n n X + tk − t Xk = ck Xk k=1 where ck = t0/n + (tk − t). Now let t = (t0 , t1 , . . . , tn ); then the moment generating function of the vector Z = (X, X1 − X, . . . , Xn − X) is ! n n n Y Y σ 2 t20 σ2 X σ 2 c2k 2 t·Z c1 X1 +···cn Xn ck Xk = exp µt0 + exp (tk − t) E[e ] = E[e ]= E[e ]= exp µck + 2 2n 2 k=1 k=1 k=1 t0 X The first factor is E[e ]. Hence X is independent of the vector (X1 − X, . . . , Xn − X) and hence X and S 2 are independent. By using the identity Xk − µ = (Xk − X) + (X − µ) we can get 2 n n 1 X 1 X X − µ) 2 2 √ X) + (X − µ) = (X − k k σ2 σ2 σ n k=1 k=1 Using moment generating functions and the independence of the two terms on the right hand side gives (n − 1)S 2 /σ 2 ∼ χ2n−1 . Method 3. Consider the transformation (x1 , . . . , xn ) → (y1 , . . . , yn ) with y1 = x − µ, y2 = x2 − x, . . . , yn = xn − x The absolute value of the Jacobian is 1/n. Also we have n n n X X X (xk − µ)2 = (xk − x)2 + n(x − µ)2 = yk2 + ny12 + (x1 − x)2 k=1 k=1 k=2 2 Pn Pn But k=1 (xk − x) = 0 and hence (x1 − x)2 = k=2 yk . Hence the density of (Y1 , . . . , Yn ) is " n #2 n X X n 1 2 1 f (y1 , . . . , yn ) = yk2 + yk exp − ny1 exp − 2 2 (2π)n/2 k=2 k=2 Because the joint density factorizes, we have shown that Y1 is independent of (Y2 , . . . , Yn ). Now (n − 1)S 2 = 2 Pn Pn Pn X)2 = (X1 − X)2 + k=2 Yk2 = + k=2 Yk2 . Hence X is independent of S 2 . k=2 Yk Pn k=1 (Xk − 5.13 Distribution of quadratic forms of normal random variables—relation to the χ2 distribution. Suppose X is an n-dimensional random vector with the non-singular N (µ, Σ) distribution. By proposition (5.7b) on page 68, know there exists a non-singular matrix Q such that QT Q = P and Z = Q(X − µ) ∼ N (0, I). Hence Pwe n T Z Z = i=1 Zi2 ∼ χ2n . Hence (X − µ)T QT Q(X − µ) ∼ χ2n We have shown that If X ∼ N (µ, Σ) then (X − µ)T Σ−1 (X − µ) ∼ χ2n We can generalize this result: Proposition(5.13a). Suppose X is an n-dimensional random vector with the non-singular N (µ, Σ) distribution. Then the random variable XT Σ−1 X has the non-central χ2n distribution with non-centrality parameter λ = µT Σ−1 µ. Proof. Take Z = Q−1 X where Σ = QQT . Then Z ∼ N (µZ , I) where µZ = Q−1 µ. Also XT Σ−1 X = XT (Q−1 )T Q−1 X = ZT Z = Z12 + · · · + Zn2 Hence XT Σ−1 X has the non-central χ2n distribution with non-centrality parameter µTZ µZ = µT Σ−1 µ. Page 72 §6 Aug 1, 2018(17:54) Bayesian Time Series Analysis 5.14 Independence of size and shape for the multivariate lognormal. See Chapter1:§22 on page 47. We say that the random vector X = (X1 , . . . , Xn ) ∼ logN (µ, Σ) iff ln(X) = ( ln(X1 ), . . . , ln(Xn ) ) ∼ N (µ, Σ). Proposition(5.14a). Suppose X = (X1 , . . . , Xn ) ∼ logN (µ, Σ). Suppose further that g1 : (0, ∞)n → (0, ∞) denotes the size variable g1 (x) = (x1 · · · xn )1/n Then g1 (X) is independent of every shape vector z(X) iff there exists c ∈ R such that cov[Yj , Y1 + · · · + Yn ] = c for all j = 1, 2, . . . , n, where Y = (Y1 , . . . , Yn ) = (ln(X1 ), . . . , ln(Xn ) ). Proof. By proposition(22.3b) on page 48,we need only prove g1 (X) is independent of one shape function. Consider the shape function z ∗ (x) = 1, x2/x1 , . . . , xn/x1 . Now g1 (X) is independent of z ∗ (X) iff (X1 · · · Xn )1/n is independent of 1, X2/X1 , . . . , Xn/X1 . This occurs iff Y1 +· · ·+Yn is independent (Y2 − Y1 , . . . , Yn − Y1 ). But the Y ’s are normal; hencePby proposition(5.8a) page 68, this occurs iff Pof Pon n n n cov[Yi − Y1 , j=1 Yj ] = 0 for i = 2, 3, . . . , n; and this occurs iff cov[Yi , j=1 Yj ] = cov[Y1 , j=1 Yj ] for i = 2, 3, . . . , n. This result implies many others. For example, suppose X1 , X2 , . . . , Xn are independent random variables with Xj ∼ logN (µj , σ 2 ) for j = 1, 2, . . . , n. Then Xj (X1 X2 · · · Xn )1/n is independent of max{X1 , X2 , . . . , Xn } and X1 + X2 + · · · + Xn (X1 X2 · · · Xn )1/n is independent of etc. max{X1 , X2 , . . . , Xn } Proposition(5.14a) leads to the following characterization of the lognormal distribution. Suppose X1 , X2 , . . . , Xn are independent positive non-degenerate random variables. Suppose g1 (0, ∞)n → (0, ∞) denotes the size variable g1 (x) = (x1 · · · xn )1/n Then there exists a shape vector z(X) which is independent of g1 (X) iff there exists σ > 0 such that every Xj ∼ logN (µj , σ 2 ). Proposition(5.14b). Proof. ⇐ Let Yj = ln(Xj ). Then Yj ∼ N (µj , σ 2 ); also Y1 , . . . , Yn are independent. Hence cov[Yj , Y1 + · · · + Yn ] = σ 2 for j = 1, 2, . . . , n. Hence result by previous proposition. ⇒ By proposition(22.3b), if there exists one shape vector which is independent of g1 (X), then all shape vectors are inde pendent of g1 (X). Hence 1, X2/X1 , . . . , Xn/X1 is independent of g1 (X) = (X1 · · · Xn )1/n . Hence Yk − Y1 is independent of Y1 + · · · + Yn for k = 2, . . . , n. Hence, by the Skitovich-Darmois theorem—see proposition (10.6b), every Yk is normal. 6 Exercises (exs-multivnormal.tex) 1. (a) Suppose the random vector X has the N (µ, Σ) distribution. Show that X − µ ∼ N (0, Σ). (b) Suppose X1 , . . . , Xn are independent with distributions N (µ1 , σ 2 ), . . . , N (µn , σ 2 ) respectively. Show that the random vector X = (X1 , . . . , Xn )T has the N (µ, σ 2 I) distribution where µ = (µ1 , . . . , µn )T . (c) Suppose X ∼ N (µ, Σ) where X = (X1 , . . . , Xn )T . Suppose further that X1 , . . . , Xn are uncorrelated. Show that X1 , . . . , Xn are independent. (d) Suppose X and Y are independent n-dimensional random vectors with X ∼ N (µX , ΣX ) and Y ∼ N (µY , ΣY ). Show that X + Y ∼ N (µX + µY , ΣX + ΣY ). 2. Suppose X ∼ N (µ, Σ) where µ= −3 1 4 ! and (a) Are (X1 , X3 ) and X2 independent? (b) Are X1 − X3 and X1 − 3X2 + X3 independent? (c) Are X1 + X3 and X1 − 2X2 − 3X3 independent? Σ= 4 0 −1 0 −1 5 0 0 2 ! 2 Multivariate Continuous Distributions §6 Page 73 Aug 1, 2018(17:54) 3. From linear regression. Suppose a is an n × m matrix with n ≥ m and rank(a) = m. Hence a has full rank. (a) Show that the m × m matrix aT a is invertible. (b) Suppose the n-dimensional random vector X has the N (µ, σ 2 I) distribution. Let B = (aT a)−1 aT and Y = BX m×n n×1 Show that Y ∼ N Bµ, σ 2 (aT a)−1 4. Suppose the 5-dimensional random vector Z = (Y, X1 , X2 , X3 , X4 ) is multivariate normal with finite expectation E[Z] = (1, 0, 0, 0, 0) and finite variance var[Z] = Σ where 1 1/2 1/2 1/2 1/2 1/2 1 1/2 1/2 1/2 Σ = 1/2 1/2 1 1/2 1/2 1 /2 1/2 1/2 1 1/2 1/2 1/2 1/2 1/2 1 1 Show that E[Y |X1 , X2 , X3 , X4 ] = 1 + 5 (X1 + X2 + X3 + X4 ). 5. Suppose X = (X1 , X2 , X3 ) has a non-singular multivariate normal distribution with E[Xj ] = µj and var[Xj ] = σj2 for j = 1, 2 and 3. Also ! 1 ρ12 ρ13 corr[X] = ρ12 1 ρ23 ρ13 ρ23 1 (a) Find E[X1 |(X2 , X3 )] and var[X1 |(X2 , X3 ). (b) Find E[(X1 , X2 )|X3 ] and var[(X1 , X2 )|X3 ]. 6. Continuation of proposition(5.14a). (a) Show that c ≥ 0. (b) Show that the size variable g1 (X) is independent of every shape vector z(X) iff the n-dimensional vector (1, . . . , 1 ) is an eigenvector of Σ. (c) Suppose c = 0. Show that (X1 · · · Xn )1/n is almost surely constant. 7. The Helmert matrix of order n. Consider the n × n matrix A with the following rows: 1 v1 = √ (1, 1, . . . , 1) n 1 v2 = √ (1, −1, 0, . . . , 0) 2 1 v3 = √ (1, 1, −2, 0, . . . , 0) 6 and in general for k = 2, 3, . . . , n: 1 (1, 1, . . . , 1, −(k − 1), 0, . . . , 0) k(k − 1) √ where the vector vk starts with (k − 1) terms equal to 1/ k(k − 1) and ends with (n − k) terms equal to 0. Check that A is orthogonal. This matrix is used in §5.12 on page 70. vk = √ 8. Suppose X = (X1 , . . . , Xn ) has the multivariate normal distribution with E[Xj ] = µj , var[Xj ] = σ 2 and corr[Xj , Xk ] = ρ|j−k| for all j and k in {1, 2, . . . , n}. Hence X ∼ N (µ, Σ) where 1 ρ ρ2 · · · ρn−1 µ 1 ρ · · · ρn−2 ρ . µ = .. and Σ = σ 2 . . . .. .. . .. .. . . . µ n−1 n−2 n−3 ρ ρ ρ ··· 1 Show that the sequence {X1 , X2 , . . . , Xn } forms a Markov chain. Page 74 §7 Aug 1, 2018(17:54) Bayesian Time Series Analysis 7 The bivariate t distribution 7.1 The bivariate t-distribution with equal variances. One possible version of the bivariate t-density is −(ν+2)/2 x2 − 2ρxy + y 2 1 p 1+ f(X,Y ) (x, y) = (7.1a) ν(1 − ρ2 ) 2π 1 − ρ2 for ν > 0, ρ ∈ (−1, 1), x ∈ R and y ∈ R. If x denotes the 2 × 1 vector (x, y), then an alternative expression which is equivalent to equation(7.1a) is −(ν+2)/2 ν (ν+2)/2 p ν + xT C−1 x 2π 1 − ρ2 1 ρ and C = . ρ 1 fX (x) = where C−1 = 1 1 − ρ2 1 −ρ −ρ 1 (7.1b) This distribution is called the tν (0, C) distribution. We shall see below in §7.3 on page 75 that C = corr[X] and X and Y have equal variances. 7.2.Characterization of the bivariate tν (0, C) distribution. The univariate tν -distribution is the distribution of p W/ν where Z ∼ N (0, 1), W ∼ χ2 and Z and W are independent. The generalisation to 2 dimensions is: Z ν Proposition(7.2a). Suppose Z = (Z1 , Z2 ) ∼ N (0, C) where 1 C= ρ ρ 1 and ρ ∈ (−1, 1) Suppose further that W ∼ χ2ν and Z and W are independent. Define X = (X, Y ) by Z X= (W/ν)1/2 Then X = (X, Y ) has the tν (0, C) density given in (7.1a). Proof. The density of (Z1 , Z2 , W ) is 2 h wi z − 2ρz1 z2 + z22 wν/2−1 1 p exp − 1 exp − ν 2 ν/2 2(1 − ρ ) 2 2 Γ( 2 ) 2π 1 − ρ2 .p .p W/ν , Y = Z W/ν and W = W . This is a 1 − 1 Consider the transformation to (X, Y, W ) where X = Z1 2 transformation and the absolute value of the Jacobian is ∂(x, y, w) ν ∂(z1 , z2 , w) = w Hence w wν/2 w x2 − 2ρxy + y 2 p − + 1 f(X,Y,W ) (x, y, w) = f (z1 , z2 , w) = ν exp ν 2 ν(1 − ρ2 ) 2 2 +1 πνΓ( ν2 ) 1 − ρ2 h wα i wν/2 x2 − 2ρxy + y 2 p = ν exp − where α = +1 (7.2a) 2 ν(1 − ρ2 ) 2 2 +1 πνΓ ν 1 − ρ2 f(Z1 ,Z2 ,W ) (z1 , z2 , w) = 2 Now using the integral of the χ2n density is 1 gives Z ∞ h xi n x 2 −1 exp − dx = 2n/2 Γ n/2 2 0 which implies ν2 +1 Z ∞ ν 2 ν2 +1 ν ν tα 2 ν t 2 exp − dt = Γ +1 = Γ 2 α 2 2 α 2 0 Integrating the variable w out of equation(7.2a) gives −(ν+2)/2 1 x2 − 2ρxy + y 2 p f(X,Y ) (x.y) = 1+ for (x, y) ∈ R2 ν(1 − ρ2 ) 2π 1 − ρ2 which is equation(7.1a) above. 2 Multivariate Continuous Distributions §8 Page 75 Aug 1, 2018(17:54) 7.3 Properties of the bivariate tν (0, C) distribution. • The marginal distributions. Both X and Y have t-distributions with ν degrees of freedom. The proof of this is left to exercise 1 on page 78. • Moments. E[X] = E[Y ] = 0 and var[X] = var[Y ] = ν/(ν − 2) for ν > 2. The correlation is corr[X, Y ] = ρ and the covariance is cov[X, Y ] = ρν/(ν − 2). The proof of these results is left to exercise 2 on page 78. It follows that ν ν 1 ρ var[X] = = C and corr[X] = C ρ 1 ν−2 ν−2 • If ρ = 0, then equation(7.1a) becomes −(ν+2)/2 x2 + y 2 1 1+ f(X,Y ) (x, y) = 2π ν Note that f(X,Y ) (x, y) 6= fX (x)fY (y) and hence X and Y are not independent even when ρ = 0. 7.4 Generalisation to non-equal variances. Suppose T1 = aX and T2 = bY where a 6= 0 and b 6= 0 and X = (X, Y ) ∼ tν (0, C). Thus 2 ν ν a 0 T1 a abρ = X and Σ = var[T] = T= = R 2 0 b T2 ν − 2 abρ b ν−2 where 4 R= a2 abρ abρ b2 and −1 R 1 = 2 2 a b (1 − ρ2 ) b2 −abρ −abρ a2 and |R−1 | = 1 a2 b2 (1 The absolute value of the Jacobian is |ab|. Substituting in equation(7.1a) on page 74 gives −(ν+2)/2 b2 t21 − 2ρabt1 t2 + a2 t22 1 p 1+ fT (t) = νa2 b2 (1 − ρ2 ) 2π|ab| 1 − ρ2 −(ν+2)/2 −(ν+2)/2 tT R−1 t ν (ν+2)/2 1 1 + ν + tT R−1 t = = 1/2 1/2 ν 2π|R| 2π|R| ν This is the tν (0, R) distribution. Note that var[T] = ν−2 R. − ρ2 ) 8 The multivariate t distribution 8.1 The density of the multivariate t-distribution, tν (0, I). If we put ρ = 0 in equation(7.1b) we see that if T ∼ tν (0, I) then −(ν+2)/2 ν (ν+2)/2 fT (t) = ν + tT t 2π Generalizing to p-dimensions leads to the following definition. Definition(8.1a). The p-dimensional random vector T has the t-distribution tν ( 0 , I ) iff T has density p×1 p×1 p×p 1 f (t) ∝ (ν+p)/2 ν + tT t where ν ∈ R and ν > 2. An alternative expression is: f (t1 , . . . , tp ) = 4 κ ν+ t21 + · · · + t2p (ν+p)/2 In general, the inverse of the 2 × 2 symmetric matrix 1 a c b −c is c b ab − c2 −c a provided ab 6= c2 . Page 76 §8 Aug 1, 2018(17:54) Bayesian Time Series Analysis The constant of proportionality, κ, can be determined by integration. Integrating out tp gives Z ∞ Z ∞ dtp dtp = 2κ f (t1 , . . . , tp−1 ) = κ (ν+p)/2 2 + · · · + t2 (ν+p)/2 0 −∞ ν + t2 + · · · + t2 ν + t p p 1 1 Z ∞ dtp 2 2 = 2κ (ν+p)/2 where α = ν + t1 + · · · + tp−1 0 α + t2p Z ∞ dtp 2κ = (ν+p)/2 (ν+p)/2 α 0 1 + t2p /α √ Z ∞ √ √ dx 2κ α = (ν+p)/2 √ (ν+p)/2 where x = tp ν + p − 1/ α α ν+p−1 0 1 + x2 /(ν + p − 1) Using the standard result that √ −(n+1)/2 Z ∞ √ nΓ 1/2 Γ n/2 t2 2 1+ dt = nB(1/2, n/2) = n Γ (n+1)/2 0 implies κ √ πΓ( (ν+p−1)/2) α(ν+p−1)/2 Γ( (ν+p)/2) √ κ π Γ( (ν+p−1)/2) 1 = 2 (ν+p) Γ( /2) [ν + t1 + · · · + t2p−1 ](ν+p−1)/2 f (t1 , . . . , tp−1 ) = By induction f (t1 ) = κπ (p−1)/2 Γ( (ν+1)/2) (ν+1)/2 Γ( (ν+p)/2) ν + t21 and so κ= ν ν/2 Γ( (ν+p)/2) π p/2 Γ( ν/2) It follows that the density of the p-dimensional tν (0, I) is f (t) = 1 ν ν/2 Γ( (ν+p)/2) π p/2 Γ( ν/2) ν + tT t (ν+p)/2 (8.1a) 8.2 Characterization of the tν (0, I) distribution. Suppose Z1 , Z2 , . . . , Zp are i.i.d. with the N (0, 1) distribution and W has the χ2ν distribution. Suppose further that Z = (Z1 , Z2 , . . . , Zp ) and W are independent. Define T = (T1 , T2 , . . . , Tp ) by Z T= (W/ν)1/2 Then T has the density in equation(8.1a). Proposition(8.2a). Proof. See exercise 3 on page 78. 8.3 Properties of the tν (0, I) distribution. • TT T/p has the F (p, ν) distribution—see exercise 6 on page 78. • The contours of the distribution are ellipsoidal (the product of independent t distributions does not have this property). • The marginal distribution of an r-dimensional subset of T has the tν (0, I) distribution. In particular, each Ti has the tν distribution. These results follow immediately from the characterization in §8.2. • E[T] = 0 and var[T] = E[TTT ] = Finally, corr[T] = I. ν ν−2 I for ν > 2. (Because W ∼ χ2ν implies E[1/W ] = 1/(ν − 2).) 2 Multivariate Continuous Distributions §8 Page 77 Aug 1, 2018(17:54) 8.4 The p-dimensional t-distribution: tν (m, C). Here C is real, symmetric, positive definite p × p matrix. The Cholesky decomposition implies there exists a real and nonsingular L with C = LLT . Let V = m + L T where T ∼ tν (0, I) p×1 p×1 p×p p×1 ν ν Then E[V] = m and var(V) = Lvar(T)LT = ν−2 LLT = ν−2 C. See exercise 4 on page 78 for the proof of the result TT T = (V − m)T C−1 (V − m) (8.4a) It follows that V has density: ν ν/2 Γ (ν+p)/2 κ f (v) = and |L| = |C|1/2 (8.4b) (ν+p)/2 where κ = p/2 Γ ν/ T −1 π 2 |L| ν + (v − m) C (v − m) A random variable which has the density given in equation(8.4b) is said to have the tν (m, C) distribution. Definition(8.4a). Suppose C is real, symmetric, positive definite p × p matrix and m is a p × 1 vector in Rp . Then the p-dimensional random vector V has the tν (m, C) distribution iff V has the density 1 f (v) ∝ (ν+p)/2 ν + (v − m)T C−1 (v − m) It follows that ν E[V] = m and var[V] = C ν−2 and the constant of proportionality is given in equation(8.4b). 8.5 Linear transformation of the tν (m, C) distribution. Suppose T ∼ tν (m, C); thus m is the mean vector and ν ν−2 C is the covariance matrix of the random vector T. Suppose V = a + AT where A is non-singular. ν It follows that T = A−1 (V − a), E[V] = a + Am and var[V] = ν−2 ACAT . T Let m1 = a + Am and C1 = ACA . Then V has the tν (m1 , C1 ) distribution—see exercise 5 on page 78. 8.6 Characterization of the tν (m, C) distribution. Proposition(8.6a). Suppose Z has the non-singluar multivariate normal distribution N (0, Σ) and W has the χ2ν distribution. Suppose further that Z and W are independent. Then T = m + Z/(W/ν)1/2 has the tν (m, Σ) distribution. Proof. Because Z has a non-singular distribution, Σ is positive definite and there exists a symmetric non-singular Q with Σ = QQ. Let Y = Q−1 Z. Then var[Y] = Q−1 var[Z](Q−1 )T = I. So Y ∼ N (0, I). Hence Y ∼ tν (0, I) T1 = p W/ν Using §8.5 gives T = m + QT1 ∼ tν (m, Σ) as required. 8.7 Summary. • The bivariate t-distribution tν (0, R). This has E[T] = 0 and var[T] = fT (t) = ν ν−2 R. The density is −(ν+2)/2 ν (ν+2)/2 ν + tT R−1 t 1/2 2π|R| Particular case: −(ν+2)/2 t21 − 2ρt1 t2 + t22 ν 1 ρ 1+ where var[T] = fT (t) = p ν(1 − ρ2 ) ν−2 ρ 1 2π 1 − ρ2 ν • The p-dimensional t-distribution tν (m, R). This has E[T] = m and var[T] = ν−2 R. The density is ν ν/2 Γ ν+p 1 f (t) = p/2 ν2 π Γ 2 |R|1/2 ν + (v − m)T R−1 (v − m) (ν+p)/2 1 • Characterization of the t-distribution. Suppose Z ∼ N (0, Σ) and W has the χ2ν distribution. Suppose further that Z and W are independent. Then T = m + Z/(W/ν)1/2 has the tν (m, Σ) distribution. Page 78 §9 Aug 1, 2018(17:54) 9 Exercises Bayesian Time Series Analysis (exs-t.tex.tex) 1. Suppose T has the bivariate t-density given in equation(7.1a) on page 74. Show that both marginal distributions are the tν -distribution and hence have density given in equation(16.1b) on page 34: −(ν+1)/2 1 t2 √ f (t) = 1 + for t ∈ R. ν B( 1/2, ν/2) ν 2. Suppose T has the bivariate t-density given in equation(7.1a) on page 74 and ν > 2. (a) Find E[X] and var[X]. (b) Find cov[X, Y ] and corr[X, Y ]. 3. Prove proposition(8.2a) on page 76: Suppose Z1 , Z2 , . . . , Zp are i.i.d. with the N (0, 1) distribution and W has the χ2ν distribution. Suppose further that Z = (Z1 , Z2 , . . . , Zp ) and W are independent. Define T = (T1 , T2 , . . . , Tp ) by T = Z (W/ν)1/2 . Then T has the following density f (t) = ν ν/2 Γ( (ν+p)/2) 1 π p/2 Γ( ν/2) ν + tT t(ν+p)/2 4. Prove equation(8.4a) on page 77: TT T = (V − m)T C−1 (V − m). 5. See §8.5 on page 77. Suppose T ∼ tν (m, C) and V = a + AT where A is non-singular. Prove that V ∼ tν (m1 , C1 ) where m1 = a + Am and C1 = ACAT . 6. Suppose the p-variate random vector T has the tν (0, I) distribution. Show that TT T/p has the F (p, ν) distribution. APPENDIX Answers Chapter 1 Section 3 on page 7 (exs-basic.tex) 1. (a) A = £1,000 × 1.04 × (1 + V1 ) × (1 + V2 ) = £1,000 × 1.04 × (1.04 + U1 )(1.04 + 2U2 ). Hence E[A] = £1 000 × 1.043 = £1,124.864 or £1,124.86. (b) For this case 1,000 1,000 1 1 C= and E[C] = E E 1.04(1.04 + U1 )(1.04 + 2U2 ) 1.04 1.04 + U1 1.04 + 2U2 Now 0.01 Z 0.01 1 du = 50(ln 1.05 − ln 1.03) E = 50 = 50 ln(1.04 + u) 1.04 + U1 −0.01 1.04 + u −0.01 0.01 Z 0.01 1 du 50 E = 25(ln 1.06 − ln 1.02) = 50 = ln(1.04 + 2u) 1.04 + 2U2 1.04 + 2u 2 −0.01 −0.01 Hence E[C] = 889.133375744 or £889.13. 2. Clearly −2a < W < 2a. For w ∈ (−2a, 2a) we have Z fW (w) = fX (x)fY (w − x) dx Now −a < x < a and −a < w − x < a; hence w − a < x < w + a. Hence Z min(a,w+a) 1 fW (w) = fX (x)fY (w − x) dx = 2 [min(a, w + a) − max(−a, −a + w)] 4a max(−a,−a+w) 1 |w| (2a − w)/4a2 if w > 0 = = 1− (w + 2a)/4a2 if w < 0 2a 2a 1/2a −2a 0 2a Figure(1.0a). The shape of the triangular density 3. Clearly 0 ≤ Y < 1; also dy dx (wmf/triangulardensity,60mm,21mm) = 4x3 = 4y 3/4 . X fX (x) X fX (x) X 1 1 fY (y) = = = = 3/4 3/4 3/4 dy/dx| | 4y 8y 4y x x x 4. Now (X − 1)2 ≥ 0; hence X 2 + 1 ≥ 2X. Because X > 0 a.e., we have X + 1/X ≥ 2 a.e. Hence result. 5. For parts (a) and (b): Z ∞ Z ∞ Z ∞Z ∞ Z ∞Z t rxr−1 [1 − F (x)] dx = rxr−1 f (t) dt dx = rxr−1 f (t)dx dt = tr f (t) dt = E[X] 0 x=0 t=x t=0 x=0 t=0 6. (a) Jensen’s Inequality is as follows: suppose X is a random variable with a finite expectation and φ : R → R is a convex function. Then φ (E[X]) ≤ E [φ(X)]. In particular, suppose φ(x) = 1/x, then φ is a convex function on (0, ∞). Hence if X is positive random variable with finite expectation, then 1/E[X] ≤ E[1/X]. Trivially, the result is still true if E[X] = ∞. Hence E 1/Sn ≥ 1/(nµ). (b) Z ∞ Z ∞ Z ∞ n 1 E[e−tX ] dt = E[e−t(X1 +···+Xn ) ] dt = E e−t(X1 +···+Xn ) dt = E S n 0 0 0 by using the Fubini-Tonelli theorem that the order of integration can be changed for a non-negative integrand. 7. (a) The arithmetic mean-geometric mean inequality gives √ x1 + · · · + xn ≥ n x1 · · · xn for all x1 > 0, . . . , xn > 0. n Hence 1 1 ≤ 1/n 1/n x1 + · · · + xn nx1 · · · xn Using independence gives " #!n 1 1 1 E ≤ E 1/n Sn n X1 Bayesian Time Series Analysis by R.J. Reed Aug 1, 2018(17:54) Answers Page 79 Page 80 Answers 1§3 Aug 1, 2018(17:54) Bayesian Time Series Analysis Now for x > 0 we have 1 1/x if 0 < x ≤ 1; 1 if x ≥ 1. x1/n Hence E[1/Sn ] is finite. (b) Because they have identical distributions, E[X1 /Sn ] = · · · = E[Xn /Sn ]. Hence Sn X1 + · · · + Xn X1 Sj X1 + · · · + Xj X1 j 1=E =E = nE Hence E =E = jE = Sn Sn Sn Sn Sn Sn n √ 8. (a) Recall |cov[X1 , X2 ]| ≤ var[X1 ] var[X2 ]; hence cov[ X/Y , Y ] is finite. Hence E[X] = E[ X/Y ] E[Y ]. Also 2 2 2 E[( X/Y )X] = E[( X /Y 2 )Y ] = E[ X /Y 2 ] E[Y ] because X /Y 2 is independent of Y . Hence 2 2 0 = cov[ X/Y , X] = E[( X/Y )X] − E[ X/Y ] E[X] = E[ X /Y 2 ] E[Y ] − {E[ X/Y ]} E[Y ] = var[ X/Y ]E[Y ] As E[Y ] > 0, it follows that var[ X/Y ] = 0 as required. (b) Clearly ln( Y /X ) = ln(Y ) − ln(X) is independent of ln(X). Using characteristic functions, φln(Y /X) (t) φln(X) (t) = φln(Y ) (t). Also ln( X/Y ) = ln(X) − ln(Y ) is independent of ln(Y ). Hence φln(X/Y ) (t) φln(Y ) (t) = φln(X) (t). Hence φln(Y /X) (t)φln(X/Y ) (t) = 1. But for any characteristic function |φ(t)| ≤ 1. Hence |φln(Y /X) (t)| = 1 everywhere. This implies1 ln(Y /X) is constant almost everywhere and this establishes the result. 9. Now 2 2 E Y − Yb = E Y − E(Y |X) + E(Y |X) − Yb =E h Y − E(Y |X) 2 i ≤ + 2E h i 2 Y − E(Y |X) E(Y |X) − Yb + E E(Y |X) − Yb By equation(1.1a) on page 3 and the law of total expectation, the first term is E[var(Y |X)]. Applying the law of total expectation to theh second term gives i n h io 2E Y − E(Y |X) E(Y |X) − Yb = 2E E Y − E(Y |X) E(Y |X) − Yb |X n o = 2E E(Y |X) − Yb × 0 = 0 Hence E Y − Yb 2 = E[var(Y |X)] + E E(Y |X) − Yb 2 which is minimized when Yb = E(Y |X). 10. (a) For the second result, just use E(XY |X) = XE(Y |X) = aX + bX 2 and take expectations. Clearly cov[X, Y ] = b var[X]. Then E(Y |X) = a + bX = a + bµX + b(X − µX ) = µY + b(X − µX ). Then use b = cov(X, Y )/var(X) = i2 h 2 2 Y (X − µX ) = and b = ρσY /σX . Finally E Y − E(Y |X) = E Y − µY − ρ σσX ρσY /σX . (b) var E(Y |X) = b2 σX Y cov[X, Y ] = σY2 + ρ2 σY2 − 2ρ2 σY2 as required. σY2 + ρ2 σY2 − 2ρ σσX (c) We have µX = c + dµY and µY = a + bµX . Hence µX = (c + ad)/(1 − bd) and µY = (a + bc)/(1 − bd). Now E[XY ] = cµY + dE[Y 2 ] and E[XY ] = aµX + bE[X 2 ]. Hence cov[X, Y ] = dvar[Y ] and cov[X, Y ] = bvar[X]. 2 Y = b/d. Finally ρ = cov[X, Y ]/(σX σY ) = d σσX and hence ρ2 = d2 b/d = bd. Hence σY2 /σX 2 11. Let g(a, b) = E ( Y − a − bX ) = E[Y 2 ] − 2aµY + a2 − 2bE[XY ] + b2 E[X 2 ] + 2abµX . Hence we need to solve ∂g(a, b) ∂g(a, b) = −2µY + 2a + 2bµX = 0 and = −2E[XY ] + 2bE[X 2 ] + 2aµX = 0 ∂a ∂b This gives E[XY ] − µX µY σY σY b= =ρ and a = µY − bµX = µY − ρ µX 2 2 σX σX E[X ] − µX 12. Z 1 2 2 6 (x + y)2 dy = (x + 1)3 − x3 = (3x2 + 3x + 1) for x ∈ [0, 1]. fX (x) = 7 7 0 7 Similarly 2 fY (y) = (3y 2 + 3y + 1) for y ∈ [0, 1]. 7 Hence 3(x + y)2 3(x + y)2 fX|Y (x|y) = 2 and fY |X (y|x) = 2 for x ∈ [0, 1] and y ∈ [0, 1]. 3y + 3y + 1 3x + 3x + 1 and so the best predictor of Y is 2 2 1 Z 1 x y 3 3 y 3 y 4 E[Y |X = x] = 2 (x2 + 2xy + y 2 )y dy = 2 + 2x + 3x + 3x + 1 0 3x + 3x + 1 2 3 4 0 2 2 3 x 2x 1 1 = 2 + + = 6x + 8x + 3 3x + 3x + 1 2 3 4 4(3x2 + 3x + 1) 1 See for example, pages 18–19 in [L UKACS(1970)] and exercise 4 on page 298 in [A SH(2000)]. Appendix Aug 1, 2018(17:54) 9 14 , (b) Now µX = µY = 2 σX = σY2 = E[X 2 2 E[X 2 ] = 9 101 ] − 14 2 = 210 Z 1Z 1 7 E[XY ] = 6 (3x4 + 3x3 + x2 )dx = 0 199 2940 . = Also Z 1 (Z 2 xy(x + y) dxdy = y y=0 0 0 Z − 81 196 R1 2 7 1 = y=0 y 2y 2 y 3 + + 4 3 2 dy = Answers 1§3 Page 81 2 7 3 5x 5 x=1 + 43 x4 + 13 x3 x=0 = ) 1 2 x(x + y) dx Z 1 dy = x=0 y y=0 2 7 3 5 + 3 4 1 2y y 2 + + 4 3 2 + 1 3 = 101 210 and dy 1 2 1 17 + + = 8 9 8 36 2 9 5 5 2940 25 and cov[X, Y ] = 17 42 − 142 = − 588 and ρ = − 588 × 199 = − 199 . Hence the best linear predictor is σY 9 144 25 µY + ρ (X − µX ) = (1 − ρ) + ρX = − X σX 14 199 199 (c) See figure (1.0a) below. Hence E[XY ] = 17 42 0.75 Best linear predictor Best predictor 0.70 0.65 0.60 0.0 0.2 0.4 0.6 0.8 1.0 Figure(1.0a). Plot of best predictor (solid line) and best linear predictor (dashed line) for exercise 12. (wmf/exs-bestlin,72mm,54mm) n! j−1 (1 − x)n−j . Recall B(j, n − (j−1)!(n−j)! x 1 j−1 f (x) = B(j,n−j+1) x (1 − x)n−j for x ∈ (0, 1) 13. (a) Using equation(2.3b) on page 5 gives the density f (x) = Γ(j)Γ(n−j+1) Γ(n+1) (j−1)!(n−j)! . n! j + 1) = = Hence the density of Xj:n is which is the Beta(j, n − j + 1) distribution. (b) E[Xj:n ] = j/(n + 1) by using the standard result of the expectation of a Beta distribution. 14. ⇐ The joint density of (X1:2 , X2:2 ) is g(y1 , y2 ) = 2f (y1 )f (y2 ) = 2λ2 e−λ(y1 +y2 ) for 0 < y1 < y2 . Now consider the ∂(w,y) transformation to (W, Y ) = (X2:2 − X1:2 , X1:2 ). The absolute value of the Jacobian is ∂(y = | − 1| = 1. Hence 1 ,y2 ) f(W,Y ) (w, y) = 2λ2 e−λ(w+y+y) = 2λe−2λy λe−λw = fY (y)fW (w) where the density of X1:2 is fY (y) = 2λe−2λy . The fact that the joint density is the product of the marginal densities implies W and Y are independent. (x+y) ⇒ P[X2:2 − X1:2 > y|X1:2 = x] = P[X2:2 > x + y|X1:2 = x] = 1−F 1−F (x) and this is independent of x. Taking x = 0 gives 1 − F (x + y) = (1 − F (x))(1 − F (y)) and F is continuous. Hence there exists λ > 0 with F (x) = 1 − e−λx . 15. By equation(2.2a) on page 4, the density of the vector is (X1:n , X2:n , . . . , Xn:n ) is g(x1 , . . . , xn ) = n!f (x1 ) · · · f (xn ) for 0 ≤ x1 ≤ x2 · · · ≤ xn . The transformation to (Y1 , Y2 , . . . , Yn ) has Jacobian with absolute value ∂(y1 , . . . , yn ) 1 ∂(x1 , . . . , xn ) = y n−1 1 Hence for y1 ≥ 0 and 1 ≤ y2 ≤ · · · ≤ yn , the density of the vector (Y1 , Y2 , . . . , Yn ) is h(y1 , . . . , yn ) = n!y1n−1 f (y1 )f (y1 y2 )f (y1 y3 ) · · · f (y1 yn ) (b) Integrating yn from yn−1 to ∞ gives h(y1 , . . . , yn−1 ) = n!y1n−2 f (y1 )f (y1 y2 )f (y1 y3 ) · · · f (y1 yn−1 ) 1 − F (y1 yn−1 ) Then integrating yn−1 over yn−2 to ∞ gives 2 1 − F (y1 yn−2 ) n−3 h(y1 , . . . , yn−2 ) = n!y1 f (y1 )f (y1 y2 )f (y1 y3 ) · · · f (y1 yn−2 ) 2 and by induction [1 − F (y1 y2 )]n−2 (n − 2)! h(y1 ) = nf (y1 ) [1 − F (y1 )]n−1 h(y1 , y2 ) = n!y1 f (y1 )f (y1 y2 ) as required. Page 82 Answers 1§5 Aug 1, 2018(17:54) Bayesian Time Series Analysis Chapter 1 Section 5 on page 14 (exs-uniform.tex) 1. n Z b−a 2 a+b 1 v n dv x− dx = a−b 2 b − a a 2 n n+1 n+1 (b − a) 1 − (−1)n+1 (b − a) − (a − b) = = 2n+1 (b − a)(n + 1) (n + 1)2n+1 E[X − µ)n ] = Z 1 b−a b 2. P[− ln X ≤ x] = P[ln X ≥ −x] = P[X ≥ e−x ] = 1 − e−x . Hence Y ∼ exponential (1). R1 3. (a) Distribution function: FZ (z) = P[XY ≤ z] = 0 P[X ≤ z/y] dy. But P[X ≤ z/y] = 1 if y ≤ z and = z/y if z < y. Rz R1 Hence FZ (z) = 0 dy + z z/y dy = z(1 − ln z) for 0 < z < 1. Density: By differentiating FZ (z) we get fz (z) = − ln z for 0 < z < 1. Alternatively, consider the transformation ∂(z,v) Z = XY and V = Y . Then 0 < Z < 1 and 0 < V < 1. The absolute value of the Jacobian is ∂(x,y) = y = v. Hence f(Z,V ) (z, v) = f(X,Y ) ( z/v, v)/v. Hence Z 1 Z 1 fX ( z/v) dv fZ (z) = dv = = − ln z for 0 < z < 1. v v=0 v=z v Pn (b) Now − ln Pn = j=1 Yj where Y1 , . . . , Yn are i.i.d. with the exponential (1) distribution, after using the result in exercise 2. Hence Z = − ln Pn ∼ Gamma(n, 1) and Z has density z n−1 e−z /Γ(n) for z > 0. Transforming back to Pn shows the density of Pn is f (x) = (− ln x)n−1 /Γ(n) = (ln 1/x)n−1 /Γ(n) for x ∈ (0, 1). 4. (a) First n = 2. Now the density of (X1 , X2 ) is 1 for 0 < x2 < x1 < 1. f(X1 ,X2 ) (x1 , x2 ) = fX2 |X1 (x2 |x1 )fX1 (x1 ) = x1 Z 1 1 1 fX2 (x2 ) = for 0 < x2 < 1. f(X1 ,X2 ) (x1 , x2 ) dx1 = − ln x1 = ln x2 x1 =x2 x1 =x2 Assume true for n − 1; to prove for n. Now Xn ∼ U (0, Xn−1 ); hence 1 f(Xn−1 ,Xn ) (xn−1 , xn ) = fXn |Xn−1 (xn |xn−1 )fXn−1 (xn−1 ) = fX (xn−1 ) xn−1 n−1 n−2 1 ln 1/xn−1 for 0 < xn < xn−1 < 1. = xn−1 (n − 2)! and hence Z 1 fXn (xn ) = xn−1 =xn n−2 n−1 ln 1/xn−1 ln 1/xn dxn−1 = (n − 2)! (n − 1)! 1 xn−1 for 0 < xn < 1. As required. (b) Now Xn = Xn−1 Z where Z ∼ U (0, 1). Similarly, by induction, Xn = X1 Z1 · · · Zn−1 which is the product of n random variables with the U (0, 1) distribution. Hence result by part (b) of exercise 3 5. H(X) = − Z a b 1 ln b−a 1 b−a dx = ln(b − a) 6. (a) Z 1 b min{b, v} − max{0, v − a} fX (v − y)dy = b ab 0 0 1 by using fX (v − y) = /a when 0 < v − y < a; i.e. when v − a < y < v. Suppose a < b. Then if 0 < v < a; v/ab fV (v) = 1/b if a < v < b; (a + b − v)/ab if b < v < a + b. (b) Now −a < W < b. Z b Z 1 b min{a, b − w} − max{0, −w} fW (w) = fY (y + w)fX (y) dy = fY (y + w) dy = a ab 0 0 1 by using fY (y + w) = /b when 0 < y + w < b; i.e. when −w < y < b − w. Suppose a < b. Then (a + w)/ab if −a < w < 0; fW (w) = 1/b if 0 < w < b − a; (b − w)/ab if b − a < w < b. Z fV (v) = b fX (v − y)fY (y) dy = Appendix Aug 1, 2018(17:54) fX (x) fX (x) ...... ......... ... .... .. ... ... ... ............................................................................ ... .... ...... .. . ... . . .. .. . . ..... ... . . .. ... .. ... . . ... ... ... ... . . .. . . ... ... . . . . ... . ... . . . ... ... ... . . . ... 1b . . ... .... ... . . ... ... ... . . ... ... ... . . ... . . ... .... ... . . ...... .. . . . ......................................................................................................................................................................................................... . ... . ...... ......... ... .... .. ... ... ... ............................................................................ ... .... ...... .. . ... . . .. .. . . ..... ... . . .. ... .. ... . . ... ... ... ... . . .. . . ... ... . . . . ... . ... . . . ... ... ... . . . ... 1b . . ... .... ... . . ... ... ... . . ... ... ... . . ... . . ... .... ... . . ...... .. . . . ......................................................................................................................................................................................................... . ... . / / 0 7. Now Answers 1§5 Page 83 a −a b−a 0 b b a+b x x Figure(2.0a). Plot of density of V = X + Y (left) and density of W = Y − X (right). ( a−t b−t P[V ≥ v] = P[X ≥ v]P[Y ≥ v] = ( FV (t) = fV (t) = 1− (a−t)(b−t) ab a (PICTEX) for 0 ≤ t ≤ min{a, b}; b for t ≥ min{a, b}. 0 if 0 ≤ t ≤ min{a, b}; t ≥ min{a, b}. 1 a + b − 2t ab for 0 ≤ t ≤ min{a, b}. Finally Z min{a,b} Z a P[Y > x] b−x 1 1 − a/2b if a ≤ b; dx = dx + dx = 1 − b/2a if a > b. a ab 0 0 min{a,b} a 8. Let V denote the arrival time. If you take the bus on route 2 then E[V ] = V = t0 + α + β. If you wait for a bus on route 1, then E[V ] = t0 + α + E[X2 − t0 |X2 > t0 ]. But the distribution of (X2 − t0 |X2 > t0 ) is U (0, a − t0 ), and hence E[X2 − t0 |X2 > t0 ] = (a − t0 )/2. Hence route 1 is faster if (a − t0 )/2 < β and route 2 is faster if (a − t0 )/2 > β. 9. For w ∈ (0, 1) we have ∞ ∞ X X P[W ≤ w] = P[W ≤ w, bU + V c = k] = P[U + V ≤ w + k, bU + V c = k] Z a P[V = X] = P[Y > X] = = = = k=−∞ ∞ X k=−∞ P[V ≤ w + k − U, bU + V c = k] = k=−∞ ∞ Z 1 X k=−∞ u=0 ∞ Z 1 X k=−∞ P[k ≤ u + V ≤ w + k] du = Z u=0 fV (v)dv du = v=k−u 1 k=−∞ u=0 ∞ Z 1 X k=−∞ ∞ X w+k−u ∞ Z X u=0 P[u + V ≤ w + k, k ≤ u + V < k + 1] du P[k − u ≤ V ≤ w + k − u] du In k=−∞ where Z k+w min{w + k − v, 1} − max{0, k − v} fV (v) dv In = v=k−1 k−1+w Z = v=k−1 Z k+w [1 − k + v]fV (v)dv + "Z =w Z k Z v=k−1+w v=k−1+w v=k−1 vfV (v)dv − " Z k k=−n Z "Z n+w In = w v=−n−1 k+w vfV (v)dv + v=k fV (v)dv − (k − 1) −n−1+w fV (v)dv + v=−n−1+w Z (w + k − v)fV (v)dv # k+w v=k and hence n X v=k k−1+w fV (v) dv + k+w wfV (v)dv + vfV (v)dv − " Z n Z Z fV (v) dv v=k−1 # n+w vfV (v)dv + v=n n+w n # k−1+w fV (v)dv − (−n − 1) Z # −n−1+w fV (v)dv v=−n−1 R n+w R n+w Now E|V | < ∞; hence E[V + ] < ∞. Hence n vfV (v) dv → 0 as n → ∞. In turn, this implies n n fV (v) dv → 0 as n → ∞. Similarly for V − and integrals. R∞ Pn the other two Hence P[W ≤ w] = limn→∞ k=−n In = w v=−∞ fV (v) dv = w as required. Page 84 Answers 1§5 Aug 1, 2018(17:54) Bayesian Time Series Analysis 10. (a) P[V ≥ t] = (1 − t)2 . Hence FV (t) = P[V ≤ t] = 2t − t2 and fV (t) = 2(1 − t) for 0 ≤ t ≤ 1. Also E[V ] = 1/3. FW (t) = P[W ≤ t] = t2 and fW (t) = 2t for 0 ≤ t ≤ 1. Also E[W ] = 2/3. (b) For v < w we have P[V ≤ v, W ≤ w] = P[W ≤ w] − P[W ≤ w, V > v] = w2 − (w − v)2 = v(2w − v); whilst for v > w we have P[V ≤ v, W ≤ w] = P[W ≤ w] = w2 . Hence ∂2 f(V,W ) (v, w) = P[V ≤ v, W ≤ w] = 2 for 0 ≤ v < w ≤ 1. ∂v∂w (c) For v < w we have P[W ≤ w|V ≤ v] = (2w − v)/(2 − v) and for v > w we have P[W ≤ w|V ≤ v] = w2 /(2v − v 2 ). Hence 2/(2 − v) if v < w; fW (w|V ≤ v) = 2w/(2v − v 2 ) if v > w. Z v Z 1 2 2 2v 2 1 − v2 3 − v2 2 E[W |V ≤ v] = w dw + w dw = + = 2v − v 2 w=0 2 − v w=v 3(2 − v) 2 − v 3(2 − v) 2 Note that E[W |V ≤ 1] = /3 = E[W ] and E[W |V ≤ 0] = 1/2 = E[X]. 11. (a) Without loss of generality, suppose we measure distances clockwise from some fixed origin O on the circle. Let D denote the length of the interval (X1 , X2 ). Then 0 < D < 1 and X2 − X1 if X2 > X1 ; 0 if 1 > X2 − X1 > 0; D= = X2 − X1 + 1 − X1 + X2 if X1 > X2 . 1 if −1 < X2 − X1 < 0. The first line corresponds to points in the clockwise order O → X1 → X2 and the second line to points in the clockwise order O → X2 → X1 . So for y ∈ (0, 1) we have P[D ≤ y] = P[0 ≤ X2 − X1 ≤ y] + P[X2 − X1 ≤ y − 1] = P[X1 ≤ X2 ≤ min{X1 + y, 1}] + P[X2 ≤ X1 + y − 1] "Z # "Z # Z 1 1−y 1 = y fX1 (x1 ) dx1 + (1 − x1 )fX1 (x1 ) dx1 + (x1 + y − 1)fX1 (x1 )dx1 x1 =0 x1 =1−y x1 =1−y 1 (1 − y)2 1 (1 − y)2 + y(y − 1) + − =y = y(1 − y) + y − + 2 2 2 2 as required. (b) Without loss of generality, take Q to be the origin and measure clockwise. So we have either Q → X1 → X2 or Q → X2 → X1 and both of these lead to the same probability. Consider the first case: Q → X1 → X2 . Then for t ∈ (0, 1) we have Z t Z t t2 P[L ≤ t, Q → X1 → X2 ] = P[X2 ≥ 1 − (t − x1 )]fX1 (x1 ) dx1 = (t − x1 )dx1 = 2 x1 =0 x1 =0 Similarly for Q → X2 → X1 . Hence P[L ≤ t] = t2 and fl (t) = 2t for t ∈ (0, 1). Finally, E[L] = 2/3. n n 2 2 12. For x ∈ (0, r) we have P[D > x] = 1 − x /r2 and hence P[D ≤ x] = 1 − 1 − x /r2 . Hence n−1 2nx x2 fD (x) = 2 1 − 2 for x ∈ (0, r). r r n−1 Z r Z 1 x2 2nx2 E[D] = 1 − dx = 2nrv 2 (1 − v 2 )n−1 dv r2 r2 0 0 Z 1 Γ 3/2 Γ(n) Γ 3/2 Γ(n + 1) 1/2 n−1 =r = nr u (1 − u) du = nr Γ n + 3/2 Γ n + 3/2 0 13. Now f(X1 ,X2 ) (x1 , x2 ) is constant on the disc. Hence f(X1 ,X2 ) (x1 , x2 ) = 1/πa2 for x21 + x22 ≤ a2 . Hence q Z √a2 −x2 2 a2 − x21 1 1 fX1 (x1 ) = dx1 = for −a < x1 < a. √ 2 πa2 x2 =− a2 −x21 πa 14. Let Yj = Xj −a. Then Y1 , Y2 , . . . , Yn are i.i.d. random variables with the U (0, b−a) distribution. Also P[X1 +· · ·+Xn ≤ t] = P[Y1 + · · · + Yn < −na]. Hence n X n 1 k n (−1) (x − na − kb + ka)+ for all x ∈ R and all n = 1, 2, . . . . Fn (x) = n (b − a) n! k k=0 n X n−1 1 k n fn (x) = (−1) (x − na − kb + ka)+ for all x ∈ R and all n = 2, 3, . . . . n (b − a) (n − 1)! k k=0 Appendix Aug 1, 2018(17:54) Answers 1§7 Page 85 R1 15. (a) By equation(4.4a) and the standard result 0 xα−1 (1 − x)β−1 dx = Γ(α)Γ(β)/Γ(α + β) which is derived in §14.1 on page 31, we have Z 1 Z 1 n−1 k n! E[Xk:n ] = n t (1 − t)n−k dt = tk (1 − t)n−k dt k−1 (n − k)!(k − 1)! 0 0 n! Γ(k + 1)Γ(n − k + 1) k = = (n − k)!(k − 1)! Γ(n + 2) n+1 Similarly Z 1 Z 1 n − 1 k+1 n! 2 tk+1 (1 − t)n−k dt E[Xk:n ]= n t (1 − t)n−k dt = (n − k)!(k − 1)! k − 1 0 0 n! Γ(k + 2)Γ(n − k + 1) (k + 1)k = = (n − k)!(k − 1)! Γ(n + 3) (n + 2)(n + 1) and so k(n − k + 1) var[Xk:n ] = (n + 1)2 (n + 2) (b) Just substitute into equation(2.5a) on page 5. This gives f(Xj:n ,Xk:n ) (x, y) = cxj−1 (y − x)k−j−1 (1 − y)n−k for 0 ≤ x < y ≤ 1 where c = n!/[ (j − 1)!(k − j − 1)!(n − k)! ]. (c) j(k + 1) E[Xj:n Xk:n ] = (n + 1)(n + 2) j k j(n − k + 1) j(k + 1) − = (n + 1)(n + 2) n + 1 n + 1 (n + 1)2 (n + 2) s cov[Xj:n , Xk:n ] j(n − k + 1) corr[Xj:n , Xk:n ] = p = k(n − j + 1) var[Xj:n ]var[Xk:n ] cov[Xj:n , Xk:n ] = E[Xj:n Xk:n ] − E[Xj:n ]E[Xk:n ] = Chapter 1 Section 7 on page 18 (exs-exponential.tex) 1. This is just the probability integral transformation—see §4.5 on page 13. The distribution function of an exponential (λ) is F (x) = 1 − e−λx . Inverting this gives G(y) = − ln(1 − y)/λ. Hence G(U ) ∼ exponential (λ). 2. Now P[U > t] = ( P[X > t] )2 = e−2t and hence U ∼ exponential (2). Let W = 21 Y . Then fW (w) = 2fY (2w) = 2e−2w for w > 0. Hence, if Z = X + W , then for z ≥ 0 we have Z z Z z Z z fZ (z) = fW (z − x)fX (x) dx = 2e−2z+2x e−x dx = 2e−2z ex dx = 2e−2z (ez − 1) = 2e−z − 2e−2z 0 0 2 0 −t 2 Finally, P[V ≤ t] = (P[X ≤ t]) = 1 − e = 1 − 2e + e for t ≥ 0, and hence fV (t) = 2e−t − 2e−2t for t ≥ 0. R∞ R ∞ −δt −e−µt ] dt by integration by parts. Hence E[f (X)]− 3. E[f (X)]−E[f (Y )] = 0 f (t)[µe−µt −δe−δt ] dt = − 0 df dt [e E[f (Y )] ≤ 0. Rz Rz 4. The density of Z = X + Y is fZ (z) = 0 fY (z − x)fX (x) dx = 0 λ2 e−λz dx = λ2 ze−λz for z ≥ 0. The joint density of (X, Z) is f(X,Z) (x, z) = fX (x)fY (z − x) = λ2 e−λz . Hence f(X,Z) (x, z) 1 = for x ∈ (0, z). fX|Z (x|z) = fZ (z) z This is the U (0, z) distribution with expectation z/2. Hence E[X|X + Y ] = (X+Y )/2. 5. ∂(y1 , y2 ) = 2 and f(Y1 ,Y2 ) (y1 , y2 ) = 1 λ2 e−λy2 for y2 ≥ 0, y1 ∈ R and −y2 ≤ y1 ≤ y2 . Now ∂(x1 , x2 ) 2 Hence Z y2 1 2 −λy2 fY2 (y2 ) = λ e dy1 = λ2 y2 e−λy2 for y2 ≥ 0. 2 −y2 Z ∞ 1 2 −λy2 1 fY1 (y1 ) = λ e dy2 = λe−λy1 if y1 ≥ 0 2 2 y Z 1∞ 1 2 −λy2 1 fY1 (y1 ) = λ e dy2 = λeλy1 if y1 < 0 2 −y1 2 (b) fR (r) = λe−λr . −t −2t Page 86 Answers 1§7 Aug 1, 2018(17:54) Bayesian Time Series Analysis 6. Now P[Y ≤ y, R ≤ r] = 2P[R ≤ r, X1 < X2 , X1 ≤ y] = 2P[X2 − X1 ≤ r, X1 < X2 , X1 ≤ y]. Hence Z y Z y P[z < X2 < r + z]f (z) dz = 2 P[Y ≤ y, R ≤ r] = 2 [F (r + z) − F (z)] f (z) dz 0 Z0 y Z r f (z + t)f (z) dt dz =2 z=0 t=0 Differentiating gives f(Y,R) (y, r) = 2f (y)f (r + y) for all y ≥ 0 and r ≥ 0. −λr Suppose X1 and X2 have the exponential distribution. Then f(Y,R) (y, r) = 2λ2 e−2λy e−λr = 2λe−2λy λe = fY (y)fR (r). Hence Y and R are independent. Suppose Y and R are independent. Then we have fY (y)fR (r) = 2f (y)f (r + y) for all y ≥ 0 and r ≥ 0. Let r = 0. 2 Then fY (y)fR (0) = 2 {f (y)} . But fY (y) = 2f (y) [1 − F (y)]. Hence fR (0) [1 − F (y)] = f (y) or cF (y) + f (y) = c. This differential equation can be solved by using the integrating factor ecy where c = fR (0). We have cecy F (y) + ecy f (y) = d cecy or dy {ecy F (y)} = cecy . Hence ecy F (y) = A + ecy . But F (0) = 0; hence A = −1 and F (y) = 1 − e−cy as required. R∞ R∞ 7. (a) P[min{X1 , X2 } = X1 ] = P[X2 ≥ X1 ] = x=0 P[X2 ≥ x]λ1 e−λx1 dx = x=0 λ1 e−(λ1 +λ2 )x1 dx = λ1 /(λ1 + λ2 ). R∞ R∞ (b) P[min{X1 , X2 } > t and min{X1 , X2 } = X1 ] = P[X2 > X1 > t] = x=t y=x λ1 e−λ1 x λ2 e−λ2 y dy dx = R∞ 1 e−(λ1 +λ2 )t = P[min{X1 , X2 } > t] P[min{X1 , X2 } = X1 ] as required. λ e−(λ1 +λ2 )x dx = λ1λ+λ x=t 1 2 R∞ (c) P[R > t and min{X1 , X2 } = X1 ] = P[X2 − X1 > t] = P[X2 > t + X1 ] = u=0 e−λ2 (t+u) λ1 e−λ1 u du = λ1 e−λ2 t /(λ1 + λ2 ). Similarly for P[R > t and min{X1 , X2 } = X2 ]. Hence P[R > t] = λ1 e−λ2 t + λ2 e−λ1 t /(λ1 + λ2 ) for t > 0. (d) Now P[R > t, min{X1 , X2 } > u] = P[R > t, min{X1 , X2 } > u, min{X1 , X2 }R= X1 ] + P[R > t, min{X1 , X2 } > ∞ u, min{X1 , X2 } = X2 ] = P[X2 − X1 > t, X1 > u] + P[X1 − X2 > t, X2 > u] = x1 =u P[X2 > t + x1 ]fX1 (x1 ) dx1 + R∞ P[X1 > t + x2 ]fX2 (x2 ) dx2 = λ1 e−λ2 t e−(λ1 +λ2 )u /(λ1 + λ2 ) + λ2 e−λ1 t e−(λ1 +λ2 )u /(λ1 + λ2 ). Hence x2 =u λ1 e−λ2 t λ2 e−λ1 t −(λ1 +λ2 )u P[R > t, min{X1 , X2 } > u] = + e = P[R > t] P[min{X1 , X2 } > u] λ1 + λ2 λ1 + λ2 8. We take n large enough so that αδn < 1. Suppose t ∈ [1 − αδn , 1]. Then 1 1 1< < t 1 − αδn Integrating over t from 1 − αδn to 1 gives αδn αδn < − ln(1 − αδn ) < 1 − αδn Hence 1 1 1 1 eαδn < <e< < e 1/αδn −1 and hence 1 − αδn (1 − αδn )1/αδn −1 (1 − αδn )1/αδn Hence 1 − αδn 1 < (1 − αδn )1/αδn < and hence lim (1 − αδn )1/αδn = e−1 n→∞ e e as required. R∞ R∞ 9. (a) P[Z ≤ t] = P[Y ≥ X/t] = x=0 P[Y ≥ x/t]λe−λx dx = λ x=0 e−µx/t e−λx dx = λ/(λ + µ/t) = λt/(µ + λt). (b) First note the following integration result: " √ !# √ Z ∞ Z ∞ √ α λx −(α/x+λx) dx e dx = exp − αλ √ + √ α λx 0 x=0 p Using the substitution x λ/α = e−y gives r Z ∞ r Z ∞ Z ∞ h √ √ i −y α α −(α/x+λx) y −y e dx = exp − αλ e + e e dy = e−y e−2 αλ cosh(y) dy λ y=−∞ λ y=−∞ x=0 # r "Z ∞ Z 0 √ √ α −y −2 αλ cosh(y) −y −2 αλ cosh(y) = e e dy + e e dy λ y=0 y=−∞ r Z ∞ Z ∞ √ √ α = e−y e−2 αλ cosh(y) dy + ey e−2 αλ cosh(y) dy λ y=0 y=0 r Z ∞ r √ √ α α =2 cosh(y)e−2 αλ cosh(y) dy = 2 K1 (2 αλ) λ y=0 λ Using this result we get Z ∞ Z ∞h i 1 − e−µz/x λe−λx dx F (z) = P[Z ≤ z] = P[XY ≤ z] = P[Y ≤ z/x]λe−λx dx = x=0 x=0 Z ∞ p p −(µz/x+λx) =1−λ e dx = 1 − 2 λµzK1 (2 λµz) x=0 Appendix Aug 1, 2018(17:54) Answers 1§9 Page 87 The density function can be found by differentiating the distribution function: " √ Z ∞ Z ∞ √ !# p µz 1 −(λx+µz/x) 1 x λ f (z) = λµ e dx = λµ exp − λµz √ + √ dx x x µz x λ 0 0 √ √ Using the substitution x = µzey / λ gives Z ∞ Z ∞ h p h p i i y −y f (z) = λµ exp − λµz e + e dy = λµ exp −2 λµz cosh(y) dy y=−∞ ∞ Z = 2λµ y=0 y=−∞ h i p p exp −2 λµz cosh(y) dy = 2λµK0 (2 λµz) √ √ √ 2 (c) When λ = µ we have F (z) = 1 − 2λ zK1 (2λ z). and f (z) = 2λ K0 (2λ z). ∂(y1 ,...,yn ) 10. Now f(Y1 ,...,Yn ) (y1 , . . . , yn ) = f(X1 ,...,Xn ) (x1 , . . . , xn ) ∂(x = f(X1 ,...,Xn ) (x1 , . . . , xn ) = λn e−λyn for 0 < y1 < 1 ,...,xn ) y2 < · · · < yn . Or by induction using fY1 (y1 ) = λe−λy1 for y1 > 0 and f(Y1 ,...,Yn ) (y1 , . . . , yn ) = fYn |(Y1 ,...,Yn−1 ) (yn |y1 , . . . , yn−1 )f(Y1 ,...,Yn−1 ) (y1 , . . . , yn−1 ) = λe−λ(yn −yn−1 ) λn−1 e−λyn−1 = λn e−λyn 11. By proposition(6.4a) Z1 Z2 Zk Xk:n ∼ + + ··· + where Z1 , . . . , Zk are i.i.d. with the exponential (λ) distribution. n n−1 n−k+1 Hence k k 1X 1 1X 1 E[Xk:n ] = var[Xk:n ] = λ n+1−` λ (n + 1 − `)2 `=1 `=1 Finally, var[Xj:n ] for 1 ≤ P j < k ≤ n. Pncov[Xj:n , Xk:n ] =P n n 12. Now `=1 (X` − X1:n ) = `=1 X` − nX1:n = `=1 X`:n − nX1:n = Z2 + · · · + Zn which is independent of Z1 . Hence result. Chapter 1 Section 9 on page 22 R∞ (exs-gamma.tex) ∞ −ux e−u |0 R∞ 1. (a) Integrating by parts gives Γ(x + 1) = 0 ux e−u du = + x 0 ux−1 e−u du = xΓ(x). (b) Γ(1) = R ∞ −u 1 2 t ; hence Γ( 1/2) = e du = 1. (c) Use parts (a) and (b) and induction. (d) Use the transformation u = 2 √ R ∞ − 1 t2 R0∞ −1/2 −u √ u e du = 2 0 e 2 dt = π. The final equality follows because the integral over (−∞, ∞) of the 0 √ standard normal density is 1. (e) Use induction. For n = 1 we have Γ( 3/2) = π/2 which is the right hand side. Now 1.3.5 . . . (2n − 1).(2n + 1) √ 2n + 1 Γ(n + 1/2) = Γ(n + 1 + 1/2) = (n + 1/2)Γ(n + 1/2) = π as required. 2 2n+1 2. E[Y ] = n/α and E[1/X] = α/(m − 1) provided m > 1. Hence E[ Y /X ] = n/(m − 1) provided m > 1. 3. fX (x) = xn−1 e−x /Γ(n). At x = n − 1, fX (x) = (n − 1)n−1 e−(n−1) /Γ(n). Hence result by Stirling’s formula. 4. The simplest way is to use the moment generating function: MX+Y (t) = E[et(X+Y ) ] = E[etX ]E[etY ] = 1/(1 − t/α)n1 +n2 which is the mgf of a Gamma(n1 + n2 , α) distribution. Alternatively, Z t Z αn1 +n2 e−αt t fX+Y (t) = fX (t − u)fY (u) du = (t − u)n1 −1 un2 −1 du Γ(n1 )Γ(n2 ) u=0 0 Z αn1 +n2 tn1 +n2 −1 e−αt Γ(n1 )Γ(n2 ) αn1 +n2 tn1 +n2 −1 e−αt 1 (1 − w)n1 −1 wn2 −1 dw = = Γ(n1 )Γ(n2 ) Γ(n1 )Γ(n2 ) Γ(n1 + n2 ) w=0 αn1 +n2 tn1 +n2 −1 e−αt = Γ(n1 + n2 ) where we have used the transformation w = u/t. 5. (a) Clearly u ∈ R and 0 < v < 1; also x = uv and y = u(1 − v). Hence n+m n+m−1 −αu m−1 ∂(x, y) u e v (1 − v)n−1 = u and f(U,V ) (u, v) = α = fU (u)fV (v) ∂(u, v) Γ(n)Γ(m) where U ∼ Gamma(n + m, α) and V ∼ Beta(m, n). (1+v)2 (b) Clearly u ∈ R and v ∈ R; also x = u/(1 + v) and y = uv/(1 + v). Hence ∂(u,v) and ∂(x,y) = u u αn+m un+m−2 e−αu v n−1 αn+m un+m−1 e−αu v n−1 = (1 + v)2 (1 + v)m+n−2 Γ(n)Γ(m) (1 + v)m+n Γ(n)Γ(m) n+m n+m−1 −αu n−1 α u e Γ(n + m) v = = fU (u)fV (v) Γ(n + m) Γ(n)Γ(m) (1 + v)m+n where U ∼ Gamma(n + m, α) and 1/(1 + V ) ∼ Beta(m, n). f(U,V ) (u, v) = Page 88 Answers 1§9 Aug 1, 2018(17:54) Bayesian Time Series Analysis R R 6. Now etX dP ≥ {X>x} etX dP ≥ etx P[X ≥ x]. Hence, for all x ≥ 0 and all t < α we have P[X ≥ x] ≤ e−tx E[etX ]. Hence e−tx P[X ≥ x] ≤ inf e−tx E[etX ] = αn inf t<α t<α (α − t)n By differentiation, the infimum occurs at t = α − n/x. Hence en−αx 2n P[X ≥ x] ≤ αn and P X ≥ ≤ 2n e−n as required. (n/x)n α 7. Let V = X + Y . Then f(X,V ) (x, v) f(X,Y ) (x, v − x) = fX|V (x|v) = fV (v) fV (v) α(αx)m−1 e−αx α (α(v − x))n−1 e−α(v−x) Γ(m + n) Γ(m) Γ(n) α(αv)m+n−1 e−αv Γ(m + n) xm−1 (v − x)n−1 for 0 < x < v. = Γ(m)Γ(n) v m+n−1 = Hence for v > 0 we have E[X|X + Y = v] = Γ(m + n) Γ(m)Γ(n) Z v x=0 1 xm (v − x)n−1 Γ(m + n) dx = m+n−1 v Γ(m)Γ(n) Z v x m x=0 v 1− x n−1 dx v Γ(m + n) Γ(m + n) Γ(m + 1)Γ(n) mv v tm (1 − t)n−1 dt = v = Γ(m)Γ(n) t=0 Γ(m)Γ(n) Γ(m + n + 1) m+n R∞ R∞ 8. (a) Now fX (x) = 0 fX|Y (x|y)fY (y) dy = αn y=0 y n e−(α+x)y dy/Γ(n) = nαn /(x + α)n+1 for x > 0. E[X] can be found by integration; or E[X|Y ] = 1/Y and hence E[X] = E E[X|Y ] = E[ 1/Y ] = α/(n − 1). (b) Now f(X,Y ) (x, y) = αn y n e−(x+α)y /Γ(n). Hence fY |X (y|x) = f(X,Y ) (x, y)/fX (x) = y n (x + α)n+1 e−(x+α)y /Γ(n + 1) for x > 0 and y > 0. 9. f(X|Y1 ,...,Yn ) (x|y1 , . . . , yn ) ∝ f (y1 , . . . , yn , x) = f (y1 , . . . , yn |x)fX (x) n Y = xe−xyi λe−λx = λxn e−x(λ+y1 +···+yn ) Z = i=1 Using the integral of the Gamma density gives Z ∞ λxn e−x(λ+y1 +···+yn ) dx = 0 and hence λn! (λ + y1 + · · · + yn )n+1 xn e−x(λ+y1 +···+yn ) (λ + y1 + · · · + yn )n+1 n! which is the Gamma(n + 1, λ + y1 + · · · + yn ) distribution. Hence E[X|Y1 = y1 , . . . , Yn = yn ] = (n + 1)/(λ + y1 + · · · + yn ). 10. Now φX (t) = E[eitX ] = α/(α − it). Using Y = αX − 1 gives φY (t) = E[eitY ] = e−it E[eitαX ] = e−it /(1 − it). Hence |φY (t)| = 1/(1 + t2 )1/2 . Choose k > 2 and then |φY (t)|k ≤ 1/(1 + t2 ) which is integrable. Also, for |t| ≥ δ we have |φY (t)| ≤ 1/(1 + δ 2√ )1/2 < 1 for all δ > 0. It follows that Sn / n √ has a bounded continuous density with density fn which satisfies limn→∞ fn (z) = n(z). Now √ Sn / n = (αGn − n)/ n. Hence √ √ n n+z n fn (z) = fGn and hence the required result. α α 11. Let Lt = SK − SK−1 . Note that {K = 1} = {S1 ≥ t}. Then for x < t we have ∞ ∞ X X P[Lt ≤ x] = P[Lt ≤ x, K = n] = P[Sn − Sn−1 ≤ x, K = n] f(X|Y1 ,...,Yn ) (x|y1 , . . . , yn ) = n=1 = = ∞ X ∞ X n=2 P[Sn−1 < t < Sn−1 + y]αe−αy dy = y=0 ∞ Z X n=2 But P[Sn − Sn−1 ≤ x, Sn−1 < t ≤ Sn ] = n=2 ∞ Z x X n=2 = n=2 t v=t−x Z P[Xn ≤ x, Sn−1 < t ≤ Sn−1 + Xn ] ∞ Z X n=2 t x fSn−1 (v)αe−αy dy dv = y=t−v ∞ Z X n=2 v=t−x n=2 fSn−1 (v) = α. Hence and Z y=0 P∞ P[Lt ≤ x] = 1 − e−αx − αxe−αx x fLt (x) = α2 xe−αx t fSn−1 (v)αe−αy dv dy v=t−y −α(t−v) e − e−αx fSn−1 (v) dv Appendix Aug 1, 2018(17:54) If x > t then P[Lt ≤ x] = ∞ X n=1 P[Lt ≤ x, K = n] = P[t < X1 ≤ x] + = e−αt − e−αx + = e−αt − e−αx + = e−αt − e−αx + = e−αt − e−αx + = e−αt − e−αx + =1−e and fLt (x) = α(1 + αt)e−αx . −αx Z ∞ X n=2 ∞ X n=2 P[Sn − Sn−1 ≤ x, K = n] P[Sn − Sn−1 ≤ x, Sn−1 < t ≤ Sn ] P[Xn ≤ x, Sn−1 < t ≤ Sn−1 + Xn ] n=2 ∞ Z x X n=2 ∞ Z X n=2 P[Sn−1 < t < Sn−1 + y]αe−αy dy y=0 t Z ∞ Z X x fSn−1 (v)αe−αy dy dv y=t−v v=0 t n=2 v=0 −αx −α(t−v) e − e−αx fSn−1 (v) dv t 2 2 −αx α x e Z ∞ α(1 + αt)xe−αx dx dx + x=t x=0 Z = −αt2 e−αt + 2α and so E[Lt ] > 1/α. (b) ∞ X − αte E[Lt ] = = Answers 1§11 Page 89 t xe−αx dx + α(1 + αt) Z ∞ xe−αx dx x=t 0 2 − e−αt α P[Wt ≤ x] = P[SK ≤ t + x] = ∞ X n=1 = P[t < X1 ≤ t + x] + = e−αt − e−α(x+t) + = e−αt − e−α(x+t) + P[Sn ≤ t + x, K = n] ∞ X P[Sn ≤ t + x, Sn ≥ t, Sn−1 < t] n=2 ∞ Z t X n=2 0 ∞ Z t X n=2 0 fSn−1 (y) P[t − y ≤ Xn ≤ t + x − y] dy fSn−1 (y) e−α(t−y) − e−α(x+t−y) dy −αx =1−e So Wt has the same exponential distribution as the original Xj —this is the “lack of memory” property. 12. (a) Use the transformation y = λxb . Then Z ∞ Z ∞ n−1 −y y e f (x) dx = dy = 1 Γ(n) 0 0 (b) Straightforward. (c) Use the transformation y = xb . Chapter 1 Section 11 on page 26 (exs-normal.tex) 1. Using the transformation v = −t gives Z −x Z x Z ∞ 1 1 1 √ exp − t2/2 dt = − √ exp − v2/2 dv = √ exp − v2/2 dv = 1 − Φ(x) Φ(−x) = 2π 2π 2π −∞ ∞ x 2. (140 − µ)/σ = Φ−1 (0.3) = −0.5244005 and (200 − µ)/σ = 0.2533471. Hence σ = 77.14585 and µ = 180.4553. 3. We can take Y = 0 with probability 1/2 and Y = Z with probability 1/2, where Z ∼ N (0, 1). Hence E[Y n ] = 21 E[Z n ] = 0 if n is odd and 12 E[Z n ] = n!/(2(n+2)/2 ( n/2)!) if n is even.. 4. (a) Clearly we hmust have c > 0; = i also a > 0 is necessaryin order to ensure fX (x) can integrate to 1. Now Q(x)2 (x−µ) b2 b 2 b b 2 b2 b2 2 a(x − a x) = a (x − 2a ) − 4a2 and hence fX (x) = c exp − 4a2 exp −a(x − 2a ) = c exp − 4a2 exp − 2σ2 2 b 1 b where µ = 2a and σ 2 = 2a . Because fX (x) integrates to 1, we must also have c exp − 4a = σ√12π . This answers (a). 2 b 1 (b) X ∼ N 2a , σ 2 = 2a . Page 90 Answers 1§11 Aug 1, 2018(17:54) Bayesian Time Series Analysis 5. Clearly X/σ and Y /σ are i.i.d. with the N (0, 1) distribution. Hence (X 2 + Y 2 )/σ 2 ∼ χ22 = Γ(1, 1/2) which is the exponential ( 1/2) distribution with density 12 e−x/2 for x > 0. Hence X 2 + Y 2 ∼ exponential ( 1/2σ2 ) with expectation 2σ 2 . (b) Clearly X1/σ, . . . , Xn/σ are i.i.d. with the N (0, 1) distribution. Hence Z 2 /σ 2 ∼ χ2n = Gamma( n/2, 1/2). Hence Z 2 ∼ Gamma( n/2, 1/2σ2 ). 6. f(X|Y1 ,...,Yn ) (x|y1 , . . . , yn ) ∝ f (y1 , . . . , yn , x) = f (y1 , . . . , yn |x)fX (x) n Y (yi − x)2 1 (x − µ)2 1 √ exp − √ exp − = 2σ 2 2σ12 σ 2π σ 2π i=1 1 Pn Pn 2 2x i=1 y1 2µx µ2 nx2 x2 i=1 yi ∝ exp − − 2 + − 2 − 2+ 2σ 2σ 2 2σ 2σ12 2σ12 2σ1 Pn 2 2 x i=1 y1 µx nx x ∝ exp − 2 − 2+ 2 2 2σ σ σ1 2σ1 Pn 2 yi αx 1 µ n = exp − + βx where α = 2 + 2 and β = 2 + i=12 2 σ σ1 σ σ1 α ∝ exp − (x − β/α)2 2 Hence the distribution of (X|Y1 = y1 , . . . , Yn = yn ) is N ( β/α, σ 2 = 1/α). Note that α, the precision of the result is the sum of the (n + 1)-precisions. Also, the mean is a weighted average of the input means: P β µ σ12 /n + ( yi )σ 2 = α σ12 /n + σ 2 q 2 2 2 2 7. (a) fY (y) = σ√22π e−y /2σ = σ1 π2 e−y /2σ for y ∈ (0, ∞). q R q h q q 2 2 ∞ 2 2 ∞ (b) E[|X|] = σ1 π2 0 xe−x /2σ dx = σ1 π2 −σ 2 e−x /2σ = σ1 π2 σ 2 = σ π2 0 8. (a) FY (x) = P[Y ≤ x] = P[−x ≤ X ≤ x] = Φ(x) − Φ(−x). Hence fY (x) = fX (x) + fX (−x); hence r µx 1 2 (x2 + µ2 ) (x − µ)2 1 (x + µ)2 √ fY (x) = √ exp − exp − + exp − = cosh 2σ 2 2σ 2 πσ 2 2σ 2 σ2 2πσ 2 2πσ 2 1 x −x by using cosh x = 2 (e + e ). Z ∞ Z ∞ 1 (x − µ)2 (x + µ)2 E[Y ] = √ x exp − dx + x exp − dx 2σ 2 2σ 2 2πσ 2 0 0 Z ∞ Z ∞ 1 (x − µ)2 (x + µ)2 √ = (x − µ) exp − dx + (x + µ) exp − dx + A 2σ 2 2σ 2 2πσ 2 0 0 1 µ2 µ2 =√ σ 2 exp − 2 + σ 2 exp − 2 2σ 2σ 2πσ 2 where Z ∞ Z ∞ µ (x − µ)2 (x + µ)2 A= √ dx − dx exp − exp − 2σ 2 2σ 2 2πσ 2 0 0 "Z # Z ∞ ∞ µ i h µ i h µ µσ −y 2 /2 −y 2 /2 −Φ − = µ 1 − 2Φ − =√ e dy − e dy = µ Φ σ σ σ 2πσ 2 −µ/σ µ/σ Hence r h µ i 2 µ2 E[Y ] = σ exp − 2 + µ 1 − 2Φ − π 2σ σ 2 2 Clearly var[Y ] = var[|X|] = E[X 2 ] − {E[|X|} = var[X] + {E[X]} − µ2Y = σ 2 + µ2 − µ2Y . (c) The mgf, E[etY ], is Z ∞ Z ∞ (x + µ)2 1 (x − µ)2 √ dx + exp tx − dx = exp tx − 2σ 2 2σ 2 2πσ 2 0 0 2 2 2 2 Z ∞ Z ∞ 1 σ t (x − µ − σ 2 t)2 σ t (x + µ − σ 2 t)2 √ exp + µt exp − dx + exp − µt exp − dx 2 2σ 2 2 2σ 2 2πσ 2 0 0 2 2 h µ i h µ i σ t σ 2 t2 = exp + µt 1 − Φ − − σt + exp − µt 1 − Φ − σt 2 σ 2 σ Hence the cf is h h µ i µ i σ 2 t2 σ 2 t2 itY φY (t) = E[e ] = exp − + iµt 1 − Φ − − iσt + exp − − iµt 1 − Φ − iσt 2 σ 2 σ Appendix Aug 1, 2018(17:54) Answers 1§11 Page 91 9. Z ∞ Z µ (x − µ)2 1 (x − µ)2 n dx + (µ − x) dx exp − (x − µ)n exp − E |X − µ|n = √ 2σ 2 2σ 2 2πσ 2 µ −∞ 2 Z ∞ 2 t n n =√ t σ exp − dt 2 2π 0 Z ∞ σn σn n+1 n+1 2n/2 σ n v (n−1)/2 exp(− v/2) dv = √ 2(n+1)/2 Γ =√ = √ Γ 2 2 π 2π 0 2π −s2 −t2 itZ1 isZ2 itZ1 +isZ2 i(t+s)X i(t−s)Y − 21 (t+s)2 − 21 (t−s)2 e =e e = E[e ]E[e ]. 10. (a) Now E[e ] = E[e ]E[e ]=e X−Y 2 2 2 1 1 (b) Let X1 = X+Y 2 , Then X1 ∼ N (0, σ = /2). Let X2 = 2 , Then X2 ∼ N (0, σ = /2). Let Z1 = 2X1 and Z2 = 2X22 . Now X1 and X2 are independent by part (a); hence Z1 and Z2 are √ independent. Hence Z1 and Z2 √ are i.i.d. χ21 = Gamma( 1/2, 1/2) with c.f. 1/ √1 − 2it. Hence −Z2 has the c.f. 1/√ 1 + 2it. Because Z1 and Z2 are independent, the c.f. of 2XY = Z1 − Z2 is 1/ 1 + 4t2 . Hence the c.f. of XY is 1/ 1 + t2 . (c) Z ∞ Z ∞ 1 2 1 itXY ityX E[e ]= E[e ]fY (y) dy = E[eityX ] √ e− 2 y dy 2π −∞ −∞ Z ∞ Z ∞ 2 1 2 2 1 2 1 2 1 1 1 e− 2 t y e− 2 y dy = √ e− 2 y (1+t ) dy = √ =√ 2π −∞ 2π −∞ 1 + t2 √ √ 2 (d) Now X = σX1 and Y = σY1 where the c.f. of X1 Y1 is 1/ 1 + t . Hence the c.f. of XY is 1/ 1 + σ 4 t2 . (e) Take σ = 1. Then the m.g.f. is Z ∞ Z ∞ 2 1 1 tXY tyX E[e ]fY (y) dy = E[etyX ] √ e− 2 (y−µ) dy E[e ]= 2π −∞ −∞ µ2 Z Z ∞ ∞ − 2 2 1 2 2 1 1 2 1 e 2 =√ eµty+ 2 t y e− 2 (y−µ) dy = √ e− 2 y (1−t )+µy(1+t) dy 2π −∞ 2π −∞ µ2 Z ∞ (1 − t2 ) e− 2 2µy exp − = √ y2 − dy 2 1−t 2π −∞ " 2 # Z ∞ 1 µ2 µ2 (1 + t) (1 − t2 ) µ √ = exp − + exp − y− dy 2 2(1 − t) 2 1−t 2π −∞ 2 µ t 1 √ = exp 1−t 1 − t2 2 For the general case E[etXY ] = E[etσ X1 Y1 ] where X1 and Y1 are i.i.d. N ( µ/σ, 1) and hence µ2 t 1 iµ2 t 1 tXY itXY √ √ E[e ] = exp and the c.f. is E[e ] = exp 2 4 2 1 − tσ 2 1 − itσ 1−σ t 1 + σ 4 t2 2 11. Use the previous question. In both cases, the c.f. is 1/(1 + t ). 12. (a) Now Z ∞ Z 2 2 1 ∞ b b du = 1 − 2 + 2 + 1 exp − 12 u2 + ub 2 du exp − 12 u2 + ub 2 2 0 u u 0 Consider the integral Z ∞ b b2 1 2 I1 = + 1 exp − u + du 2 2 u u2 0 dz The transformation u → z with z = u − ub is a 1 − 1 transformation: (0, ∞) → (−∞, ∞). Also du = 1 + ub2 . Hence Z ∞ √ 1 2 I1 = e−b e− 2 z dz = e−b 2π −∞ Now consider the integral ∞ 2 b 1 − 2 exp − 12 u2 + ub 2 du u 0 √ √ Consider the transformation z = u + ub . This is a 1 − 1 transformation (0, b) → (∞, 2 b) and a 1 − 1 transformation √ √ ( b, ∞) → (2 b, ∞). Hence Z √b Z ∞ ! Z 2√b Z ∞ 2 1 2 1 2 b 1 b I2 = + √ 1 − 2 exp − 2 u2 + u2 eb e− 2 z dz + √ eb e− 2 z dz = 0 du = u 2 b 0 b ∞ Z I2 = as required. (b) Just use the transformation u = |a|v in part (a) and then set b1 = b/|a|. Page 92 Answers 1§13 Aug 1, 2018(17:54) Bayesian Time Series Analysis Chapter 1 Section 13 on page 30 (exs-logN.tex) 1. Let S4 denote the accumulated value at time t = 4 and let s0 denote the initial amount invested. Then S4 = s0 (1 + I1 )(1 + P4 I2 )(1 + I3 )(1 + I4 ) and ln(S4 /s0 ) = j=1 ln(1 + Ij ) 1 2 Recall that if If Y ∼ lognormal(µ, σ 2 ) then Z = ln Y ∼ N (µ,pσ 2 ). Also E[Y ] = E[eZ ] = eµ+ 2 σ and var[Y ] = 2 2 2 e2µ+σ (eσ − 1). Hence eσ = 1 + var[Y ]/E[Y ]2 and eµ = E[Y ]/ 1 + var[Y ]/E[Y ]2 or µ = ln E[Y ] − σ 2 /2. Using mean=1.08 and variance=0.001 gives µ1 = 0.0765325553785 and σ12 = 0.000856971515297. Using mean=1.06 and variance=0.002 gives µ2 = 0.0573797028389 and σ22 = 0.00177841057009. Hence ln(S4 /s0 ) ∼ N (2µ1 + 2µ2 , 2σ12 + 2σ22 ) = N (0.267824516435, 0.00527076417077). We qwant 0.95 = P[S5 > 5000] = P[ln(S5 /s0 ) > ln(5000/s0 )] = P[Z > (ln(5000/s0 ) − (2µ1 + 2µ2 ))/ 2σ12 + 2σ22 ] Hence ln(5000/s0 ) − (2µ1 + 2µ2 ) ln(5000/s0 ) − (2µ1 + 2µ2 ) q q 0.05 = Φ and so = Φ−1 (0.05) 2 2 2 2 2σ1 + 2σ2 2σ1 + 2σ2 q Hence ln(5000/s0 ) = (2µ1 + 2µ2 ) + Φ−1 (0.05) 2σ12 + 2σ22 = 0.148408095871. q −1 2 2 Hence s0 = 5000 exp −(2µ1 + 2µ2 ) − Φ (0.05) 2σ1 + 2σ2 = 4310.39616086 or £4,310.40. 2 2 2. (a) Let Z = 1/X. Then ln(Z) = − ln(X) ∼ (b) Let Z = cX b . Then N (−µ, σ ). Hence Z ∼ logN (−µ, σ ). ln(Z) = ln(c) + b ln(X) ∼ N ln(c) + bµ, b2 σ 2 . Hence Z ∼ logN ln(c) + bµ, b2 σ 2 . 3. (a) Now X ∼ logN (µ, σ 2 ); hence ln(X) ∼ N (µ, σ 2 ). Hence ln(GMX ) = E[ln(X)] = µ; hence GMX = eµ . 2 (b) Now ln(GVX ) = var[ln(X)] = σ 2 . Hence GVX = eσ and GSDX = eσ . µ µ 4. The median is e because P[X < e ] = P[ln(X) < µ] = 1/2. Hence the median equals the geometric mean. The mean 1 2 is eµ+ 2 σ by equation(12.3a) on page 28. For the mode, we need to differentiate the density function which is c (ln(x) − µ)2 fX (x) = exp − for x > 0. x 2σ 2 Hence c c 2(ln(x) − µ) (ln(x) − µ)2 dfX (x) = − 2− exp − dx x x 2xσ 2 2σ 2 2 2 1 2 µ µ+ σ which equals 0 when x = eµ−σ . Clearly mode = eµ−σ < median = e < mean = e 2 . (b) Lower quartile: 0.25 = P[X < q1 ] = P[ln(X) < ln(q1 )] = Φ ln(q1σ)−µ and hence ln(q1σ)−µ = −0.6744898 and hence 5. 6. 7. 8. q1 = eµ−0.6744898σ . Similarly for the upper quartile, q3 =eµ+0.6744898σ . ln(αp )−µ . Hence ln(αp ) = µ + σβp as required. (c) p = P[X ≤ αp ] = P[ln(X) ≤ ln(αp )] = Φ σ Pn Pn Pn Pn Pn (a) Let Z = X1 · · · Xn . Then ln(Z) = i=1 ln(Xi ) ∼ N µi , i=1 σi2 . Hence Z ∼ logN µi , i=1 σi2 . i=1 i=1 Pn 1/n 2 2 (b) Let Z = (X Q1 n· · · Xani) . Then ln(Z) =Pn i=1 ln(Xi )/n ∼ N (µ, Pσn /n). Hence Pn Z2 ∼2 logN (µ, σ /n). Pn (c) Let Z = i=1 Xi . Then ln(Z) = i=1 ai ln(Xi ) ∼ N i=1 ai µi , i=1 ai σi . Hence mn = i=1 ai µi and qP n 2 2 sn = i=1 ai σi . Let Z = X1 /X2 . Then Z ∼ logN (µ1 − µ2 , σ12 + σ22 ) by using the previous 2 questions. 2 2 1 2 We know that α = eµ+ 2 σ and β = e2µ+σ (eσ −1 ). Hence α2 β α α2 or µ = ln p and eµ = p =p σ 2 = ln 1 + 2 α 1 + β/α2 β + α2 β + α2 Now for x ∈ (0, k) we have 1 (ln(x) − µ)2 √ f (x|X < k) = exp − 2σ 2 P[X < k]σx 2π 2 √ 1 (ln(x) − µ) ln(k) − µ = exp − where α = σ 2πΦ xα 2σ 2 σ and hence Z 1 k (ln(x) − µ)2 E[X|X < k] = exp − dx α 0 2σ 2 2 1 2 2 Using the transformation w = ln(x) − µ − σ 2 /σ gives dw dx = xσ and (ln(x) − µ) /σ = (w + σ) . Hence 2 2 Z Z 1 (ln(k)−µ−σ )/σ (w + σ)2 σ (ln(k)−µ−σ )/σ w2 σ 2 E[X|X < k] = exp − xσ dw = exp − + + µ dw α −∞ 2 α −∞ 2 2 2 Φ ln(k)−µ−σ σ σ2 = eµ+ 2 ln(k)−µ Φ σ Appendix Aug 1, 2018(17:54) Answers 1§15 Page 93 1 2 The other result is similar or use E[X|X < k]P[X < k] + E[X|X > k]P[X > k] = E[X] = eµ+ 2 σ . 9. (a) Z x (ln(u) − µ)2 1 1 du G(x) = exp −jµ − j 2 σ 2 uj √ exp − 2 2σ 2 uσ 2π 0 Z x 1 (ln(u) − µ − jσ 2 )2 √ = exp − du as required. 2σ 2 0 uσ 2π Setting j = 1 in part (a) shows that xfX (x) = E[X]fX1 (x) where X1 ∼ logN (µ + σ 2 , σ 2 ). (b) Z ∞Z ∞ Z ∞Z u (v − u)fX (u)fX (v) dvdu (u − v)fX (u)fX (v) dvdu + 2E[X]γX = u=0 v=u u=0 v=0 Z ∞Z v Z ∞Z u (v − u)fX (u)fX (v) dvdu (u − v)fX (u)fX (v) dvdu + = v=0 u=0 u=0 v=0 Z ∞Z u (u − v)fX (u)fX (v) dvdu =2 u=0 v=0 Z ∞ Z ∞ Z u vfX (v)dv fX (u)du =2 uFX (u)fX (u)du − 2 v=0 u=0 u=0 Z ∞ Z ∞ FX (u)fX1 (u)du − FX1 (u) fX (u)du where X1 ∼ logN (µ + σ 2 , σ 2 ). = 2E[X] u=0 u=0 = 2E[X] P[X ≤ X1 ] − P[X1 ≤ X] = 2E[X] P[ X/X1 ≤ 1] − P[ X1/X ≤ 1] But X/X1 ∼ logN (−σ 2 , 2σ 2 ) and P[ X1/X ≤ 1] = P[X ≥ X1 ] = 1 − P[X < X1 ]. Hence γX = 2P[Y ≤ 1] − 1 where Y ∼ logN (−σ 2 , 2σ 2 ). σ −1 as required. = 2Φ √ 2 Chapter 1 Section 15 on page 33 (exs-betaarcsine.tex.tex) 1. Z Γ(α + β) 1 m+α−1 Γ(α + β)Γ(m + α) x (1 − x)β−1 dx = Γ(α)Γ(β) 0 Γ(α)Γ(m + α + β) 2. X has density fX (x) = αxα−1 . Let Y = − ln(X). Then Y ∈ (0, ∞) and X = e−Y and . dy fX (x) dx = αe−y(α−1) × e−y = αe−αy for y ∈ (0, ∞), as required. E[X m ] = dx dy = −x. Hence fY (y) = dy = (1 + y)2 . Hence 3. Let Y = X/(1 − X); then Y ∈ (0, ∞). Also X = Y /(1 + Y ), 1 − X = 1/(1 + Y ) and dx α−1 β−1 α−1 dx x (1 − x) 1 1 y fY (y) = fX (x) = = for y ∈ (0, ∞), as required. 2 dy B(α, β) (1 + y) B(α, β) (1 + y)α+β 4. (a) Suppose α > 0 and β > 0. Then for all x > 0 we have Z ∞ xα−1 xα B(α + 1, β − 1) α fX (x) = Hence E[X] = dx = = α+β B(α, β) (1 + x)α+β B(α, β) (1 + x) B(α, β) β − 1 0 R∞ α −α−β using 0 x (1 + x) dx = B(α + 1, β − 1) for all α > −1 and β > 1. (b) Similarly, for all β > 2 we have Z ∞ α(α + 1) xα+1 B(α + 2, β − 2) = E[X 2 ] = dx = α+β B(α, β) (1 + x) B(α, β) (β − 1)(β − 2) 0 Hence var[X]. (c) Suppose α ≤ 1. Then fX (x) ↑ as x ↓ and the mode is at 0. Now suppose α > 1 and let g(x) = xα−1 /(1 + x)α+β . Then g 0 (x) = ( α − 1 − x(1 + β) ) /x(1 + x). Hence g 0 > 0 for x < (α − 1)/(1 + β) and g 0 > 0 when x > (α − 1)/(1 + β). So the mode is at (α − 1)/(1 + β). dy dy 2 2 α+1 (d) Let Y = 1/X; hence | | = 1/x . Hence f (y) = f (x)/| | = x f (x) = x B(α, β)(1 + x)α+β = Y X X dx dx β−1 α+β y B(β, α)(1 + y) as required. (Note that B(α, β) = B(β, α).) 1 (e) Let V = X/Y and W = Y ; then ∂(v,w) ∂(x,y) = y . Now f(X,Y ) (x, y) = and hence f(X,Y ) (x, y) xn1 −1 y n2 e−(x+y) v n1 −1 wn1 +n2 −1 e−w(1+v) xn1 −1 y n2 −1 e−(x+y) = and f(V,W ) (v, w) = = Γ(n1 )Γ(n2 ) Γ(n1 )Γ(n2 ) Γ(n1 )Γ(n2 ) ∂(v,w) ∂(x,y) Z ∞ v n1 −1 v n1 −1 Γ(n1 + n2 ) fV (v) = wn1 +n2 −1 e−w(1+v) dw = as required. Γ(n1 )Γ(n2 ) w=0 Γ(n1 )Γ(n2 ) (1 + v)n1 +n2 (f) Just use X/Y = (2X)/(2Y ) and part (e). Page 94 Answers 1§17 Aug 1, 2018(17:54) Bayesian Time Series Analysis 5. Throughout x ∈ [0, 1] and we take arcsin(x) ∈ [0, π2 ]. Let y = arcsin(x); then sin(y) = x and sin(−y) = −x. Hence arcsin(−x) = − arcsin(x). Let y = π2 − arccos(x). Then sin(y) = x and hence arccos(x) + arcsin(x) = π2 . Now sin(y) = x and 1 − 2x2 = cos2 (y) − sin2 (y) = cos(2y); hence 21 arccos(1 − 2x2 ) = y = arcsin(x). √ Combining gives 2 arcsin( x) = arccos(1 − 2x) = π2 − arcsin(1 − 2x) = π2 + arcsin(2x − 1). 6. (a) Let Y = kX + m. Hence fX (x) fX (x) 1 fY (y) = dy = the density of an arcsin(m + ak, m + bk) distribution. = √ k π (y − m − ak)(bk + m − y) | dx | (b) Let Y = X 2 ; then Y ∈ (0, 1). Also fX (x) fX (x) fX (x) 1 1 √ fY (y) = 2 dy = 2 = = = √ as required. 2 2|x| |x| π y(1 − y) |x|π 1 − x | dx | (c) Let Y = sin(X). Then Y ∈ (−1, 1). Also fX (x) fX (x) 2 1 1 1 fY (y) = 2 dy = 2 = = p = √ as required. | cos(x)| 2π | cos(x)| π 1 − y 2 π (1 − y)(1 + y) | dx | Let Y = sin(2X). Then Y ∈ (−1, 1). Also 1 1 1 1 fX (x) fX (x) = = p = √ as required. fY (y) = 4 dy = 4 2 2| cos(2x)| π | cos(2x)| π (1 − y)(1 + y) | dx | π 1−y Let Y = − cos(2X). Then Y ∈ (−1, 1). Also fX (x) 1 1 1 1 fX (x) = = p = √ fY (y) = 4 dy = 4 as required. 2| sin(2x)| π | sin(2x)| π 1 − y 2 π (1 − y)(1 + y) | dx | 7. (a) Now V = X + Y has a triangular distribution on (−2π, 2π) with density: |v| 1 1− for −2π < v < 2π. fV (v) = 2π 2π Let Z = sin(V ) = sin(X + Y ). Then Z ∈ (−1, 1). Also X fV (v) X fV (v) 1 |v| 1 X fZ (z) = =√ 1− = dz | cos(v)| 2π | dv | 1 − z 2 2π v v v −1 The 4 values of v leading to z are sin−1 and −2π + sin−1 (z) and −π − sin−1 (z) P(z), π − sin (z), which are both√positive, 2 which are both negative. This gives v |v| = 4π. Hence fZ (z) = 1/π 1 − z . (b) Now −Y ∼ Uniform(−π, π). Hence result follows by part (a). Chapter 1 Section 17 on page 37 0 0 0 (Y12 (exs-tCauchyF.tex.tex) Yn2 )/σ 2 . 0 0 χ2n . Also Then Y ∼ + ··· + X Z=p ∼ tn Y 0 /n 2. Use the transformation from (X, Y ) to (W, Z) where X W =Y and Z=p Y /n Hence w ∈ (0, ∞) and z ∈ R and √ √ ∂(w, z) = √n = √ n ∂(x, y) y w Now √ ∂(x, y) = f(X,Y ) (x, y) √w f(W,Z) (w, z) = f(X,Y ) (x, y) ∂(w, z) n √ n/2−1 −y/2 2 1 y e w √ = √ e−x /2 n/2 n 2 Γ n/2 2π 2 1 e−z w/2n w(n−1)/2 e−w/2 = 1/2 (n+1)/2 1/2 π 2 n Γ n/2 But Z ∞ Γ n+1 (n−1)/2 −αw 2 w e dw = (n+1)/2 α 0 Hence Γ n+1 1 1 z2 2 (n+1)/2 fZ (z) = 1/2 (n+1)/2 1/2 where α = 1+ 2 n π 2 n Γ n/2 α −(n+1)/2 Γ n+1/2 1 z2 √ = 1/2 1+ as required. n n π Γ n/2 1. Let X = X/σ . Then X ∼ N (0, 1). Let Y = Appendix 3. First, let x = (t − α)/s: Aug 1, 2018(17:54) Answers 1§17 Page 95 n/2 n/2 Z ∞ 1 1 dt = s dx 1 + (t − α)2 /s2 1 + x2 −∞ −∞ But from equation(16.1b) on page 34, we know that −(n+1)/2 Z ∞ √ t2 1+ dt = B( 1/2, n/2) n n −∞ √ Letting x = t/ n gives Z ∞ (1 + x2 )−(n+1)/2 dx = B 1/2, n/2 Z ∞ −∞ 4. 5. 6. 7. 8. 9. 10. and hence the result. Pn P Proof 1. Now Y = Z12 + · · · + Zn2 ; hence i=1 Zi2 /n−→1 asq n → ∞ by the Weak Law of Large Numbers. Using the Pn √ √ P 2 simple inequality | a − 1| = |a − 1|/| a + 1| < |a − 1| gives i=1 Zi /n−→1 as n → ∞. One of the standard results by Slutsky(see for example p285 in Probability and Stochastic Processes by Grimmett and Stirzaker) is: D P D if Zn −→Z as n → ∞ and Yn −→c as n →√∞ where c 6= 0 then Zn /Yn −→Z/c as n → ∞. Hence result. n −n Proof 2. Stirling’s formula is n! ∼ n e 2πn as n → ∞. Using this we can show that 1 1 √ =√ lim as n → ∞. n→∞ B( 1/2, n/2) n 2π Also −(n+1)/2 2 t2 lim 1 + = e−t /2 as n → ∞. n→∞ n √ Now W = X/Y = ( X/σ)/( Y /σ). Hence we can take σ = 1 without loss of generality. (a) Now W = X/Y = X/ Z where X ∼ N (0, 1) and Z ∼ χ21 and X and Z are independent. Hence W ∼ t1 = γ1 , the Cauchy density. 2 (b) As for part (a). (c) The folded Cauchy ∂u density which 2is fW (w) = 2/π(1 + w ) for w > 0. ∂u P = (a) Let W = tan(U ). Then fW (w) = fU (u) ∂w = 1/π(1 + w ). (b) Let W = tan(U ). Then fW (w) = u fU (u) ∂w (2/2π) × 1/(1 + w2 ) as required. h i |t| n t t (a) φ(t) = E[eitY ] = E[ei n (X1 +···Xn ) ] = E[ei n X1 ]n = e−s n = e−s|t| as required. (b) Use proposition(2.7a) √ D on page 6. Hence 2 n Mn /(πs) =⇒ N (0, 1) as n → ∞. Hence Mn is asymptotically normal with mean 0 and variace π 2 s2 /(2n). (a) φ2X (t) = E[eit2X ] = E[ei(2t)X ] = e−2s|t| . This is the Cauchy γ2s distribution. (b) φX+Y (t) = E[eit(X+Y ) ] = E[eit(aU +cU +bV +dV ) ] = E[eit(a+c)U ]E[eit(b+d)V ] = e−s(a+c)|t| e−s(b+d)|t| = e−s(a+b+c+d)|t| which is the Cauchy distribution γs(a+b+c+d) . We have Y = R sin(Θ) and X = R cos(Θ). Also ∂(x,y) ∂(r,θ) = r. Hence ∂(x, y) = 1 e−(x2 +y2 )/2 r = 1 re−r2 /2 f(R,Θ) (r, θ) = f(X,Y ) (x, y) ∂(r, θ) 2π 2π −r 2 /2 Hence Θ is uniform on (−π, π), R has density re for r > 0 and R and Θ are independent. If W = R2 then the density of W is fW (w) = 12 e−w/2 for w > 0; this is the χ22 distribution. Let X = tan(Θ); hence Θ ∈ (− π/2, π/2). Then P[Θ ≤ θ] = P[X ≤ tan(θ)] = 1/2 + θ/π and fΘ (θ) = 1/π. Hence Θ has the uniform distribution on (− π/2, π/2). Now 2X/(1 − X 2 ) = tan(2Θ). So we want the distribution of W = tan(Y ) where Y has the uniform distribution on (−π, π). Hence X dw 1 =2 1 as required. fW (w) = fY (y) dy 2π 1 + w2 y 11. Now X = b tan Θ and so 0 < X < ∞. For x > 0 we have P[X ≤ x] = P[|b tan Θ| ≤ x] = P[| tan Θ| ≤ x/b] = 2/π tan−1 x/b. Differentiating gives 2 b fX (x) = π b2 + x 2 (b) P[ 1/X ≤ x] = P[X ≥ 1/x] = 1 − π2 tan−1 1/bx and this has density π2 1+bb2 x2 for x > 0. 12. (a) Clearly f ≥ 0. Using the transformation y = (1 + x2 )1/2 tan t gives Z ∞ Z ∞ Z π/2 1 1 1 cos t 1 f (x, y) dy = dy = dt = 2 2 2 3/2 2π y=−∞ (1 + x + y ) 2π t=−π/2 1 + x π(1 + x2 ) y=−∞ which is the standard Cauchy distribution. This answers parts (a) and(b). (c) The absolute value of the Jacobian of the transformation is ∂(x,y) ∂(r,θ) = r. Hence r f(R,Θ) (r, θ) = for r > 0 and θ ∈ (0, 2π). 2π(1 + r2 )3/2 Hence R and Θ are independent. Θ ∼ U (0, 2π) and R has density fR (r) = r/(1 + r2 )3/2 for r > 0. Page 96 Answers 1§17 Aug 1, 2018(17:54) Bayesian Time Series Analysis 13. Use the transformation from (X, Y ) to (V, W ) where X/m nX V = = and W =Y Y /n mY Hence v ∈ (0, ∞) and w ∈ (0, ∞) and ∂(v, w) n n ∂(x, y) = my = mw Now ∂(x, y) = f(X,Y ) (x, y) mw f(V,W ) (v, w) = f(X,Y ) (x, y) ∂(v, w) n 15. 16. 17. xm/2−1 e−x/2 y n/2−1 e−y/2 mw n 2m/2 Γ m/2 2n/2 Γ n/2 = (mwv/n)m/2−1 e−mvw/2n wn/2−1 e−w/2 mw n 2m/2 Γ m/2 2n/2 Γ n/2 = w mv v m/2−1 (m/n)m/2 w(m+n)/2−1 e− 2 (1+ n ) 2(m+n)/2 Γ(m/2)Γ(n/2) R∞ wk−1 e−αw dw = Γ(k)/αn with α = 21 (1 + mv n ) and integrating out w gives m+n m/2−1 m/2 Γ 2 Γ m+n v (m/n) v m/2−1 mm/2 nn/2 2 fV (v) = m = as required. n (m+n)/2 Γ( 2 )Γ( n2 ) 2(m+n)/2 α(m+n)/2 Γ( m 2 )Γ( 2 ) (n + mv) Now Z = X12 /Y12 where X1 = X/σ and Y1 √ = Y /σ are i.i.d. N (0, 1). Now X12 and Y12 are i.i.d. χ21 . Hence Z has the F1,1 distribution which has density fZ (z) = 1/π z(1 + z) for z > 0. Now F = (nX)/(mY ); hence E[F ] = nE[X]E[1/Y ]/m = nmE[1/Y ]/m = nE[1/Y ] = n/(n − 2). (a) By definition(18.6a) on page 41. (b) Using §8.3 on page 19 gives 2α1 X1 ∼ Gamma(n1 , 1/2) = χ22n1 and 2α2 X2 ∼ Gamma(n2 , 1/2) = χ22n2 . Hence result by definition(18.6a). Let Y = nX/m(1 − X); then Y ∈ (0, ∞). Also X = mY /(n + mY ), 1 − X = n/(n + mY ) and m dy 1 n dy = = and hence n dx (1 − x)2 dx m(1 − x)2 Hence dx xm/2−1 (1 − x)n/2−1 m(1 − x)2 y m/2−1 mm/2 nn/2 = fY (y) = fX (x) = as required. m n n dy B( 2 , 2 ) n (my + n)(m+n)/2 B m 2 , 2 Using 14. = 0 dy 18. (a) This is the reverse of exercise 17.2 (b) Let Y = αX/β; then | dx | = α/β. Now for x ∈ (0, ∞) we have α β α−1 α β α−1 (2α) (2β) x 1 α β x 1 = fX (x) = B(α, β) 2α+β [αx + β]α+β B(α, β) [αx + β]α+β fX (x) αα−1 β β+1 (βy/α)α−1 1 y α−1 1 fY (y) = dy = = for y ∈ (0, ∞). B(α, β) β α+β [1 + y]α+β B(α, β) [1 + y]α+β | dx | Or use X = (Y /2α)/(Z/2β) = (βY )/(αZ) where Y ∼ χ22α , Z ∼ χ22β and Y and Z are independent. Hence X ∼ Beta0 (α, β) by part (f) of 4 on page 34. 19. Now W = nX/(mY ); hence mW/n = X/Y where X ∼ χ2m , Y ∼ χ2n and X and Y are independent. Hence by part (f) of exercise 4 on page 34, X/Y ∼ Beta0 ( m/2, n/2). P 20. Proof 1. Now W = (X/m)/(Y /n). Now Y = Z12 + · · · + Zn2 where Z1 , . . . , Zn are i.i.d. N (0, 1). Hence Y /n−→1 as D n → ∞. Hence, by Slutsky’s theorem (see answer to exercise 4 on page 95), mW = X/(Y /n)−→χ2m as n → ∞. Proof 2. Now Γ( m+n ) mm/2 nn/2 wm/2−1 fW (w) = m 2 n for w ∈ (0, ∞). Γ( 2 )Γ( 2 ) [mw + n](m+n)/2 lim fW (w) = n→∞ 2 n/2 Γ( m+n mm/2 wm/2−1 2 )n lim n→∞ Γ( n )[mw + n](m+n)/2 Γ( m 2 ) 2 = Γ( m+n mm/2 wm/2−1 1 2 ) lim m n n→∞ (1 + mw/n)n/2 Γ( )[mw + n]m/2 Γ( 2 ) 2 = Γ( m+n mm/2 wm/2−1 −mw/2 2 ) e lim m n n→∞ Γ( 2 ) Γ( 2 )[mw + n]m/2 Suppose X has density fX and Y = α(X) where the function α has an inverse. Let fY denote the density of Y . Then if Z has density fY then α−1 (Z) has density fX . This follows because P[α−1 (Z) ≤ z] = P[Z ≤ α(z)] = P[Y ≤ α(z)] = P[α(X) ≤ α(z)] = P[X ≤ z]. Appendix Aug 1, 2018(17:54) = Answers 1§21 Page 97 mm/2 wm/2−1 −mw/2 1 e Γ( m 2m/2 2 ) √ 1 by using Stirling’s formula: n! ∼ 2π nn+ 2 e−n as n → ∞. Finally, convergence in densities implies convergence in distribution (see page 252 in [F ELLER(1971)]). Chapter 1 Section 19 on page 41 (exs-noncentral.tex.tex) 1. Use moment generating functions or characteristic functions. n Y 1 λt 1 λj t E[etZ ] = = exp exp 1 − 2t 1 − 2t (1 − 2t)k/2 (1 − 2t)kj /2 j=1 Hence Z has a non-central χ2k distribution with non-centrality parameter λ where k = k1 + · · · + kn and λ = λ1 + · · · + λn . Pn Pn 2. (a) E[W ] = j=1 E[Xj2 ] = j=1 (1 + µ2j ) = n + λ (b) Suppose Z ∼ N (0, 1). Then E[Z] = E[Z 3 ] = 0, E[Z 2 ] = 1 and E[Z 4 ] = 3. Suppose X = Z + µ. Then E[X] = µ, E[X 2 ] = 1 + µ2 , E[X 3 ] = 3µ + µ3 , and E[X 4 ] = 3 + 6µ2 + µ4 . Hence var[X 2 ] = 2 + 4µ2 . Hence var[W ] = var[X12 ] + · · · + var[Xn2 ] = 2n + 4λ. 3. Rearranging equation(18.4a) on page 41 gives ∞ ∞ X X e−λ/2 ( λ/2)j e−x/2 x(n+2j−2)/2 e−λ/2 ( λ/2)j fX (x) = = fn+2j (x) j! j! 2(n+2j)/2 Γ( n/2 + j) j=0 j=0 where fn+2j (x) is the density of the χ2n+2j distribution. Hence result. Chapter 1 Section 21 on page 45 (exs-powerPareto.tex) 1. Now Y = (X − a)/h has the standard power distribution Power(α, 0, 1) which has density f (y) = αy α−1 for 0 < y < 1. R1 For j = 1, 2, . . . , we have E[Y j ] = α 0 y α+j−1 dy = α/(α + j). Hence α α α E[Y 2 ] = and hence var[Y ] = E[Y ] = α+1 α+2 (α + 1)2 (α + 2) Now X = a + hY . Hence αh αh2 2aαh αh2 E[X] = a + E[X 2 ] = + + a2 and var[X] = α+1 α+2 α+1 (α + 1)2 (α + 2) 2. P[Mn ≤ x] = (x − a)nα /hnα and so Mn ∼ Power(nα, a, h). 3. (a) FMn (x) = P[Mn ≤ x] = xn for 0 < x < 1. This is the Power(n, 0, 1) distribution with density fMn (x) = nxn−1 for 0 < x < 1. (b) P[U 1/n ≤ x] = P[U ≤ xn ] = xn for 0 < x < 1. The same distribution as for part (a). (c) Now (X − a)/h ∼ Power(α, 0, 1); hence by part (b) we have (X − a)/h ∼ U 1/α and X ∼ a + hU 1/α . Then use the binomial theorem and E[U j/α ] = α/(α + j). 4. Now X ∈ (0, h). Hence Y ∈ (ln( 1/h), ∞) and Y − ln( 1/h) ∈ (0, ∞). Now P[Y ≤ y] = P[ln(X) ≥ −y] = P[X ≥ e−y ] = 1 − e−αy /hα = 1 − e−α(y−ln(1/h)) for y > ln( 1/h). Hence the density is αe−α(y−ln(1/h)) for y > ln( 1/h), a shifted exponential. 5. Equation(2.3b) on page 5 gives n! n! k−1 n−k fk:n (x) = f (x) {F (x)} {1 − F (x)} = αxkα−1 (1 − xα )n−k (k − 1)!(n − k)! (k − 1)!(n − k)! and using the transformation v = xα gives Z 1 Z 1 1 n! n! E[Xk:n ] = αxkα (1 − xα )n−k dx = v k−1+ α (1 − v)n−k dv (k − 1)!(n − k)! 0 (k − 1)!(n − k)! 0 = 2 E[Xk:n ]= Γ(k + α1 )Γ(n − k + 1) n! Γ(k + α1 ) n! = (k − 1)!(n − k)! Γ(n + α1 + 1) (k − 1)! Γ(n + α1 + 1) n! Γ(k + α2 ) (k − 1)! Γ(n + α2 + 1) Page 98 Answers 1§21 Aug 1, 2018(17:54) Bayesian Time Series Analysis 6. (a) Rx Sn E[Y1 + · · · + Yn−1 ] (n − 1)E[Y ] n − 1 0 yf (y) dy (n − 1)α E X(n) = x = 1 + =1+ =1+ =1+ X(n) x x x F (x) α+1 as required. (b) The density of (X1:n , X2:n , . . . Xn:n ) is f (x1 , . . . , xn ) = n!αn (x1 x2 · · · xn )α−1 /hnα for 0 ≤ x1 ≤ x2 · · · ≤ xn . Consider the transformation to (W1 , W2 , . . . , Wn ) where W1 = X1:n /Xn:n , W2 = X2:n /Xn:n , . . . , Wn−1 = X(n−1):n /Xn:n and Wn = Xn:n . This has Jacobian with absolute value ∂(w1 , . . . , wn ) 1 ∂(x1 , . . . , xn ) = xn−1 n Hence for 0 < w1 < 1, . . . , 0 < wn < 1, the density of the vector (W1 , . . . , Wn ) is αn−1 α−1 (n−1)α wn h(w1 , . . . , wn ) = wnn−1 f (w1 wn )f (w2 wn ) · · · f (wn−1 wn ) = α(n−1) w1α−1 w2α−1 · · · wn−1 h Hence W1 , W2 , . . . , Wn are independent. Hence W1 + · · · + Wn−1 is independent of Wn as required. 7. (a) The distribution of X(i) give X(i+1) = x is the same as the distribution of the maximum of i independent random i variables from the density f (y)/F (x) for y ∈ (0, x); this maximum has distribution function {F (y)/F (x)} and density i−1 i if (y){F (y)} /{F (x)} for y ∈ (0, x). (a) Hence " # Z x r X(i) i E y r f (y){F (y)}i−1 dy (21.7a) X(i+1) = x = r r X(i+1) x {F (x)}i 0 ⇐ Substituting f (y) = αy α−1 /hα and FR(y) = y α /hα in the right hand side of equation(21.7a) gives iα/(iα + r) x as required. ⇒ Equation(21.7a) gives i 0 y r f (y){F (y)}i−1 dy = cxr {F (x)}i for x ∈ (0, h). Differentiating with r i−1 r−1 respect to x gives ix f (x){F (x)} = cx {F (x)}i−1 [rF (x) + xif (x)]. Hence f (x)/F (x) = cr/ix(1 − c) > 0 because c < 1. Hence result. (b) # " Z x r X(i+1) ixr f (y){F (y)}i−1 E dy (21.7b) X(i+1) = x = r X(i) {F (x)}i 0 yr and then as for part (a). 8. By equation(2.2a), the density of the vector is (X1:n , X2:n , . . . , Xn:n ) is g(x1 , . . . , xn ) = n!f (x1 ) · · · f (xn ) = n!αn (x1 x2 · · · xn )α−1 for 0 ≤ x1 ≤ x2 · · · ≤ xn . The transformation to (W1 , W2 , . . . , Wn ) has Jacobian with absolute value ∂(w1 , . . . , wn ) 1 n−2 n−1 2 and x1 = w1 w2 · · · wn ∂(x1 , . . . , xn ) = x2 · · · xn where x2 · · · xn = w2 w3 · · · wn−1 wn Hence for 0 < w1 < 1, . . . , 0 < wn < 1, the density of the vector (W1 , . . . , Wn ) is h(w1 , . . . , wn ) = n!αn xα−1 (x2 · · · xn )α = (αw1α−1 )(2αw22α−1 ) · · · (nαwnnα−1 ) = fW1 (w1 )fW2 (w2 ) · · · fWn (wn ) 1 Hence W1 , W2 , . . . , Wn are independent. Also fWk (wk ) = kαwkkα−1 which is Power(kα, 0, 1). (b) Now Xk:n = Wk Wk+1 · · · Wn ; hence (k + 1)α nα kα ··· E[Xk:n ] = EWk ]E[Wk+1 ] · · · E[Wn ] = kα + 1 (k + 1)α + 1 nα + 1 Γ(k + α1 ) n! 1 1 1 n! = · · · = (k − 1)! k + α1 k + 1 + α1 (k − 1)! Γ(n + 1 + α1 ) n + α1 2 2 Similarly for E[Xk:n ] = EWk2 ]E[Wk+1 ] · · · E[Wn2 ]. −1/α 9. (a) Just use P[Y ≤ y] = P[U ≤ y] = P[U 1/α ≥ 1/y] = P[U ≥ 1/yα ]. (b) Just use Y ∼ Pareto(α, a, x0 ) iff (Y − a)/x0 ∼ Pareto(α, 0, 1) and part (a). (c) By part (a), 1/X ∼ Pareto(α, 0, 1) iff 1/X ∼ U −1/α where U ∼ Uniform(0, 1) and hence iff X ∼ U 1/α where U ∼ Uniform(0, 1) and hence iff X ∼ Power(α, 0, 1). 10. n xα xαn n 0 0 = P[Mn > x] = P[X1 > x] = which is Pareto(αn, a, x0 ). (x − a)α (x − a)αn α+1 11. Now f (x) = αxα for x > x0 . (a) E[X] = αx0 /(α − 1) if α > 1 and ∞ if α ≤ 1. Also E[X 2 ] = αx20 /(α − 2) if 0 /x α > 2 and ∞ otherwise. Hence var[X] = αx20 /(α − 1)2 (α − 2) provided α > 2. √ α (b) The median is x0 2. The mode is x0 . (c) E[X n ] = αxn0 /(α − n) for α > Rn and ∞ otherwise. ∞ tx α+1 (d) Suppose t < 0. Then E[etX ] = x0 αxα dx. Set v = −tx; hence v > 0. Then 0 e /x Z ∞ Z ∞ α −v α+1 αx0 e (−t) dx α(−x0 t)α e−v E[etX ] = = dx = α(−x0 t)α Γ(−α, −x0 t) v α+1 (−t) v α+1 −tx0 −tx0 R∞ where Γ(s, x) = x ts−1 e−t dt is the incomplete gamma function. Hence the c.f. is E[eitX ] = α(−ix0 t)α Γ(−α, −ix0 t). Appendix Aug 1, 2018(17:54) Answers 1§21 Page 99 12. Part (a) of exercise 11 shows that E[X] is infinite. Also ∞ Z ∞ Z 1 1 ∞ −5/2 1 −2x−3/2 1 1 1 E dx = x dx = = = 3/2 X x 2x 2 1 2 3 3 1 1 −αy −α(y−ln(x0 )) 13. P[Y ≤ y] = P[ln(X) ≤ y] = P[X ≤ ey ] = 1 − xα e = 1 − e for y > ln(x0 ). Hence the density 0 is αe−α(y−ln(x0 )) for y > ln(x0 ), a shifted exponential. In particular, if X ∼ Pareto(α, 0, 1), then the distribution of Y = ln(X) is the exponential (α) distribution. R ∞ ln(x) 14. Now GMX is defined by ln(GMX ) = E[ln X]. Either use exercise 13 or directly: E[ln X] = αxα dx = 0 x0 xα+1 R 1 1 α ∞ −αy αx0 ln(x0 ) ye dy = ln(x0 ) + α and hence GMX = x0 exp α . From the answer to exercise 1§13(9) on page 93 we have, where E[X] = αx0 /(α − 1), Z ∞ Z ∞ Z u 2E[X]γX = 2 uF (u)f (u)du − 2 vf (v)dv f (u)du u=0 u=0 v=0 Z Z ∞ h ∞ Z u x α i αxα αxα αxα 0 0 0 0 v u 1− du − 2 dv du =2 α+1 u uα+1 uα+1 v=x0 v u=x0 u=x0 Z ∞ Z u Z ∞ 1 1 xα 1 2 2α 0 = 2αxα − du − 2α x dv du 0 0 α 2α α α+1 u u u=x0 u v=x0 v u=x0 = and hence 2α2 x0 2αx0 − (2α − 1)(α − 1) (2α − 1) α α−1 1 − = 2α − 1 2α − 1 2α − 1 15. Using equation(2.3b) on page 5 gives γX = n! n! α k−1 n−k f (x) {F (x)} {1 − F (x)} = α+1 (k − 1)!(n − k)! (k − 1)!(n − k)! x k−1 n! α 1 = 1− α (k − 1)!(n − k)! x(n−k+1)α+1 x 1− fk:n (x) = and hence n! E[Xk:n ] = (k − 1)!(n − k)! = n! (k − 1)!(n − k)! Z ∞ x(n−k+1)α 1 Z α 1 v n−k+1 1 v 1+ α Γ(n − k − n! = (k − 1)!(n − k)! Γ(n − 0 k−1 1 xα(n−k) k−1 1 1− α dx x (1 − v)k−1 dv = 1 α 1 α 1 xα n! (k − 1)!(n − k)! Z 0 1 1 v n−k− α (1 − v)k−1 dv + 1)Γ(k) n! Γ(n − k + 1 − α1 ) = (n − k)! Γ(n + 1 − α1 ) + 1) n! Γ(n − k + 1 − α2 ) (n − k)! Γ(n + 1 − α2 ) 16. By equation(2.2a) on page 4, the density of the vector is (X1:n , X2:n , . . . , Xn:n ) is g(x1 , . . . , xn ) = n!f (x1 ) · · · f (xn ) = n!αn / (x1 x2 · · · xn )α+1 for 1 ≤ x1 ≤ x2 · · · ≤ xn . The transformation to (W1 , W2 , . . . , Wn ) has Jacobian with absolute value ∂(w1 , . . . , wn ) 1 n−1 n−2 ∂(x1 , . . . , xn ) = x1 · · · xn−1 where x1 · · · xn−1 = w1 w2 · · · wn−1 and xn = w1 w2 · · · wn Hence for w1 > 1, . . . , wn > 1, the density of the vector (W1 , . . . , Wn ) is n!αn n!αn h(w1 , . . . , wn ) = = nα+1 (n−1)α+1 α α+1 2α+1 w α+1 (x1 · · · xn−1 ) xn w1 w2 · · · wn−1 n 2 E[Xk:n ]= = nα (n − 1)α 2α α · · · 2α+1 α+1 = fW1 (w1 )fW2 (w2 ) · · · fWn (wn ) w1nα+1 w2(n−1)α+1 wn−1 wn Hence W1 , W2 , . . . , Wn are independent. Also fWk (wk ) = (b) Now Xk:n = W1 W2 · · · Wk ; hence (n−k+1)α (n−k+1)α+1 wk which is Pareto((n − k + 1)α, 0, 1). nα (n − 1)α (n − k + 1)α ··· nα − 1 (n − 1)α − 1 (n − k + 1)α − 1 n! 1 1 1 n! Γ(n − k + 1 − α1 ) = · · · = (n − k)! n − α1 n − 1 − α1 (n − k)! Γ(n + 1 − α1 ) n − k + 1 − α1 2 2 2 2 Similarly for E[Xk:n ] = EW1 ]E[W2 ] · · · E[Wk ]. 17. Just use (X1 , . . . , Xn ) ∼ ( Y11 , . . . , Y1n ). E[Xk:n ] = EW1 ]E[W2 ] · · · E[Wk ] = Page 100 Answers 1§21 Aug 1, 2018(17:54) Bayesian Time Series Analysis α 18. Let Z = Y /X . Note that F (x) = 1 − xα 0 /x for x ≥ x0 and 0 otherwise. Then for z > 1 we have Z ∞ Z ∞ Z ∞ αxα αx2α 1 1 xα 0 0 dx = 1 − dx = 1 − α P[Z ≤ z] = F (zx)f (x) dx = 1 − α0 α α+1 α 2α+1 z x x z x 2z x0 x0 x0 For z < 1 we have Z ∞ Z ∞ xα αxα zα zα 0 1 − α0 α F (zx)f (x) dx = P[Z ≤ z] = dx = z α − = α+1 z x x 2 2 x0 /z x0 /z and hence 1 α/z α+1 if z > 1; fZ (z) = 12 α−1 if z < 1; 2 αz 19. P[M ≤ x, Y /X ≤ y] = P[M ≤ x, Y /X ≤ y, X ≤ Y ] + P[M ≤ x, Y /X ≤ y, Y < X] = P[X ≤ x, Y ≤ yX, X ≤ Y ] + P[Y ≤ x, Y ≤ yX, Y < X] P[X ≤ x, Y ≤ yX, X ≤ Y ] + P[Y ≤ x, Y < X] if y > 1; = P[Y ≤ x, Y ≤ yX] if y < 1. Rx P[v ≤ Y ≤ yv]f (v) dv + P[Y ≤ x, Y < X] if y > 1; x0 = P[Y ≤ x, Y ≤ yX] if y < 1. (Rx Rx [F (yv) − F (v)] f (v) dv + x0 [1 − F (v)] f (v) dv if y > 1; = Rxx0 1 − F ( v/y) f (v) dv if y < 1. x0 ( Z x 1 f (v) x2α 1 − y1α + 1 if y > 1; 0 where I = dv = 1 − = xα I 0 α 2xα x2α x0 v yα if y < 1. 0 ( 1 − 2y1α if y > 1; x2α 0 = 1 − 2α = P[M ≤ x] P[ Y /X ≤ y] by using exercises 10 and 18. yα x if y < 1. 2 20. Define the vector (Y1 , Y2 , . . . , Yn ) by X2:n Xn:n , . . . , Yn = X1:n X1:n Exercise 15 on page 8(with answer 1§3(15) on page 81) shows that for y1 > 0 and 1 ≤ y2 ≤ · · · ≤ yn the density of the vector (Y1 , . . . , Yn ) is h(y1 , . . . , yn ) = n!y1n−1 f (y1 )f (y1 y2 )f (y1 y3 ) · · · f (y1 yn ) 1 1 αn xαn 1 0 = n! αn+1 · · · α+1 yn y1 y2α+1 y3α+1 αnxαn 1 1 1 1 1 1 0 = αn+1 (n − 1)!αn−1 α+1 α+1 · · · α+1 = g(y1 ) (n − 1)!αn−1 α+1 α+1 · · · α+1 yn yn y1 y2 y3 y2 y3 where g is the density of Y1 = X1:n (see answer 10 above). Hence part (a). (b) R∞ Sn E[Y1 + · · · + Yn−1 ] (n − 1)E[Y ] n − 1 x yf (y) dy (n − 1)α E X(1) = x = 1 + =1+ =1+ =1+ X(1) x x x 1 − F (x) α−1 (S −X ) S n 1:n n as required. (c) The result of part (a) implies Y1 is independent of Y2 + · · · + Yn = /X1:n = /X1:n − 1. Hence Y1 is independent of Sn/X1:n as required. 21. The distribution of X(i+1) give X(i) = x is the same as the distribution of the minimum of n − i independent random on−i n 1−F (y) variables from the density f (y)/[1 − F (x)] for y ∈ (x, ∞); this minimum has distribution function 1 − 1−F (x) Y1 = X1:n , Y2 = and density (n − i)f (y){1 − F (y)}n−i−1 /{1 − F (x)}n−i for y ∈ (x, ∞). Hence " # Z ∞ r X(i+1) (n − i) E X = x = y r f (y){1 − F (y)}n−i−1 dy (i) r X(i) xr {1 − F (x)}n−i x (21.21a) α+1 α ⇐ Substituting f (x) = αxα and F (x) = 1 − xα 0 /x 0 /x in the right hand side of equation(21.21a) gives (n − ⇒ Equation(21.21a) gives i)α/((nR− i)α + r) as required. The condition α > r/(n − i) ensures the integral is finite. ∞ (n − i) x y r f (y){1 − F (y)}n−i−1 dy = cxr {1 − F (x)}n−i for x ∈ (x0 , ∞). Differentiating with respect to x gives f (x)/{1 − F (x)} = cr/[x(n − i)(c − 1)] = α/x where α = rc/[(n − i)(c − 1)] > r/(n − i). Hence result. Part (b) is similar. 22. Let X1 = ln(X) and Y1 = ln(Y ). Then X1 and Y1 are i.i.d. random variables with an absolutely continuous distribution. Also min(X1 , Y1 ) = ln [min(X, Y )] is independent of Y1 − X1 = ln(Y /X). Hence there exists λ > 0 such that X1 and Y1 have the exponential (λ) distribution. Hence X = eX1 and Y = eY1 have the Pareto(λ, 0, 1) distribution. Appendix Aug 1, 2018(17:54) Answers 1§21 Page 101 23. ⇐ By equation(2.3b) on page 5 the density of Xi:n is i−1 n−i 1 1 n! n! β i−1 n−i 1− β fi:n (t) = f (t) {F (t)} {1 − F (t)} = (i − 1)!(n − i)! (i − 1)!(n − i)! tβ+1 t tβ i−1 n! β β = t −1 for t > 1. βn+1 (i − 1)!(n − i)! t By equation(2.5a) on page 5, the joint density of (Xi:n , Xj:n ) is, where c = n!/[(i − 1)!(j − i − 1)!(n − j)!], i−1 j−1−i n−j f(i:n,j:n) (u, v) = cf (u)f (v) F (u) F (v) − F (u) 1 − F (v) i−1 j−i−1 n−j β2 1 1 1 1 = c β+1 β+1 1 − β − β u v u uβ v vβ β i−1 β j−i−1 u −1 v − uβ β2 = c β+1 β(n−j+1)+1 u v uβ uβ v β 2 i−1 β j−i−1 β = c β(j−1)+1 β(n−i)+1 uβ − 1 v − uβ for 1 ≤ u < v. u v 1 Use the transformation (T, W ) = (Xi:n , Xj:n/Xi:n ). The absolute value of the Jacobian is ∂(t,w) ∂(u,v) = | /u| = 1/t. Hence f(T,W ) (t, w) = c β2 tβn+1 wβ(n−i)+1 i−1 β j−i−1 tβ − 1 w −1 = fi:n (t)fW (w) The fact that the joint density is the product of the marginal densities implies W and Y = Xi:n are independent. ⇒ The joint density of (Xi:n , Xj:n ) is given by equation(2.5a) on page 5. The transformation to T = Xi:n , W = Xj:n /Xi:n has Jacobian with absolute value ∂(t,w) ∂(u,v) = 1/u = 1/t. Hence (T, W ) has density i−1 j−i−1 n−j f(T,W ) (t, w) = ctf (t)f (wt) F (t) F (wt) − F (t) 1 − F (wt) Now T = Xi:n has density given by equation(2.3b) on page 5: n! i−1 n−i f (t) {F (t)} {1 − F (t)} fT (t) = (i − 1)!(n − i)! Hence the conditional density is, for all t > 1 and w > 1, j−i−1 n−j f(T,W ) (t, w) (n − i)! tf (tw) F (tw) − F (t) 1 − F (tw) f(W |T ) (w|t) = = fT (t) (j − i − 1)!(n − j)! 1 − F (t) 1 − F (t) 1 − F (t) ∂q(t, w) 1 − F (tw) (n − i)! j−i−1 n−j − {1 − q(t, w)} {q(t, w)} where q(t, w) = = (j − i − 1)!(n − j)! ∂w 1 − F (t) and by independence, this must be independent of t. Hence there exists a function g(w) with ∂q(t, w) j−i−1 n−j g(w) = {1 − q(t, w)} {q(t, w)} ∂w j−i−1 ∂q(t, w) X j − i − 1 r n−j = (−1)r {q(t, w)} {q(t, w)} ∂w r r=0 j−i−1 X j − i − 1 ∂q(t, w) r+n−j = {q(t, w)} (−1)r ∂w r r=0 j−i−1 r+n−j+1 ∂ X j−i−1 {q(t, w)} = (−1)r ∂w r r+n−j+1 r=0 and hence Z g1 (w) = g(w) dw = j−i−1 X r=0 0= r+n−j+1 j−i−1 {q(t, w)} (−1)r r r+n−j+1 j−i−1 X j−i−1 r+n−j ∂q(t, w) (−1)r {q(t, w)} r ∂t r=0 ∂q(t, w) ∂q(t, w) = g(w) ∂t ∂w ∂g1 (w) = ∂t and hence ∂q(t,w) = 0 and so q(t, w) is a function of w only. Setting t = 1 shows that q(t, w) = (1 − F (tw))/(1 − F (t)) = ∂t q(1, w) = (1 − F (w))/(1 − F (1)). Hence 1 − F (tw) = (1 − F (w))(1 − F (t))/(1 − F (1)). But F (1) = 0; hence we have the following equation for the continuous function F : (1 − F (tw)) = (1 − F (t))(1 − F (w)) Page 102 Answers 1§23 Aug 1, 2018(17:54) Bayesian Time Series Analysis for all t ≥ 1 and w ≥ 1 with boundary conditions F (1) = 0 and F (∞) = 1. This is effectively Cauchy’s logarithmic functional equation. It leads to 1 F (x) = 1 − β for x ≥ 1. x Chapter 1 Section 23 on page 50 (exs-sizeshape.tex.tex) 1. Consider the size variable g1 : (0, ∞) → (0, ∞) with g(x) = x1 . The associated shape function is z1 (x) = 1, x2/x1 = (1, a). Hence z1 (x) is constant. For any other shape function z2 , we know that z2 (x) = z2 ( z(x) ) (see the proof of equation(22.3a) on page 48. Hence result. 2. (a) Consider the size variable g ∗ (x) = x1 ; the associated shape function is z ∗ (x) = 1, x2 /x1 . Hence 3 with probability 1/2; z ∗ (X) = 1 /3 with probability 1/2. If z is any other shape function, then by equation(22.3a) on page 48 we have z ∗ (X) = z ∗ ( z(X) ). Hence z(X) cannot be almost surely constant. The possible values of the 3 quantities are as follows: probability z(X) g1 (X) g2 (X) √ 1/4 1 3 ( /4, /4) 4 √3 1/4 ( 1/4, 3/4) 12 4 √ 1/4 ( 3/4, 1/4) 4 √3 1/4 ( 3/4, 1/4) 12 4 Clearly z(X) is independent of both g1 (X) and g2 (X). (b) By proposition(22.3b) on page 48 we know that g1 (X)/g2 (X) is almost surely constant. It is easy to check that √ g1 (X)/g2 (X) = 3/4. 3. ⇐ Let Yj = Xjb for j = 1, 2, . . . , n. Then Y1 , . . . , Yn are independent random variables. By proposition(22.4a), the b b shape vector 1, Y2/Y1 , . . . , Yn/Y1 is independent of Y1 + · · · + Yn . This means 1, X2 /X1b , . . . , Xn/X1b is independent of X1b + · · · + Xnb . Hence 1, X2/X1 , . . . , Xn/X1 is independent of (X1b + · · · + Xnb )1/b as required. ⇒ We are given that 1, X2/X1 , . . . , Xn/X1 is independent of (X1b + · · · + Xnb )1/b . Hence 1, X2b/X1b , . . . , Xnb/X1b is independent of X1b + · · · + Xnb . By proposition(22.4a), there exist α > 0, k1 > 0, . . . , kn > 0 such that Xjb ∼ Gamma(kj , α) andRhence Xj ∼ GGamma(kj , α,Rb). ∞ ∞ 4. (a) P[X1 < X2 ] = 0 P[X2 > x]λ1 e−λ1 x dx = 0 e−λ2 x λ1 e−λ1 x dx = λ1 /(λ1 + λ2 ). (b) The lack of memory property implies the distribution of V = X2 − X1 given X2 > X1 is the same as the distribution of X2 and the distribution of X2 − X1 given X2 < X1 is the same as the distribution of −X1 . Hence ( 2 if v ≤ 0; λ1 eλ1 v λ1λ+λ λ1 λ2 2 + fV (v|X1 < X2 ) = fV (v) = fV (v|X2 < X1 ) −λ2 v λ1 λ1 + λ2 λ1 + λ2 if v ≥ 0. λ2 e λ1 +λ2 2 (c) Now V = X2 − X1 . Hence Z fV (v) = fX2 (x + v)fX1 (x) dx = λ1 λ2 e−λ2 v Z {x:x+v>0} Hence if v ≥ 0 we have fV (v) = λ1 λ2 e−λ2 v fV (v) = λ1 λ2 e−λ2 v e−(λ1 +λ2 )x dx x=max{−v,0} Z ∞ e−(λ1 +λ2 )x dx = λ1 λ2 e−λ2 v λ1 + λ2 e−(λ1 +λ2 )x dx = λ1 λ2 eλ1 v λ1 + λ2 0 and if v ≤ 0 we have ∞ Z ∞ −v 5. Now V = Y2 − Y1 . By exercise 4, we know that λ1 λ2 eλ1 v if v ≤ 0; fV (v) = λ1 + λ2 e−λ2 v if v ≥ 0. ( and P[V ≤ v] = λ2 λ1 v λ1 +λ2 e 1 1 − λ1λ+λ e−λ2 v 2 P[U ≥ u] = e−λ1 (u−a) e−λ2 (u−a) = ea(λ1 +λ2 ) e−(λ1 +λ2 )u Now for u ≥ a and v ∈ R we have Z if v ≤ 0; if v ≥ 0. for u ≥ a. ∞ P[U ≥ u, V ≤ v] = P[Y1 ≥ u, Y2 ≥ u, Y2 − Y1 ≤ v] = = λ1 eλ1 a Z y1 =u P[u ≤ Y2 ≤ v + y1 ]fY1 (y1 ) dy1 ∞ y1 =u P[u ≤ Y2 ≤ v + y1 ]e−λ1 y1 dy1 where P[u ≤ Y2 ≤ v + y1 ] = λ2 eλ2 a 0 R v+y1 y2 =u e−λ2 y2 dy2 = eλ2 a [e−λ2 u − e−λ2 (v+y1 ) ] if v + y1 > u; if v + y1 < u. Appendix Aug 1, 2018(17:54) Hence for v ≥ 0 we have P[U ≥ u, V ≤ v] = λ1 e = λ1 e λ1 a Z Answers 1§25 Page 103 ∞ P[u ≤ Y2 ≤ v + y1 ]e−λ1 y1 dy1 y1 =u (λ1 +λ2 )a Z ∞ y1 =u e−λ1 y1 e−λ2 u − e−λ2 (v+y1 ) dy1 e−λ1 u e−(λ1 +λ2 )u λ1 e−λ2 v = λ1 e(λ1 +λ2 )a e−λ2 u − e−λ2 v = e(λ1 +λ2 )a e−(λ1 +λ2 )u 1 − λ1 λ1 + λ2 λ1 + λ2 = P[U ≥ u]P[V ≤ v] Similarly, for v ≤ 0 we have P[U ≥ u, V ≤ v] = λ1 eλ1 a = λ1 e Z ∞ P[u ≤ Y2 ≤ v + y1 ]e−λ1 y1 dy1 y1 =u−v Z ∞ (λ1 +λ2 )a y1 =u−v e−λ1 y1 e−λ2 u − e−λ2 (v+y1 ) dy1 = P[U ≥ u]P[V ≤ v] Hence the result. 6. Suppose θj > 0 and x0 > 0. Then X ∼ Pareto(α, 0, θj x0 ) iff X/θj ∼ Pareto(α, 0, x0 ) and proceed as in the proof of proposition(22.5a) on page 49. 7. Suppose θj > 0 and h > 0. Then X ∼ Power(α, 0, θj h) iff X/θj ∼ Power(α, 0, h) and proceed as in the proof of proposition(22.6a) on page 50. Chapter 1 Section 25 on page 52 8. (a) Let W = X + Y . For w > 0, Z ∞ Z fW (w) = fX (x)fY (w − x) dx = x=w For w < 0 Z ∞ αe−αx αeα(w−x) dx = α2 eαw x=w w fW (w) = −∞ (exs-other.tex) fY (y)fX (w − y) dy = α2 e−αw Z ∞ e−2αx dx = x=w Z w e2αy dy = −∞ α −αw e 2 α αw e 2 and hence fW (w) = α2 e−α|w| for w ∈ R; this is the Laplace(0, α) distribution. (b) The Laplace(0, α) distribution by part (a). 9. (a) The expectation, median and mode are all µ. Also var[X] = var[X − µ] = 2/α2 by using the representation of the Laplace as the difference of two independent exponentials given in exercise 8. (b) The distribution function is 1 −α(µ−x) e if x < µ; FX (x) = 2 1 −α(x−µ) 1 − 2e if x ≥ µ. (c) Using the representation in exercise 8 again implies the moment generating function is α α2 eµt α = 2 for |t| < α. E[etX ] = eµt E[et(X−µ) ] = eµt α − t α + t α − t2 R x α −α|y| R x α −αy Rx 10. For x > 0 we have P[|X| < x] = P[−x < X < x] = −x 2 e dy = 2 0 2 e dy = 0 αe−αy dy. Hence |X| has the exponential density αe−αx R ∞for x > 0. R∞ 11. For z > 0 we have fZ (z) = y=0 fX (z + y)fY (y) dy = λµ y=0 e−λz e−λy e−µy dy = λµe−λz /(λ + µ). R∞ R∞ For z < 0 we have fZ (z) = y=−z fX (z + y)fY (y) dy = λµe−λz y=−z e−(λ+µ)y dy = λµeµz /(λ + µ) 12. (a) Now E[etX ] = α2 eµt /(α2 − t2 ). Hence if Y = kX + b then E[etY ] = ebt E[etkX ] = ebt α2 eµkt /(α2 − k 2 t2 ) = et(kµ+b) α12 /(α12 − t2 ) where α1 = α/k. This is the mgf of the Laplace(kµ + b, α/k) distribution. (b) Now X − µ ∼ Laplace(0, α); hence α(X − µ) ∼ Laplace(0, 1). Pn (c) By part α(X − µ) ∼ Laplace(0, 1); hence α|X − µ|exponential (1); hence α i=1 |Xi − µ| ∼ Gamma(n, 1); P(b), n hence 2α i=1 |Xi − µ| ∼ Gamma(n, 1/2) = χ22n 13. Now |X − µ| ∼ exponential (α) = Gamma(1, α). Hence result by equation(16.8a) on page 37. 14. Let W = X and Z = ln( X/Y ). Now (X, Y ) ∈ (0, 1) × (0, 1). Clearly W ∈ (0, 1). Also 0 < X < X/Y ; hence eZ = X/Y > X = W . This implies: if Z > 0 then 0 < W < 1 and if Z < 0 then 0 < W < eZ . z −z 1 Then | ∂(w,z) . ∂(x,y) | = /y = e /x. Hence f(W,Z) (w, z) = y = we R 1 −z R ez 1 −z If z > 0 then fZ (z) = 0 we dw = 2 e . If z < 0 then fZ (z) = 0 we−z dw = 21 ez . The Laplace(0, 1) distribution. 15. Let Z = X(2Y − 1). Then E[etX(2Y −1) ] = 12 E[e−tX ] + 12 E[etX ] = required. 1 α 2 α−t + 1 α 2 α+t = α 2α 2 α2 −t2 = α2 α2 −t2 for |t| < α as Page 104 Answers 1§25 Aug 1, 2018(17:54) Bayesian Time Series Analysis 16. Let Y1 = (X1 + X2 )/2; Y2 = (X3 + X4 )/2, Y3 = (X1 − X2 )/2 and Y4 = (X3 − X4 )/2. Then X1 X2 = Y12 − Y32 and X3 X4 = Y22 − Y42 . Hence X1 X2 − X3 X4 = (Y12 + Y42 ) − (Y22 + Y32 ). Now t1 − t3 t2 + t4 t2 − t4 t1 + t3 E exp ( i(t1 Y1 + t2 Y2 + t3 Y3 + t4 Y4 ) ) = E exp i X1 + X2 + X3 + X4 2 2 2 2 2 2 2 (t1 + t3 ) (t1 − t3 ) (t2 + t4 ) (t2 − t4 )2 = exp − − − − 8 8 8 8 2 2 2 2 t +t +t +t = exp − 1 3 2 4 = E[eit1 Y1 ] E[eit2 Y2 ] E[eit3 Y3 ] E[eit4 Y4 ] 4 2 1 Hence Y1 , Y2 , Y3 and Y4 are i.i.d. N (0, σ = /2). Hence 2(X1 X2 − X3 X4 ) = 2(Y12 + Y42 ) − 2(Y22 + Y32 ) is equal to the difference of two independent χ22 = exponential ( 1/2) distributions which is the Laplace(0, 1/2) distribution. (b) X1 X2 + X3 X4 = (Y12 + Y22 ) − (Y32 + Y42 ) and then as for part (a). 17. Using characteristic functions. Z ∞ Z ∞ Z ∞ √ √ 2 2 1 1 E[eitY 2X ] = E[eitY 2x ]e−x dx = e− 2 2xt e−x dx = e−x(1+t ) dx = 1 + t2 0 0 0 and this is the c.f. of the Laplace(0, 1) distribution. Hence result. √ √ 1 ,y2 ) √ Using densities. Let Y1 = X and Y2 = Y 2X. Hence Y1 > 0 and Y2 ∈ R and ∂(y 2x = 2y1 . Hence ∂(x,y) = f(Y1 ,Y2 ) (y1 , y2 ) = 18. 19. 20. 21. 22. 2 f(X,Y ) (x, y) 1 −x 1 −y2 /2 1 √ =√ e √ e = √ e−y1 e−y2 /4y1 2 πy 2y1 2y1 2π 1 Using the substitution y1 = z 2 /2 and equation(11.12b) on page 27 gives Z ∞ Z ∞ y2 z2 2 1 1 1 −(y1 +y22 /4y1 ) √ e−( 2 + 2z2 ) dz = e−|y2 | fY2 (y2 ) = e dy1 = √ 2 πy1 2 2π 0 0 as required. (a) The absolute value of the Jacobian of the transformation is ∂(x, y) cos θ −r sin θ = ∂(r, θ) sin θ r cos θ = r Hence for r ∈ (0, ∞) and θ ∈ (0, 2π) we have r −r2 /2σ2 f(R,Θ) (r, θ) = e 2πσ 2 2 2 Hence Θ is uniform on (0, 2π) with density f (θ) = 1/2π and R has density fR (r) = (r/σ 2 )e−r /2σ for r > 0. (b) 2 2 1 −(x2 +y2 )/2σ2 1 1 r = e for (x, y) ∈ R2 . f(X,Y ) (x, y) = 2 e−r /2σ σ 2π r 2πσ 2 Hence X and Y are i.i.d. N (0, σ 2 ). 2 2 (a) P[R ≤ r] = 1 − e−r /2σ for r ≥ 0. (b) Using the substitution r2 = y shows Z ∞ Z ∞ 1 1 n+2 n n+1 −r 2 /2σ 2 n/2 −y/2σ 2 n/2 n E[R ] = 2 r e dr = y e dy = 2 σ Γ σ 0 2σ 2 0 2 R ∞ n−1 −λx where the last equality comes from the integral of the gamma density: 0 x e dx = Γ(n)/λn . p (c) E[R] = σ π/2; E[R2 ] = 2σ 2 and√hence var[R] = (4 − π)σ 2 /2. (d) By differentiating the density, the mode is at σ. From part (a), the median is at σ 2 ln(2). √ 2 2 2 2 P[X ≤ x] = P[ −2 ln U ≤ x/σ] = P[−2 ln U ≤ x2 /σ 2 ] = P[ln U ≥ −x2 /2σ 2 ] = P[U ≥ e−x /2σ ] = 1 − e−x /2σ as required. 2 2 2 (a) Now fR (r) = re−r /2σ /σ 2 for r > 0. Let V = R2 . Then fV (v) = e−v/2σ /2σ 2 which is the exponential (1/2σ 2 ), 2 or equivalently Gamma(n = 1, α = 1/2σ ). (b) The density of V = R2 is fV (v) = e−v/2 /2 which is the exponential (1/2), or equivalently Gamma(n = 1, α = 1/2) = χ22 . (b) Because the sum of i.i.d. exponentials is 2 a Gamma distribution, the result follows. (c) X has density f (x) = λe−λx for x > 0. Hence fY (y) = 2λye−λy X √ for y > 0. This is the Rayleigh (1/ 2λ) distribution. First consider the case µ = 0. Then we have 2 2 2 2 x 1 fX (x) = 2 e−x /2s for x > 0. Also fY |X (y|x) = √ e−y /2x for y ∈ R. s 2πx2 Hence, by using the integration result in equation(11.12b) on page 27, we get Z ∞ Z ∞ 2 2 2 2 1 1 1 √ e− 2 (x /s +y /x ) dx = e−|y|/s fY (y) = fY |X (y|x)fX (x) dx = 2 2s 2π x=0 x=0 s which is the Laplace(0, 1/s) distribution. Finally, if Y |X ∼ N (µ, σ = X) then (Y − µ)|X ∼ N (0, σ = X) and (Y − µ) ∼ Laplace(0, 1/s). Hence result. Appendix Aug 1, 2018(17:54) Answers 2§2 Page 105 23. (a) Setting β = 1 gives fX (x) = e−x/γ /γ which is the exponential (1/γ) distribution. 2 (b) Setting β = 2 gives fX (x) = 2xe−(x/γ) /γ 2 ; this is the Rayleigh distribution with σ 2 = γ 2 /2. γ 1/β−1 24. Now dw . Hence dx = β x β−1 β βwβ−1 −w/γ)β β w e for w > 0. e−(w/γ) = fW (w) = γ γ γβ β dy 25. (a) Now dx = α; hence fY (y) = βy β−1 e−y/(αγ) /(αγ)β which is the Weibull (αγ, β) distribution. (b) Using the substitution y = xβ /γ β gives Z ∞ Z ∞ β n β y n/β e−y dy = γ n Γ 1 + xn+β−1 e−(x/γ) dx = γ n E[X n ] = β γ 0 β 0 2 1 2 (c) E[X] = γΓ(1 + /β ) where β is the shape and γ is the scale. Also var[X] = γ Γ(1 + /β ) − Γ(1 + 1/β )2 The median is γ(ln 2)1/β . The mode is γ(1 − 1/β )1/β if β > 1 and 0 otherwise. (d) E[et ln X ] = E[X t ] = γ t Γ 1 + t/β . 26. (a) It is h(x) = f (x)/[1 − F (x)] = βxβ−1 /γ β . (b) Straightforward. 27. (− ln U ) has the exponential (1) distribution; hence result by exercise 24. Chapter 2 Section 2 on page 59 T T T T T T T (exs-multiv.tex) T 1. cov[a X, b Y] = E[(a X − a µX )(b Y − b µY ) ] = a cov[X, Y]b. 2. (a) Just use E[X + a] = µX + b and E[Y + d] = µY + d. T T T (b) cov[AX, BY] = E (AX − Aµ )(BY − Bµ ) = E A(X − µ )(Y − µ ) B = Acov[X, Y]BT . X Y X Y T (c)h cov[aX + bV, cY + dW] = E (aX + bV − aµX − bµY )(cY i + dW − cµY − dµW ) which equals E {a(X − µX ) + b(V − µY )} {c(Y − µY ) + d(W − µW )} (d) Similar to (c). T which equals the right hand side. 3. 1 1 1 1 2 2 1 2 3 var[X] = . . . .. .. .. 1 2 3 T T ··· ··· ··· .. . 1 2 3 .. . ··· n 4. Let Y = X − α. Then E[(X − α)(X − α) ] = E[YY ] = var[Y] + µY µTY = var[X] + (µX − α)(µx − α)T . 5. Now cov[Y − AX, X] = cov[Y, X] − Avar[X]. Now ΣX = var[X] is non-singular. Hence if we take A = cov[Y, X]Σ−1 X then cov[Y − AX, X] = 0. 6. (a) Expanding the left hand side gives AE[XXT ]BT +AµbT +aµT BT +abT = AE[XXT ]BT +(Aµ+a)(Bµ+b)T −AµµT BT . Then use equation(1.1a): Σ = E[XXT ] − µµT . (b) This is just a special case of (a). (c) Now aT X is 1 × 1 and hence aT X = XT a. This implies E[XaT X] = E[XXT ]a and hence the result. 7. Left hand side equals E[XT AT BX] + µT AT b + aT Bµ + aT b = E[XT AT BX] − µT AT Bµ + (Aµ + a)T (Bµ + b). By equation(1.8a) we have E[XT AT BX] = trace(AT BΣ) + µT AT Bµ. Then use trace(A) = trace(AT ) and trace(AB) = trace(BA) and ΣT = Σ. Hence result. Other parts are all special cases of this result. 8. Just apply proposition(1.8a) on page 57 to the random vector X − b which has expectation µ − b and variance Σ. 9. Using the representation in example(1.7d) on page 57, we have Q = XT AX where A = I − n1 1. Using equation(1.8a) on page 57 gives E[Q] = E[XT AX] = trace(AΣ) + µT Aµ. Using A = I − n1 1 gives trace(AΣ) = trace(Σ) − n1 trace(1Σ) = 2 Pn Pn α − n1 (α + 2β) = [(n − 1)α − 2β] /n. Finally µT Aµ = µT Iµ − n1 µT 1µ = k=1 µ2k − n1 k=1 µk . 10. (a) Straightforward algebraic expansionPcombined with A = AT . Pn n (b) Let b = Aµ. Then W2 = 2bT Y = 2 k=1 bk Yk and var[W2 ] = 4σ 2 k=1 b2k = 4σ 2 bT b = 4σ 2 µT A2 µ. (c) Now E[W2 ] = 0; hence cov[W1 , W2 ] = E[W1 W2 ] = 2E[YT AYµT AY]. Now µT AY is 1 × 1; hence cov[W1 , W2 ] = T 2E[µT AYYT AY] = 2E[bT Zc] YYP . P where P c = AY and Z =P P Hence cov[W1 , W2 ] = 2E[ j k bj ck Zj,k ] = 2E[ j k bj ck Yj Yk ]. Using ck = ` ak,` Y` gives cov[W1 , W2 ] = P P P Pn 2E[ j j ` bj ak,` Yj Yk Y` ] = 2µ3 k=1 bk ak,k = 2µ3 bT d proving (c). (d) var[XT AX] = var[W1 ] + var[W2 ] + 2cov[W1 , W2 ] = (µ4 − 3σ 4 )dT d + 2σ 4 trace(A2 ) + var[W2 ] + 2cov[W1 , W2 ]. Pn 1 2 T 11. Let Q = k=1 (Xk − X) . By example(1.7d) on page 57, we know that Q = X AX where A = I − n 1. Then trace(AΣ) = trace(IΣ) − n1 trace(1Σ) = nσ 2 − σ 2 [1 + (n − 1)ρ] = σ 2 (1 − ρ)(n − 1). Also µT Aµ = (XT AX) = 0. X=µ Hence result. Page 106 Answers 2§4 Aug 1, 2018(17:54) Bayesian Time Series Analysis 12. By §10.3 on page 25, we know that E[(X − µ)2 ] = σ 2 , E[(X − µ)3 ] = 0 and E[(X − µ)4 ] = 3σ 4 . (a) Now (n − 1)S 2 = XT AX where A = I − n1 1. By equation(1.9a) on page 58, var[XT AX] = (µ4 − 3σ 4 )dT d + 2σ 4 trace(A2 ) + 4σ 2 µT A2 µ + 4µ3 µT Ad = 2σ 4 trace(A2 ) + 4σ 2 µT A2 µ. Now Aµ = 0. Hence var[XT AX] = 2σ 4 trace(A2 ). Now A = I − n1 1; hence A2 = I − n1 1 − n1 1 + n12 12 = I − n2 1 + n1 1 = A. Hence trace(A2 ) = trace(A) = n − 1 and so var[XT AX] = 2σ 4 (n − 1). Pn Pn−1 (b) By example(1.8e) on page 57, 2(n−1)Q = 2 k=1 Xk2 −X12 −Xn2 −2 k=1 Xk Xk+1 and 2(n−2)E[Q] = σ 2 (2n−2). 2 2 Also var[2(n−1)Q] = (µ4 −3σ 4 )dT d+2σ 4 trace(A2 )+4σ 2 µT AP µ+4µ3 µT Ad =P2σ 4 trace(A )+4σ 2 µTP A2 µ. P Again Aµ = n n Pn n n 4 2 2 2 0. Hence var[2(n − 1)Q] = 2σ trace(A ). Now trace(A ) = j=1 [A ]j,j = j=1 k=1 aj,k ak,j = j=1 k=1 a2j,k = 6(n − 2) + 4 = 6n − 8. Hence result. 13. The second term is an odd function of x3 ; hence Z Z 1 2 1 − 1 (x21 +x22 ) 1 − 1 (x21 +x22 ) 1 2 f(X1 ,X2 ) (x1 , x2 ) = f(X1 ,X2 ,X3 ) (x1 , x2 , x3 ) dx3 = e e− 2 x3 dx3 = e 2 1/2 (2π) (2π) (2π) x3 x3 Hence X1 and X2 are independent. Similarly for the pair X1 and X3 and the pair X2 and X3 . Because X1 , X2 and X3 all have the N (0, 1) distribution, it is clear that f(X1 ,X2 ,X3 ) (x1 , x2 , x3 ) 6= fX1 (x1 )fX2 (x2 )fX3 (x3 ) Chapter 2 Section 4 on page 64 (exs-bivnormal.tex) For questions 1–8, see the method in the solution of example(3.3a) on page 62. 1. Setting Q(x, y) = 32 (x2 − xy + y 2 ) gives a1 = 2/3, a2 = − 1/3 and a3 = 2/3. Hence √ 2/3 1 − 1/3 and |P| = . Hence c = 1/2π 3. P= 2 1 /3 − /3 3 2 (b) We have cov[X1 , X2 ] = −a2 /(a1 a3 − a2 ) = 1 and hence they are not independent. 2. Setting Q(x, y) = 2(x2 + 2xy + 4y 2 ) gives a1 = 2, a2 = 2 and a3 = 8. Clearly (µ1 , µ2 ) = (0, 0). Hence 1 8 −2 2 2 and |P| = 12. Also Σ = P−1 = P= 2 8 12 −2 2 √ √ 1/2 Finally |P| = 2 3 and so k = 3/π. 3. Setting Q(x, y) = 2x2 + y 2 + 2xy − 22x − 14y + 65 gives a1 = 2, a2 = 1 and a3 = 1. Hence 1 −1 2 1 −1 and |P| = 1 and Σ = P = P= −1 2 1 1 Finally k = 2π/|P|1/2 = 2π. Now for the mean vector: setting ∂Q(x,y) = 0 and ∂Q(x,y) = 0 gives 4µ1 + 2µ2 − 22 = 0 and 2µ2 + 2µ1 − 14 = 0. Hence ∂x ∂y (µ1 , µ2 ) = (4, 3). 4. Setting Q(y1 , y2 ) = y12 + 2y22 − y1 y2 − 3y1 − 2y2 + 4 gives a1 = 1, a2 = − 1/2 and a3 = 2. Hence 1 8 2 1 − 1/2 7/4 and Σ = P−1 = P= and |P| = − 1/2 2 7 2 4 1 ,y2 ) 1 ,y2 ) Setting ∂Q(y = 0 and ∂Q(y = 0 gives 2µ1 − µ2 − 3 = 0 and 4µ2 − µ1 − 2 = 0. Hence µ1 = 2 and µ2 = 1. ∂y1 ∂y2 5. Proceed as in question 2 above. Hence integral = π/√3. 6. Setting Q(y1 , y2 ) = 61 y12 + 2y1 (y2 − 1) + 4(y2 − 1)2 gives µ1 = 0, µ2 = 1, a1 = 1/6, a2 = 1/6 and a3 = 2/3. Hence 1 1 1 8 −2 P= and |P| = 1/12 and Σ = P−1 = −2 2 6 1 4 7. Using equation (3.3b) on page 62 for P gives XT PX − X12 σ12 = (σ22 X12 − 2ρσ1 σ2 X1 X2 + σ12 X22 )/[σ12 σ22 (1 − ρ2 )] − 2 2 2 2 2 − 2ρX1 X2 /σ1 σ2 + Xp (ρ 2 /σ2 )/(1 − ρ ) = (ρX1 /σ1 − X2 /σ2 ) /(1 − ρ ). Let Z = (ρX1 /σ1 − X2 /σ2 )/ 1 − ρ2 ). Then Z is normal with E[Z] = 0 and var[Z] 2ρcov[X1 , X2 ]/σ1 σ2 )/(1 − ρ2 ) = (ρ2 + 1 − 2ρ2 )/(1 − ρ2 ) = 1. Hence Z 2 ∼ χ21 as required. 2 X12 /σ12 X12 σ12 = = E[Z 2 ] = (ρ2 + 1 − 8. (a) E[X1 |X2 ] = E[Y − αX2 |X2 ] = E[Y |X2 ] − αX2 = E[Y ] − αX2 = µ1 + αµ2 − αX2 . (b) Suppose (X1 , X2 ) has a bivariate normal distribution with µ1 = µ2 = 0. Consider the random vector (X2 , Y ) with Y = X1 − ρσ1 X2 /σ2 . This is a non-singular linear transformation and hence by proposition (3.7a) on page 63 the new vector has a bivariate normal with mean (0, 0) and variance matrix 2 σ2 0 0 (1 − ρ2 )σ12 Hence Y and X2 are independent. Hence by part (a) we have E[X1 |X2 ] = X2 ρσ1 /σ2 . Hence if V1 = X1 + µ1 and V2 = X2 + µ2 then E[V1 |V2 ] = µ1 + (V2 − µ2 )ρσ1 /σ2 . Appendix Aug 1, 2018(17:54) Answers 2§4 Page 107 9. (a) Z has the N (0, 1) distribution. (b) Using 2 fXY (x, y) fXY (x, y) 1 x − 2ρxz + z 2 = p p fXZ (x, z) = = exp − ∂(x,z) 2(1 − ρ2 ) 1 − ρ2 2π 1 − ρ2 ∂(x,y) (c) We now have ∂(u, v) ∂(x, z) = σ1 σ2 and hence the density of (U, V ) is that given in equation(3.3a) on page 61. 10. Now Z Z Z 1 itY itQ(x1 ,x2 ) p E[e ] = e e−(1−it)Q(x1 ,x2 ) dx1 dx2 fX1 X2 (x1 , x2 ) dx1 dx2 = 2πσ1 σ2 1 − ρ2 x1 x2 Define α1 and α2 by σ1 = (1 − it)1/2 α1 and σ2 = (1 − it)1/2 α2 . Hence p α1 α2 1 1 p 2πα1 α2 1 − ρ2 = E[eitY ] = = 2 σ σ 1 − it 1 2 2πσ1 σ2 1 − ρ and hence Y has an exponential(1) distribution. 2 11. (a) If |Σ| = 0 then σ12 σ22 = σ12 . The possibilities are d (i) σ12 = 0, σ1 = 0 and σ2 6= 0; φ(t) = exp(− 21 σ22 t22 ). Hence (X1 , X2 ) = (0, Z) where Z ∼ N (0, σ22 ). d (ii) σ12 = 0, σ1 6= 0 and σ2 = 0; φ(t) = exp(− 21 σ12 t21 ). (X1 , X2 ) = (Z, 0) where Z ∼ N (0, σ12 ). d (iii) σ12 = 0, σ1 = 0 and σ2 = 0; φ(t) = 1; (X1 , X2 ) = (0, 0). (iv) σ12 6= 0 and ρ = σ12 /σ1 σ2 = +1; φ(t) = E[ei(t1 X1 +t2 X2 ) ] = exp − 21 (σ1 t1 + σ2 t2 )2 . Hence if Z = σ2 X1 − σ1 X2 then a.e. setting t1 = tσ2 and t2 = tσ1 gives φZ (t) = E[ei(tσ2 X1 −tσ1 X2 ) ] = E[exp(− 21 t2 × 0)] = 1. Hence σ2 X1 = σ1 X2 . a.e. (v) σ12 6= 0 and ρ = σ12 /σ1 σ2 = −1; σ2 X1 = −σ1 X2 . d (b) (i) (X1 , X2 ) = (µ1 , Z) where Z ∼ N (µ2 , σ22 ). d a.e. d (ii) (X1 , X2 ) = (Z, µ2 ) where Z ∼ N (µ1 , σ12 ). a.e. (iii) (X1 , X2 ) = (µ1 , µ2 ). (iv) σ2 (X1 − µ1 ) = σ1 (X2 − µ2 ). (v) σ2 (X1 − µ1 ) = −σ1 (X2 − µ2 ). 12. (a) Now (X, Y ) is bivariate normal with E[X] = E[Y ] = 0 and variance matrix a21 + a22 a1 b1 + a2 b2 a1 b1 + a2 b2 b21 + b22 Hence E[Y |X] = ρσ2 X/σ1 = X(a1 b1 + a2 b2 )/(a21 + a22 ). (b) By simple algebra (a2 T1 − a1 T2 )(a2 b1 − a1 b2 ) Y − E[Y |X] = a21 + a22 2 (a2 T1 − a1 T2 )2 (a2 b1 − a1 b2 )2 Y − E[Y |X] = (a21 + a22 )2 and E (a2 T1 − a1 T2 )2 = E(a22 T12 − 2a1 a2 T1 T2 + a21 T22 ) = a21 + a22 . Hence result. 13. Clearly ρ = 0 implies (X, Y ) are independent and hence X 2 and Y 2 are independent. Conversely, suppose X 2 and Y 2 are independent. Recall the characteristic function is 1 2 1 T 2 φX (t) = exp − t Σt = exp − t1 + 2ρt1 t2 + t2 2 2 2 2 2 2 Now E[X Y ] is the coefficient of t1 t2 /4 in the expansion of φ(t). Hence E[X 2 Y 2 ] is the coefficient of t21 t22 /4 in the expansion of 1 2 2 1 2 1− t1 + 2ρt1 t2 + t22 + t1 + 2ρt1 t2 + t22 2 8 Hence E[X 2 Y 2 ] is the coefficient of t21 t22 /4 in the expansion of 2 1 2 t1 + 2ρt1 t2 + t22 8 Hence E[X 2 Y 2 ] = 2ρ2 + 1. Independence of X 2 and Y 2 implies E[X 2 ]E[Y 2 ] = 1. Hence ρ = 0. This could also be obtained by differentiating the characteristic function—but this is tedious. Here are some of the steps: 1 φ(t) = exp − f (t1 , t2 ) where f (t1 , t2 ) = t21 + 2ρt1 t2 + t22 2 ∂φ(t) 1 ∂f 1 =− exp − f (t1 , t2 ) ∂t1 2 ∂t1 2 " 2 # ∂ 2 φ(t) 1 ∂ 2 f 1 ∂f 1 exp − f (t , t ) = − + 1 2 2 ∂t21 4 ∂t1 2 ∂t21 Page 108 Answers 2§4 Aug 1, 2018(17:54) Bayesian Time Series Analysis 1 ∂ 2 φ(t) = g(t , t ) exp − f (t , t ) where g(t1 , t2 ) = t21 + ρ2 t22 − 1 + 2ρt1 t2 1 2 1 2 2 ∂t21 " 2 # ∂ 4 φ(t) ∂2g 1 ∂2f ∂f ∂g 1 ∂f 1 = − g(t1 , t2 ) 2 − + g(t1 , t2 ) exp − f (t1 , t2 ) 2 ∂t2 ∂t2 4 ∂t2 2 ∂t21 ∂t22 ∂t22 ∂t2 Setting t1 = t2 = 0 gives 2ρ2 − 21 (−1)(2) = 2ρ2 + 1 as above. 14. Let V1 = X1 /σ1 and V2 = X2 /σ2 . Then (V1 , V2 ) has a bivariate normal distribution with E[V1 ] = E[V2 ] = 0 and var[V1 ] = var[V2 ] = 1. Then " # v12 − 2ρv1 v2 + v22 1 p fV1 V 2 (v1 , v2 ) = exp − 2(1 − ρ2 ) 2π 1 − ρ2 Consider the transformation (W, V ) with W = V1 and V = V1 /V2 . The range is effectively R2 —apart from a set of measure 0. The Jacobian is ∂(w, v) 1 0 |v1 | = ∂(v1 , v2 ) 1/v2 −v1 /v 2 = v 2 2 2 and hence v 2 fV V (v1 , v2 ) fW V (w, v) = 2 1 2 where v1 = w and v2 = w/v |v1 | Hence w2 (v 2 − 2ρv + 1) |w| p fW V (w, v) = exp − 2(1 − ρ2 )v 2 2πv 2 1 − ρ2 We now integrate out w to find the density of V : Z ∞ Z ∞ w w2 (v 2 − 2ρv + 1) 1 p p fV (v) = exp − dw = w exp −αw2 dw 2 )v 2 2 2 2 2 2(1 − ρ πv 1 − ρ πv 1 − ρ 0 0 v 2 − 2ρv + 1 1 p where α = 2(1 − ρ2 )v 2 2πv 2 α 1 − ρ2 p 1 − ρ2 = 2 π(v − 2ρv + 1) Now Z = X1 /X2 = V (σ1 /σ2 ) which is a 1 − 1 transformation. Hence fZ (z) = σ2 fV (v)/σ1 and the result follows. 15. The idea is to linearly transform (X1 , X2 ) to (V1 , V2 ) so that V1 and V2 are i.i.d. N (0, 1). In general, Σ = AAT and if we set V = A−1 X then var[V] = A−1 Σ(A−1 )T = I and so V1 and V2 are independent. p p p 1 p 1 p a b Suppose A = with a = 1 + ρ + 1 − ρ and b = 1 + ρ − 1 − ρ . Then a2 − b2 = 1 − ρ2 , b a 2 2 a2 + b2 = 1 and 2ab =ρ. Hence AAT = Σ. aX1 − bX2 −bX1 + aX2 1 a −b Also A−1 = 2 . So let V1 = and V2 = . a − b2 −b a a2 − b2 a2 − b2 Then E[V1 ] = E[V2 ] = 0, var[V1 ] = var[V2 ] = 1 and cov(V1 , V2 ) = 0. As (V1 , V2 ) is bivariate normal, this implies that V1 and V2 are i.i.d. N (0, 1). Hence V12 + V22 ∼ χ22 . But X 2 + X22 − 2ρX1 X2 V12 + V22 = 1 1 − ρ2 as required 16. Recall 2 1 x 1 2ρxy y 2 p f (x, y) = + exp − − 2(1 − ρ2 ) σ12 σ1 σ2 σ22 2πσ1 σ2 1 − ρ2 By using the transformation v = x/σ1 and w = y/σ2 which has ∂(v,w) ∂(x,y) = σ1 σ2 , we get Z ∞Z ∞ Z ∞Z ∞ 1 (x2 − 2ρxy + y 2 ) p P[X ≥ 0, Y ≥ 0] = f (x, y) dxdy = exp − dxdy (4.16a) 2(1 − ρ2 ) x=0 y=0 x=0 y=0 2π 1 − ρ2 Z 0 Z ∞ Z 0 Z ∞ 1 (x2 − 2ρxy + y 2 ) p f (x, y) dxdy = dxdy (4.16b) P[X ≤ 0, Y ≥ 0] = exp − 2(1 − ρ2 ) x=−∞ y=0 x=−∞ y=0 2π 1 − ρ2 Now use polar coordinates: x = r cos θ and y = r sin θ. Hence tan θ = y/x and r2 = x2 + y 2 . Also ∂(x,y) ∂(r,θ) = r. Hence 2 Z π/2 Z ∞ 1 r (1 − ρ sin 2θ) p P[X ≥ 0, Y ≥ 0] = r exp − drdθ 2(1 − ρ2 ) 2π 1 − ρ2 θ=0 r=0 Z π/2 Z ∞ 1 1 − ρ sin 2θ p = r exp[−αr2 ] drdθ where α = 2 2(1 − ρ2 ) 2π 1 − ρ θ=0 r=0 = Appendix Aug 1, 2018(17:54) Answers 2§6 Page 109 ∞ Z π/2 Z π/2 exp[−αr2 ] 1 1 1 p p dθ = dθ 2 2 −2α 2α 2π 1 − ρ θ=0 2π 1 − ρ θ=0 r=0 p Z 1 − ρ2 π/2 dθ = 2π 1 − ρ sin 2θ θ=0 Similarly, where the transformation θ → θ − π/2 is used in the last equality p p Z Z dθ dθ 1 − ρ2 π 1 − ρ2 π/2 P[X ≤ 0, Y ≥ 0] = = 2π 1 − ρ sin 2θ 2π 1 + ρ sin 2θ θ=π/2 θ=0 = dt We now use the transformation t = tan θ. Hence dθ = sec2 θ = 1 + t2 . Also sin 2θ = 2t/(1 + t2 ), cos 2θ = (1 − t2 )/(1 + t2 ) 2 and tan 2θ = 2t/(1 − t ). Hence P[X ≥ 0, Y ≥ 0] p p p Z Z Z 1 − ρ2 ∞ dt 1 − ρ2 ∞ dt 1 − ρ2 ∞ dt = = = 2 2 2 2 2 2π 2π 2π t=0 t − 2ρt + 1 t=0 (t − ρ) + (1 − ρ ) t=−ρ t + (1 − ρ ) p Z ∞ 1 dx 1 1 − ρ2 t x −1 p by using the standard result = = tan−1 + c tan p 2 2 2 2 2π x +a a a 1−ρ 1 − ρ t=−ρ " # ρ 1 1 1 π + tan−1 p = + sin−1 ρ as required. = 2 2π 2 4 2π 1−ρ The transformation (X, Y ) → (−X, −Y ) shows that P[X ≥ 0, Y ≥ 0] = P[X ≤ 0, Y ≤ 0]. Chapter 2 Section 6 on page 72 1. 2. 3. 4. 5. (exs-multivnormal.tex) (a) The characteristic function of X is φX (t) = exp itT µ − 21 tT Σt . Hence the characteristic function of Y = X − µ is φY (t) = E exp(itT Y) = E exp(itT X) exp(−itT µ) = exp − 21 tT Σt . Hence Y ∼ N (0, Σ). (b) E[exp(itT X)] = E[exp(it1 X1 )] · · · E[exp(itn Xn )] = exp(it1 µ1 − 21 σ 2 t21 ) · · · exp(itn µn − 21 σ 2 t2n ) = exp(itT µ − 1 2 T 2 σ t It) as required. Pn Pn (c) Σ = diag[d1 , . . . , dn ]. Hence φX (t) = exp i i=1 ti di − 12 i=1 t2i di = exp(it1 µ1 − 12 t21 d1 ) · · · exp(itn µn − 12 t2n dn ) which means that X1 , . . . , Xn are independent with distributionsN (µ1 , d1 ), . . . , N (µn , dn ) respectively. (d) φZ (t) = φX (t)φY (t) = exp itT (µX + µY ) − 21 tT (ΣX + ΣY )t (a) The vectors (X1 , X3 ) and X2 are independent iff the 2 × 1 matrix cov[(X1 , X3 ), X2 ] = 0 (by property of the multivariate normal). But cov[X1 , X2 ] = 0 and cov[X3 , X2 ] = 0. Hence result. (b) cov[X1 − X3 , X1 − 3X2 + X3 ] = var[X1 ] − 3cov[X1 , X2 ] + cov[X1 , X3 ] − cov[X3 , X1 ] + 3cov[X3 , X2 ] − var[X3 ] = 4 − 0 − 1 + 1 + 0 − 2 = 2. Hence not independent. (c) cov[X1 + X3 , X1 − 2X2 − 3X3 ] = 4 − 0 + 3 − 1 + 0 − 6 = 0. Hence independent. (a) Suppose x ∈ Rm with aT ax = 0; then xT aT ax = 0. Hence (ax)T (ax) = 0; hence ax = 0. Hence x1 α1 +· · ·+xm αm = 0 where α1 , . . . , αm are the columns of a. But rank(a) = m. Hence x = 0. Hence rank(aT a) = m and so aT a is invertible. (b) First note that if G is any symmetric non-singular matrix, then GG−1 = I; hence (G−1 )T G = I; hence G−1 = (G−1 )T . Clearly E[Y] = Bµ and var[Y] = σ 2 BBT = σ 2 (aT a)−1 aT a(aT a)−1 = σ 2 (aT a)−1 as required. By proposition(5.9a) on page 69, we know there exists a0 , a1 , a2 , a3 and a4 in R such that E[Y |X1 , X2 , X3 , X4 ] = a0 + a1 X1 + a2 X2 + a3 X3 + a4 X4 . Taking expectations gives a0 = 1. Also, taking expectations of E[Y X1 |X1 , X2 , X3 , X4 ] = X1 EY |X1 , X2 , X3 , X4 ] = X1 + a1 X12 + a2 X1 X2 + a3 X1 X3 + a4 X1 X4 gives 1/2 = a1 + 1/2(a2 +a3 +a4 ). Similarly 1/2 = a2 + 1/2(a1 +a3 +a4 ), 1/2 = a3 + 1/2(a1 +a2 +a4 ), and 1/2 = a4 + 1/2(a1 +a2 +a3 ). Subtracting shows that a1 = a2 = a3 = a4 ; indeed this was obvious from the symmetry in Σ. Combining this result with 1/2 = a + 1/2(a + a + a ) gives a = a = a = a = 1/5. 1 2 3 4 1 2 3 4 (a) The matrix of regression coefficients is −1 ρ23 σ2 σ3 σ22 −1 A ΣZ = [ ρ12 σ1 σ2 ρ13 σ1 σ3 ] ρ23 σ2 σ3 σ32 1×2 2×2 1 −ρ23 σ2 σ3 σ32 = [ ρ12 σ1 σ2 ρ13 σ1 σ3 ] 2 2 σ22 σ2 σ3 (1 − ρ223 ) −ρ23 σ2 σ3 σ1 = [ σ3 (ρ12 − ρ13 ρ23 ) σ2 (ρ13 − ρ12 ρ23 ] σ2 σ3 (1 − ρ223 ) T Using the general result that E[Y|Z] = µY − AΣ−1 (Z − µZ ) and var[Y|Z] = ΣY − AΣ−1 Z A gives Z σ1 X2 − µ2 X3 − µ3 E[X1 |(X2 , X3 )] = µ1 − (ρ12 − ρ13 ρ23 ) + (ρ13 − ρ12 ρ23 ) σ2 σ3 (1 − ρ223 ) 2 2 ρ − 2ρ12 ρ23 ρ13 + ρ13 var[X1 |(X2 , X3 )] = σ12 1 − 12 1 − ρ223 Page 110 Answers 2§9 Aug 1, 2018(17:54) Bayesian Time Series Analysis (b) The matrix of regression coefficients is ρ13 σ1 σ3 1 −1 A ΣZ = ρ23 σ2 σ3 σ32 2×1 1×1 T Using the general result that E[Y|Z] = µY − AΣ−1 (Z − µZ ) and var[Y|Z] = ΣY − AΣ−1 Z A gives Z 1 ρ13 σ1 σ3 (X3 − µ3 ) µ1 E[(X1 , X2 )|X3 ] = − 2 µ2 σ3 ρ23 σ2 σ3 (X3 − µ3 ) 1 ρ213 σ12 σ32 ρ13 ρ23 σ1 σ2 σ32 σ12 ρ12 σ1 σ2 var[(X1 , X2 )|X3 ] = − ρ223 σ22 σ32 ρ12 σ1 σ2 σ22 σ32 ρ13 ρ23 σ1 σ2 σ32 (1 − ρ213 )σ12 (ρ12 − ρ13 ρ23 )σ1 σ2 = (ρ12 − ρ13 ρ23 )σ1 σ2 (1 − ρ223 )σ22 6. Let 1 denote the n × 1-dimensional vector with every entry equal to 1. (a) Now c = cov[Yj , Y1 + · · · + Yn ] is the sum of the j th row of Σ. Adding over all n rows gives nc = 1T Σ1 = var[1Y] = var[Y1 + · · · + Yn ] ≥ 0. (b) Suppose there exists c ∈ R such that cov[Yj , Y1 + · · · + Yn ] = c for all j = 1, 2, . . . , n. Then Σ1 = c1 and hence 1 is an eigenvector of Σ. The converse is similar. (c) We are given that cov[Yj , Y1 + · · · + Yn ] = 0 for j = 1, 2, . . . , n. Consider the random vector Z = (Y1 , . . . , Yn , Y1 + · · · + Yn ) Because every linear combination `T Z of the components of Z has a univariate normal distribution, it follows that Z has a multivariate normal distribution. Also cov[Y, Y1 + · · · + Yn ] = 0. Hence Y = (Y1 , . . . , Yn ) is independent of Y1 + · · · + Yn . But Y1 + · · · + Yn is a function of Y = (Y1 , . . . , Yn ); hence Y1 + · · · + Yn is almost surely constant3 . Hence (X1 · · · Xn )1/n is almost surely constant. 7. Clearly v1 · v1 = v2 · v2 = · · · = vn · vn . Also v1 · vk = 0 for k = 2, . . . , n. Also v2 · vk = 0 for k = 3, . . . , n. And so on. Hence result. 8. We need to find the conditional density fXn |(Xn−1 ,···,X2 ,X1 ) (xn |(xn−1 , . . . , x2 , x1 ). This is the density of −1 T N µY + AΣ−1 where Y = {Xn } and Z = (Xn−1 , . . . , X1 ). Z (z − µZ ), ΣY − AΣZ A The matrix of regression coefficients is 1 ρ ρ2 · · · ρn−2 −1 1 ρ · · · ρn−3 1 ρ A Σ−1 = σ 2 [ ρ ρ2 · · · ρn−1 ] 2 . . . .. .. Z .. .. σ .. . 1×(n−1) (n−1)×(n−1) . n−2 n−3 n−4 ρ ρ ρ ··· 1 1 −ρ 0 0 ··· 0 0 0 2 −ρ 0 ··· 0 0 0 −ρ 1 + ρ 2 0 −ρ 1 + ρ −ρ · · · 0 0 0 1 = [ ρ ρ2 · · · ρn−1 ] . .. .. .. .. .. .. .. 1 − ρ2 . . . . . . . .. 0 0 0 0 · · · −ρ 1 + ρ2 −ρ 0 0 0 0 ··· 0 −ρ 1 = [ρ 0 0 ··· 0 0] Thus the distribution of Xn given (Xn−1 , . . . , X1 ) is N (µn +ρ(Xn−1 −µn−1 ), σ 2 (1−ρ2 )), proving the Markov property. Chapter 2 Section 9 on page 78 1. Use the transformation x22 = (ν + Z ∞ p 2 2π 1 − ρ fY (y) = 1+ −∞ 2 2 (exs-t.tex.tex) 2 1)(x − ρy) /(1 − ρ )(ν + y ). Then −(ν+2)/2 −(ν+2)/2 Z ∞ x2 − 2ρxy + y 2 (x − ρy)2 + (y 2 + ν)(1 − ρ2 ) dx = dx ν(1 − ρ2 ) ν(1 − ρ2 ) −∞ −(ν+2)/2 r Z ∞ y2 x22 (1 − ρ2 )(ν + y 2 ) = 1+ 1+ dx2 ν ν+1 ν+1 −∞ and hence −(ν+2)/2 √ y2 1+ B( 1/2, (ν+1)/2) ν + 1 ν −(ν+1)/2 B( 1/2, (ν+1)/2) √ y2 = ν 1+ 2π ν 1 fY (y) = 2π r ν + y2 ν+1 and then use 3 By definition, the random variables X and Y are independent iff the generated σ-fields σ(X) and σ(Y ) are independent. If Y = f (X), then σ(Y ) ⊆ σ(X) and hence for every A ∈ σ(Y ) we have A is independent of itself and so P(A) = 0 or P(A) = 1. Hence the distribution function of Y satisfies FY (y) = 0 or 1 for every y ∈ R and hence Y is almost surely constant. Appendix Answers 2§9 Page 111 Aug 1, 2018(17:54) 2π ν 2. (a) Provided ν > 1, E[X] = 0. Provided ν > 2, var[X] = ν/(ν − 2). (b) By the usual characterization of the t-distribution, E[XY ] = E[Z1 Z2 ]ν/E[1/W ] = ρν/(ν − 2). Hence corr[X, Y ] = ρ. 3. Now ν/2−1 h wi 1 1 T w f(Z1 ,...,Zp ,W ) (z1 , . . . , zp , w) = exp − z z exp − ν 2 2 (2π)p/2 2ν/2 Γ( 2 ) 1/2 p/2 p/2 T T Use the transformation T = Z (W/ν) ; this has Jacobian ν /w . Also, Z Z = W T T/ν. Hence h w wν/2−1 w i wp/2 f(T1 ,...,Tp ,W ) (t1 , . . . , tp , w) = (p+ν)/2 p/2 ν exp − tT t − 2ν 2 ν p/2 2 π Γ 2 w(p+ν)/2−1 w tT t = (p+ν)/2 p/2 p/2 ν exp − 1+ 2 ν 2 π ν Γ 2 Integrating out w gives Z ∞ 1 w tT t (p+ν)/2−1 w fT (t) = (p+ν)/2 p/2 p/2 ν exp − 1+ dw 2 ν 2 π ν Γ 2 0 B( 1/2, (ν+1)/2) B( 1/2, ν/2) = Now let y = (1 + tT t/ν)w. This leads to ∞ h yi (p+ν)/2−1 dy y exp − 2 2(p+ν)/2 π p/2 ν p/2 Γ 2 (1 + tT t/ν)(p+ν)/2 0 Z ∞ h yi νν (p+ν)/2−1 = (p+ν)/2 p/2 ν y exp − dy 2 2 π Γ 2 (ν + tT t)(p+ν)/2 0 p + ν νν (p+ν)/2 × 2 Γ = (p+ν)/2 p/2 ν 2 2 π Γ 2 (ν + tT t)(p+ν)/2 p+ν ν ν Γ 2 = p/2 ν π Γ 2 (ν + tT t)(p+ν)/2 T 4. Now C = LL and C is symmetric. Hence LT C = CLT ; hence C−1 LT = LT C−1 . Of course C−1 is also symmetric; hence LC−1 = C−1 L. Using (V − m)T = TT LT gives (V − m)T L = TT LT L = TT C and hence (V − m)T LC−1 = TT and T = C−1 LT (V − m). Hence TT T = (V − m)T C−1 (V − m). 5. First, note that T −1 T −1 (v − m1 )T C−1 1 (v − m1 ) = (v − a − Am) C1 (v − a − Am) = (At − Am) C1 (At − Am) fT (t) = Then use 1 Z ν T −1 = (t − m)T AT C−1 1 A(t − m) = (t − m) C (t − m) 1 fV (v) ∝ (ν+p)/2 ν + (t − m)T C−1 (v − m) 6. This follows from proposition(8.2a) on page 76. Page 112 Answers 2§9 Aug 1, 2018(17:54) Bayesian Time Series Analysis APPENDIX References [ACZ ÉL (1966)] ACZ ÉL , J. (1966). Lectures on functional equations and their applications. Academic Press. ISBN 0-12-043750-3. Cited on page 16 in 1§6.3. [A RNOLD et al. (2008)] A RNOLD , B., BALAKRISHNAN , N. & NAGARAJA , H. (2008). A first course in order statistics. SIAM. ISBN 0-89871-648-9. Cited on pages 6 in 1§2.7 and 6 in proposition 1(2.7a). [A SH (2000)] A SH , R. B. (2000). Real analysis and probability. 2nd ed. Academic Press. ISBN 0-12-065202-1. Cited on page 80 in answer to exercise 8 in 1§3. [B ILLINGSLEY (1995)] B ILLINGSLEY, P. (1995). Probability and measure. John Wiley. ISBN 0-471-00710-2. Cited on pages 28 in proposition 1(12.5a) and 29 in proposition 1(12.5b). [C RAWFORD (1966)] C RAWFORD , G. B. (1966). Characterization of geometric and exponential distributions. Annals of Mathematical Statistics, 37 1790–1795. Cited on pages 47 in exercise 22 in 1§21; 49 in proposition 1(22.5a) and 50 in proposition 1(22.6a). [DALLAS (1976)] DALLAS , A. (1976). Characterizing the Pareto and power distributions. Annals of the Institute of Statistical Mathematics, 28 491–497. Cited on pages 42 in 1§20.2; 43 in 1§20.2; 44 in 1§20.4; 44 in 1§20.4; 45 in exercise 7 in 1§21 and 46 in exercise 21 in 1§21. [DAVID & BARTON (1962)] DAVID , F. & BARTON , D. (1962). Combinatorial chance. Griffin. ISBN 0-85264-057-9. Cited on page 14 in 1§4.7. [F ELLER (1968)] F ELLER , W. (1968). An introduction to probability theory and its applications. Volume I . 3rd ed. John Wiley. ISBN 0-471-257087. Cited on pages 10 in 1§4.2; 11 in 1§4.2 and 16 in proposition 1(6.3a). [F ELLER (1971)] F ELLER , W. (1971). An introduction to probability theory and its applications. Volume II . 2nd ed. John Wiley. ISBN 0-471-257095. Cited on pages 14 in 1§4.7; 15 in exercise 9 in 1§5; 15 in exercise 11 in 1§5; 21 in 1§8.7 and 97 in answer to exercise 20 in 1§17. [F RISTEDT & G RAY (1997)] F RISTEDT, B. & G RAY, L. G. (1997). A modern approach to probability theory. Springer. ISBN 0-8176-3807-5. Cited on pages 6 in example 1(2.7b) and 66 in proposition 2(5.4a). [G ALAMBOS & KOTZ (1978)] G ALAMBOS , J. & KOTZ , S. (1978). Characterizations of probability distributions. Springer. ISBN 3-540-08933-0. Cited on pages 8 in exercise 14 in 1§3 and 16 in 1§6.4. [JAMES (1979)] JAMES , I. R. (1979). Characterization of a family of distributions by the independence of size and shape variables. Annals of Statistics, 7 869–881. Cited on pages 47 in 1§22.1; 51 in exercise 6 in 1§23 and 51 in exercise 7 in 1§23. [K AGAN et al. (1973)] K AGAN , A., L INNIK , Y. & R AO , C. (1973). Characterization problems in mathematical statistics. John Wiley. ISBN 0-471-45421-4. Cited on page 26 in proposition 1(10.6b). [K UCZMA (2009)] K UCZMA , M. (2009). An introduction to the theory of functional equations and inequalities. 2nd ed. Birkhaüser. ISBN 978-3-7643-8748-8. Cited on page 16 in 1§6.3. [L UKACS (1955)] L UKACS , E. (1955). A characterization of the gamma distribution. Annals of Mathematical Statistics, 26 319–324. Cited on page 21 in proposition 1(8.8a). Bayesian Time Series Analysis by R.J. Reed Aug 1, 2018(17:54) Answers Page 113 Page 114 Answers 2§9 [L UKACS (1970)] Aug 1, 2018(17:54) Bayesian Time Series Analysis L UKACS , E. (1970). Characteristic functions. 2nd ed. Griffin. ISBN 0-85264-170-2. Cited on page 80 in answer to exercise 8 in 1§3. [M ARSAGLIA (1989)] M ARSAGLIA , G. (1989). The X+Y, X/Y characterization of the gamma distribution. In G LESER , L., P ERLMAN , M. D., P RESS , S. J. & S AMPSON , A. R. (eds.), Contributions to probability and statistics. Essays in honor of Ingram Olkin.. ISBN 1-4612-8200-4. Cited on page 21 in proposition 1(8.8a). [M ATHAI & P EDERZOLI (1977)] M ATHAI , A. M. & P EDERZOLI , G. (1977). Characterizations of the normal probability law. John Wiley. ISBN 0-85226-558-1. Cited on page 26 in 1§10.6. [M ORAN (2003)] M ORAN , P. (2003). An introduction to probability theory. 2nd ed. Oxford University Press. ISBN 0-19-853242-3. Cited on page 26 in proposition 1(10.6a). [M OSSIMAN (1970)] M OSSIMAN , J. E. (1970). Size allometry: size and shape variables with characterizations of the lognormal and generalized gamma distributions. Journal of the American Statistical Association, 65 930– 945. Cited on page 47 in 1§22.1. [PATEL & R EAD (1996)] PATEL , J. K. & R EAD , C. B. (1996). Handbook of the normal distribution. 2nd ed. Dekker. ISBN 0-82479-342-0. Cited on page 26 in 1§10.6. [R AO (1973)] R AO , C. R. (1973). Linear statistical inference and its applications. 2nd ed. John Wiley. ISBN 0-471-21875-8. Cited on page 56 in 2§1.4. [S AATY (1981)] S AATY, T. (1981). Modern nonlinear equations. Dover. ISBN 0-486-64232-1. Cited on page 16 in 1§6.3. [S RIVASTAVA (1965)] S RIVASTAVA , M. S. (1965). A characterization of Pareto’s distribution and (k + 1)xk /θk+1 . Annals of Mathematical Statistics, 36 361–362. Cited on pages 43 in 1§20.2 and 44 in 1§20.4. [W HITWORTH (1901)] W HITWORTH , W. (1901). Choice and chance. 5th ed. George Bell. ISBN 1-332-81083-7. Cited on page 14 in 1§4.7. [W INKELBAUER (2014)] arXiv:1209.4340v2. Cited on page 25 in 1§10.3. W INKELBAUER , A. (2014). Moments and absolute moments of the normal distribution.