Solution of homework assignment 1 Problem 1. 1. Expectation: A EXt = 2π Z 2π cos(ω0 t + φ)dφ = 0, ∀t 0 because the integral of cos x over a period equals 0. Covariance: EXt Xt+h Z 2π 1 A2 cos(ω0 t + φ) cos(ω0 (t + h) + φ)dφ = 2π 0 " Z A2 2π 1 1 = cos(ω0 t + φ + ω0 (t + h) + φ) + cos(ω0 t + φ − ω0 (t + h) − φ)dφ 2π 0 2 2 i 2 Z 2π h A = cos(2ω0 t + 2φ + ω0 h) + cos(ω0 h) dφ 4π 0 A2 cos(ω0 h). = 2 Here the formula 2 cos α cos β = cos(α + β) + cos(α − β) is used. The process is stationary because EXt Xt+h depends only on h. 2. Expectation EXt = Eξ = 0, ∀t. Covariance: EXt Xt+h = Eξ 2 = 1, ∀h, t. Thus the process is stationary. Problem 2. 1. We have E(Y − c)2 = E(Y − c ± µ)2 = E(Y − µ)2 + (c − µ)2 , which is minimized with respect to c by the choice c = c∗ = µ. 2. Use the same argument, adding and subtracting E(Y |X): 2 2 2 E[ Y − f (X) ± E(Y |X) |X] = E[ Y − E(Y |X) |X] + f (X) − E(Y |X) , which is minimized by f (X) = E[Y |X]. 3. First write E[Y − f (x)]2 = E{E[Y − f (X)|X]2 } and then proceed as in the previous items. Problem 3. 1-2. First two items exactly as in Problem 2, where one needs to substitute Y = Xn+1 and vector (X1 , . . . , Xn ) instead of X. 1 3. If X1 , X2 . . . are i.i.d. then according to item 2 the best predictor is E[Xn+1 |X1 , . . . , Xn ] = EXn+1 = µ, and the minimum mean squared error of this predictor is E(Xn+1 − µ)2 = var(Xn+1 ). Problem 4. a. Xt = a + bZt−1 + cZt−3 . The mean function EXt = a + bEZt−1 + cEZt−3 = a, ∀t. The autocovariance function cov(Xt , Xt+h ) = E[(Xt − a)(Xt+h − a)] = E[(bZt−1 + cZt−3 )(bZt−1+h + cZt−3+h )] = b2 E(Zt−1 Zt−1+h ) + bcE(Zt−1 Zt−3+h ) + cbE(Zt−3 Zt−1+h ) + c2 E(Zt−3 Zt−3+h ) 2 2 σ (b + c2 ), h = 0, σ 2 cb, |h| = 2, = 0, otherwise. Hence the process is stationary. b. Xt = Z1 cos(ωt) + Z2 sin(ωt). The mean fuction E(Xt ) = E(Z1 ) cos(ωt) + E(Z2 ) sin(ωt) = 0, ∀t. The autocovariance function cov(Xt , Xt+h ) = E(Xt Xt+h ) = E{[Z1 cos(ωt) + Z2 sin(ωt)][Z1 cos(ω(t + h)) + Z2 sin(ω(t + h))]} = E(Z12 ) cos(ωt) cos(ω(t + h)) + E(Z22 ) sin(ωt) sin(ω(t + h)) = σ 2 [cos(ωt) cos(ω(t + h)) + sin(ωt) sin(ω(t + h))] = σ 2 cos(ωh). Here we have used the fact that Z1 and Z2 are independent, i.e. E(Z1 Z2 ) = 0, and the well–known formula cos(α − β) = cos(α) cos(β) + sin(α) sin(β). Thus the process is stationary. c. Xt = Zt cos(ωt) + Zt−1 sin(ωt). The mean function E(Xt ) = E(Zt ) cos(ωt) + E(Zt−1 ) sin(ωt) = 0, ∀t. Let’s compute cov(Xt , Xt+1 ). We have cov(Xt , Xt+1 ) = E(Xt Xt+1 ) = E{[Zt cos(ωt) + Zt−1 sin(ωt)][Zt+1 cos(ω(t + 1)) + Zt sin(ω(t + 1))]} = E(Zt2 ) cos(ωt) sin(ω(t + 1)) = σ 2 cos(ωt) sin(ω(t + 1)) 2 Because cov(Xt , Xt+1 ) depends on t, the process is non–stationary. d. Xt = a + bZ0 . The process is stationary, EXt = a, cov(Xt , Xt+h ) = b2 σ 2 , ∀h. e. Xt = Z0 cos(ωt). The process is non-stationary because var(Xt ) = E(Z02 ) cos2 (ωt) = σ 2 cos2 (ωt) depends on t. f. Xt = Zt Zt−2 . The mean function E(Xt ) = E(Zt Zt−2 ) = 0, ∀t. The autocovariance function cov(Xt , Xt+h ) = E(Xt Xt+h ) = = E(Zt Zt−2 Zt+h Zt+h−2 ) 4 σ , h=0 0, otherwise Here we have used the fact that {Zt } are independent. In fact, {Xt } is a white noise process! Problem 5. a. To show that Xt is not an IID noise it is sufficient to note that distributions of Xt for odd and even t are not the same. For even t, Xt = Zt ∼ N (0, 1), while for odd t, Xt is distributed as the 2 2 2 “normalized” χ2 (1) random variable. Indeed, Zt−1 ∼ χ2 (1); hence EZt−1 = 1 and var(Zt−1 ) = 2 and 2 E 12 (Zt−1 − 1)2 = 1. Thus, for all t, EXt = 0 and var(Xt ) = 1, however with different distributions. To show that {Xt } is a white noise sequence we note that Xt and Xt−h are independent for |h| ≥ 2 so EXt Xt+h = 0 for all |h| ≥ 2. Let us compute EXt Xt+1 . If t is even then EXt Xt+1 = EZt √12 (Zt2 − 1) = √1 (EZ 3 t 2 − EZt ) = 0. If t is odd then EXt Xt+1 = E √1 (Z 2 t−1 2 − 1)Zt+1 = 0. This shows that Xt is white noise. b. If n even then n + 1 is odd, Xn+1 = √1 (Z 2 n 2 − 1), Xn = Zn , and X1 , . . . , Xn−1 are determined via Zt with 1 ≤ t ≤ n − 1. Therefore, by independence of {Zt }, if n even, E[Xn+1 |X1 , . . . , Xn ] = E √1 (Z 2 n 2 − 1)|Zn = √1 (Z 2 n 2 − 1) = √1 (X 2 n 2 − 1). Now assume that n is odd; then n + 1 even, Xn+1 = Zn+1 and Xn+1 is independent of X1 , . . . , Xn . Thus, for odd n E[Xn+1 |X1 , . . . , Xn ] = EZn+1 = 0. Therefore the best predictor of Xn+1 on the basis of X1 , . . . , Xn is 0 if n is odd, and is even. 3 √1 (X 2 n 2 − 1) is n Problem 6. a. We have 2 MSE(a) = E(Xn+k − aXn )2 = E(Xn+k − 2aXn+k Xn + a2 Xn2 ) = γX (0) − 2aγX (k) + a2 γX (0). Minimizing the last expression w.r.t. a, we obtain a∗ = γX (k)/γX (0) = ρX (k). b. Substituting a = a∗ in the above expression for MSE(a) we have MSE(a∗ ) = γX (0)(1 − ρ2X (k)). Problem 7. Pp Let mt = k=0 ck tk ; then ∆mt = p X ck tk − k=0 = p−1 X p X ct (t − 1)k = k=0 p X ck tk − (t − 1)k k=0 ck tk − (t − 1)k + cp [tp − (t − 1)p ] k=0 Using the binomial formula we obtain tp − (t − 1)p = tp − p X p j=0 j (−1)p−j tj = − p−1 X p j=0 j (−1)p−j tj , which is a polynomial of degree p − 1 in t. Thus ∆mt is a polynomial of degree p − 1. Problem 8. We have Vt = ∆∆12 Xt = ∆∆12 (a + bt) + ∆∆12 st + ∆∆12 Yt = ∆(st − st−12 ) + ∆(Yt − Yt−12 ) = Yt − Yt−1 − Yt−12 + Yt−13 . Hence by stationarity of Yt E∆∆12 Xt = E(Yt − Yt−1 − Yt−12 + Yt−13 ) = 2µ − 2µ = 0, 4 and γV (h) = E(Vt Vt+h ) = E{(Yt − Yt−1 − Yt−13 + Yt−12 )(Yt+h − Yt+h−1 − Yt+h−13 + Yt+h−12 )} = γY (h) − γY (h − 1) − γY (h − 13) − γY (h − 12) −γY (h + 1) + γY (h) + γY (h − 12) + γY (h − 11) −γY (h + 13) + γY (h + 12) + γY (h) − γY (h + 1) +γY (h + 12) − γY (h + 11) − γY (h − 1) + γY (h) = 4γY (h) − 2γY (h − 1) − 2γY (h + 1) + 2γY (h − 12) + 2γY (h − 12) −γY (h − 11) − γY (h + 11) − γY (h + 13) − γY (h − 13). Problem 9. We have Wt = q X 1 {a + b(t + j) + Yt+j } 2q + 1 j=−q = a+ = q X 1 a + bt + Yt+j 2q + 1 j=−q q q X X b 1 (t + j) + Yt+j 2q + 1 j=−q 2q + 1 j=−q Therefore EWt = a + bt; hence Wt is not stationary. cov(Wt , Wt+h ) = q q X X 1 E{Yt+j Yt+k+h } (2q + 1)2 j=−q k=−q = 1 (2q + 1)2 q q X X γY (k − j + h) j=−q k=−q Note that even though Wt is non-stationary, cov(Wt , Wt+h ) does not depend on t. 5