University of Illinois Problem Set #2: Solutions Page 1 of 2 ECE 361 Spring 2001 Problems and Solutions: 1. Q(x) = p for x ≈ –2•ln p minus a correction term as given by (26.2.22) or (26.2.23) of Abramowitz and Stegun. Note that if we approximate Q(x) by its upper bound 0.5•exp(–x2 /2), we get x ≈ –2•ln 2p which is slightly smaller than –2•ln p (Prove this!). The approximations in Abramowitz and Stegun are not too useful in the sense that evaluating Q(•) at their spproximate solution can give a p-value that is off by as much as 4.5×10–4 ; not too helpful if p = 10–5 !! Note that Q(y) = 10 –5 for y ≈ 4.8 using ≈ –2•ln p or 4.65 using ≈ –2•ln 2p. Both numbers are larger than the “actual” value Q(4.26504) = 10–5 which I found iteratively using the 1/(x 2π)•exp(–x2 /2) approximation to Q(x). Hence, Q( 2x) = 10 –5 for x = 9.10 or 9.6dB. Since binary PSK has error probability Q(2Eb /N0 ) as you learned in ECE 359, we have just discovered that an SNR of 9.6 dB is required to achieve an error probability of 10–5 . 2T cos(πtT) ↔ rect(f/T)cos(πf/T) (what happened to s(–f)?) This is not quite π 1 – (2tT) 2 cos(απt/T) Tπ what we need: replace the constant T by the constant α/T to get ↔ rect(fT/α)cos(πfT/α). 1 – (2αt/T)2 2α X(f) is a cosinusoidal function of amplitude X(0) = (Tπ/2α) > T (because π > 2 and α ≤ 1) and support [–α/2T,α/2T] ⊂ [–1/2T,1/2T] for α < 1. Area under X(f) = x(0) = 1. (b),(c) The two functions are as shown in the leftmost figure below. Now, x(t)y(t)↔ X(f)✳Y(f) = Z(f) where ✳ 2.(a) By duality, we have that ∞ denotes convolution. Z(f) = ∞ ∫ X(λ)Y(f–λ)dλ = ∫ X(λ)Y(λ–f)dλ since Y(f) is an even function. Thus, we -∞ -∞ see that “flipping” is not necessary; just “sliding” Y to the right by f suffices. Since X and Y are even functions of f, so must Z be an even function of f (why?). Now, if 0 ≤ f ≤ (1–α)/2T, then Y(λ–f) is as π/2 +α/2T shown in the second figure below, and we get Z(f) = ∫ (Tπ/2α)cos(πλT/α)Tdλ = (T/2) ∫ cos x dx = T. –π/2 –α/2T On the other hand, if (1–α)/2T ≤ f ≤ (1+α)/2T, then Y(λ–f) is as shown in the third figure below where the left end point of Y(λ–f) is f–1/2T (be sure you understand why this is so!), and thus we get π/2 +α/2T Z(f) = ∫ (Tπ/2α)cos(πλT/α)Tdλ = (T/2) ∫ cos x dx = (T/2)(1 – sin β) where β = π(fT/α–1/2α). β f–1/2T Finally, if f > (1+α)/2T, then Y(λ–f) is as shown in the right hand figure below and hence Z(f) = 0. X(f) Y(f) Τπ/2α T α 1 2T 2T −1 2T −α 2T −α 2T Putting the results together, we get the following result (cf. Blahut, p. 28) T, πT(|f| – 1/2T) Z(f) = [1 – sin ( )], α 0, T 2 ∞ 3.(a) ∞ 1– α 0 ≤ |f| ≤ 2T , 1– α 1+α ≤ |f| ≤ 2T , 2T 1+α |f| > 2T . ∞ ∞ M M M M * ⌠ M 2 ∑si (t) dt = ∑si (t) ∑sj (t) dt = ∑ ∫|si (t)| dt + ∑ ∑ ∫si (t)[sj (t)]*dt i=1 –∞ i=1 i ≠j j=1 –∞ ⌡i=1 ⌡ i=1 j=1 ⌠ M –∞ 2 –∞ University of Illinois Problem Set #2: Solutions Page 2 of 2 M = MEs + (b) M ∑ ∑ ECE 361 Spring 2001 ρij = ME s + M(M–1) –ρ where –ρ is the average of the M(M–1) inner products. But i=1 i ≠j j=1 the integral must be nonnegative (why?) and hence MEs + M(M–1) –ρ ≥ 0, that is, –ρ ≥ – E s /(M–1). ∞ ∞ ∞ 2 M M M * ⌠ ⌠ M |ai si (t)|2 dt > 0 because each integral is ai si (t) dt = ai si (t) aj sj (t) dt = i=1 –∞ j=1 ⌡i=1 ⌡ i=1 –∞ –∞ M nonnegative, and at least one is strictly positive. Hence, the sum ai si (t) cannot be identically zero. i=1 ∞ ∞ ∞ ∞ ∞ – – |qi (t)|2dt = |si (t) – s(t)| 2 dt = |si (t)|2dt + |s(t)| 2 dt – si (t)[–s(t)]* + [s i (t)]*–s(t)dt –∞ –∞ –∞ –∞ –∞ 1 1 1 = Es + M 2 E s – 2 Es = Es 1 – because of the orthogonality of the signals si (t). It follows from M M M part (a) that some inner product of the qi (t)’s must be at least as large as [E s (M–1)/M]/(M–1) = –Es /M. It is easy to show that the sum of the qi (t) is 0 and thus the signals cannot be independent. ∞ ∞ 1 1 1 σij = qi (t)[qj (t)]*dt = [si (t) – –s(t)][sj (t) – –s(t)]*dt = 0 + M 2 E s – 2 Es = – Es for all i ≠ j. M M M –∞ –∞ Hence, all the inner products have value equal to the minimum possible average value of the inner product. ∑ ∑ ∑ ∫ ∑ ∑ (c) ∫ ∫ ∫ ∫ ( (d) ) ∫ ∫ T/2 4.(a) E0 = ∫ ∫|s0(t)|2dt = ||s0||2 = A2T. –T/2 T/2 E1 = ∫|s1(t)|2dt = ||s1||2 = A2T. This latter is obtained via a standard –T/2 integral that occurs so often that you should just memorize the following result! The energy delivered by a sinusoidal signal over any time interval that is a multiple of one quarter of the period is T/2 0.5×(amplitude)2 ×(time interval). Note also that 〈s0 ,s 1 〉 = 2A2 cos(πt/T)dt = 2 ∫ –T/2 2A2 T/π. The signal outputs are thus A 2 T – 2 2A2 T/π if a 0 is transmitted and 2 2A2 T/π – A 2 T if a 1 is transmitted. T/2 (b) σ2 = ∫ N0 N N | s0 (t)–s1 (t)|2 dt = 0 ||s0 –s1 ||2 = 0 2 –T/2 2 2 [||s0||2 + ||s1 ||2 – 2 〈s0 ,s 1 〉] = N0 A2 T(1–2 2/π). If a 0 is transmitted, z is a Gaussian random variable with mean A2 T – 2 2A2 T/π = A2 T(1–2 2/π) and variance N0 A2 T(1–2 2/π). Since the threshold is (E0 –E1 )/2 = 0, Pe,0 = P{z < 0 | 0 transmitted} = Q A2 T(1–2 2/π) = Q ||s0 –s1 || exactly as obtained in class. 2N N0 2 T(1–2 2/π) 0 N A 0 ||s –s || Similarly, Pe,1 = P e = Q 0 1 also. 2N0 = Q A2 T(1–2 2/π) T/2 (c) The difference now is that 〈s0 ,s 1 〉 = ∫ 2A2 sin(πt/T)dt = 0, that is, the signals are orthogonal. Hence, –T/2 the signal outputs are ±A2 T while the noise variance is N 0 A2 T. Hence, Pe = (d) Q A2 T = Q N0 as you might remember from a previous course (or Problem Set #1). Since 1 – 2 2/π < 1, the error probability in part (b) is larger than the error probability on part (c) Eb N0