Sums of independent variables
A cab is assigned after a random delay and then drives a random time to the rider. Knowing how each part is distributed, how is the total distributed, so that the dispatcher can say how likely a pickup within eleven minutes is?
Where you are 5 in Several random variables
Why this exists
Linearity gives the mean of a sum (Expectation, linearity and indicators, Result 2), and for independent terms the variances add (Variance and covariance, Result 5; Combining random variables). A dispatcher needs more. Jasmine assigns a cab after a delay X, and it then drives a time Y to the rider; promising a pickup within 11 minutes needs P(X + Y \le 11), which the mean and variance do not determine. Samir, at an ambulance desk, wants the chance that the third call of a shift comes within six minutes, a sum of three gaps.
For independent variables the distribution of the sum comes from those of the terms by one operation, convolution. Repeated, it shows why a few families keep their shape under addition, the starting point of Moment generating functions and the central limit theorem.
Definition
In plain words: to find the chance that X + Y equals s, run over every way of splitting s as x + (s - x), and add up the chances that X = x and Y = s - x together. For independent variables each of those chances is a product.
Discrete. Let p and q be pmfs. Their convolution is the function
the sum running over the countably many x with p(x) > 0.
Densities. Let f and g be densities. Their convolution is
an integral of a nonnegative function, so a number or +\infty. The substitution x \mapsto s - x (re-indexing, for sums) shows f * g = g * f.
The integrand is nonzero only where x lies in the support of f and s - x in the support of g. Finding that set of x for each s is the real work of a convolution; it changes as s passes the ends of the supports, so f * g usually comes in pieces.
Gamma. For a whole number n \ge 1 and a rate \lambda > 0, T \sim \operatorname{Gamma}(n, \lambda) when
and f_T(t) = 0 for t \le 0. \operatorname{Gamma}(1, \lambda) is \operatorname{Exponential}(\lambda). Other shapes than whole numbers use the gamma function in place of (n - 1)!; this lesson needs only whole n.
Intuition
Picture the pair (X, Y) in the plane. The event \{X + Y \le s\} is the half-plane below the line x + y = s, and as s grows the line sweeps across the support. The density of the sum at s measures how much joint density the line meets. For independent terms the joint density is f_X(x) f_Y(y) (Joint distributions, Result 4(a)), which along the line is f_X(x) f_Y(s - x); integrating it gives the convolution.
Figure 1 shows Example 2, where (X, Y) is uniform on a rectangle. Each line x + y = s cuts a chord from the rectangle, and f_S(s) is proportional to its length: short near the corners, longest across the middle. So the density of the sum rises, stays level, and falls (Figure 2).
A second view conditions on Y: given Y = y, the sum is at most s exactly when X \le s - y, and independence leaves X unchanged. Averaging over Y by the tower rule of Conditional expectation gives P(X + Y \le s) = E(F_X(s - Y)), the first formula of Result 2.
Results
Result 1 (discrete convolution). Let X and Y be independent and discrete, with pmfs p_X and p_Y. Then S = X + Y is discrete with pmf p_X * p_Y:
If X and Y take only whole-number values, P(S = s) = \sum_{k=0}^{s} p_X(k) \, p_Y(s - k) for s = 0, 1, 2, \dots.
Result 2 (convolution of densities). Let X and Y be independent with densities f_X and f_Y, and S = X + Y. For every real s,
So F_S is continuous, and f_X * f_Y is a density of S whenever it is continuous except at finitely many points of each bounded interval, as in every example here (measure theory removes that proviso).
Result 3 (families closed under sums). Let X_1, \dots, X_n be independent and S_n = X_1 + \dots + X_n.
(a) If X_i \sim \operatorname{Binomial}(m_i, p), with one p for all, then S_n \sim \operatorname{Binomial}(m_1 + \dots + m_n, p).
(b) If X_i \sim \operatorname{Poisson}(\lambda_i), then S_n \sim \operatorname{Poisson}(\lambda_1 + \dots + \lambda_n).
(c) If every X_i \sim \operatorname{Exponential}(\lambda), with one \lambda, then S_n \sim \operatorname{Gamma}(n, \lambda), and for t \ge 0
(d) If X_i \sim N(\mu_i, \sigma_i^2), then S_n \sim N(\mu_1 + \dots + \mu_n, \ \sigma_1^2 + \dots + \sigma_n^2), a The normal distribution again.
The right side in (c) is P(N \le n - 1) for N \sim \operatorname{Poisson}(\lambda t), so a gamma tail can be read from Poisson probabilities.
Why it is true
Result 1. S takes countably many values, sums of values of X and Y. The event \{S = s\} is the disjoint union of the events \{X = x, \ Y = s - x\} over the values x of X, apart from an event of probability 0 on which X takes no value of positive probability. Countable additivity and the product rule of Result 4 of Independence of events and variables give the sum. For whole-number values, p_X(k) \, p_Y(s - k) = 0 unless 0 \le k \le s.
Result 2. By Result 4(a) of Joint distributions, (X, Y) has joint density f_X(x) f_Y(y), and by Result 2 there F_S(s) is its integral over the half-plane x \le s - y. Integrating first in x (Fubini, for a nonnegative integrand), then substituting x = u - y,
The inner integral of the first line is F_X(s - y), the first formula. Exchanging the order in the second line gives \int_{-\infty}^{s} \int f_Y(y) f_X(u - y) \, dy \, du, and the inner integral is (f_Y * f_X)(u) = (f_X * f_Y)(u). Since F_S(s) = \int_{-\infty}^{s} h with h = f_X * f_Y \ge 0, F_S is continuous, so P(S = a) = 0 and P(a \le S \le b) = F_S(b) - F_S(a) = \int_a^b h, the definition of a density.
Result 3. By Result 5 of Independence of events and variables, S_{n-1}, a function of X_1, \dots, X_{n-1}, is independent of X_n. So by induction on n it is enough to add one independent term: S_n = S_{n-1} + X_n, with S_{n-1} of the stated form.
(a) Let m = m_1 + \dots + m_{n-1}, r = m_n and q = 1 - p. By Result 1,
and the sum is \binom{m + r}{s} p^s q^{m + r - s}: choosing s of m + r objects is choosing k of the first m and s - k of the other r, for some k (Counting models).
(b) The case of two terms is Exercise 16 of Discrete distributions: binomial, geometric, Poisson, proved there from Result 1 and the binomial theorem.
(c) For n = 1 there is nothing to prove. For n \ge 2, S_{n-1} \sim \operatorname{Gamma}(n - 1, \lambda), and for s > 0 the integrand of the convolution is nonzero only for 0 < x < s, where e^{-\lambda x} e^{-\lambda (s - x)} = e^{-\lambda s} does not depend on x:
the \operatorname{Gamma}(n, \lambda) density; for s \le 0 the convolution is 0. It is continuous, so Result 2 applies. For the tail, let G(t) be the right side, so G(0) = 1 and G(t) \to 0 as t \to \infty. Differentiating, the second sum below is the first without its last term, so
Hence 1 - G(t) = \int_0^t f_{S_n}, which is P(S_n \le t).
(d) Subtracting a constant shifts a density: P(X - \mu \le t) = F_X(t + \mu). So X_i - \mu_i \sim N(0, \sigma_i^2), and S_n minus \mu_1 + \dots + \mu_n is the sum of these. It is therefore enough to convolve N(0, a^2) with N(0, b^2). Let c^2 = a^2 + b^2. Expanding both sides confirms
With \tau = ab/c, the integral of e^{-(x - m)^2/(2\tau^2)} over the line is \displaystyle \tau \sqrt{2\pi}, since the N(m, \tau^2) density has total 1. So
the N(0, c^2) density, continuous, so Result 2 applies. Adding back the means shifts it to N(\mu_1 + \mu_2, c^2).
Worked examples
Cabs at two stands
Jasmine needs four cabs for a group pickup. The numbers X free at the north stand and Y free at the south stand are independent, with
| k | p_X(k) | p_Y(k) |
|---|---|---|
| 0 | 0.2 | 0.1 |
| 1 | 0.5 | 0.3 |
| 2 | 0.3 | 0.4 |
| 3 | 0 | 0.2 |
Find the pmf of S = X + Y and P(S \ge 4), and check E(S).
SolutionBy Result 1 with whole-number values, p_S(s) = \sum_{k=0}^{s} p_X(k) \, p_Y(s - k). For s = 2 the splits are 0 + 2, 1 + 1 and 2 + 0, so p_S(2) = 0.2(0.4) + 0.5(0.3) + 0.3(0.1) = 0.26. In full,
| s | 0 | 1 | 2 |
|---|---|---|---|
| p_S(s) | 0.02 | 0.11 | 0.26 |
| s | 3 | 4 | 5 |
| p_S(s) | 0.33 | 0.22 | 0.06 |
with total 1. So P(S \ge 4) = 0.22 + 0.06 = 0.28. From the table, E(S) = 0.11 + 0.52 + 0.99 + 0.88 + 0.30 = 2.8, and linearity agrees: E(X) + E(Y) = 1.1 + 1.7 = 2.8.
Assign, then drive
Jasmine’s cab is assigned after X \sim \operatorname{Uniform}(0, 2) minutes and then drives Y \sim \operatorname{Uniform}(4, 10) minutes to the rider, independently. Find the density of the wait S = X + Y, P(S \le 11), and the median of S.
SolutionThe integrand f_X(x) f_Y(s - x) is \tfrac12 \cdot \tfrac16 = \tfrac{1}{12} when 0 < x < 2 and 4 < s - x < 10, that is s - 10 < x < s - 4, and 0 otherwise. So f_S(s) is \tfrac{1}{12} times the length of the overlap of (0, 2) with (s - 10, s - 4), the chord of Figure 1. The overlap is (0, s - 4) for 4 < s < 6, all of (0, 2) for 6 \le s \le 10, and (s - 10, 2) for 10 < s < 12:
and 0 elsewhere, the trapezoid of Figure 2. It is continuous, so Result 2 makes it the density of S. Then
so P(S \le 11) = \tfrac{23}{24} \approx 0.958. Next, \displaystyle F_S(6) = \int_4^6 \frac{s - 4}{12} \, ds = \tfrac16, and F_S then rises by \tfrac16 per minute, so \displaystyle F_S(m) = \tfrac16 + \frac{m - 6}{6} = \tfrac12 at m = 8 minutes.
Answer, then assign
At Samir’s ambulance desk a call waits X \sim \operatorname{Exponential}(2) minutes to be answered and then Y \sim \operatorname{Exponential}(1) minutes until a unit is assigned, independently. Find the density of S = X + Y, its mode (the point where the density is largest), and P(S > 3).
SolutionFor s > 0 the integrand is nonzero only for 0 < x < s, so by Result 2
and f_S(s) = 0 for s \le 0; this is continuous, as it tends to 0 at 0. Its derivative -2e^{-s} + 4e^{-2s} = 2e^{-2s}(2 - e^s) is positive before s = \ln 2 and negative after, so the mode is \ln 2 \approx 0.693 minutes. Finally
The rates differ, so S is not gamma: Result 3(c) needs one rate.
The third call of a shift
Calls reach Samir’s desk with independent gaps, each exponential with mean 4 minutes. Let T be the time from the start of the shift to the third call. Find the density of T, its mean and standard deviation, and P(T \le 6).
SolutionEach gap is \operatorname{Exponential}(\tfrac14) (Expectation, linearity and indicators, Result 5), and T is the sum of three, so by Result 3(c) T \sim \operatorname{Gamma}(3, \tfrac14):
By linearity E(T) = 3 \cdot 4 = 12 minutes, and as the gaps are independent with variance 16 (Variance and covariance, Results 5 and 7), \operatorname{Var}(T) = 48 and \displaystyle \sigma_T = 4\sqrt3 \approx 6.93 minutes. The tail formula with \lambda t = \tfrac32 gives
Figure 3 compares the times to the first, second and third calls.
Door to door
For a booked ride, the wait X for pickup is N(9, 4) and the ride Y is N(22, 12), in minutes, independently. Find P(X + Y > 36), and the time that the door-to-door total stays under with probability 0.95.
SolutionBy Result 3(d), S = X + Y \sim N(31, 16), so \sigma_S = 4. With \Phi the standard normal distribution function,
As \Phi(1.645) \approx 0.95, the total stays under 31 + 1.645 \cdot 4 \approx 37.58 minutes with probability 0.95. The mean and variance alone give only Chebyshev’s bound P(|S - 31| \ge 5) \le \tfrac{16}{25} (Markov and Chebyshev inequalities, Result 3).
Where the first call fell
Let X_1 and X_2 be independent \operatorname{Exponential}(\lambda), the gaps before the first and second calls, and S = X_1 + X_2. Prove that given S = v > 0, the time X_1 of the first call is \operatorname{Uniform}(0, v). So if Samir’s second call came at minute 10, the first is equally likely to have come at any time in the first 10 minutes.
SolutionBy Result 4(a) of Joint distributions, (X_1, X_2) has joint density \lambda^2 e^{-\lambda(x_1 + x_2)} on the open quadrant D = (0, \infty)^2. The map (x_1, x_2) \mapsto (u, v) = (x_1, x_1 + x_2) takes D one-to-one onto E = \{(u, v) : 0 < u < v\}, with inverse x_1 = u, x_2 = v - u, whose Jacobian is 1 \cdot 1 - 0 \cdot (-1) = 1. By Result 7 there, (X_1, S) has joint density \lambda^2 e^{-\lambda v} on E. By Result 3 there,
the \operatorname{Gamma}(2, \lambda) density of Result 3(c). For v > 0 the conditional density of Result 5 there is
the \operatorname{Uniform}(0, v) density, whatever the rate.
Boundaries
Independence is needed. If X \sim \operatorname{Uniform}(0, 1) and Y = 1 - X, then Y is also \operatorname{Uniform}(0, 1), but X + Y = 1 has no density at all, while independent copies have a density on (0, 2) by Result 2.
The parameters must match. Result 3(a) needs one p and Result 3(c) one rate; Example 3 adds two exponentials with different rates and gets a density that is not gamma. Most families are not closed at all, uniforms (Example 2) and geometrics (Exercise 4) among them.
Mixed sums need their own formula. If one term is discrete and the other has a density, neither Result 1 nor Result 2 applies; the sum has a density that is a weighted sum of shifted copies (Exercise 13).
Where this goes next
Repeated convolution is heavy by hand. The moment generating function turns it into multiplication, M_{X + Y} = M_X M_Y for independent terms, and recovers Result 3 in a line. Convolving one density with itself many times and rescaling approaches the normal curve, the The central limit theorem, proved. A random sum, as in Result 5 of Conditional expectation, averages the n-fold convolutions with weights P(N = n). The The bivariate normal is a dependent pair for which every aX + bY is still normal.
Exercises
Discrete sums
- Jasmine’s log gets X fare entries in a quiet quarter hour, equally likely to be 1, 2 or 3, and Y cancellation entries, independent of X, with P(Y = 0) = \tfrac12, P(Y = 1) = \tfrac13 and P(Y = 2) = \tfrac16. Find the pmf of the number of entries X + Y, and P(X + Y \ge 4).
- The north fleet has 12 cabs and the south fleet 8, and each cab is on duty with probability 0.7, independently. Find the distribution of the number T on duty and P(T \ge 16).
- Medical calls reach Samir’s desk at \operatorname{Poisson}(2.5) per hour and fire calls at \operatorname{Poisson}(0.5) per hour, independently. Find the probability of no calls in an hour, of at most 2 calls, and of exactly one fire call given 4 calls in all.
- Each radio call Samir makes reaches a driver with probability \tfrac13, independently, and he needs two drivers, found one after the other. Let T be the total number of calls. Find P(T = n), P(T = 4) and E(T).
Densities
- Let X \sim \operatorname{Uniform}(0, 2) and Y \sim \operatorname{Exponential}(1) be independent. Find the density of S = X + Y and P(S > 3).
- Let X \sim \operatorname{Exponential}(\lambda) and Y \sim \operatorname{Exponential}(\mu) be independent with \lambda \ne \mu. Show that S = X + Y has density \displaystyle \frac{\lambda\mu}{\mu - \lambda} (e^{-\lambda s} - e^{-\mu s}) for s > 0, and find P(S \le 1) when \lambda = 1 and \mu = 3.
- Cab requests reach Jasmine with independent exponential gaps of mean 2 minutes. Let T be the time to the fourth request. Find its density, mean and standard deviation, and P(T > 10).
- Let X \sim \operatorname{Uniform}(0, 1), and let Y, independent of X, have density 2y on (0, 1). Find the density of S = X + Y, P(S \le 1) and E(S).
- Two ambulances head to the same scene, arriving after X \sim N(11, 9) and Y \sim N(14, 16) minutes, independently. Find P(Y < X) and P(X + Y > 30).
- Samir’s call handling over a shift totals N(40, 36) minutes, independently over shifts. Find the distribution of the total over 5 shifts and the probability that it exceeds 220 minutes.
- One call in four at Samir’s desk needs a transfer that adds exactly 2 minutes; the others need none. The call then takes Y \sim \operatorname{Exponential}(1) minutes to handle, independent of the transfer. With X \in \{0, 2\} the transfer time and S = X + Y, find P(S > 3) by conditioning on X, and the density of S.
Proofs
- Let X and Y be independent with densities. Prove that X - Y has density \int_{-\infty}^{\infty} f_X(s + y) \, f_Y(y) \, dy, under the proviso of Result 2.
- Let X take the values x_1, \dots, x_m with probabilities p_1, \dots, p_m, and let Y, independent of X, have a density g. Prove that X + Y has the density \sum_j p_j \, g(s - x_j).
- Let T_r be the sum of r independent \operatorname{Geometric}(p) variables, and q = 1 - p. Prove that P(T_r = n) = \binom{n - 1}{r - 1} p^r q^{n - r} for n = r, r + 1, \dots.
- Let X and Y be independent with densities, and f_X(x) \le M for all x. Prove that (f_X * f_Y)(s) \le M for all s: adding an independent term never raises the peak of a density.
- Let X \sim \operatorname{Poisson}(\lambda) and Y \sim \operatorname{Poisson}(\mu) be independent. Prove that given X + Y = n, \displaystyle X \sim \operatorname{Binomial}(n, \frac{\lambda}{\lambda + \mu}).
- Let X \sim N(\mu_X, \sigma_X^2) and Y \sim N(\mu_Y, \sigma_Y^2) be independent, and a, b nonzero constants. Prove that aX + bY \sim N(a\mu_X + b\mu_Y, \ a^2\sigma_X^2 + b^2\sigma_Y^2).
- Give X and Y, each N(0, 1), for which X + Y is not N(0, 2).
Answers
- p_{X+Y} is \tfrac16, \tfrac{5}{18}, \tfrac13, \tfrac16, \tfrac{1}{18} at 1, 2, 3, 4, 5, and P(X + Y \ge 4) = \tfrac29. By Result 1, p(s) = \tfrac13 \sum_k p_Y(s - k) over k = 1, 2, 3; for s = 3 that is \tfrac13 (\tfrac16 + \tfrac13 + \tfrac12).
- T \sim \operatorname{Binomial}(20, 0.7) by Result 3(a), and P(T \ge 16) = \sum_{k=16}^{20} \binom{20}{k} 0.7^k \, 0.3^{20 - k} \approx 0.2375.
- e^{-3} \approx 0.0498, \tfrac{17}{2} e^{-3} \approx 0.4232 and \tfrac{125}{324} \approx 0.3858. The total is \operatorname{Poisson}(3) by Result 3(b), so P(\le 2) = e^{-3}(1 + 3 + \tfrac92). By Exercise 16 the fire calls given 4 in all are \operatorname{Binomial}(4, \tfrac16), so the last is 4 \cdot \tfrac16 (\tfrac56)^3.
- P(T = n) = (n - 1) \tfrac19 (\tfrac23)^{n - 2} for n \ge 2, P(T = 4) = \tfrac{4}{27} and E(T) = 6. T is the sum of two independent \operatorname{Geometric}(\tfrac13) variables; in Result 1 each of the n - 1 terms is (\tfrac23)^{k-1} \tfrac13 (\tfrac23)^{n-k-1} \tfrac13. By linearity E(T) = 3 + 3.
- f_S(s) = \tfrac12 (1 - e^{-s}) for 0 < s < 2, f_S(s) = \tfrac12 (e^2 - 1) e^{-s} for s \ge 2, and 0 for s \le 0; P(S > 3) = \tfrac12 (e^2 - 1)e^{-3} \approx 0.1590. The integrand \tfrac12 e^{-(s - x)} is nonzero for 0 < x < \min(s, 2).
- For s > 0, \int_0^s \lambda e^{-\lambda x} \mu e^{-\mu(s - x)} \, dx = \lambda\mu e^{-\mu s} \int_0^s e^{(\mu - \lambda)x} \, dx, which is the stated density. With \lambda = 1, \mu = 3 it is \tfrac32 (e^{-s} - e^{-3s}), and P(S \le 1) = 1 - \tfrac32 e^{-1} + \tfrac12 e^{-3} \approx 0.4731.
- T \sim \operatorname{Gamma}(4, \tfrac12) with \displaystyle f_T(t) = \frac{t^3}{96} e^{-t/2} for t > 0; mean 8 minutes, variance 16, standard deviation 4 minutes. With \lambda t = 5, P(T > 10) is e^{-5}(1 + 5 + \tfrac{25}{2} + \tfrac{125}{6}), that is \tfrac{118}{3} e^{-5} \approx 0.2650.
- f_S(s) = s^2 for 0 < s \le 1, f_S(s) = 2s - s^2 for 1 < s < 2, 0 elsewhere; P(S \le 1) = \tfrac13 and E(S) = \tfrac76. For s \le 1, \int_0^s 2(s - x) \, dx = s^2; for s > 1, x runs over (s - 1, 1) and the integral is 1 - (s - 1)^2. E(S) = \tfrac12 + \tfrac23.
- P(Y < X) = \Phi(-0.6) \approx 0.2743 and P(X + Y > 30) = 1 - \Phi(1) \approx 0.1587. -X has density f_X(-t), which is N(-11, 9), so Y - X \sim N(3, 25) by Result 3(d), and P(Y - X < 0) = \Phi(\tfrac{-3}{5}). Also X + Y \sim N(25, 25).
- N(200, 180) by Result 3(d), and the probability is \displaystyle 1 - \Phi(20/\sqrt{180}) \approx 0.0680, as \displaystyle 20/\sqrt{180} \approx 1.49.
- P(S > 3) = \tfrac34 e^{-3} + \tfrac14 e^{-1} \approx 0.1293, since by the The law of total probability and independence it is \tfrac34 P(Y > 3) + \tfrac14 P(Y > 1). By Exercise 13, f_S(s) = \tfrac34 e^{-s} for 0 < s < 2, f_S(s) = \tfrac34 e^{-s} + \tfrac14 e^{-(s - 2)} for s \ge 2, and 0 for s \le 0.
- P(-Y \le t) = \int_{-t}^{\infty} f_Y(y) \, dy = \int_{-\infty}^{t} f_Y(-u) \, du, so -Y has density f_Y(-u); it is independent of X (Result 5 of Independence of events and variables). Result 2 gives X - Y the density \int f_X(x) f_Y(x - s) \, dx, and x = s + y turns it into the stated form.
- Let G = F_Y. By the The law of total probability and independence, P(X + Y \le s) = \sum_j p_j G(s - x_j), and substituting y = u - x_j, G(s - x_j) = \int_{-\infty}^{s} g(u - x_j) \, du. So F_{X+Y}(s) = \int_{-\infty}^{s} h with h(u) = \sum_j p_j g(u - x_j), a finite sum of shifts of g, hence continuous except at finitely many points of each bounded interval. The end of the proof of Result 2 makes h a density.
- Induction on r; r = 1 is the geometric pmf. T_r is independent of the next term (Result 5 of Independence of events and variables), so by Result 1 P(T_{r+1} = n) = \sum_{k=r}^{n-1} \binom{k - 1}{r - 1} p^r q^{k - r} \cdot p q^{n - k - 1}, which is p^{r+1} q^{n - r - 1} \sum_{k=r}^{n-1} \binom{k-1}{r-1}. The sum counts the r-element subsets of \{1, \dots, n - 1\} by their largest element k, so it is \binom{n - 1}{r}, as claimed.
- (f_X * f_Y)(s) is \int f_X(s - y) f_Y(y) \, dy, at most M \int f_Y(y) \, dy = M, as f_Y has total 1.
- For 0 \le k \le n, \displaystyle P(X = k \mid X + Y = n) = \frac{P(X = k) P(Y = n - k)}{P(X + Y = n)} by independence, with the denominator from Result 3(b). The factors e^{-\lambda - \mu} cancel, leaving \displaystyle \frac{n!}{k!(n - k)!} \cdot \frac{\lambda^k \mu^{n - k}}{(\lambda + \mu)^n}, the \displaystyle \operatorname{Binomial}(n, \frac{\lambda}{\lambda + \mu}) pmf.
- For a > 0, P(aX \le t) = F_X(t/a); for a < 0 it is 1 - F_X(t/a). Either way the derivative is \displaystyle \frac{1}{|a|} f_X(t/a), which is \displaystyle \frac{1}{|a| \sigma_X \sqrt{2\pi}} e^{-(t - a\mu_X)^2/(2a^2\sigma_X^2)}, the N(a\mu_X, a^2\sigma_X^2) density. Likewise for bY, and aX, bY are independent (Result 5 of Independence of events and variables), so Result 3(d) finishes.
- X \sim N(0, 1) and Y = -X, which is N(0, 1) by Exercise 17 with a = -1. Then X + Y = 0 with probability 1, which has no density and is not N(0, 2). Result 3(d) fails because X and Y are not independent.