Densities and cumulative distribution functions
A waiting time can be any number in an interval, so it lands on each single value with probability zero. How can it still carry probabilities, and which one function holds all of them?
Where you are 5 in Random variables
Why this exists
Discrete distributions: binomial, geometric, Poisson described a random variable by its pmf, the mass p_X(x) = P(X = x) at each value. That fails for a quantity read on a continuous scale, such as the minutes until the next boat reaches a marina’s fuel dock. At most n values can each have probability \tfrac{1}{n} or more, so at most countably many values carry positive mass, and for a wait none does. The probability is spread along the line instead, thicker in some places than others.
The school lesson Density curves: area, mean and median read proportions as areas under a curve. Here that picture becomes a definition, and calculus does the work: integrals give probabilities, improper integrals handle unbounded ranges, and the The fundamental theorem of calculus ties the density to the distribution function F_X(x) = P(X \le x), the one description that works for every random variable.
Definition
A random variable X is continuous, or has a density, when there is a function f_X : \mathbb{R} \to [0, \infty) such that for all real a \le b
Such an f_X is a probability density function, or density, of X. Throughout, a density is continuous except at finitely many points of each bounded interval, and integrals are Riemann integrals, improper (Improper integrals with an infinite integrand) where an end is infinite or f_X is unbounded. The values where f_X > 0 form the support of X. The distribution function is F_X(x) = P(X \le x), as for every random variable (Result 2 of Discrete distributions: binomial, geometric, Poisson).
For 0 < q < 1, a q-quantile of X is a number x with F_X(x) = q; the \tfrac12-quantile is a median, and the \tfrac14- and \tfrac34-quantiles are the quartiles, as in Percentiles. When F_X is continuous, each q has one by the intermediate value theorem, since F_X tends to 0 on the left and 1 on the right; it is unique where F_X is strictly increasing.
Uniform. X \sim \operatorname{Uniform}(a, b), for a < b, when \displaystyle f_X(x) = \frac{1}{b - a} for a \le x \le b and 0 otherwise: subintervals of [a, b] of equal length are equally likely, the flat shape of Uniform and bimodal shapes.
Exponential. T \sim \operatorname{Exponential}(\lambda), for a rate \lambda > 0, when f_T(t) = \lambda e^{-\lambda t} for t \ge 0 and 0 for t < 0. It models the wait for an event that comes at a steady rate \lambda per unit of time, with no memory of how long it has been awaited (Result 6, Example 6).
Intuition
Read f_X as probability per unit length. For small h > 0, at a point x where f_X is continuous,
which Result 3 makes exact as a limit. So f_X(x) says how thickly probability lies near x; it is not a probability, and in Example 1 it reaches \tfrac{16}{9}. Only areas are probabilities.
F_X(x) is the area under f_X to the left of x. Figure 1 shows both for Example 1: where the density is high, F_X climbs steeply; where it is zero, F_X is flat. In Discrete distributions: binomial, geometric, Poisson the distribution function was a staircase that climbed only by jumps. With a density it climbs without a single jump, and a variable can also mix the two (Example 5).
Results
Result 1 (single values). If X has a density, then P(X = x) = 0 for every real x. Hence for a < b the probabilities P(a < X < b), P(a < X \le b), P(a \le X < b) and P(a \le X \le b) all equal \int_a^b f_X(x) \, dx.
Result 2 (density and distribution function). If X has density f_X, then for every real x
Result 3 (the slope of F_X). If X has density f_X, then F_X is continuous everywhere, and F_X'(x) = f_X(x) at every x where f_X is continuous.
Result 4 (from F_X back to f_X). Let the distribution function F_X of X be continuous, and differentiable with a continuous derivative except at finitely many points. Then X has the density f_X = F_X', given any value (say 0) at the exceptional points.
Result 5 (every density occurs). Every f \ge 0 of the kind above with \int_{-\infty}^{\infty} f = 1 is the density of some random variable. This is stated without proof: building one needs Lebesgue measure, the subject of measure-theoretic probability.
Result 6 (uniform and exponential). If X \sim \operatorname{Uniform}(a, b), then F_X(x) = 0 for x < a, \displaystyle F_X(x) = \frac{x - a}{b - a} for a \le x \le b, and F_X(x) = 1 for x > b. If T \sim \operatorname{Exponential}(\lambda), then F_T(t) = 1 - e^{-\lambda t} for t \ge 0, so P(T > t) = e^{-\lambda t}, and for all s, t \ge 0
Exercise 16 proves the converse, as Example 6 of Discrete distributions: binomial, geometric, Poisson did for the geometric: a positive waiting time with this property is exponential.
Why it is true
Result 1. The definition allows a = b = x, and gives P(X = x) = \int_x^x f_X(t) \, dt = 0. For a < b, \{a \le X \le b\} is the disjoint union of \{a < X < b\}, \{X = a\} and \{X = b\}, so by finite additivity removing either end changes no probability.
Result 2. The events \{-n \le X \le x\} grow with n, and their union is \{X \le x\}, because each X(\omega) is a real number and exceeds -n once n is large. Continuity (Result 8 of Probability axioms and sample spaces) gives
Since f_X \ge 0, \int_{-c}^{x} f_X is nondecreasing in c, so its limit along whole numbers is its limit as c \to \infty, the improper integral \int_{-\infty}^{x} f_X. The events \{-n \le X \le n\} grow to \Omega, and the same argument gives total 1. Finally P(X > x) = 1 - F_X(x), and splitting the total integral at x leaves \int_x^\infty f_X.
Result 3. By Result 2 of Discrete distributions: binomial, geometric, Poisson, F_X is continuous from the right and its jump at x is F_X(x) - F_X(x^-) = P(X = x), which is 0 by Result 1. So F_X is continuous. Now let f_X be continuous at x. For small h \ne 0, Result 2 gives F_X(x + h) - F_X(x) = \int_x^{x + h} f_X(t) \, dt, and \displaystyle f_X(x) = \frac{1}{h} \int_x^{x+h} f_X(x) \, dt, so
which tends to 0 as h \to 0 because f_X is continuous at x. This is the first part of the fundamental theorem of calculus, proved in The fundamental theorem of calculus, proved.
Result 4. F_X is nondecreasing, so F_X' \ge 0 wherever it exists. Take c < d with no exceptional point in (c, d). For c < c' < d' < d, F_X' is continuous on [c', d'], and the second part of the fundamental theorem gives F_X(d') - F_X(c') = \int_{c'}^{d'} F_X'(x) \, dx. As c' \to c^+ and d' \to d^- the left side tends to F_X(d) - F_X(c), because F_X is continuous, so the right side converges to the (possibly improper) integral \int_c^d F_X' and equals it. The exceptional points cut any [a, b] into finitely many such pieces, and adding gives F_X(b) - F_X(a) = \int_a^b F_X'. Since F_X has no jump at a, P(X = a) = 0, so P(a \le X \le b) = P(a < X \le b) = F_X(b) - F_X(a).
Result 6. For a \le x \le b, \displaystyle F_X(x) = \int_a^x \frac{dt}{b - a} = \frac{x - a}{b - a}, which is 1 at x = b, so the density has total 1. For the exponential and t \ge 0,
which tends to 1 as t \to \infty. Since \{T > s + t\} \subseteq \{T > s\}, the Finding a conditional probability gives P(T > s + t \mid T > s) = e^{-\lambda (s + t)} / e^{-\lambda s} = e^{-\lambda t}.
Worked examples
How empty the tank is
When a boat pulls in to Layla’s fuel dock, the fraction X of its tank that is empty has density f_X(x) = c x^2 (1 - x) for 0 \le x \le 1, and 0 otherwise. Find c and F_X, then P(X \le \tfrac12) and P(X > \tfrac34). Where is the density largest?
SolutionBy Result 2 the density has total 1:
so c = 12. For 0 \le x \le 1, F_X(x) = \int_0^x 12 (t^2 - t^3) \, dt = 4x^3 - 3x^4, with F_X = 0 to the left of 0 and 1 to the right of 1. Then
Next, f_X'(x) = 12x(2 - 3x) is positive on (0, \tfrac23) and negative on (\tfrac23, 1), so the largest value is f_X(\tfrac23) = 12 \cdot \tfrac49 \cdot \tfrac13 = \tfrac{16}{9}. A density above 1 is no contradiction: an interval of length 0.01 centered at \tfrac23 has probability about 0.018 (Figure 1).
A barge due before 9:40
The fuel barge that refills Layla’s tanks arrives at a time uniformly distributed between 9:00 and 9:40. Let U be its arrival in minutes after 9:00. Find P(U \le 15). At 9:20 it has not come: find the probability that it arrives by 9:30, and the distribution function of U given U > 20.
SolutionU \sim \operatorname{Uniform}(0, 40), so by Result 6, \displaystyle F_U(x) = \frac{x}{40} for 0 \le x \le 40, and \displaystyle P(U \le 15) = \frac{15}{40} = \frac38. Since \{20 < U \le 30\} \subseteq \{U > 20\}, the Finding a conditional probability gives
In the same way \displaystyle P(U \le x \mid U > 20) = \frac{x - 20}{20} for 20 \le x \le 40, the distribution function of \operatorname{Uniform}(20, 40). This wait remembers: the chance of arrival within the next 10 minutes was P(U \le 10) = \tfrac14 at 9:00 and is \tfrac12 at 9:20.
Waiting for the next boat
Diego models the time T, in minutes from now, until the next boat reaches the fuel dock as \operatorname{Exponential}(0.05). Find P(T > 30), P(10 < T \le 20) and the median of T. If the dock then stays quiet for 15 minutes, find the probability that it stays quiet for more than 10 minutes longer.
SolutionBy Result 6, P(T > t) = e^{-0.05 t} for t \ge 0. So P(T > 30) = e^{-1.5} \approx 0.2231, and
the shaded area in Figure 2. The median m solves e^{-0.05 m} = \tfrac12, so m = 20 \ln 2 \approx 13.86 minutes.
Quiet for 15 minutes is T > 15, and more than 10 minutes longer is T > 25. Result 6 gives
The quiet quarter hour changes nothing, unlike the barge of Example 2.
How long a guest boat stays
Diego models the length X, in hours, of a visit to the guest dock by F_X(x) = 0 for x < 1 and \displaystyle F_X(x) = 1 - \frac{1}{x^2} for x \ge 1. Show that Result 4 applies, find a density, check its total, and find P(X > 3), P(2 < X \le 4) and the median.
SolutionBoth formulas give 0 at x = 1, so F_X is continuous. Away from 1 its derivative is 0 for x < 1 and \displaystyle \frac{2}{x^3} for x > 1, both continuous; at x = 1 the slopes from the left and right are 0 and 2, a corner, which Result 4 allows. So X has density \displaystyle f_X(x) = \frac{2}{x^3} for x > 1 and 0 for x \le 1. As a check,
Then \displaystyle P(X > 3) = 1 - F_X(3) = \frac19, and \displaystyle P(2 < X \le 4) = F_X(4) - F_X(2) = \frac14 - \frac{1}{16} = \frac{3}{16}. The median solves \displaystyle \frac{1}{m^2} = \frac12, so \displaystyle m = \sqrt{2} \approx 1.414 hours, about 1 hour 25 minutes.
A dock that is sometimes free
When a boat reaches Layla’s fuel dock, the dock is free with probability 0.4 and the boat waits 0 minutes; otherwise it waits. The wait W, in minutes, has F_W(x) = 0 for x < 0 and F_W(x) = 1 - 0.6 e^{-x/5} for x \ge 0. Show that W has neither a density nor a pmf. Find P(0 < W \le 5), P(W > 10) and the median of W.
SolutionAt 0 the jump is F_W(0) - F_W(0^-) = 0.4, so P(W = 0) = 0.4 by Result 2 of Discrete distributions: binomial, geometric, Poisson, which Result 1 forbids for a variable with a density. F_W is continuous at every x \ne 0, so P(W = x) = 0 there, and by countable additivity every countable set S has P(W \in S) \le 0.4; so W is not discrete either. It is mixed: mass 0.4 at 0, and the other 0.6 spread over (0, \infty) with density 0.12 e^{-x/5}, whose integral over (0, \infty) is 0.12 \cdot 5 = 0.6 (Figure 3).
The distribution function still answers every question:
Since F_W(0) = 0.4 < \tfrac12, the median m is positive and solves 0.6 e^{-m/5} = \tfrac12, so m = 5 \ln 1.2 \approx 0.9116 minutes.
The exponential as a limit of geometric waits
Fix \lambda > 0. Split each hour into n > \lambda equal slots; in each slot a boat enters the harbor with probability \displaystyle \frac{\lambda}{n}, independently of the other slots. Let \displaystyle N_n \sim \operatorname{Geometric}(\frac{\lambda}{n}) be the number of slots up to and including the first with a boat, and \displaystyle T_n = \frac{N_n}{n} the time in hours to the end of that slot. Prove that F_{T_n}(t) \to 1 - e^{-\lambda t} for every t \ge 0, and F_{T_n}(t) = 0 for t < 0.
SolutionSince \displaystyle T_n \ge \frac1n > 0, F_{T_n}(t) = 0 for t < 0. Let t \ge 0. N_n is a whole number, so \{T_n > t\} = \{N_n > \lfloor nt \rfloor\}, and Result 4 of Discrete distributions: binomial, geometric, Poisson gives \displaystyle P(T_n > t) = (1 - \frac{\lambda}{n})^{\lfloor nt \rfloor}. Because nt - 1 < \lfloor nt \rfloor \le nt and \displaystyle 0 < 1 - \frac{\lambda}{n} < 1, a larger exponent gives a smaller power:
As in Result 5 of Discrete distributions: binomial, geometric, Poisson, \displaystyle n \ln(1 - \frac{\lambda}{n}) \to -\lambda, so \displaystyle (1 - \frac{\lambda}{n})^{nt} = e^{t \cdot n \ln(1 - \lambda/n)} \to e^{-\lambda t}. The right side is this times \displaystyle (1 - \frac{\lambda}{n})^{-1}, which tends to 1. By the squeeze theorem P(T_n > t) \to e^{-\lambda t}, so F_{T_n}(t) \to 1 - e^{-\lambda t}, the \operatorname{Exponential}(\lambda) distribution function. Each T_n is discrete, yet the limit has a density.
Boundaries
Not every random variable has a density. Discrete variables have a pmf instead, mixed ones like Example 5 have neither, and some variables have a continuous distribution function yet no density (one is built from the Cantor function; stated without proof, it is shown in measure-theoretic probability). The distribution function exists for every random variable and always works.
A density’s value at a point is not a probability. Changing f_X at finitely many points changes no integral, so X has many densities, and f_X(x) may exceed 1.
Probability zero is not impossibility. Every value of a wait with a density has probability 0, yet the wait takes some value. Countable additivity adds only countably many events, and the line holds uncountably many values (Countable and uncountable sets).
There is no uniform distribution on an unbounded interval. A density equal to a constant c on [0, \infty) has total 0 or \infty, never 1.
Lack of memory is a modeling claim. The exponential suits waits for events at a steady rate unaffected by the past; a barge on a schedule (Example 2) or a part that wears out is not exponential.
Where this goes next
Expectation, linearity and indicators defines the mean of a variable with a density by an integral against f_X, in place of the weighted sum of Discrete distributions: binomial, geometric, Poisson. Functions of a random variable uses the distribution function to find the density of a function of X. Joint distributions gives two variables one density f_{X,Y}, with probabilities as double integrals over regions of the plane, and shows that it splits into a product exactly when the variables are independent in the sense of Independence of events and variables. The The normal distribution and its standard form have a distribution function with no elementary formula, so its values come from tables and software.
Exercises
Densities
- X has density \displaystyle f_X(x) = c \sqrt{x} for 0 \le x \le 4, and 0 otherwise. Find c, F_X, P(X \le 1) and the median of X.
- The error X, in centimeters, of a reading from the marina’s tide gauge has density f_X(x) = c e^{-|x|} for all real x. Find c, P(|X| > 2) and P(-1 < X \le 3).
- Find c so that \displaystyle f(x) = \frac{c}{x^2} for x \ge 3, and 0 otherwise, is a density, and find P(X > 6) for a variable with this density. Then show that no c makes \displaystyle g(x) = \frac{c}{x} for x \ge 1, and 0 otherwise, a density.
- X has density \displaystyle f_X(x) = \frac{c}{1 + x^2} for all real x. Find c, F_X, P(X > 1) and P(|X| \le 1).
- Which of these are densities? (a) f(x) = 5 for 0 \le x \le 0.2, else 0. (b) f(x) = x - \tfrac12 for 0 \le x \le 2, else 0. (c) f(x) = \tfrac{3}{26} x^2 for 1 \le x \le 3, else 0. (d) f(x) = e^x for x \le 0, else 0.
Distribution functions
- X has F_X(x) = 0 for x < 0 and F_X(x) = 1 - (1 + x) e^{-x} for x \ge 0. Find a density, P(X > 2) and P(1 < X \le 2).
- X has F_X(x) = 0 for x < 0, F_X(x) = 0.2 + 0.3x for 0 \le x < 2, and F_X(x) = 1 for x \ge 2. Find P(X = 0), P(X = 2), P(0 < X < 2) and P(X \le 1). Does X have a density?
- X has density f_X(x) = 2x e^{-x^2} for x \ge 0, and 0 otherwise. Find F_X and the three quartiles of X.
Uniform and exponential
- Diego drops a fender off a 60-meter pier at a point uniformly distributed along it. Find the probability that it lands within 12 meters of an end. Given that it lands beyond the 30-meter mark, find the probability that it lands beyond the 50-meter mark.
- X \sim \operatorname{Uniform}(a, b) with F_X(5) = 0.25 and F_X(11) = 0.75. Find a, b and P(X > 12).
- The time T, in hours, until the next gust above 25 knots at the harbor mouth is \operatorname{Exponential}(0.2). Find P(T \le 2), P(T > 5), and the time t with P(T > t) = 0.1.
- T is exponential and P(T > 4) = 0.36. Find \lambda, P(T > 8) and P(T > 2).
- Let T \sim \operatorname{Exponential}(\lambda). Show that P(s < T \le s + t \mid T > s) = P(T \le t) for s, t \ge 0, and evaluate it for \lambda = 0.25, s = 3, t = 4.
Proofs
- Let X have a density and let C be a countable set of reals. Prove that P(X \in C) = 0. Deduce that X is a rational number with probability 0.
- Let X have density f_X, continuous at x_0. Prove that \displaystyle \frac{1}{2h} P(x_0 - h \le X \le x_0 + h) \to f_X(x_0) as h \to 0^+.
- Let T satisfy P(T > 0) = 1 and P(T > s + t) = P(T > s) \, P(T > t) for all s, t \ge 0. Prove that T \sim \operatorname{Exponential}(\lambda) for some \lambda > 0.
- Let T \sim \operatorname{Exponential}(\lambda) and N = \lceil T \rceil, the number of the unit time period in which the event comes. Prove that N \sim \operatorname{Geometric}(1 - e^{-\lambda}), and find its parameter when \lambda = 0.5.
Answers
- c = \tfrac{3}{16}; F_X(x) = \tfrac18 x^{3/2} for 0 \le x \le 4 (0 before, 1 after); P(X \le 1) = \tfrac18; median 4^{2/3} \approx 2.520. The total is \tfrac{2c}{3} \cdot 4^{3/2} = \tfrac{16c}{3}, and the median solves x^{3/2} = 4.
- c = \tfrac12; P(|X| > 2) = e^{-2} \approx 0.1353; P(-1 < X \le 3) = 1 - \tfrac12 (e^{-1} + e^{-3}) \approx 0.7912. By symmetry the total is 2c \int_0^\infty e^{-x} \, dx = 2c. Then F_X(x) = \tfrac12 e^x for x < 0 and 1 - \tfrac12 e^{-x} for x \ge 0, and the last answer is F_X(3) - F_X(-1).
- c = 3 and P(X > 6) = \tfrac12: \int_3^\infty c x^{-2} \, dx = \tfrac{c}{3} and \int_6^\infty 3 x^{-2} \, dx = \tfrac36. For g: c < 0 makes g negative, c = 0 gives total 0, and for c > 0, \displaystyle \int_1^b \frac{c}{x} \, dx = c \ln b \to \infty.
- c = \tfrac{1}{\pi}; F_X(x) = \tfrac12 + \tfrac{1}{\pi} \arctan x; P(X > 1) = \tfrac14; P(|X| \le 1) = \tfrac12. \displaystyle \int_{-b}^{b} \frac{dx}{1 + x^2} = 2 \arctan b \to \pi; F_X(1) = \tfrac34 and F_X(-1) = \tfrac14.
- (a) Yes: total 5 \cdot 0.2 = 1; values above 1 are allowed. (b) No: its total is \left[ \tfrac{x^2}{2} - \tfrac{x}{2} \right]_0^2 = 1, but it is negative on [0, \tfrac12). (c) Yes: total \tfrac{3}{26} \cdot \tfrac{27 - 1}{3} = 1. (d) Yes: total \int_{-\infty}^0 e^x \, dx = 1.
- f_X(x) = x e^{-x} for x > 0, 0 for x < 0; P(X > 2) = 3e^{-2} \approx 0.4060; P(1 < X \le 2) = 2e^{-1} - 3e^{-2} \approx 0.3298. F_X is continuous (0 at x = 0), and for x > 0, F_X'(x) = -e^{-x} + (1 + x) e^{-x}, so Result 4 applies.
- P(X = 0) = 0.2, P(X = 2) = 1 - 0.8 = 0.2, P(0 < X < 2) = F_X(2^-) - F_X(0), which is 0.8 - 0.2 = 0.6, and P(X \le 1) = 0.5. No density: by Result 1 a variable with one has no jumps, and this F_X jumps at 0 and at 2.
- F_X(x) = 1 - e^{-x^2} for x \ge 0, from \int_0^x 2t e^{-t^2} \, dt. A q-quantile solves e^{-x^2} = 1 - q, so \displaystyle x = \sqrt{-\ln(1 - q)}: \displaystyle \sqrt{\ln \tfrac43} \approx 0.5364, \displaystyle \sqrt{\ln 2} \approx 0.8326, \displaystyle \sqrt{\ln 4} \approx 1.1774.
- \tfrac25 and \tfrac13. The position is \operatorname{Uniform}(0, 60); the two end pieces have total length 24, giving \tfrac{24}{60}; and P(X > 50 \mid X > 30) = \tfrac{10/60}{30/60}.
- a = 2, b = 14, P(X > 12) = \tfrac16. From \displaystyle \frac{5 - a}{b - a} = \tfrac14 and \displaystyle \frac{11 - a}{b - a} = \tfrac34, subtracting gives \displaystyle \frac{6}{b - a} = \tfrac12, so b - a = 12 and 5 - a = 3. Then P(X > 12) = \tfrac{2}{12}.
- P(T \le 2) = 1 - e^{-0.4} \approx 0.3297; P(T > 5) = e^{-1} \approx 0.3679; t = 5 \ln 10 \approx 11.51 hours, from e^{-0.2 t} = 0.1.
- \lambda = \tfrac12 \ln \tfrac53 \approx 0.2554, from e^{-4\lambda} = 0.36, that is, 4\lambda = \ln \tfrac{25}{9}. Then P(T > 8) = (e^{-4\lambda})^2 = 0.1296 and P(T > 2) = (e^{-4\lambda})^{1/2} = 0.6.
- \{s < T \le s + t\} is \{T > s\} with \{T > s + t\} removed, so the conditional probability is 1 - P(T > s + t \mid T > s) = 1 - P(T > t) = P(T \le t) by Result 6. Here 1 - e^{-1} \approx 0.6321.
- List C as c_1, c_2, \dots (finitely many if C is finite). The events \{X = c_i\} are disjoint with union \{X \in C\}, so countable additivity and Result 1 give P(X \in C) = \sum_i P(X = c_i) = 0. The rationals form a countable set (Countable sets), so P(X \in \mathbb{Q}) = 0.
- By Results 1 and 2, P(x_0 - h \le X \le x_0 + h) = \int_{x_0 - h}^{x_0 + h} f_X(t) \, dt. Subtracting 2h f_X(x_0) = \int_{x_0 - h}^{x_0 + h} f_X(x_0) \, dt and dividing by 2h, the difference between \displaystyle \frac{1}{2h} P(x_0 - h \le X \le x_0 + h) and f_X(x_0) is at most \sup_{|t - x_0| \le h} |f_X(t) - f_X(x_0)| in size, which tends to 0 by continuity at x_0.
- Let G(t) = P(T > t) = 1 - F_T(t), nonincreasing and continuous from the right (Result 2 of Discrete distributions: binomial, geometric, Poisson), with G(0) = 1 and G(t) \to 0 as t \to \infty. Induction gives G(\tfrac{m}{n}) = G(\tfrac1n)^m for whole numbers m \ge 0, n \ge 1, and so G(1) = G(\tfrac1n)^n. If G(1) = 0, then G(\tfrac1n) = 0 for all n, against G(\tfrac1n) \to G(0) = 1. If G(1) = 1, then G(k) = 1 for all whole k, against G(k) \to 0. So 0 < G(1) < 1; put \lambda = -\ln G(1) > 0. Then G(\tfrac1n) = G(1)^{1/n} = e^{-\lambda / n}, so G(q) = e^{-\lambda q} for every rational q \ge 0. For real t \ge 0, take rationals q_k decreasing to t; continuity from the right gives G(t) = e^{-\lambda t}. Since P(T \le 0) = 0, F_T(t) = 0 for t < 0. So F_T is the exponential distribution function, and Result 4 gives the density \lambda e^{-\lambda t} for t > 0.
- P(T \le 0) = 0, so N \ge 1. For k = 1, 2, \dots, \{N = k\} = \{k - 1 < T \le k\}, so by Result 6, P(N = k) = e^{-\lambda (k - 1)} - e^{-\lambda k} = (e^{-\lambda})^{k - 1} (1 - e^{-\lambda}), the \operatorname{Geometric}(1 - e^{-\lambda}) pmf. For \lambda = 0.5 the parameter is 1 - e^{-0.5} \approx 0.3935.