Joint distributions

Two numbers read off the same catch, such as how much of a crate is cod and how much is haddock, are tied together. What single object holds every probability about the pair, and how do the distribution of each one alone, and of one given the other, come out of it?

Where you are 5 in Several random variables

Why this exists

A crate at Sofia’s fishery is part cod, part haddock and part other fish. The cod fraction X and the haddock fraction Y each have a density (Densities and cumulative distribution functions), but those two densities cannot give P(Y > 2X), or say how the haddock fraction is distributed once the cod has been weighed. That needs the distribution of the pair.

For discrete variables it is the joint table of Joint and marginal probability. Here a pair of continuous variables gets one function on the plane whose Double integrals as volume over regions are probabilities.

Definition

The joint distribution function of random variables X and Y is F_{X,Y}(x, y) = P(X \le x, \ Y \le y), the comma meaning “and”. X and Y are jointly continuous, with joint density f_{X,Y} : \mathbb{R}^2 \to [0, \infty), when for all real x and y

F_{X,Y}(x, y) = \int_{-\infty}^{x} \int_{-\infty}^{y} f_{X,Y}(s, t) \, dt \, ds.

A joint density is assumed continuous except on finitely many segments and graphs of continuous functions in each bounded region. For a nonnegative function an improper double integral equals either iterated integral, finite or infinite; this form of Fubini’s theorem (Iterated integrals and Fubini's theorem) is used without proof.

The densities f_X of X alone and f_Y of Y alone are then the marginal densities. For y with f_Y(y) > 0, the conditional density of X given Y = y is

f_{X \mid Y}(x \mid y) = \frac{f_{X,Y}(x, y)}{f_Y(y)},

and P(X \in A \mid Y = y) means \int_A f_{X \mid Y}(x \mid y) \, dx; exchanging letters defines f_{Y \mid X}. For n variables the definitions use n-fold integrals, and the results carry over.

Intuition

Picture f_{X,Y} as a surface over the plane enclosing volume 1; the probability of a region is the volume above it. Near a point where f_{X,Y} is continuous,

\begin{aligned}&P(x \le X \le x + h, \ y \le Y \le y + k) \\ &\quad \approx f_{X,Y}(x, y) \, hk,\end{aligned}

so f_{X,Y} is probability per unit area. The marginal f_X(x) totals the surface along the vertical line at x. A conditional density keeps one slice, rescaled to area 1, and independence says every slice has the same shape.

0.5 1 0.5 1 0
x + y = 1
y = 2x
\tfrac13
x
y
Figure 1The support of the joint density of Example 1. The shaded region is the event \{Y > 2X\}, and the volume under 3(x + y) above it is P(Y > 2X) = \tfrac13. A vertical slice at x runs only up to y = 1 - x, so the slices change length and X and Y are dependent.

Results

Result 1 (rectangles and margins). For any X and Y, with F = F_{X,Y}, a < b and c < d,

\begin{aligned}&P(a < X \le b, \ c < Y \le d) \\ &\quad = F(b, d) - F(a, d) - F(b, c) + F(a, c),\end{aligned}

and F_X(x) = \lim_{y \to \infty} F(x, y), F_Y(y) = \lim_{x \to \infty} F(x, y).

Result 2 (regions). If (X, Y) has joint density f, then \iint_{\mathbb{R}^2} f \, dA = 1, and P((X, Y) \in D) = \iint_D f \, dA for every rectangle D, with or without its edges, and every region D bounded by finitely many segments and graphs of continuous functions. The rectangle case is proved below; the general region is stated without proof; measure theory proves it. So a bounded segment or graph, of zero area, has probability 0.

Result 3 (marginals). If (X, Y) has joint density f, then

\begin{aligned}f_X(x) &= \int_{-\infty}^{\infty} f(x, y) \, dy, \\ f_Y(y) &= \int_{-\infty}^{\infty} f(x, y) \, dx.\end{aligned}

Result 4 (independence). (a) If X and Y are independent with densities f_X and f_Y, then f_X(x) f_Y(y) is a joint density of (X, Y). (b) If (X, Y) has joint density g(x) h(y) for all real x, y, with g, h \ge 0, then X and Y are independent, with densities that are constant multiples of g and h. Such a density is positive exactly on the product set \{g > 0\} \times \{h > 0\}, so a density positive only on a triangle never factors.

Result 5 (conditional densities). Let (X, Y) have joint density f.

(a) For f_Y(y) > 0, f_{X \mid Y}(\cdot \mid y) is nonnegative with total 1.

(b) If f_Y is continuous at y with f_Y(y) > 0, and G(t) = \int_{-\infty}^{x} f(s, t) \, ds is continuous at t = y, then as h \to 0^+

\begin{aligned}&P(X \le x \mid y \le Y \le y + h) \\ &\quad \to \int_{-\infty}^{x} f_{X \mid Y}(s \mid y) \, ds.\end{aligned}

(c) f(x, y) = f_{X \mid Y}(x \mid y) \, f_Y(y) where f_Y(y) > 0, and, reading the integrand as 0 where f_Y(y) = 0,

f_X(x) = \int_{-\infty}^{\infty} f_{X \mid Y}(x \mid y) \, f_Y(y) \, dy.

(d) Mixed pairs. Let N be discrete, let \Lambda have density f_\Lambda, and let a model give for each \lambda a pmf p_{N \mid \Lambda}(n \mid \lambda), meaning that

P(N = n, \ \Lambda \le t) = \int_{-\infty}^{t} p_{N \mid \Lambda}(n \mid \lambda) f_\Lambda(\lambda) \, d\lambda.

Then P(N = n) is this integral over the whole line, and if P(N = n) > 0, given N = n the variable \Lambda has density

f_{\Lambda \mid N}(\lambda \mid n) = \frac{p_{N \mid \Lambda}(n \mid \lambda) \, f_\Lambda(\lambda)}{P(N = n)}.

Result 6 (means). Let (X, Y) have joint density f, and let g be continuous except on finitely many segments and graphs in each bounded region. If \iint |g| f \, dA converges, then

E(g(X, Y)) = \iint_{\mathbb{R}^2} g(x, y) \, f(x, y) \, dA.

This two-variable form of Result 1 of Expectation, linearity and indicators is stated without proof; measure-theoretic probability proves it in general. From it, if X and Y have finite means, E(X + Y) = E(X) + E(Y), and if they are also independent with densities, E(XY) = E(X) E(Y).

Result 7 (change of variables). Let (X, Y) have joint density f, zero outside an open set D. Let T map D one-to-one onto an open set E, with inverse (u, v) \mapsto (x(u, v), y(u, v)) having continuous partial derivatives and Jacobian

J = \frac{\partial x}{\partial u} \frac{\partial y}{\partial v} - \frac{\partial x}{\partial v} \frac{\partial y}{\partial u} \ne 0

on E. Then (U, V) = T(X, Y) has joint density f(x(u, v), y(u, v)) \, |J| on E, and 0 off E. As no integral sees a set of zero area, D may omit one (Example 4).

Why it is true

Result 1. The rectangle formula opens the proof of Result 3 of Independence of events and variables. The events \{X \le x, \ Y \le n\} grow with n to \{X \le x\}, so continuity (Result 8 of Probability axioms and sample spaces) gives F_X(x) = \lim_n F(x, n), the limit as y \to \infty since F is nondecreasing in y.

Results 2 and 3. Let Q(x, y) = \{s \le x, \ t \le y\}, so \iint_{Q(x, y)} f \, dA = F(x, y). The rectangle (a, b] \times (c, d] is Q(b, d) with Q(a, d) and Q(b, c) removed and their overlap Q(a, c) put back, so the integral over it is the combination in Result 1, its probability. The quadrants Q(n, n) grow to the plane, so the total is 1. For Result 3, Result 1 and Fubini give

\begin{aligned}F_X(x) &= \lim_{n \to \infty} F(x, n) \\ &= \int_{-\infty}^{x} \int_{-\infty}^{\infty} f(s, t) \, dt \, ds,\end{aligned}

so f_X is a density of X, and likewise f_Y. Then P(X = a) = P(Y = c) = 0 (Densities and cumulative distribution functions, Result 1), so the edges of a rectangle carry no probability.

Result 4. Independence means F(x, y) = F_X(x) F_Y(y) for all x, y. (a) Then F(x, y) = \int_{-\infty}^{x} f_X \int_{-\infty}^{y} f_Y, which is \int_{-\infty}^{x} \int_{-\infty}^{y} f_X(s) f_Y(t) \, dt \, ds. (b) Let G = \int g and H = \int h. By Fubini GH = 1, so both are positive and finite, and Result 3 gives f_X = gH and f_Y = Gh. So F(x, y) = \int_{-\infty}^{x} g \int_{-\infty}^{y} h, which is \displaystyle \frac{F_X(x)}{H} \cdot \frac{F_Y(y)}{G} = F_X(x) F_Y(y).

Result 5. (a) By Result 3, \int f(x, y) \, dx = f_Y(y). (b) By Fubini, P(X \le x, \ y \le Y \le y + h) = \int_y^{y + h} G(t) \, dt and P(y \le Y \le y + h) = \int_y^{y + h} f_Y(t) \, dt. Divided by h these tend to G(y) and f_Y(y), as both are continuous at y, and G(y) / f_Y(y) is the right side. (c) The product is the definition. If f_Y(t) = 0, then 0 \le \int_{-\infty}^{x} f(s, t) \, ds \le f_Y(t) = 0, so setting f to 0 on the lines y = t with f_Y(t) = 0 leaves F unchanged; Result 3 for the new density is the formula. (d) Let t \to \infty; then P(\Lambda \le t \mid N = n) is P(N = n, \ \Lambda \le t) / P(N = n) by the Finding a conditional probability.

Result 6. Since |x + y| \le |x| + |y|, and Fubini with Result 3 gives \iint |x| f \, dA = E(|X|) < \infty, the formula applies to x + y. Splitting x into positive and negative parts, Fubini gives \iint x f \, dA = E(X), likewise for y. For independent X and Y, Result 4(a) gives the density f_X(x) f_Y(y), and the same splitting gives \iint xy f_X(x) f_Y(y) \, dA = E(X) E(Y).

Result 7. For a region B \subseteq E as in Result 2, \{(U, V) \in B\} = \{(X, Y) \in T^{-1}(B)\}, so by the change of variables theorem (Change of variables and the Jacobian), used without proof,

\begin{aligned}P((U, V) \in B) &= \iint_{T^{-1}(B)} f \, dA \\ &= \iint_B f(x(u, v), y(u, v)) \, |J| \, du \, dv.\end{aligned}

As P((U, V) \notin E) = 0, taking B the part of E in a quadrant gives F_{U,V}.

Worked examples

Example 1

A crate of cod and haddock

In a crate at Sofia’s fishery, X is the fraction of the weight that is cod and Y the fraction that is haddock. Their joint density is f(x, y) = c(x + y) on the triangle x \ge 0, y \ge 0, x + y \le 1, and 0 elsewhere. Find c and the marginal densities, show that X and Y are dependent, and find P(Y > 2X).

Solution

By Result 2 the total is 1. For fixed x in [0, 1], y runs from 0 to 1 - x:

\begin{aligned}\int_0^{1 - x} (x + y) \, dy &= x(1 - x) + \tfrac12 (1 - x)^2 \\ &= \tfrac12 (1 - x^2),\end{aligned}

and \int_0^1 \tfrac12 (1 - x^2) \, dx = \tfrac13, so c = 3. By Result 3, f_X(x) = \tfrac32 (1 - x^2) for 0 \le x \le 1, and by symmetry f_Y(y) = \tfrac32 (1 - y^2) for 0 \le y \le 1.

No crate is more than half cod and more than half haddock, so P(X > \tfrac12, \ Y > \tfrac12) = 0. But P(X > \tfrac12) = \int_{1/2}^1 \tfrac32 (1 - x^2) \, dx = \tfrac{5}{16}, likewise for Y, and the product \tfrac{25}{256} is not 0. By Result 3 of Independence of events and variables, X and Y are dependent.

The event \{Y > 2X\} is shaded in Figure 1; y = 2x meets x + y = 1 at (\tfrac13, \tfrac23). For 0 \le x \le \tfrac13, y runs from 2x to 1 - x, and 3 \left[ xy + \tfrac12 y^2 \right]_{2x}^{1 - x} = \tfrac32 - \tfrac{27}{2} x^2, so

\begin{aligned}P(Y > 2X) &= \int_0^{1/3} \left( \tfrac32 - \tfrac{27}{2} x^2 \right) dx \\ &= \tfrac12 - \tfrac16 = \tfrac13.\end{aligned}
Example 2

Haddock, given the cod

For the crates of Example 1, find the conditional density of Y given X = x. Given that a crate is half cod, find the probability that more than a quarter of it is haddock, and compare with P(Y > \tfrac14). Find \operatorname{Cov}(X, Y).

Solution

For 0 \le x < 1, f_X(x) > 0, and for 0 \le y \le 1 - x

\begin{aligned}f_{Y \mid X}(y \mid x) &= \frac{3(x + y)}{\tfrac32 (1 - x^2)} \\ &= \frac{2(x + y)}{1 - x^2},\end{aligned}

and 0 for other y. At x = \tfrac12 it is \tfrac83 (\tfrac12 + y) for 0 \le y \le \tfrac12, so

\begin{aligned}P(Y > \tfrac14 \mid X = \tfrac12) &= \tfrac83 \int_{1/4}^{1/2} \left( \tfrac12 + y \right) dy \\ &= \tfrac83 \left( \tfrac18 + \tfrac{3}{32} \right) = \tfrac{7}{12}.\end{aligned}

By Result 5(b) this is the limit of P(Y > \tfrac14 \mid \tfrac12 \le X \le \tfrac12 + h). Without the cod reading, P(Y > \tfrac14) = \int_{1/4}^1 \tfrac32 (1 - y^2) \, dy, which is \tfrac{81}{128} \approx 0.633, larger than \tfrac{7}{12} \approx 0.583: half a crate of cod leaves less room for haddock.

Next, E(X) = \int_0^1 x \cdot \tfrac32 (1 - x^2) \, dx = \tfrac38, and E(Y) = \tfrac38 by symmetry. Result 6 gives

\begin{aligned}E(XY) &= \int_0^1 \int_0^{1 - x} 3xy(x + y) \, dy \, dx \\ &= \int_0^1 \left( \tfrac32 x^2 (1 - x)^2 + x (1 - x)^3 \right) dx \\ &= \tfrac{1}{20} + \tfrac{1}{20} = \tfrac{1}{10},\end{aligned}

so \operatorname{Cov}(X, Y) = \tfrac{1}{10} - \tfrac{9}{64} = -\tfrac{13}{320} \approx -0.0406 (Variance and covariance): more cod, less haddock.

Example 3

Which comes first

Let X and Y be independent with densities f_X and f_Y. Prove that P(X < Y) = \int_{-\infty}^{\infty} F_X(y) \, f_Y(y) \, dy, and that P(X < Y) = \tfrac12 when X and Y also have the same distribution.

Solution

By Result 4(a), (X, Y) has joint density f_X(x) f_Y(y). The region \{x < y\} is bounded by a line, so by Result 2 and Fubini, with x integrated first,

P(X < Y) = \int_{-\infty}^{\infty} f_Y(y) \int_{-\infty}^{y} f_X(x) \, dx \, dy,

and the inner integral is F_X(y). Exchanging letters, P(Y < X) = \int F_Y(x) f_X(x) \, dx, the same integral when X and Y share one density. The line y = x is a countable union of bounded segments, each of probability 0 by Result 2, so P(X = Y) = 0. The three events partition \Omega, so 2 P(X < Y) = 1.

Example 4

Where the net lands

Lucas throws a cast net toward a buoy at the origin, and its center lands at (X, Y), in meters east and north, with X and Y independent N(0, 4). Let R be its distance from the buoy and \Theta its direction, in (0, 2\pi) counterclockwise from east. Find the joint density of (R, \Theta), show that R and \Theta are independent, and find P(R \le 1), the median of R, and the probability that the net lands northeast of the buoy and more than 4 meters from it.

Solution

By Result 4(a) and the The normal distribution density, \displaystyle f(x, y) = \frac{1}{8\pi} e^{-(x^2 + y^2)/8}. Let D be the plane without the ray \{y = 0, \ x \ge 0\}, of zero area. The map (x, y) \mapsto (r, \theta) takes D one-to-one onto E = (0, \infty) \times (0, 2\pi), with inverse x = r \cos \theta, y = r \sin \theta and

\begin{aligned}J &= \cos \theta \cdot r \cos \theta - (-r \sin \theta) \sin \theta \\ &= r,\end{aligned}

the r \, dr \, d\theta of Double integrals in polar coordinates. By Result 7, on E,

\begin{aligned}f_{R,\Theta}(r, \theta) &= \frac{r}{8\pi} e^{-r^2/8} \\ &= \frac{r}{4} e^{-r^2/8} \cdot \frac{1}{2\pi}.\end{aligned}

With both factors 0 off E, Result 4(b) makes R and \Theta independent, with \Theta \sim \operatorname{Uniform}(0, 2\pi) and \displaystyle f_R(r) = \frac{r}{4} e^{-r^2/8} for r > 0. Integrating, P(R > r) = e^{-r^2/8}. So P(R \le 1) = 1 - e^{-1/8} \approx 0.1175; the median solves e^{-m^2/8} = \tfrac12, so \displaystyle m = \sqrt{8 \ln 2} \approx 2.355 meters; and northeast is 0 < \Theta < \tfrac{\pi}{2}, so by independence

P(0 < \Theta < \tfrac{\pi}{2}, \ R > 4) = \tfrac14 e^{-2} \approx 0.0338.
Example 5

A good day or a lucky hour

The catch rate on Lucas’s longline changes from day to day: on a given day it is \Lambda fish per hour, where \Lambda \sim \operatorname{Exponential}(\tfrac12). Given \Lambda = \lambda, the number N of fish caught in the first hour is \operatorname{Poisson}(\lambda). Find the pmf of N. Given N = 3, find the conditional density of \Lambda and P(\Lambda > 2 \mid N = 3).

Solution

This is a mixed pair as in Result 5(d), with f_\Lambda(\lambda) = \tfrac12 e^{-\lambda/2} for \lambda > 0 and the Poisson pmf of Discrete distributions: binomial, geometric, Poisson. For n = 0, 1, 2, \dots

\begin{aligned}P(N = n) &= \int_0^\infty e^{-\lambda} \frac{\lambda^n}{n!} \cdot \tfrac12 e^{-\lambda/2} \, d\lambda \\ &= \frac{1}{2 \, n!} \int_0^\infty \lambda^n e^{-3\lambda/2} \, d\lambda.\end{aligned}

Integrating by parts n times gives \displaystyle \int_0^\infty \lambda^n e^{-a\lambda} \, d\lambda = \frac{n!}{a^{n+1}} for a > 0, so P(N = n) = \tfrac12 (\tfrac23)^{n+1} = \tfrac13 (\tfrac23)^n. So N + 1 \sim \operatorname{Geometric}(\tfrac13), and P(N = 3) = \tfrac{8}{81}. Result 5(d) gives, for \lambda > 0,

\begin{aligned}f_{\Lambda \mid N}(\lambda \mid 3) &= \frac{\tfrac{1}{12} \lambda^3 e^{-3\lambda/2}}{8/81} \\ &= \tfrac{27}{32} \lambda^3 e^{-3\lambda/2}.\end{aligned}

Substituting u = \tfrac32 \lambda, and using \int_a^\infty u^3 e^{-u} \, du = e^{-a}(a^3 + 3a^2 + 6a + 6), again by parts,

\begin{aligned}P(\Lambda > 2 \mid N = 3) &= \tfrac16 \int_3^\infty u^3 e^{-u} \, du \\ &= 13 e^{-3} \approx 0.6472.\end{aligned}

Before the hour, P(\Lambda > 2) = e^{-1} \approx 0.3679. Three fish shift belief toward a good day (Figure 2), the continuous form of the updating in Conditioning and Bayes' rule, rigorously.

0 2 4 6 8 0.25 0.5 prior
\text{given } N = 3
Fish per hour
\lambda
Figure 2The density of the catch rate \Lambda before the first hour (prior, dashed) and given a catch of N = 3 fish, from Example 5. The shaded area is P(\Lambda > 2 \mid N = 3) = 13 e^{-3} \approx 0.6472; before the catch the same event had probability e^{-1} \approx 0.3679.

Boundaries

Two variables with densities need not have a joint density. The pair (X, X) lies with probability 1 on the line y = x, of zero area (Exercise 12). Results 2 to 7 need a joint density.

Uncorrelated is not independent. A point uniform on a disk has \operatorname{Cov}(X, Y) = 0, yet X and Y are dependent (Exercise 13).

Conditioning on a value depends on the approach. For independent \operatorname{Uniform}(0, 1) variables, Result 7 gives X uniform given Y - X = 0, but density 2x given Y / X = 1, though both say Y = X. A conditional density belongs to a variable, not to one event of probability 0 (the Borel paradox).

Where this goes next

The mean of f_{Y \mid X}(\cdot \mid x) is the E(Y \mid X = x) of Conditional expectation. Result 2 over the regions \{x + y \le s\} gives the distribution of a sum, the convolution of Sums of independent variables. The most used joint density, with elliptical level curves, is the The bivariate normal.

Exercises

Joint densities and regions

  1. (X, Y) has joint density f(x, y) = c(2x + y) for 0 \le x \le 1, 0 \le y \le 1, and 0 elsewhere. Find c, P(X \le \tfrac12, \ Y \le \tfrac12) and P(X + Y \le 1).
  2. (X, Y) has joint density f(x, y) = c e^{-y} for 0 < x < y, and 0 elsewhere. Find c, both marginal densities and P(Y > 2X).
  3. F_{X,Y}(x, y) = \tfrac12 xy(x + y) for 0 \le x, y \le 1; it is 0 if x < 0 or y < 0, and otherwise equals its value at (\min(x, 1), \min(y, 1)). Find P(\tfrac12 < X \le 1, \ \tfrac12 < Y \le 1) and F_X.
  4. Fish surface at a point (X, Y) uniform over a circular pen of radius 10 meters centered at the origin, with joint density \displaystyle \frac{1}{100\pi} on the disk. Find f_X, P(X > 6) and P(X^2 + Y^2 \le 25).

Marginals and independence

  1. (X, Y) has joint density f(x, y) = c x y^2 for 0 \le x \le 2, 0 \le y \le 1, and 0 elsewhere. Find c, decide whether X and Y are independent, and find the marginal densities and P(X < 2Y).
  2. Decide whether X and Y are independent when their joint density is (a) 8xy for 0 \le x \le y \le 1; (b) 4xy for 0 \le x, y \le 1; (c) x + y for 0 \le x, y \le 1; each 0 elsewhere.
  3. Sofia’s boat returns after X \sim \operatorname{Uniform}(0, 2) hours and the ice truck arrives after Y \sim \operatorname{Exponential}(1) hours, independently. Find the probability that the truck arrives first.

Conditional densities and transformations

  1. For the density of Exercise 2, find f_{X \mid Y}(x \mid y), f_{Y \mid X}(y \mid x), P(X \le 1 \mid Y = 4) and P(Y > 3 \mid X = 1).
  2. A haul fills a fraction X \sim \operatorname{Uniform}(0, 1) of a boat’s hold, and given X = x, the fraction Y of the hold filled with salmon is \operatorname{Uniform}(0, x). Find the joint density, f_Y, P(Y > \tfrac12) and f_{X \mid Y}(x \mid y).
  3. The catch rate \Lambda on a trap line, in fish per hour, is \operatorname{Uniform}(0, 3), and given \Lambda = \lambda the catch N in one hour is \operatorname{Poisson}(\lambda). Find P(N = 0) and P(\Lambda > 1 \mid N = 0).
  4. X and Y are independent \operatorname{Exponential}(1), and R = Y / X. Use the map (x, y) \mapsto (u, r) = (x, y/x) to find the joint density of (X, R), then f_R and P(R \le 2).

Proofs

  1. Prove that if (X, Y) has a joint density, then P(X = Y) = 0. Deduce that if X has a density, (X, X) has no joint density.
  2. (X, Y) is uniform on the unit disk, with density \displaystyle \frac{1}{\pi} there. Prove that \operatorname{Cov}(X, Y) = 0 but X and Y are dependent.
  3. Let (X, Y) have joint density f, let a, b > 0, and let c, d be constants. Prove from the definition, without Result 7, that (aX + c, \ bY + d) has joint density \displaystyle \frac{1}{ab} f\left( \frac{u - c}{a}, \frac{v - d}{b} \right).
  4. Let (X, Y) have joint density f, and suppose f_{X \mid Y}(x \mid y) = g(x) for all x and every y with f_Y(y) > 0. Prove that X and Y are independent and that g is a density of X.
Answers
  1. c = \tfrac23; P(X \le \tfrac12, \ Y \le \tfrac12) = \tfrac18; P(X + Y \le 1) = \tfrac13. The total is c(1 + \tfrac12). Over [0, \tfrac12]^2, \iint 2x \, dA = \tfrac18 and \iint y \, dA = \tfrac{1}{16}. Over the triangle, \iint x \, dA = \iint y \, dA = \tfrac16.
  2. c = 1; f_X(x) = e^{-x} for x > 0; f_Y(y) = y e^{-y} for y > 0; P(Y > 2X) = \tfrac12. Last, P(Y > 2X) = \int_0^\infty e^{-2x} \, dx.
  3. \tfrac38, and F_X(x) = \tfrac12 x(x + 1) for 0 \le x \le 1. By Result 1 the rectangle has probability 1 - \tfrac38 - \tfrac38 + \tfrac18, and F_X(x) = F(x, 1).
  4. \displaystyle f_X(x) = \frac{\sqrt{100 - x^2}}{50\pi} for |x| \le 10; \displaystyle P(X > 6) = \frac{100 \arccos 0.6 - 48}{100\pi} \approx 0.1424; P(X^2 + Y^2 \le 25) = \tfrac14. \{X > 6\} is a sector of angle 2 \arccos 0.6 minus a triangle of area 6 \cdot 8.
  5. c = \tfrac32; independent; f_X(x) = \tfrac{x}{2} on [0, 2], f_Y(y) = 3y^2 on [0, 1]; P(X < 2Y) = \tfrac35. The density is \tfrac{x}{2} \cdot 3y^2 on a product set (Result 4(b)), and \int_0^1 3y^2 \int_0^{2y} \tfrac{x}{2} \, dx \, dy = \int_0^1 3y^4 \, dy.
  6. (a) Dependent: P(X > \tfrac12, \ Y < \tfrac12) = 0, while P(X > \tfrac12) = \tfrac{9}{16} and P(Y < \tfrac12) = \tfrac{1}{16}, from f_X(x) = 4x(1 - x^2) and f_Y(y) = 4y^3. (b) Independent: 4xy = 2x \cdot 2y on the square. (c) Dependent: f_X(x) = x + \tfrac12, so P(X \le \tfrac12) = P(Y \le \tfrac12) = \tfrac38, but P(X \le \tfrac12, \ Y \le \tfrac12) = \tfrac18 \ne \tfrac{9}{64}.
  7. \displaystyle P(Y < X) = \frac{1 + e^{-2}}{2} \approx 0.5677. By Example 3 with letters exchanged, P(Y < X) is \int F_Y(x) f_X(x) \, dx = \int_0^2 \tfrac12 (1 - e^{-x}) \, dx.
  8. f_{X \mid Y}(x \mid y) = \tfrac1y for 0 < x < y; f_{Y \mid X}(y \mid x) = e^{-(y - x)} for y > x; P(X \le 1 \mid Y = 4) = \tfrac14; P(Y > 3 \mid X = 1) = e^{-2} \approx 0.1353. Divide e^{-y} by y e^{-y} and by e^{-x}.
  9. f(x, y) = \tfrac1x for 0 < y < x < 1; f_Y(y) = -\ln y for 0 < y < 1; \displaystyle P(Y > \tfrac12) = \frac{1 - \ln 2}{2} \approx 0.1534; \displaystyle f_{X \mid Y}(x \mid y) = \frac{1}{x \ln(1/y)} for y < x < 1. The joint density is f_X(x) f_{Y \mid X}(y \mid x) by Result 5(c), and \int_y^1 \tfrac{dx}{x} = -\ln y.
  10. \displaystyle P(N = 0) = \frac{1 - e^{-3}}{3} \approx 0.3167 and \displaystyle P(\Lambda > 1 \mid N = 0) = \frac{e^{-1} - e^{-3}}{1 - e^{-3}} \approx 0.3348. By Result 5(d), P(N = 0) = \int_0^3 \tfrac13 e^{-\lambda} \, d\lambda, and given N = 0, \Lambda has density \displaystyle \frac{e^{-\lambda}}{1 - e^{-3}} on (0, 3).
  11. f_{X,R}(u, r) = u e^{-u(1 + r)} for u, r > 0; \displaystyle f_R(r) = \frac{1}{(1 + r)^2} for r > 0; P(R \le 2) = \tfrac23. The inverse is x = u, y = ur, with J = 1 \cdot u - 0 \cdot r = u, and \int_0^\infty u e^{-u(1 + r)} \, du = (1 + r)^{-2}.
  12. The line y = x is the union of the countably many segments \{(t, t) : n \le t \le n + 1\}, n an integer, each of zero area and so of probability 0 by Result 2; countable additivity gives P(X = Y) = 0. If (X, X) had a joint density, P(X = X) would be 0, but it is 1.
  13. The disk and the density are unchanged by (x, y) \mapsto (-x, y), which preserves area, while x and xy change sign, so by Result 6 E(X) = E(XY) = 0 and \operatorname{Cov}(X, Y) = 0. But 0.8^2 + 0.8^2 > 1, so P(X > 0.8, \ Y > 0.8) = 0, while P(X > 0.8) and P(Y > 0.8) are positive. By Result 3 of Independence of events and variables, X and Y are dependent.
  14. Since a, b > 0, the event \{aX + c \le u, \ bY + d \le v\} is \displaystyle \{X \le \frac{u - c}{a}, \ Y \le \frac{v - d}{b}\}, of probability \int_{-\infty}^{(u - c)/a} \int_{-\infty}^{(v - d)/b} f(s, t) \, dt \, ds. Substitute \displaystyle t = \frac{\tau - d}{b} and \displaystyle s = \frac{\sigma - c}{a}, so \displaystyle dt \, ds = \frac{d\tau \, d\sigma}{ab} and the upper limits become v and u. With \displaystyle \tilde f(\sigma, \tau) = \frac{1}{ab} f\left( \frac{\sigma - c}{a}, \frac{\tau - d}{b} \right), the probability is \int_{-\infty}^{u} \int_{-\infty}^{v} \tilde f(\sigma, \tau) \, d\tau \, d\sigma, the definition of a joint density for (aX + c, \ bY + d).
  15. Let \Gamma(x) = \int_{-\infty}^{x} g(s) \, ds. If f_Y(t) > 0, then \int_{-\infty}^{x} f(s, t) \, ds = f_Y(t) \Gamma(x) by the definition of f_{X \mid Y}; if f_Y(t) = 0, both sides are 0, since 0 \le \int_{-\infty}^{x} f(s, t) \, ds \le f_Y(t). By Fubini, F(x, y) = \int_{-\infty}^{y} f_Y(t) \Gamma(x) \, dt, which is \Gamma(x) F_Y(y). Letting y \to \infty, Result 1 gives F_X = \Gamma, so g is a density of X and F(x, y) = F_X(x) F_Y(y), which is independence.
math.wiki · CC-BY-SA-4.0