The bivariate normal

Maya and Kevin ski the same course on the same snow, so their times rise and fall together. Is there a joint model for two normal variables, fixed by a handful of numbers, in which every sum, difference and prediction is again normal?

Where you are 5 in Several random variables

Why this exists

A single normal variable is fixed by its mean and standard deviation. In a biathlon sprint Maya and Kevin ski the same course on the same snow, so their times are dependent. Kevin’s first lap hints at his last; Maya’s horizontal and vertical shot errors in a crosswind lean together. These are questions about a joint distribution.

The bivariate normal is fixed by five numbers: two means, two standard deviations and a correlation. In it every combination aX + bY + c and every conditional distribution is normal, and zero correlation means independence.

Definition

In plain words: after each variable is standardized, the joint density falls off as e to minus a quadratic linking the two through one number, the correlation \rho.

Let \mu_X, \mu_Y be real, \sigma_X, \sigma_Y > 0 and -1 < \rho < 1. Write \displaystyle u = \frac{x - \mu_X}{\sigma_X}, \displaystyle v = \frac{y - \mu_Y}{\sigma_Y} and

Q(x, y) = \frac{u^2 - 2\rho uv + v^2}{1 - \rho^2}.

The pair (X, Y) has the bivariate normal distribution with parameters \mu_X, \mu_Y, \sigma_X^2, \sigma_Y^2, \rho when it has the joint density

f(x, y) = \frac{e^{-Q(x, y)/2}}{2\pi \sigma_X \sigma_Y \sqrt{1 - \rho^2}}

for all real x, y. We also say X and Y are jointly normal. With means 0 and standard deviations 1 the pair is standard bivariate normal with correlation \rho. \Phi is the distribution function of N(0, 1), as in The standard normal distribution, and \displaystyle \varphi(z) = \frac{1}{\sqrt{2\pi}} e^{-z^2/2} its density.

Completing the square in v gives the square identity:

u^2 - 2\rho uv + v^2 = (1 - \rho^2) u^2 + (v - \rho u)^2.

So Q \ge 0, with Q = 0 only at (\mu_X, \mu_Y), where f is largest.

Intuition

2 2 u v
v = u
v = 0.6u
Figure 1Contours of the standard bivariate normal density with \rho = 0.6. The shaded ellipse holds half the probability and the outer one 90 percent. The long axis is the dashed line v = u; the line of conditional means v = 0.6u is flatter and runs through the leftmost and rightmost points of every contour.

The density depends on (x, y) only through Q, so it is constant along each curve Q = c, an ellipse centered at the means (Result 6). Figure 1 draws two for \rho = 0.6. They lean along v = u: a large u tends to come with a large v.

The solid line v = 0.6u is not the long axis. By the square identity, for fixed u the density is symmetric in v about v = \rho u, so this line bisects every vertical chord of every contour and passes through the leftmost and rightmost points (the dots). It is the conditional mean E(V \mid U = u) of Result 3. Since |\rho| < 1 it is flatter than the long axis: a first lap one standard deviation fast predicts a last lap only \rho standard deviations fast. This pull toward the center is regression toward the mean.

\rho = -0.7
\rho = 0
\rho = 0.7
Figure 2The same two contours, holding 50 and 90 percent, for \rho = -0.7, 0 and 0.7 in standardized coordinates. The sign of \rho sets the lean, its size the narrowness.

Negative \rho leans the ellipses the other way; \rho = 0 makes them circles, and the pair independent (Result 4).

Results

Throughout, (X, Y) is bivariate normal with the parameters of the definition, \displaystyle U = \frac{X - \mu_X}{\sigma_X} and \displaystyle V = \frac{Y - \mu_Y}{\sigma_Y}.

Result 1 (two independent standard normals). (a) If Z_1 and Z_2 are independent N(0, 1), the pair

\begin{aligned} X &= \mu_X + \sigma_X Z_1, \\ Y &= \mu_Y + \sigma_Y \big(\rho Z_1 + \sqrt{1 - \rho^2} \, Z_2\big) \end{aligned}

is bivariate normal with parameters \mu_X, \mu_Y, \sigma_X^2, \sigma_Y^2, \rho; in particular f has integral 1. (b) Conversely, if (X, Y) is bivariate normal, then Z_1 = U and \displaystyle Z_2 = \frac{V - \rho U}{\sqrt{1 - \rho^2}} are independent N(0, 1), and X, Y are given by the formulas in (a).

Result 2 (linear combinations). For constants a, b, c with a and b not both 0,

\begin{aligned} aX + bY + c \sim N\big(&a\mu_X + b\mu_Y + c, \\ &a^2\sigma_X^2 + 2ab\rho\sigma_X\sigma_Y + b^2\sigma_Y^2\big), \end{aligned}

and this variance is positive. In particular X \sim N(\mu_X, \sigma_X^2), Y \sim N(\mu_Y, \sigma_Y^2) and \operatorname{Cov}(X, Y) = \rho \sigma_X \sigma_Y, so \rho is the correlation \rho(X, Y) of Variance and covariance.

Result 3 (conditional distributions). For every real x, the conditional density of Y given X = x is a normal density, with

\begin{aligned} E(Y \mid X = x) &= \mu_Y + \rho \tfrac{\sigma_Y}{\sigma_X} (x - \mu_X), \\ \operatorname{Var}(Y \mid X = x) &= \sigma_Y^2 (1 - \rho^2). \end{aligned}

So E(Y \mid X), a straight line in X, is the best predictor of Y from X, with mean squared error \sigma_Y^2 (1 - \rho^2). Given Y = y, likewise, X is normal with mean \displaystyle \mu_X + \rho \frac{\sigma_X}{\sigma_Y}(y - \mu_Y) and variance \sigma_X^2 (1 - \rho^2).

Result 4 (independence). X and Y are independent exactly when \rho = 0.

Result 5 (linear maps). (a) Let Z_1, Z_2 be independent N(0, 1), S = m_1 + \alpha Z_1 + \beta Z_2 and T = m_2 + \gamma Z_1 + \delta Z_2 with \alpha\delta - \beta\gamma \ne 0. Then (S, T) is bivariate normal. (b) Hence if ad - bc \ne 0, then (aX + bY + e, \ cX + dY + g) is bivariate normal, with means and variances from Result 2 and covariance ac\sigma_X^2 + (ad + bc)\rho\sigma_X\sigma_Y + bd\sigma_Y^2.

Result 6 (elliptical contours). For c > 0 the contour Q(x, y) = c is an ellipse centered at (\mu_X, \mu_Y). In the coordinates u, v its axes lie along v = u and v = -u, with half-lengths \displaystyle \sqrt{(1 + \rho)c} and \displaystyle \sqrt{(1 - \rho)c}. The region Q \le c has area \displaystyle \pi c \, \sigma_X \sigma_Y \sqrt{1 - \rho^2}, and

P\big(Q(X, Y) \le c\big) = 1 - e^{-c/2}.
Why it is true

Result 1. The map z_1 = u, \displaystyle z_2 = \frac{v - \rho u}{\sqrt{1 - \rho^2}} takes the plane one to one onto itself. Its inverse x = \mu_X + \sigma_X z_1, \displaystyle y = \mu_Y + \sigma_Y(\rho z_1 + \sqrt{1 - \rho^2} \, z_2) has Jacobian \displaystyle J = \sigma_X\sigma_Y\sqrt{1 - \rho^2}, and the map itself has Jacobian \displaystyle \frac1J. By the square identity Q = z_1^2 + z_2^2, so

\begin{aligned} f\big(x(z), y(z)\big) \, J &= \frac{e^{-(z_1^2 + z_2^2)/2}}{2\pi} \\ &= \varphi(z_1) \varphi(z_2). \end{aligned}

(b) Result 7 of Joint distributions gives (Z_1, Z_2) this density, and Result 4(b) there makes Z_1, Z_2 independent N(0, 1); solve for X and Y. (a) Independent Z_1, Z_2 have the density \varphi(z_1)\varphi(z_2) (Result 4(a) there), and Result 7 applied to the map back gives (X, Y) the density \varphi(z_1)\varphi(z_2)/J = f, which therefore has integral 1.

Result 2. By Result 1(b), aX + bY + c = m + \alpha Z_1 + \beta Z_2 with m = a\mu_X + b\mu_Y + c, \alpha = a\sigma_X + b\rho\sigma_Y and \displaystyle \beta = b\sigma_Y\sqrt{1 - \rho^2}. Each of \alpha Z_1, \beta Z_2 is a constant or a strictly monotone function of one of Z_1, Z_2, so the rectangle test (Result 3 of Independence of events and variables) makes them independent. Results 2, 3 and 5 of Moment generating functions give the combination M(t) = e^{mt} e^{(\alpha^2 + \beta^2) t^2/2}, and \alpha^2 + \beta^2 = a^2\sigma_X^2 + 2ab\rho\sigma_X\sigma_Y + b^2\sigma_Y^2. This is positive: \beta = 0 forces b = 0, and then \alpha = a\sigma_X \ne 0. So M is the N(m, \alpha^2 + \beta^2) function, and Result 4 there (uniqueness, stated without proof in that lesson) names the distribution.

Taking (a, b) = (1, 0) and (0, 1) gives the marginals. With a = b = 1 the variance is \sigma_X^2 + 2\rho\sigma_X\sigma_Y + \sigma_Y^2, while Result 4 of Variance and covariance makes it \sigma_X^2 + 2\operatorname{Cov}(X, Y) + \sigma_Y^2, so \operatorname{Cov}(X, Y) = \rho\sigma_X\sigma_Y.

Result 3. Put m(x) = \mu_Y + \rho\sigma_Y u and s^2 = \sigma_Y^2(1 - \rho^2), so that \displaystyle \frac{(v - \rho u)^2}{1 - \rho^2} = \frac{(y - m(x))^2}{s^2}. The square identity factors the density:

\begin{aligned} f(x, y) &= \frac{e^{-u^2/2}}{\sigma_X \sqrt{2\pi}} \\ &\quad \cdot \frac{e^{-(y - m(x))^2/(2s^2)}}{s\sqrt{2\pi}}. \end{aligned}

The second factor is the N(m(x), s^2) density in y, with integral 1, so integrating over y (Result 3 of Joint distributions) leaves f_X(x) equal to the first factor, and \displaystyle f_{Y \mid X}(y \mid x) = \frac{f(x, y)}{f_X(x)} is the second. Result 4 of Conditional expectation makes E(Y \mid X) the best predictor, with mean squared error E(\operatorname{Var}(Y \mid X)) = s^2. The density is symmetric in the roles of X and Y.

Result 4. If \rho = 0, then Q = u^2 + v^2 and f is a function of x times a function of y, so Result 4(b) of Joint distributions gives independence. If X and Y are independent, \operatorname{Cov}(X, Y) = 0 by Result 5 of Variance and covariance, so \rho = 0 by Result 2.

Result 5. (a) Put \displaystyle s_S = \sqrt{\alpha^2 + \beta^2} and \displaystyle s_T = \sqrt{\gamma^2 + \delta^2}, both positive as \alpha\delta - \beta\gamma \ne 0. The rotation \displaystyle W_1 = \frac{\alpha Z_1 + \beta Z_2}{s_S}, \displaystyle W_2 = \frac{-\beta Z_1 + \alpha Z_2}{s_S} has an inverse of Jacobian 1 and keeps z_1^2 + z_2^2 = w_1^2 + w_2^2, so by Result 7 of Joint distributions W_1, W_2 are independent N(0, 1). Substituting, S = m_1 + s_S W_1 and

\begin{aligned} T = m_2 &+ \tfrac{\alpha\gamma + \beta\delta}{s_S} W_1 \\ &+ \tfrac{\alpha\delta - \beta\gamma}{s_S} W_2. \end{aligned}

Expanding checks (\alpha^2 + \beta^2)(\gamma^2 + \delta^2) = (\alpha\gamma + \beta\delta)^2 + (\alpha\delta - \beta\gamma)^2, so \displaystyle r = \frac{\alpha\gamma + \beta\delta}{s_S s_T} has r^2 < 1 and \displaystyle |\alpha\delta - \beta\gamma| = s_S s_T \sqrt{1 - r^2}. Replacing W_2 by -W_2 if needed (same density), \displaystyle T = m_2 + s_T\big(rW_1 + \sqrt{1 - r^2} \, W_2\big); apply Result 1(a). (b) By Result 1(b) both combinations have the form of (a), with determinant \displaystyle (ad - bc)\sigma_X\sigma_Y\sqrt{1 - \rho^2} \ne 0; the covariance is Result 3 of Variance and covariance.

Result 6. Rotate the (u, v) axes by 45^\circ: \displaystyle u = \frac{p - q}{\sqrt2}, \displaystyle v = \frac{p + q}{\sqrt2}. Then u^2 + v^2 = p^2 + q^2 and 2uv = p^2 - q^2, so Q = c becomes

\frac{p^2}{(1 + \rho)c} + \frac{q^2}{(1 - \rho)c} = 1,

an ellipse with axes along v = u and v = -u and area \displaystyle \pi\sqrt{(1 + \rho)c}\sqrt{(1 - \rho)c} = \pi c\sqrt{1 - \rho^2}; scaling to x, y multiplies areas by \sigma_X\sigma_Y. By Result 1(b) and the square identity, Q(X, Y) = Z_1^2 + Z_2^2, and Result 2 of Joint distributions in polar coordinates gives

\begin{aligned} P(Z_1^2 + Z_2^2 \le c) &= \int_0^{2\pi}\!\!\int_0^{\sqrt c} \frac{e^{-r^2/2}}{2\pi} \, r \, dr \, d\theta \\ &= 1 - e^{-c/2}. \end{aligned}

Worked examples

Example 1

Maya against Kevin on the same snow

In a sprint, Maya’s ski time X and Kevin’s ski time Y, in minutes, are bivariate normal with \mu_X = 24.0, \sigma_X = 0.5, \mu_Y = 24.4, \sigma_Y = 0.6 and \rho = 0.75. Find the probability that Maya is faster, and that she is faster by more than half a minute. Compare with independent times.

Solution

Let D = X - Y. Result 2 with a = 1, b = -1 gives

\begin{aligned} D &\sim N(-0.4, \ 0.25 - 2(0.75)(0.5)(0.6) + 0.36) \\ &= N(-0.4, 0.16), \end{aligned}

with standard deviation 0.4. So \displaystyle P(D < 0) = \Phi(\frac{0.4}{0.4}) = \Phi(1) \approx 0.8413, and \displaystyle P(D < -0.5) = \Phi(\frac{-0.1}{0.4}) = \Phi(-0.25) \approx 0.4013. For independent times \operatorname{Var}(D) = 0.61 and \displaystyle P(D < 0) = \Phi(0.4/\sqrt{0.61}) \approx 0.6957: the shared snow cancels in the difference.

Example 2

Kevin’s last lap, given his first

Kevin’s first-lap time X and last-lap time Y, in seconds, are bivariate normal with \mu_X = 300, \sigma_X = 8, \mu_Y = 312, \sigma_Y = 10 and \rho = 0.6. His first lap took 290 seconds. Find the conditional distribution of his last lap, and the probability it is under 300 seconds before and after the first lap.

Solution

By Result 3, given X = 290, Y is normal with mean \displaystyle 312 + 0.6 \cdot \frac{10}{8}(290 - 300), which is 312 - 7.5 = 304.5, and variance 10^2(1 - 0.6^2) = 64, standard deviation 8. Hence

P(Y < 300 \mid X = 290) = \Phi(-0.5625) \approx 0.2869.

Before the first lap, Y \sim N(312, 100) by Result 2, and P(Y < 300) = \Phi(-1.2) \approx 0.1151. The first lap was 1.25 standard deviations fast, the predicted last lap only 0.75 = 0.6 \times 1.25: regression toward the mean.

Example 3

Reading a shot group from its density

In a crosswind, Maya’s standing shot lands X cm right of and Y cm above the center of the target, where (X, Y) has joint density

f(x, y) = k \exp\Big(-\frac{x^2 - 2xy + 4y^2 - 2x - 4y}{6}\Big).

Find \mu_X, \mu_Y, \sigma_X, \sigma_Y, \rho, the constant k, and the conditional distribution of X given Y = y.

Solution

The exponent must be \displaystyle -\frac{Q}{2} plus a constant. Expanding (x - h)^2 - 2(x - h)(y - m) + 4(y - m)^2, the x and y coefficients are -2h + 2m and 2h - 8m; matching -2 and -4 gives h = 2, m = 1, and

\begin{aligned} &x^2 - 2xy + 4y^2 - 2x - 4y \\ &\quad = (x - 2)^2 - 2(x - 2)(y - 1) \\ &\qquad + 4(y - 1)^2 - 4. \end{aligned}

So \mu_X = 2, \mu_Y = 1. Matching coefficients with \displaystyle \frac{Q}{2},

\begin{aligned} \frac{1}{2(1 - \rho^2)\sigma_X^2} &= \frac16, \\ \frac{1}{2(1 - \rho^2)\sigma_Y^2} &= \frac46, \\ \frac{\rho}{(1 - \rho^2)\sigma_X\sigma_Y} &= \frac26. \end{aligned}

The first two give (1 - \rho^2)\sigma_X^2 = 3 and (1 - \rho^2)\sigma_Y^2 = \tfrac34, so (1 - \rho^2)\sigma_X\sigma_Y = \tfrac32, and the third gives \rho = \tfrac13 \cdot \tfrac32 = \tfrac12. Then \sigma_X = 2 and \sigma_Y = 1. The normalizing factor is \displaystyle \frac{1}{2\pi \cdot 2 \cdot 1 \cdot \sqrt3/2}, and the -4 leaves a factor e^{4/6}, so \displaystyle k = \frac{e^{-2/3}}{2\pi\sqrt3} \approx 0.0472.

By Result 3 with the roles exchanged, X given Y = y has mean 2 + \tfrac12 \cdot 2(y - 1) and variance 4 \cdot \tfrac34: it is N(y + 1, 3).

Example 4

Where nine shots in ten land

For the shot group of Example 3, find the contour holding 90 percent of the shots and the area inside it, and the fraction inside Q = 1. Of the points (5, 2.5) and (5, -0.5), equally far from the center, which lies inside the 90 percent contour?

Solution

By Result 6, 1 - e^{-c/2} = 0.9 gives c = 2\ln 10 \approx 4.605, and the area is

\begin{aligned} \pi c \, \sigma_X\sigma_Y\sqrt{1 - \rho^2} &= \pi(2\ln 10)(2)(1)\big(\tfrac{\sqrt3}{2}\big) \\ &= 2\sqrt3 \, \pi \ln 10 \approx 25.06 \end{aligned}

square centimeters. The contour Q = 1 holds only 1 - e^{-1/2} \approx 0.3935 of the shots.

At (5, 2.5), u = v = 1.5, so \displaystyle Q = \frac{2.25 - 2.25 + 2.25}{0.75} = 3: inside, on the contour holding 1 - e^{-3/2} \approx 0.7769. At (5, -0.5), u = 1.5 and v = -1.5, so \displaystyle Q = \frac{3(2.25)}{0.75} = 9: outside, with only e^{-9/2} \approx 0.0111 of shots beyond that contour.

Example 5

What the first lap cannot predict

Let (X, Y) be bivariate normal and W = Y - E(Y \mid X). Prove that W \sim N(0, \sigma_Y^2(1 - \rho^2)), that W is independent of X, and that \operatorname{Var}(Y) = \rho^2\sigma_Y^2 + \sigma_Y^2(1 - \rho^2) splits accordingly.

Solution

By Result 3, E(Y \mid X) = \mu_Y + \rho\sigma_Y U, and by Result 1(b), \displaystyle Y = \mu_Y + \sigma_Y(\rho Z_1 + \sqrt{1 - \rho^2} \, Z_2) with Z_1 = U. So

W = \sigma_Y\sqrt{1 - \rho^2} \, Z_2,

and Results 2, 4 and 5 of Moment generating functions give W \sim N(0, \sigma_Y^2(1 - \rho^2)). As X = \mu_X + \sigma_X Z_1 and W are increasing functions of Z_1 and of Z_2, the rectangle test (Result 3 of Independence of events and variables) passes from (Z_1, Z_2) to (X, W). Then Y = E(Y \mid X) + W is a sum of independent terms, so variances add (Result 5 of Variance and covariance), with \operatorname{Var}(\mu_Y + \rho\sigma_Y U) = \rho^2\sigma_Y^2.

Example 6

Both stages slower than usual

Maya’s prone and standing range times X and Y are bivariate normal with \rho = \tfrac12. Show that for any \rho the probability P(X > \mu_X, \ Y > \mu_Y) is \displaystyle \frac14 + \frac{\arcsin\rho}{2\pi}, and find the probability that both stages are slower than her means.

Solution

By Result 1(b), X > \mu_X exactly when Z_1 > 0, and Y > \mu_Y exactly when \displaystyle \rho Z_1 + \sqrt{1 - \rho^2} \, Z_2 > 0. Write (Z_1, Z_2) = (R\cos\Theta, R\sin\Theta) with R > 0 and -\pi < \Theta \le \pi, ignoring the origin, which has probability 0. By the polar integral of Result 6, a sector of angle w and radius n has probability \displaystyle \frac{w}{2\pi}(1 - e^{-n^2/2}), which tends to \displaystyle \frac{w}{2\pi}: \Theta is uniform.

Let \displaystyle \alpha = \arcsin\rho \in (-\frac{\pi}{2}, \frac{\pi}{2}), so \displaystyle \cos\alpha = \sqrt{1 - \rho^2} and \displaystyle \rho Z_1 + \sqrt{1 - \rho^2} \, Z_2 = R\sin(\Theta + \alpha). The condition Z_1 > 0 puts \Theta in \displaystyle (-\frac{\pi}{2}, \frac{\pi}{2}), where \Theta + \alpha \in (-\pi, \pi) and \sin(\Theta + \alpha) > 0 exactly when \Theta > -\alpha. The arc \displaystyle (-\alpha, \frac{\pi}{2}) has angle \displaystyle \frac{\pi}{2} + \alpha, so

P(X > \mu_X, \ Y > \mu_Y) = \frac14 + \frac{\arcsin\rho}{2\pi}.

With \rho = \tfrac12, \displaystyle \arcsin\tfrac12 = \frac{\pi}{6} and the probability is \displaystyle \frac14 + \frac{1}{12} = \frac13, against \displaystyle \frac14 for independent stages.

Boundaries

Normal margins are not enough. Let X \sim N(0, 1), c > 0, and Y = X when |X| \le c, Y = -X otherwise. Then Y \sim N(0, 1), yet the pair is not bivariate normal (Exercise 17 for c = 1). At c \approx 1.538, where \displaystyle E(X^2 \mathbf{1}_{\{|X| \le c\}}) = \frac12, E(XY) = 0: X and Y are normal and uncorrelated yet dependent, since |Y| = |X|. Result 4 needs joint normality.

Correlation \pm 1. At \rho = \pm 1 Result 1(a) gives \displaystyle Y = \mu_Y \pm \frac{\sigma_Y}{\sigma_X}(X - \mu_X), a pair on a line with no joint density (Exercise 18). The wider definition “every aX + bY is normal” admits it as a degenerate case.

Straight lines and ellipses are special. For a general joint density E(Y \mid X = x) bends (Conditional expectation). Result 6 covers regions bounded by contours only; a disk centered at the means has no elementary formula unless \sigma_X = \sigma_Y and \rho = 0.

Where this goes next

Linear combinations of n independent standard normals give the multivariate normal, fixed by a mean vector and a covariance matrix, with ellipsoids whose axes are eigenvectors of that matrix (The spectral theorem). The The central limit theorem, proved has a two-dimensional form with a bivariate normal limit. Result 3 is the model behind Least squares regression and the Sampling distribution of the slope.

Exercises

Sums and differences

  1. On one lap, Kevin’s ski time X and range time Y, in seconds, are bivariate normal with \mu_X = 420, \sigma_X = 12, \mu_Y = 35, \sigma_Y = 4 and \rho = -0.25. Find the distribution of his lap time X + Y, P(X + Y < 440), and the correlation of X + Y with X.
  2. (X, Y) is standard bivariate normal with correlation 0.3. Find the distribution of 2X - Y and P(2X - Y > 3).
  3. (X, Y) is bivariate normal with \sigma_X = 3, \sigma_Y = 2 and \rho = 0.5. Find the constant b for which X - bY is independent of Y.
  4. In a pursuit, Maya’s time X and Kevin’s time Y, in minutes, are bivariate normal with \mu_X = 26.0, \mu_Y = 26.3 and \sigma_X = \sigma_Y = 0.4. What correlation \rho makes P(X < Y) = 0.9?

Conditional distributions and densities

  1. Kevin’s prone and standing range times X and Y, in seconds, are bivariate normal with \mu_X = 30, \sigma_X = 3, \mu_Y = 34, \sigma_Y = 5 and \rho = 0.4. Given X = 27, find the conditional distribution of Y and P(Y > 36 \mid X = 27).
  2. For the pair in Exercise 5, write E(Y \mid X) and E(X \mid Y) as straight lines, and find the product of their slopes.
  3. (X, Y) has joint density \displaystyle k e^{-\frac23(x^2 + xy + y^2)}. Find \sigma_X, \sigma_Y, \rho, k and the conditional distribution of Y given X = x.
  4. (X, Y) has joint density k e^{-(5x^2 - 6xy + 2y^2)/2}. Find \operatorname{Var}(X), \operatorname{Var}(Y), \rho and k.
  5. Let \sigma_Y = 10. How large must |\rho| be for the best predictor of Y from X to have root mean squared error 5?

Contours and quadrants

  1. Kevin’s prone shots land X cm right of and Y cm above the center, bivariate normal with means 0, \sigma_X = 0.6, \sigma_Y = 0.5 and \rho = 0.2. Find the area inside the contour holding half his shots, and the fraction inside the contour Q = 4.
  2. For the standard bivariate normal with correlation -\tfrac12, find P(X > 0, \ Y > 0) and P(X > 0, \ Y < 0) exactly.
  3. Maya’s first and second lap times are bivariate normal with \rho = 0.8. Find the probability that both laps are faster than their means, and that both are on the same side of their means.
  4. In a mixed relay, Maya’s leg time X and Kevin’s leg time Y, in seconds, are bivariate normal with \mu_X = \mu_Y = 1{,}200, \sigma_X = \sigma_Y = 30 and \rho = 0.4. Their total was 2{,}460 seconds. Find the conditional distribution of Maya’s leg time given the total.

Proofs

  1. Let (X, Y) be bivariate normal. Prove that X + Y and X - Y are independent if \sigma_X = \sigma_Y, and not independent if \sigma_X \ne \sigma_Y.
  2. Let (X, Y) be standard bivariate normal with correlation \rho. Using the tower rule and Result 3, but not Result 2, prove that E(XY) = \rho.
  3. Prove that for a bivariate normal pair and all real s, t, E(e^{sX + tY}) = e^{m + w/2}, where m = s\mu_X + t\mu_Y and w = s^2\sigma_X^2 + 2st\rho\sigma_X\sigma_Y + t^2\sigma_Y^2.
  4. Let X \sim N(0, 1), and Y = X if |X| \le 1, Y = -X if |X| > 1. Prove that Y \sim N(0, 1) and that (X, Y) is not bivariate normal.
  5. Let (X, Y) be bivariate normal, (a, b) \ne (0, 0) and ad - bc = 0. Prove that (aX + bY, \ cX + dY) has no joint density.
Answers
  1. X + Y \sim N(455, 136), P(X + Y < 440) \approx 0.0992, correlation \approx 0.9432. By Result 2 the variance is 144 + 2(-0.25)(12)(4) + 16 = 136, and \displaystyle P = \Phi(-15/\sqrt{136}). By Result 5(b), \operatorname{Cov}(X + Y, X) = 144 - 12 = 132; divide by \displaystyle 12\sqrt{136}.
  2. 2X - Y \sim N(0, 3.8) and P(2X - Y > 3) \approx 0.0619. Result 2: 4 - 2(2)(0.3) + 1 = 3.8, and \displaystyle P = 1 - \Phi(3/\sqrt{3.8}).
  3. b = \tfrac34. The map has determinant 1, so (X - bY, Y) is bivariate normal (Result 5(b)) and independence means zero covariance (Result 4): 3 - 4b = 0.
  4. \rho \approx 0.8288. X - Y \sim N(-0.3, 0.32(1 - \rho)), so \displaystyle 0.3/\sqrt{0.32(1 - \rho)} = \Phi^{-1}(0.9) \approx 1.2816, giving 1 - \rho \approx 0.1712.
  5. Y \mid X = 27 \sim N(32, 21) and P(Y > 36 \mid X = 27) \approx 0.1914. Mean \displaystyle 34 + 0.4 \cdot \frac53(-3) = 32, variance 25(0.84) = 21, and \displaystyle P = 1 - \Phi(4/\sqrt{21}).
  6. E(Y \mid X) = 34 + \tfrac23(X - 30) and E(X \mid Y) = 30 + 0.24(Y - 34), with product of slopes 0.16 = \rho^2. By Result 3 the slopes are \displaystyle \rho\frac{\sigma_Y}{\sigma_X} and \displaystyle \rho\frac{\sigma_X}{\sigma_Y}.
  7. \sigma_X = \sigma_Y = 1, \rho = -\tfrac12, \displaystyle k = \frac{1}{\pi\sqrt3} \approx 0.1838, and Y \mid X = x \sim N(-\tfrac{x}{2}, \tfrac34). No linear terms, so both means are 0. The x^2 and y^2 coefficients give \sigma_X = \sigma_Y and 2(1 - \rho^2)\sigma_X^2 = \tfrac32; the xy coefficient, \displaystyle \frac{\rho}{(1 - \rho^2)\sigma_X^2} = -\frac23, gives 2\rho = -1, so \sigma_X^2 = 1. Then use Result 3.
  8. \operatorname{Var}(X) = 2, \operatorname{Var}(Y) = 5, \displaystyle \rho = \frac{3}{\sqrt{10}} \approx 0.9487 and \displaystyle k = \frac{1}{2\pi}. Matching gives (1 - \rho^2)\sigma_X^2 = \tfrac15 and (1 - \rho^2)\sigma_Y^2 = \tfrac12, so \displaystyle (1 - \rho^2)\sigma_X\sigma_Y = \frac{1}{\sqrt{10}}, and the xy coefficient \displaystyle \frac{\rho}{(1 - \rho^2)\sigma_X\sigma_Y} = 3 gives \rho. Then \displaystyle \sigma_X\sigma_Y\sqrt{1 - \rho^2} = 1.
  9. \displaystyle |\rho| = \frac{\sqrt3}{2} \approx 0.866. By Result 3, \displaystyle 10\sqrt{1 - \rho^2} = 5, so \rho^2 = \tfrac34.
  10. About 1.28 square centimeters, and 1 - e^{-2} \approx 0.8647. Half the shots: c = 2\ln 2 (Result 6); area \displaystyle \pi(2\ln 2)(0.6)(0.5)\sqrt{0.96}.
  11. P(X > 0, Y > 0) = \tfrac16 and P(X > 0, Y < 0) = \tfrac13. Example 6 gives \displaystyle \frac14 - \frac{1}{12}, and P(X > 0) = \tfrac12.
  12. Both faster: \displaystyle \frac14 + \frac{\arcsin 0.8}{2\pi} \approx 0.3976; same side: \displaystyle \frac12 + \frac{\arcsin 0.8}{\pi} \approx 0.7952. Q is unchanged by (u, v) \mapsto (-u, -v), so both below matches both above (Example 6).
  13. X \mid X + Y = 2{,}460 \sim N(1{,}230, 270), standard deviation \approx 16.43. (X, X + Y) is bivariate normal (Result 5(b)) with variances 900 and 2{,}520 and covariance 1{,}260, so correlation squared 0.7; Result 3 gives mean \displaystyle 1{,}200 + \frac12(60) and variance 900(0.3).
  14. The map (x, y) \mapsto (x + y, x - y) has determinant -2, so the pair is bivariate normal by Result 5(b), with covariance \sigma_X^2 + 0 \cdot \rho\sigma_X\sigma_Y - \sigma_Y^2. By Result 4 they are independent exactly when \sigma_X^2 = \sigma_Y^2, that is \sigma_X = \sigma_Y.
  15. By the proof of Result 3 and symmetry, f_X = f_Y = \varphi, so E(X^2) = E(Y^2) = 1 and \displaystyle E(|XY|) \le \frac12 E(X^2 + Y^2) is finite. The tower rule and taking out what is known (Results 1 and 2 of Conditional expectation) give E(XY) = E(X \, E(Y \mid X)) = E(\rho X^2) = \rho by Result 3.
  16. If s = t = 0 both sides are 1. Otherwise Result 2 gives sX + tY \sim N(m, w), and the left side is its moment generating function at 1, which is e^{m + w/2} by Result 5 of Moment generating functions.
  17. Split on |X| \le 1: P(Y \le y) is P(X \le y, |X| \le 1) plus P(-X \le y, |-X| > 1). As -X has the distribution of X, the second term is P(X \le y, |X| > 1), and the sum is \Phi(y). But X + Y = 0 whenever |X| > 1, so P(X + Y = 0) \ge 2(1 - \Phi(1)) > 0. A bivariate normal pair would make X + Y normal with positive variance (Result 2), hence with a density and P(X + Y = 0) = 0 (Result 1 of Densities and cumulative distribution functions).
  18. As ad = bc and (a, b) \ne (0, 0), (c, d) = \lambda(a, b) for some \lambda (\displaystyle \lambda = \frac{c}{a} if a \ne 0; if a = 0, then c = 0 and \displaystyle \lambda = \frac{d}{b}). So T = cX + dY equals \lambda S with S = aX + bY, and (S, T) lies on the line t = \lambda s. A joint density g would give each segment \{t = \lambda s, \ |s| \le n\}, of area 0, probability \iint g \, dA = 0 (Result 2 of Joint distributions), and by continuity of probability the whole line probability 0, not 1.
math.wiki · CC-BY-SA-4.0