Why Five Parameters Matter
A bivariate normal model describes two numerical random variables jointly. It is useful when a problem concerns a pair of measurements and their association, rather than either measurement in isolation. For example, a hypothetical quality-control study might model the calibration errors of two sensors installed in the same device. The model has five parameters: one mean and one variance for each variable, and a correlation that describes their linear association.
Those parameters do more than summarize the two measurements separately. Together, they determine the joint density: where it is centered, how widely it spreads in each direction, and how its density contours are oriented. The support of a nonsingular bivariate normal model is the entire two-dimensional plane, even when values in some parts of the plane have very low density.
\(f_{X,Y}(x,y)=\dfrac{1}{2\pi\sigma_X\sigma_Y\sqrt{1-\rho^2}}\exp\!\left[-\dfrac{1}{2(1-\rho^2)}\left\{\left(\dfrac{x-\mu_X}{\sigma_X}\right)^2-2\rho\left(\dfrac{x-\mu_X}{\sigma_X}\right)\left(\dfrac{y-\mu_Y}{\sigma_Y}\right)+\left(\dfrac{y-\mu_Y}{\sigma_Y}\right)^2\right\}\right]\).
This tutorial treats only the nonsingular case \(-1<\rho<1\). The boundary cases are distinct models.
The standardization inside the density puts the two variables on comparable scales. Define \(z_X=(x-\mu_X)/\sigma_X\) and \(z_Y=(y-\mu_Y)/\sigma_Y\). The expression in braces is then a quadratic function of \(z_X\) and \(z_Y\). Thus, a complete parameter specification needs both means, both variances (or standard deviations), and the correlation. Giving only the separate means and variances does not determine the joint density.
Worked Example: Completing a Parameter Specification
A hypothetical paired-sensor model assigns calibration error \(X\), in millimeters, mean \(0.5\) and variance \(4\), and calibration error \(Y\), also in millimeters, mean \(-1\) and variance \(9\). Their correlation is \(0.25\). State the full parameter specification and covariance, and describe the model’s support.
Step 1: Identify the standard deviations. The given values are variances, so \(\sigma_X=\sqrt{4}=2\) mm and \(\sigma_Y=\sqrt{9}=3\) mm.
Step 2: Convert correlation to covariance. Using \(\operatorname{Cov}(X,Y)=\rho\sigma_X\sigma_Y\), \[ \operatorname{Cov}(X,Y)=0.25(2)(3)=1.5\text{ mm}^2. \] The covariance is positive, consistent with the positive correlation.
Step 3: State the model. The parameters are \((\mu_X,\mu_Y,\sigma_X^2,\sigma_Y^2,\rho)=(0.5,-1,4,9,0.25)\), with covariance \(1.5\text{ mm}^2\). Since \(-1<0.25<1\), this is within the nonsingular case considered here. Its support is all pairs \((x,y)\in\mathbb{R}^2\).
Check. Dividing the covariance by the product of the standard deviations gives \(1.5/(2\cdot3)=0.25\), the specified correlation. The model is an idealized probability model; full support does not mean every possible pair is equally plausible.
How the Parameters Shape the Joint Density
The means locate the center of the density at \((\mu_X,\mu_Y)\). Increasing \(\sigma_X\) spreads the distribution in the \(X\)-direction, while increasing \(\sigma_Y\) spreads it in the \(Y\)-direction. The correlation affects how the variables vary together. Positive correlation associates above-mean values of one variable with above-mean values of the other, and below-mean values with below-mean values. Negative correlation associates an above-mean value of one with a below-mean value of the other. Correlation zero removes this linear association, but a claim about independence requires care beyond simply reading a correlation value.
For a fixed positive value \(c\), a density contour is a set of points where the quadratic expression in the exponent equals \(c\): \[ \left(\frac{x-\mu_X}{\sigma_X}\right)^2 -2\rho\left(\frac{x-\mu_X}{\sigma_X}\right)\left(\frac{y-\mu_Y}{\sigma_Y}\right) +\left(\frac{y-\mu_Y}{\sigma_Y}\right)^2 =c(1-\rho^2). \] For the nonsingular model these contours are ellipses centered at the means. They describe equal-density locations, not boundaries of the support. The density is positive at every finite point in the plane.
The orientation of the ellipses depends on the variances as well as on the correlation. When the variances differ, the angle \(\theta\) of a principal axis relative to the \(X\)-axis satisfies \(\tan(2\theta)=2\rho\sigma_X\sigma_Y/(\sigma_X^2-\sigma_Y^2)\), with the axis selected according to the corresponding eigenvector. If the variances are equal and \(\rho\ne0\), the axes are at \(45^\circ\) to the coordinate axes (the sign of \(\rho\) determines which diagonal is the major axis). If the variances are equal and \(\rho=0\), the contours are circles. Therefore, it is not generally correct to say that moving \(\rho\) closer to zero makes the ellipses less tilted. For unequal variances, the orientation approaches the coordinate axes as \(|\rho|\) decreases toward zero; with equal variances, every nonzero correlation gives diagonal axes.
Worked Example: Evaluating a Joint Density
Suppose \((X,Y)\) has a nonsingular bivariate normal distribution with means \(50\) and \(100\), variances \(16\) and \(9\), and correlation \(0.5\). Find the joint density at \((x,y)=(54,97)\). Interpret what the result represents.
Step 1: Find the standard deviations and standardized coordinates. \(\sigma_X=\sqrt{16}=4\) and \(\sigma_Y=\sqrt{9}=3\). Therefore, \[ z_X=\frac{54-50}{4}=1,\qquad z_Y=\frac{97-100}{3}=-1. \]
Step 2: Evaluate the quadratic expression. \[ z_X^2-2\rho z_Xz_Y+z_Y^2 =1-2(0.5)(1)(-1)+1=3. \] Since \(1-\rho^2=1-0.25=0.75\), the exponent is \[ -\frac{3}{2(0.75)}=-2. \]
Step 3: Evaluate the density. \[ f_{X,Y}(54,97) =\frac{1}{2\pi(4)(3)\sqrt{0.75}}e^{-2} =\frac{e^{-2}}{12\pi\sqrt{3}} \approx 0.0021. \] The density is about \(0.0021\) per product unit of \(X\) and \(Y\). It is not the probability that \(X=54\) and \(Y=97\); for continuous variables, a single point has probability zero.
Check. The standardized coordinates are one standard deviation above the \(X\)-mean and one below the \(Y\)-mean. With positive correlation, that opposite-direction pairing has lower density than a pairing with both standardized coordinates positive and of similar size. The exponent calculation reflects that relationship.
Why the Quadratic Expression Has Elliptical Contours
A short algebraic rearrangement makes the shape of the density easier to understand. Write \(z_X=(x-\mu_X)/\sigma_X\) and \(z_Y=(y-\mu_Y)/\sigma_Y\), and call the quadratic expression \(Q\). Completing the square gives
That last expression would double-count a term, so use the following exact completion instead:
Expanding the right side gives \(z_X^2-2\rho z_Xz_Y+\rho^2z_Y^2+(1-\rho^2)z_Y^2=Q\). Because \(1-\rho^2>0\), this expression is positive except at \(z_X=z_Y=0\). Setting \(Q\) equal to a positive constant therefore gives a closed, elliptical contour around the standardized center. This also explains why the nonsingular density has no finite cutoff: every finite pair of standardized coordinates has a finite exponent and a positive density.
Worked Example: Interpreting Contour Orientation
Compare the contour orientation for two hypothetical bivariate normal models. In the first, both variances are \(9\) and the correlation is \(0.6\). In the second, the variances are \(25\) and \(4\), and the correlation is again \(0.6\). Find the principal-axis orientation in each case and explain why the same correlation does not force the same angle.
Model 1: Equal variances. Here \(\sigma_X=\sigma_Y=3\). The variances match and \(\rho\ne0\), so the contour axes are at \(45^\circ\) and \(135^\circ\). Since the correlation is positive, the major axis follows the positive-slope diagonal, at \(45^\circ\) to the \(X\)-axis.
Model 2: Unequal variances. Here \(\sigma_X=5\) and \(\sigma_Y=2\). Use the orientation equation: \[ \tan(2\theta)=\frac{2(0.6)(5)(2)}{25-4}=\frac{12}{21}=\frac{4}{7}. \] Thus \(2\theta=\arctan(4/7)\approx29.74^\circ\), giving \(\theta\approx14.87^\circ\). Since the \(X\)-variance is larger, the major axis is the one closer to the \(X\)-axis.
Conclusion and check. The first model’s axes are diagonal; the second model’s major axis is about \(14.87^\circ\) from the \(X\)-axis. The correlation gives the standardized association, but the variances also affect the orientation in the original coordinate system. Reporting both angles to two decimals, the second model’s major-axis angle is \(14.87^\circ\).
Common Parameter and Geometry Traps
The most frequent parameter error is confusing variance with standard deviation. The formula for covariance uses standard deviations: \(\operatorname{Cov}(X,Y)=\rho\sigma_X\sigma_Y\), not \(\rho\sigma_X^2\sigma_Y^2\). The latter has the wrong units and is generally numerically incorrect. A second trap is treating a correlation as though it alone determines the contour orientation. It does not: the two variances matter, and equal variances have a special geometry. A third is mistaking a high-density region or an ellipse contour for the support. The nonsingular model assigns positive density throughout \(\mathbb{R}^2\), even far from its center.
A reliable exam approach is to write the parameter tuple first, convert variances to standard deviations before calculating covariance or correlation, and then interpret the geometry in the original units. State explicitly that the model is nonsingular when using the density above. These steps keep parameter calculations separate from contour interpretation and are aligned with the SOA Exam P (Probability) syllabus.
Check Your Understanding
Work through each original problem, then compare your result with the answer in parentheses.
- A nonsingular bivariate normal pair has means \(12\) and \(-3\), variances \(16\) and \(9\), and covariance \(6\). Find its correlation and state the five-parameter specification. (Answer: \(\rho=0.5\); \((12,-3,16,9,0.5)\).)
- A hypothetical pair of engineering deviations has standard deviations \(2\) and \(5\), and correlation \(-0.3\). Find the covariance, including its sign, and explain what the sign indicates. (Answer: \(-3\); the deviations tend to vary in opposite directions.)
- A bivariate normal pair has equal variances and correlation \(-0.8\). State the orientation of its principal axes and explain whether reducing the magnitude of the correlation to \(0.2\) changes that orientation. (Answer: The axes are at \(45^\circ\) and \(135^\circ\), with the major axis on the negative-slope diagonal; at correlation \(-0.2\), the axes retain those orientations.)
- For a standard bivariate normal pair with correlation \(0.5\), evaluate the joint density at \((0,1)\). (Answer: \(\dfrac{e^{-2/3}}{\pi\sqrt{3}}\approx0.0943\).)
- A nonsingular bivariate normal model has mean pair \((4,7)\), variances \(1\) and \(36\), and correlation \(0\). State its support and describe the contour orientation. (Answer: Support is all of \(\mathbb{R}^2\); contours are axis-aligned ellipses, with a wider spread in the \(Y\)-direction.)