Bivariate Normal Parameterization Pitfalls

Why Parameterization Errors Matter

A bivariate normal problem may give means, variances, standard deviations, covariance, or correlation. These quantities are related, but they are not interchangeable. Confusing covariance with correlation or treating a variance as a standard deviation can produce a wrong conditional distribution, an impossible correlation, or even a negative variance. The safest approach is to identify each parameter, check that the values are compatible, and only then apply a conditional-normal formula.

The formulas below assume $X$ and $Y$ are jointly bivariate normal. Having two normally distributed marginal variables alone does not guarantee that their pair is jointly bivariate normal, so marginal normality by itself is not sufficient for these conditional formulas.

Definition: For a jointly bivariate normal pair $(X,Y)$, write $E[X]=\mu_X$, $E[Y]=\mu_Y$, $\operatorname{Var}(X)=\sigma_X^2>0$, and $\operatorname{Var}(Y)=\sigma_Y^2>0$. Its covariance is $\operatorname{Cov}(X,Y)$, and its correlation is $$ \rho=\frac{\operatorname{Cov}(X,Y)}{\sigma_X\sigma_Y}. $$ Equivalently, $\operatorname{Cov}(X,Y)=\rho\sigma_X\sigma_Y$. Correlation is dimensionless and must satisfy $-1\leq \rho\leq 1$; covariance carries the product of the units of $X$ and $Y$.

In this notation, $X\sim N(\mu_X,\sigma_X^2)$ and $Y\sim N(\mu_Y,\sigma_Y^2)$, where the second parameter is a variance. Thus, if a problem states a marginal standard deviation, square it to get the variance; if it states a variance, take its square root to get the standard deviation.

Conditional Distributions: Keep the Parameters Straight

For a jointly bivariate normal pair with positive marginal variances and $-1<\rho<1$, the conditional distributions are normal. The conditional mean adjusts the mean of the target variable according to how far the conditioning value is from its own mean. The conditional variance is the target variance multiplied by $1-\rho^2$.

Theorem: If $(X,Y)$ is jointly bivariate normal and $-1<\rho<1$, then $$ Y\mid X=x\sim N\left(\mu_Y+\rho\frac{\sigma_Y}{\sigma_X}(x-\mu_X),\ \sigma_Y^2(1-\rho^2)\right), $$ and $$ X\mid Y=y\sim N\left(\mu_X+\rho\frac{\sigma_X}{\sigma_Y}(y-\mu_Y),\ \sigma_X^2(1-\rho^2)\right). $$ The second parameter in each conditional $N$ expression is the conditional variance, not the conditional standard deviation. These formulas require joint bivariate normality. When $|\rho|=1$, the pair is degenerate and the conditional variance is zero; use the deterministic relationship between the variables rather than treating the pair as a nonsingular bivariate normal model.

The covariance form is often useful when a problem supplies covariance directly. For example, the conditional mean of $Y$ given $X=x$ can also be written as $\mu_Y+\frac{\operatorname{Cov}(X,Y)}{\sigma_X^2}(x-\mu_X)$, and its conditional variance as $\sigma_Y^2-\frac{\operatorname{Cov}(X,Y)^2}{\sigma_X^2}$. These are the same formulas, because $\operatorname{Cov}(X,Y)=\rho\sigma_X\sigma_Y$. Check which quantities are actually given before choosing a form.

Why the Conditional Formula Holds

A short standardization argument explains both the conditional mean and conditional variance. Define standardized variables $U=(X-\mu_X)/\sigma_X$ and $V=(Y-\mu_Y)/\sigma_Y$. They each have mean zero and variance one, and their correlation is $\rho$. Form the residual $W=V-\rho U$. It has zero covariance with $U$, and its variance is $1-\rho^2$:

\operatorname{Cov}(U,W)=\operatorname{Cov}(U,V)-\rho\operatorname{Var}(U)=\rho-\rho=0, \qquad \operatorname{Var}(W)=1-2\rho^2+\rho^2=1-\rho^2.

Because $U$ and $W$ are linear combinations of a jointly normal pair, they are jointly normal. For jointly normal variables, zero covariance implies independence. Therefore, conditioning on $U$ leaves the distribution of $W$ unchanged, with mean zero and variance $1-\rho^2$. Since $V=\rho U+W$, fixing $X=x$ fixes $U=(x-\mu_X)/\sigma_X$; consequently, $V$ has conditional mean $\rho(x-\mu_X)/\sigma_X$ and conditional variance $1-\rho^2$. Transforming back from $V$ to $Y=\mu_Y+\sigma_Y V$ gives the stated conditional mean and variance. This derivation also shows why the factor for conditional variance is $1-\rho^2$, not $1-\rho$.

Worked Example: Convert Covariance Before Conditioning

A hypothetical monitoring system records two jointly bivariate normal measurements, $X$ and $Y$. Suppose $X\sim N(40,16)$, $Y\sim N(70,9)$, and $\operatorname{Cov}(X,Y)=6$. Find the correlation and the conditional distribution of $Y$ given $X=44$. State the conditional standard deviation as well.

Method. First read the second parameters as variances. Thus $\sigma_X=\sqrt{16}=4$ and $\sigma_Y=\sqrt{9}=3$. Convert the supplied covariance to correlation:

\rho=\frac{6}{(4)(3)}=\frac{6}{12}=0.5.

The value is within the permitted interval $[-1,1]$. For $Y\mid X=44$, the conditional mean is

70+0.5\frac{3}{4}(44-40) =70+0.375(4) =70+1.5 =71.5.

The conditional variance and standard deviation are

9(1-0.5^2)=9(0.75)=6.75, \qquad \sqrt{6.75}\approx 2.5981.

Therefore, $Y\mid X=44\sim N(71.5,6.75)$, with conditional standard deviation approximately $2.5981$. Check: the conditioning value $44$ is above $\mu_X=40$, and the positive correlation shifts the conditional mean of $Y$ above $70$, as expected. The conditional variance is positive and less than the original variance $9$. The trap avoided is using the covariance $6$ as if it were a correlation.

Worked Example: Use Marginal Standard Deviations, Not Variances

Suppose a hypothetical pair of jointly bivariate normal quantities satisfies $X\sim N(100,25)$, $Y\sim N(20,9)$, and $\rho=-0.4$. Find the conditional distribution of $X$ given $Y=23$.

Method. The marginal variances are $25$ and $9$, so the marginal standard deviations are $\sigma_X=5$ and $\sigma_Y=3$. The covariance is

\operatorname{Cov}(X,Y)=(-0.4)(5)(3)=-6.

Using the conditional formula for $X\mid Y=y$, the conditional mean is

100+(-0.4)\frac{5}{3}(23-20) =100-\frac{2}{3}(3) =100-2 =98.

The conditional variance is

$$ 25(1-(-0.4)^2)=25(1-0.16)=25(0.84)=21. $$

Hence $X\mid Y=23\sim N(98,21)$, with conditional standard deviation $\sqrt{21}\approx 4.5826$. Check: $23$ is above the mean of $Y$, and the negative correlation shifts the conditional mean of $X$ below $100$, consistent with the result. The trap avoided is multiplying $\rho$ by the marginal variances $25$ and $9$ to calculate covariance; covariance uses the marginal standard deviations.

Check Compatibility Before Calculating

Correlation cannot be chosen independently of the marginal variances and covariance. By the definition of correlation, or equivalently the Cauchy–Schwarz inequality for covariance, any valid pair must satisfy

\left|\operatorname{Cov}(X,Y)\right|\leq \sigma_X\sigma_Y.

If this condition fails, there is no joint distribution with the stated parameters—not just no bivariate normal distribution. In particular, substituting an impossible correlation into a conditional variance formula may create a negative answer. That negative value is a warning that the inputs are inconsistent, not a variance to interpret.

Worked Example: Detect an Impossible Covariance

A proposed model describes two jointly bivariate normal measurements $R$ and $S$ with $\operatorname{Var}(R)=4$, $\operatorname{Var}(S)=9$, and $\operatorname{Cov}(R,S)=7$. Determine whether these parameters are valid. If not, state the largest possible absolute covariance for these marginal variances.

Method. The standard deviations are $\sigma_R=\sqrt{4}=2$ and $\sigma_S=\sqrt{9}=3$. The largest possible absolute covariance is therefore

\sigma_R\sigma_S=(2)(3)=6.

But the proposed covariance has absolute value $7$, which exceeds $6$. Equivalently, its implied correlation would be

\rho=\frac{7}{(2)(3)}=\frac{7}{6}\approx 1.1667,

outside the required interval $[-1,1]$. Thus no valid bivariate normal pair has all the proposed parameters, and no conditional distribution should be calculated from them. The largest possible absolute covariance is $6$. Check: at the boundary values $+6$ or $-6$, the correlation would be $+1$ or $-1$, respectively. The trap avoided is treating an out-of-range correlation as a usable input to the conditional formula.

Worked Example: Keep Conditional Variance Separate From Standard Deviation

Suppose $X$ and $Y$ are jointly bivariate normal with $X\sim N(12,9)$, $Y\sim N(30,16)$, and $\rho=0.5$. Find $P(Y>35\mid X=15)$.

Method. Here $\sigma_X=3$ and $\sigma_Y=4$. The conditional mean is

30+0.5\frac{4}{3}(15-12) =30+2(3) =32.

The conditional variance is $16(1-0.5^2)=16(0.75)=12$, so the conditional standard deviation is $\sqrt{12}$, not $12$. Standardize the requested threshold using that standard deviation:

P(Y>35\mid X=15) =1-\Phi\left(\frac{35-32}{\sqrt{12}}\right) =1-\Phi(0.8660) \approx 0.1932.

The probability is approximately $0.1932$, to four decimal places. Check: the threshold $35$ is above the conditional mean $32$, so an upper-tail probability below $0.5$ is appropriate. The trap avoided is using the conditional variance $12$ as though it were the standard deviation in the $z$-score.

A Reliable Parameter Check

Before calculating a conditional probability from a bivariate normal model, identify the information supplied and keep its role explicit. A short check prevents most parameterization errors:

Label each marginal quantity.
Record whether the problem gives a variance or a standard deviation. Convert to $\sigma_X$ and $\sigma_Y$ before using a formula involving standard deviations.

Reconcile covariance and correlation.
Use $\operatorname{Cov}(X,Y)=\rho\sigma_X\sigma_Y$. Check that the implied correlation lies between $-1$ and $1$.

Use the formula for the variable being predicted.
For $Y\mid X=x$, the mean adjustment contains $\sigma_Y/\sigma_X$ and the conditional variance begins with $\sigma_Y^2$. Reverse those roles for $X\mid Y=y$.

Separate variance from standard deviation.
Compute the conditional variance first. Take its square root only when a standard deviation is needed, such as in a standardized probability calculation.

These checks are especially useful when a problem gives mixed information—for example, one marginal variance, one standard deviation, and either covariance or correlation. A numerical result is not automatically meaningful just because the arithmetic can be carried out. The stated parameters must first describe a possible model, and the conditional formula must match the direction of conditioning.

Key takeaway: In a bivariate normal calculation, covariance equals correlation times the two marginal standard deviations. Conditional variance is the target marginal variance times $1-\rho^2$. Verify the parameter values, the direction of conditioning, and the distinction between variance and standard deviation before evaluating a probability.

Check Your Understanding

For each problem, identify the parameterization carefully before calculating.

A jointly bivariate normal pair has $\operatorname{Var}(X)=36$, $\operatorname{Var}(Y)=4$, and $\operatorname{Cov}(X,Y)=-6$. Find its correlation. (Answer: $\rho=-0.5000$)
A jointly bivariate normal pair has $X\sim N(8,16)$, $Y\sim N(10,25)$, and $\rho=0.3$. Find the conditional distribution of $Y$ given $X=12$. (Answer: $Y\mid X=12\sim N(11.5,22.75)$)
A proposed jointly bivariate normal pair has marginal variances $9$ and $16$, and covariance $-13$. Is this parameter set valid? (Answer: No; the implied correlation is $-13/12$, outside $[-1,1]$)
A jointly bivariate normal pair has $X\sim N(50,100)$, $Y\sim N(8,4)$, and $\rho=-0.6$. Find the conditional distribution of $X$ given $Y=10$. (Answer: $X\mid Y=10\sim N(44,64)$)
A jointly bivariate normal pair has marginal variances $25$ and $49$, with covariance $35$. Determine its correlation and whether it is nonsingular. (Answer: $\rho=1$; the pair is degenerate, not nonsingular)

What You'll Learn