Joint Lower-Tail Probabilities from the Bivariate Normal CDF

Why Joint Lower-Tail Probabilities Need Two Thresholds

Suppose two measurements are jointly normally distributed, and a question asks for the probability that both are below specified values. Each threshold must be put on the scale of its own variable, and the dependence between the variables must be retained. Once those steps are complete, the probability is a value of the standard bivariate normal cumulative distribution function.

This is not generally the product of two one-variable probabilities. That shortcut is valid when the two jointly normal variables are independent, which occurs when their correlation is zero. For nonzero correlation, the bivariate normal CDF accounts for how the two measurements tend to vary together or in opposite directions.

Definition: Let $Z_1$ and $Z_2$ be standard normal variables with correlation $\rho$, where $-1<\rho<1$. Their standard bivariate normal CDF is $$ \Phi_2(h,k;\rho)=P(Z_1\le h,\ Z_2\le k). $$ It is the probability that the first component is at most $h$ and the second is at most $k$, jointly. The order of the thresholds matters: $h$ applies to $Z_1$, and $k$ applies to $Z_2$.

For a nondegenerate standard bivariate normal pair, the joint density used to define this CDF is

\phi_2(z_1,z_2;\rho) = \frac{1}{2\pi\sqrt{1-\rho^2}} \exp\!\left( -\frac{z_1^2-2\rho z_1z_2+z_2^2}{2(1-\rho^2)} \right).

Thus $\Phi_2(h,k;\rho)$ is the integral of this density over all points with $z_1\le h$ and $z_2\le k$. Except in special cases, the value is obtained from a bivariate normal table or numerical CDF function rather than an elementary formula. A calculator or software must use the same correlation and threshold order as the model.

Standardizing the Pair

Let $X$ and $Y$ be jointly normally distributed, with means $\mu_X,\mu_Y$, positive standard deviations $\sigma_X,\sigma_Y$, and correlation $\rho$. Define

Z_1=\frac{X-\mu_X}{\sigma_X}, \qquad Z_2=\frac{Y-\mu_Y}{\sigma_Y}.

Then $Z_1,Z_2$ are standard normal with correlation $\rho$. For thresholds $a,b$, standardization gives

P(X\le a,\ Y\le b) = \Phi_2\!\left( \frac{a-\mu_X}{\sigma_X}, \frac{b-\mu_Y}{\sigma_Y}; \rho \right).

The two standardized thresholds are calculated separately, using the corresponding mean and standard deviation. Because each standard deviation is positive, the direction of each inequality is unchanged. The correlation is unchanged as well: subtracting constants and dividing by positive constants does not change the correlation between the standardized variables.

Confirm the model.
Check that the pair is jointly normal and record both means, both positive standard deviations, and the correlation.

Standardize each cutoff.
Calculate $h=(a-\mu_X)/\sigma_X$ and $k=(b-\mu_Y)/\sigma_Y$, matching each cutoff to its own variable.

Evaluate the CDF.
Use $\Phi_2(h,k;\rho)$, with the first threshold first and the second threshold second.

Check the result.
The joint probability cannot exceed either marginal lower-tail probability. If $\rho=0$, it must equal the product of those marginals.

Why the Standardization Formula Holds

The formula follows by translating the original event into the event for the standardized pair. Since $\sigma_X>0$,

X\le a \quad\Longleftrightarrow\quad \frac{X-\mu_X}{\sigma_X} \le \frac{a-\mu_X}{\sigma_X}.

Likewise, since $\sigma_Y>0$,

Y\le b \quad\Longleftrightarrow\quad \frac{Y-\mu_Y}{\sigma_Y} \le \frac{b-\mu_Y}{\sigma_Y}.

The two statements hold together exactly when both standardized inequalities hold. Therefore

\begin{aligned} P(X\le a,\ Y\le b) &=P\!\left( \frac{X-\mu_X}{\sigma_X}\le \frac{a-\mu_X}{\sigma_X}, \frac{Y-\mu_Y}{\sigma_Y}\le \frac{b-\mu_Y}{\sigma_Y} \right)\\ &=\Phi_2\!\left( \frac{a-\mu_X}{\sigma_X}, \frac{b-\mu_Y}{\sigma_Y};\rho \right). \end{aligned}

The last equality uses the definition of the standard bivariate normal CDF and the fact that the standardized pair has correlation $\rho$. Joint normality is essential: knowing only the two marginal normal distributions and their correlation does not, in general, determine a joint distribution.

Worked Example: Two Hypothetical Service Measurements

In a hypothetical monitoring system, daily response time $X$ and processing time $Y$ are jointly normal. Suppose $X\sim N(72,10^2)$, $Y\sim N(40,8^2)$, and their correlation is $0.5$. Find the probability that $X\le72$ and $Y\le48$. Give the probability to four decimal places.

Method. Standardize both cutoffs and evaluate the bivariate CDF. The event uses the response-time cutoff first and the processing-time cutoff second.

For $X$, the standardized cutoff is

h=\frac{72-72}{10}=0.

For $Y$, it is

k=\frac{48-40}{8}=\frac{8}{8}=1.

The standardized pair has correlation $0.5$, so

P(X\le72,\ Y\le48)=\Phi_2(0,1;0.5)\approx0.4688.

The marginal probabilities are $P(X\le72)=\Phi(0)=0.5000$ and $P(Y\le48)=\Phi(1)\approx0.8413$. The joint result, $0.4688$, is below both marginal probabilities, as it must be. The calculation avoids treating the two measurements as independent.

Worked Example: Both Measurements Below One Standard Deviation Above Their Means

A hypothetical pair of jointly normal production measurements has $U\sim N(120,15^2)$, $V\sim N(80,5^2)$, and correlation $0.5$. Find $P(U\le135,\ V\le85)$, to four decimal places.

Method. Convert each cutoff to a standard-normal scale, then use the joint CDF. For the first measurement,

h=\frac{135-120}{15}=\frac{15}{15}=1.

For the second measurement,

k=\frac{85-80}{5}=\frac{5}{5}=1.

The requested probability is therefore

P(U\le135,\ V\le85) = \Phi_2(1,1;0.5) \approx0.7452.

As a check, each marginal probability is $\Phi(1)\approx0.8413$, so $0.7452$ is below both. It is also greater than the independence benchmark $\Phi(1)^2\approx0.7079$, consistent with the positive association in this jointly normal model.

Worked Example: Negative Correlation and a Joint Lower Tail

For a hypothetical jointly normal pair $R,S$, suppose $R\sim N(30,4^2)$, $S\sim N(50,10^2)$, and the correlation is $-0.5$. Calculate the probability that $R\le34$ and $S\le60$, to four decimal places.

Method. Standardize each cutoff and keep the negative correlation in the CDF. The first standardized cutoff is

h=\frac{34-30}{4}=\frac{4}{4}=1,

and the second is

k=\frac{60-50}{10}=\frac{10}{10}=1.

Thus

P(R\le34,\ S\le60) = \Phi_2(1,1;-0.5) \approx0.6865.

Both marginal probabilities are $\Phi(1)\approx0.8413$, which bounds the joint probability from above. The result is below the independence value, approximately $0.7079$, in line with the negative correlation. The correlation does not change either marginal probability, but it does affect the joint probability.

Common Mistakes and Numerical Checks

A frequent error is to multiply the two marginal probabilities even when the correlation is nonzero. For the first worked example, that incorrect calculation would be

P(X\le72)P(Y\le48) = \Phi(0)\Phi(1) \approx0.5000(0.8413)=0.4207.

The correct probability is $\Phi_2(0,1;0.5)\approx0.4688$. The product assumes independence; the stated correlation is $0.5$, so that assumption is not justified. For a jointly normal pair, zero correlation does imply independence, but that fact does not apply when the correlation is nonzero.

Other errors are often easier to catch before using a CDF table or software:

Swapping the thresholds. $\Phi_2(h,k;\rho)$ applies $h$ to the first variable and $k$ to the second. Keep the variable order consistent throughout.
Using a variance in place of a standard deviation. If $X\sim N(\mu,\sigma^2)$, standardize using $\sigma$, not $\sigma^2$.
Dropping the correlation or changing its sign. Standardization preserves the original correlation; it does not make the variables independent.
Ignoring marginal bounds. The joint event is a subset of either individual event, so $P(X\le a,Y\le b)\le P(X\le a)$ and $P(X\le a,Y\le b)\le P(Y\le b)$.
Applying the method without a joint normal model. Marginal normality alone is not enough to justify a bivariate normal CDF calculation.

For continuous variables, using $<$ rather than $\le$ at either cutoff does not change the probability, because an individual cutoff has probability zero. This does not alter the need to use the correct joint model and correlation.

Key takeaway: For a jointly normal pair with positive standard deviations, standardize each lower cutoff separately, retain the pair’s correlation, and evaluate $\Phi_2(h,k;\rho)$. Use a product of marginal probabilities only when independence is justified.

Check Your Understanding

For each problem, identify the standardized thresholds before evaluating or simplifying the probability. Round numerical probabilities to four decimal places.

A hypothetical jointly normal pair has $A\sim N(18,3^2)$, $B\sim N(40,5^2)$, and correlation $0$. Find $P(A\le21,\ B\le35)$. (Answer: $0.5000\times0.1587=0.0793$)
A hypothetical jointly normal pair has means $10$ and $25$, standard deviations $2$ and $5$, and correlation $0.5$. Find the probability that the first variable is at most $10$ and the second is at most $30$. (Answer: $\Phi_2(0,1;0.5)\approx0.4688$)
For a jointly normal pair with correlation $0.5$, both variables have mean $0$ and standard deviation $1$. Find the probability that both are at most $1$. (Answer: $\Phi_2(1,1;0.5)\approx0.7452$)
For a jointly normal pair with correlation $-0.5$, both variables have mean $0$ and standard deviation $1$. Find the probability that both are at most $1$. (Answer: $\Phi_2(1,1;-0.5)\approx0.6865$)
A hypothetical jointly normal pair has $M\sim N(100,20^2)$, $N\sim N(60,4^2)$, and correlation $0.3$. Find the probability that $M\le80$ and $N\le64$, leaving the result as a standard bivariate CDF evaluation. (Answer: $\Phi_2(-1,1;0.3)$)

What You'll Learn