Boundary Cases at Perfect Positive and Negative Correlation

Why Perfect Correlation Is a Boundary Case

For two jointly normal variables with correlation strictly between \(-1\) and \(1\), the pair has variation in two directions: knowing one variable does not determine the other exactly. At correlation \(1\) or \(-1\), that changes. Once one variable is known, the other is fixed by an exact linear equation. The pair lies on a line rather than spreading across the plane.

This boundary case is useful in joint probability questions because it replaces a two-variable problem with a one-variable one. The essential work is to identify the line, translate every condition in the event onto that line, and find the resulting interval for one normal variable. The method is aligned with the SOA Exam P (Probability) syllabus.

The assumptions matter. Both standard deviations must be positive for the correlation and the ratio in the relationship to be defined. If either variable is constant, its correlation with the other variable is not defined. At perfect correlation, the joint distribution is singular: it has no ordinary density spread over the plane, so a bivariate normal density formula that requires an inverse covariance matrix cannot be used.

Why the Pair Must Lie on a Line

Standardize the two variables by setting \(U=(X-\mu_X)/\sigma_X\) and \(V=(Y-\mu_Y)/\sigma_Y\). Each has mean 0 and variance 1, and their correlation is \(\rho\). The variance of \(V-\rho U\) is

When \(\rho=1\) or \(\rho=-1\), this variance is zero. A random variable with variance zero equals its mean with probability 1; here \(E[V-\rho U]=0\), so \(V-\rho U=0\) with probability 1. Therefore \(V=\rho U\) with probability 1. Substituting the definitions of \(U\) and \(V\) gives the stated line equation. This derivation also explains why the sign of the correlation determines the direction of the line: a positive sign means the standardized variables move together, while a negative sign means they move in opposite directions.

Write \(\Phi(z)=P(Z\le z)\) for the standard normal distribution function, where \(Z\sim N(0,1)\). If the event reduces to \(a<X\le b\), with \(a<b\), then

The same calculation applies to other combinations of strict or non-strict endpoints. Since \(X\) has a continuous normal distribution, any particular endpoint has probability zero, so changing an endpoint from strict to non-strict does not change the answer. Do not, however, discard an endpoint condition before checking whether the inequalities leave an interval, a single point, or no possible values.

Interpreting the Geometry and Avoiding Traps

The line equation is not merely a convenient approximation. Under the stated assumptions it holds with probability 1. Thus, for example, at \(\rho=-1\), an event such as \(X>c\) and \(Y>d\) may describe an empty part of the line even though each separate condition has positive probability. The joint probability is determined by their intersection on the line, not by the product or overlap of two areas in the plane.

A common incorrect approach is to use a nonsingular bivariate normal formula at \(\rho=\pm1\). Such formulas often include a factor involving \(\sqrt{1-\rho^2}\) in a denominator; at the boundary this quantity is zero. The resulting expression is not a valid density calculation. The correction is to use the exact linear relationship and compute a univariate normal probability.

Another trap is to standardize each variable separately and then treat the resulting conditions as independent. Standardization changes location and scale, but it does not remove dependence. At perfect correlation, the standardized variables satisfy \(V=U\) for \(\rho=1\) and \(V=-U\) for \(\rho=-1\), exactly. They are maximally dependent, not independent.

A reliable calculation keeps enough precision in normal probabilities until the final subtraction. For instance, use \(\Phi(1/3)-\Phi(-0.5)\approx0.3220212\), which rounds to 0.3220 at four decimal places. Subtracting already-rounded values can occasionally change the final rounded result. State the requested precision, retain unrounded or sufficiently precise values during the calculation, and round only once at the end.

The boundary relationship is especially helpful when several linear expressions in \(X\) and \(Y\) appear in one event: after substitution, every condition is a restriction on the same variable. It is also a useful diagnostic. If the translated inequalities demand both \(X<r\) and \(X\ge s\) with \(s\ge r\), the event is impossible except possibly at a single endpoint; a continuous normal variable assigns that endpoint probability zero.