When a Sum Has No Standard Distribution

When a Sum Needs Its Own Distribution

Independent sums often match familiar distribution families, but independence alone does not guarantee that they do. If the individual distributions do not combine into a recognizable standard form, the sum can still be handled directly: condition on one variable, or equivalently integrate the product of the component densities over the values that can produce the specified sum.

A useful example is the sum of a uniform variable and an exponential variable. The density changes form at the upper endpoint of the uniform variable, so using a single formula over the whole range is a common source of errors. The key is to identify which values of the uniform variable are possible for a given sum.

The exponential convention matters: \(\lambda\) is the rate, so \(E[Y]=1/\lambda\). The assumptions of independence and the stated supports are essential to the convolution below. Since both variables are nonnegative, \(S\) cannot be negative. For a specified positive value \(s\), the contribution from \(X=x\) requires \(Y=s-x\geq 0\), as well as \(0\leq x\leq a\). Thus \(x\) must lie between \(\max(0,s-a)\) and \(\min(a,s)\).

Deriving the Piecewise Density

For independent continuous variables, the density of their sum at \(s\) is the integral of \(f_X(x)f_Y(s-x)\) over all feasible \(x\). In this model, the feasible interval changes when \(s\) passes \(a\). If \(0<s<a\), every \(x\) from 0 to \(s\) is allowed. If \(s\geq a\), the possible values run from \(s-a\) to \(a\).

The first piece follows by integrating over \(0\leq x\leq s\):

For \(s\geq a\), the lower limit becomes \(s-a\), so the integral is instead

This calculation explains the change in form: below \(a\), increasing \(s\) increases the range of feasible uniform values; at and above \(a\), the feasible interval has fixed length \(a\) and shifts to the right. The two formulas agree at \(s=a\), where each gives \((1-e^{-\lambda a})/a\). The density is positive for \(s>0\), zero for \(s\leq 0\), and the support—the closure of the set where the density is positive—is \([0,\infty)\).

As a check, the total probability in the upper piece is

Integrating the lower piece from 0 to \(a\) gives \(1-(1-e^{-\lambda a})/(a\lambda)\). The two probabilities sum to 1, as required.

The Cumulative Distribution Function

Integrating the density gives a convenient CDF for probability calculations. For \(0<s<a\), integrate the lower density from 0 to \(s\). For \(s\geq a\), it is often simpler to integrate the upper tail from \(s\) to infinity and subtract from 1.

For example, when \(s\geq a\), the tail integral is

Both CDF pieces give the same value at \(s=a\), so there is no jump there. Because \(S\) is continuous, the probability of any single exact value is zero; interval probabilities are found by subtracting CDF values.

Common Mistake: Using the Wrong Piece

The formulas are separated by the threshold \(a\), not by whether \(s\) is “close” to the uniform interval. In particular, \(0<s<a\) requires the lower density piece, and \(s\geq a\) requires the upper piece. The integration limits change at that threshold, so the formulas cannot be swapped.

For the second example, a wrong calculation would apply the upper density formula at \(s=0.75\):

This is not the density at 0.75: the upper formula was derived for \(s\geq 1\), but \(0.75<1\). The correct lower-piece calculation is \(1-e^{-1.5}\approx 0.7769\). The problem is not that a density exceeds 1—densities can do so—but that the formula was used outside its valid range.

Why Direct Convolution Is Useful

The sum \(X+Y\) is determined by its distribution even when it is not identified as one of the familiar named families. Convolution gives a direct route to its density, and the CDF then gives tail and interval probabilities. This is especially valuable when a problem supplies simple component distributions but asks about a sum whose distribution does not match a standard closure pattern.

The practical steps are consistent across such calculations:

This uniform-exponential calculation illustrates why the support and the integration limits should be identified before any algebra. A plausible-looking exponential expression can still be wrong if it comes from integrating over values that cannot occur.