Joint Densities For Continuous Losses

Why Joint Densities Matter

When two losses are modeled together, an event of interest may involve both at once: their combined amount may exceed a threshold, or one loss may be large while the other stays in a specified range. For continuous losses, probability is assigned to regions of possible pairs, not to individual points. A joint density provides the rule for calculating that probability.

The key practical step is to translate the event into a region for the pair $(X,Y)$, where $X$ and $Y$ are continuous losses. The density is then integrated over that region. Incorrect limits can produce a plausible-looking answer that describes a different event, so the support of the model and the boundaries of the event must both be identified before integrating.

Definition: A joint probability density function of continuous random variables $X$ and $Y$ is a nonnegative function $f_{X,Y}(x,y)$ such that, for every suitable region $R$ in the plane, $$ P\big((X,Y)\in R\big)=\iint_R f_{X,Y}(x,y)\,dA. $$ Here $dA$ denotes area, equivalently $dx\,dy$ or $dy\,dx$ when the limits describe the same region. The density must integrate to one over the entire plane: $$ \iint_{\mathbb{R}^2} f_{X,Y}(x,y)\,dx\,dy=1. $$

The support is the set of pairs where the density can be positive. Outside the support, the density is zero. A density is probability per unit area; its value at one point is not the probability of that exact pair. For a continuous pair with a joint density, any single point—and, more generally, a boundary consisting of curves with zero area—has probability zero.

For example, if the event is $a<X\le b$ and $c<Y\le d$, its region is a rectangle, and the probability is

P(a<X\le b,\ c<Y\le d) = \int_a^b\int_c^d f_{X,Y}(x,y)\,dy\,dx.

For a nonrectangular event, one or more integration limits usually depend on the other variable. Choosing the order of integration that gives the clearest limits is often the simplest approach.

Why Integrating the Density Gives Probability

The probability interpretation follows from how density describes probability locally. Divide a region $R$ into small cells. If the density changes little within a cell of area $\Delta A$, the probability in that cell is approximately $f_{X,Y}(x,y)\Delta A$, using a representative point $(x,y)$ in the cell. Adding these approximations across the cells gives a sum of density times area. As the cells become smaller, that sum approaches the double integral over $R$.

The same reasoning applies to the whole support. The total probability must be one, so the integral of the density over the support must be one. This normalization condition is both a defining requirement and a useful check on a proposed density.

P\big((X,Y)\in R\big) = \lim_{\text{cell areas}\to 0} \sum_{\text{cells in }R} f_{X,Y}(x,y)\,\Delta A = \iint_R f_{X,Y}(x,y)\,dA.

Worked Probability Calculations

Worked Example: A Sum Threshold on a Rectangular Support

Suppose two hypothetical losses $X$ and $Y$, measured in thousands of dollars, have a constant joint density on $0\le x\le4$, $0\le y\le6$, and zero density elsewhere. Find the probability that $X+Y\le4$.

Method. First verify the density, then integrate it over the part of its support satisfying the event. The support is a rectangle of area $4(6)=24$, so the constant density is $1/24$. It is nonnegative and its integral over the support is $24(1/24)=1$.

The event $X+Y\le4$ means $y\le4-x$. Within the support, $x$ ranges from 0 to 4, and for each such $x$, $y$ ranges from 0 to $4-x$. Therefore,

\begin{aligned} P(X+Y\le4) &=\int_0^4\int_0^{4-x}\frac{1}{24}\,dy\,dx\\ &=\int_0^4\frac{4-x}{24}\,dx\\ &=\frac{1}{24}\left[4x-\frac{x^2}{2}\right]_0^4\\ &=\frac{1}{24}(16-8) =\frac{1}{3} \approx 0.3333. \end{aligned}

The region is a right triangle with area $4(4)/2=8$, so the constant-density check gives $8/24=1/3$, as required. The common mistake is to integrate over the whole rectangular support, which gives 1 and answers a different question. The sloping event boundary $y=4-x$ must restrict the integration region.

Worked Example: A Nonconstant Density Near the Origin

Let $X$ and $Y$ be hypothetical continuous losses in normalized units with density $f_{X,Y}(x,y)=x+y$ for $0\le x\le1$, $0\le y\le1$, and zero elsewhere. Find $P(X+Y\le 1/2)$.

Method. Check normalization and integrate the density over the event region. The density is nonnegative on the square, and

\int_0^1\int_0^1(x+y)\,dy\,dx = \int_0^1\left(x+\frac12\right)dx = \frac12+\frac12 =1.

For $X+Y\le1/2$, $x$ runs from 0 to $1/2$, and $y$ runs from 0 to $1/2-x$. Thus,

\begin{aligned} P(X+Y\le1/2) &=\int_0^{1/2}\int_0^{1/2-x}(x+y)\,dy\,dx\\ &=\int_0^{1/2}\left(x(1/2-x)+\frac{(1/2-x)^2}{2}\right)dx\\ &=\int_0^{1/2}\left(\frac18-\frac{x^2}{2}\right)dx\\ &=\left[\frac{x}{8}-\frac{x^3}{6}\right]_0^{1/2}\\ &=\frac{1}{16}-\frac{1}{48} =\frac{1}{24} \approx 0.0417. \end{aligned}

The integration limits describe a triangle inside the square. A common error is to use $0\le y\le1/2$ for every $x$ between 0 and $1/2$; that includes pairs with $x+y>1/2$. The upper limit must decrease with $x$.

Worked Example: Combining Two Boundaries

Suppose hypothetical losses $X$ and $Y$, in normalized units, have density $f_{X,Y}(x,y)=x/4$ on the square $0\le x\le2$, $0\le y\le2$, and zero elsewhere. Find the probability that $Y>1$ and $X+Y>3$.

Method. Represent both conditions in the support, then integrate over their intersection. The density is nonnegative, and its total integral is

\int_0^2\int_0^2\frac{x}{4}\,dy\,dx = \int_0^2\frac{x}{2}\,dx = \left[\frac{x^2}{4}\right]_0^2 =1.

The inequality $X+Y>3$ requires $x>3-y$. Since $x\le2$, this can occur only when $y>1$. Therefore, the separate condition $Y>1$ does not further restrict the region. For $1<y\le2$, $x$ ranges from $3-y$ to 2:

\begin{aligned} P(Y>1,\ X+Y>3) &=\int_1^2\int_{3-y}^2\frac{x}{4}\,dx\,dy\\ &=\frac18\int_1^2\left(4-(3-y)^2\right)dy\\ &=\frac18\int_1^2(4-t^2)\,dt \qquad (t=3-y)\\ &=\frac18\left[4t-\frac{t^3}{3}\right]_1^2\\ &=\frac18\left(\frac{16}{3}-\frac{11}{3}\right) =\frac{5}{24} \approx 0.2083. \end{aligned}

The region is the upper-right triangle above the line $x+y=3$ in the square. The common mistake is to integrate over all $1<y\le2$ and $0\le x\le2$, applying only $Y>1$. That larger region includes pairs that fail $X+Y>3$; the lower limit $3-y$ enforces the second condition.

Setting Limits Reliably

A dependable procedure is to separate the model from the event. First write down the support and the density, including where the density is zero. Then translate each event condition into a boundary in the $(x,y)$-plane. Only after identifying the intersection of the support and event should you choose integration limits.

Check the density.
Confirm nonnegativity on the support and that the integral over the support is one.

Sketch or describe the region.
Include both the support boundaries and every inequality defining the event.

Choose an integration order.
State the outer-variable range and the inner-variable limits as functions of the outer variable, if needed.

Integrate and check.
Evaluate the integral, confirm the result is between zero and one, and check that the limits describe the intended region.

For a constant density, area provides a useful independent check: the event probability is the event-region area divided by the support area. For a nonconstant density, area alone is not enough; the density weights different parts of the region differently. In either case, the final probability must lie between zero and one.

Common Exam Trap

A joint density value is not a point probability. In the first example, the density at every point inside the support is $1/24$, but $P(X=2,Y=1)=0$, not $1/24$. The density has units of inverse area, while a probability is dimensionless. Probability arises only after integrating the density over a region with positive area.

Another frequent error is to use limits from the support alone and forget the event. A rectangular support does not imply that every event probability is a rectangular integral. For example, a sum constraint creates a diagonal boundary. Expressing that boundary explicitly before integrating is a simple way to avoid including pairs that do not meet the condition.

Key takeaway: A joint density assigns probability through area: integrate it over the intersection of the support and the event region. Correctly identifying that region is as important as evaluating the integral.

Check Your Understanding

For each problem, identify the event region before evaluating its probability.

Let $X,Y$ have constant density on $0\le x\le3,\ 0\le y\le5$, and zero elsewhere. Find $P(X+Y\le3.5)$. (Answer: $0.4000$)
Let $f_{X,Y}(x,y)=x+y$ on the unit square and zero elsewhere. Find $P(X+Y>1)$. (Answer: $0.6667$)
Let $(X,Y)$ have constant density on the triangle $x\ge0,\ y\ge0,\ x+y\le2$, and zero elsewhere. Find $P(X>1)$. (Answer: $0.2500$)
Let $f_{X,Y}(x,y)=x/4$ on $0\le x,y\le2$, and zero elsewhere. Find $P(X+Y>3)$. (Answer: $5/24\approx0.2083$)
On $x\ge0,\ y\ge0,\ x+y\le1$, let $f_{X,Y}(x,y)=2(2x+y)$, and let it be zero elsewhere. Find $P(X+Y\le1/2)$. (Answer: $0.1250$)

What You'll Learn