Why the Shape of a Joint Distribution Matters
A joint distribution describes which pairs of values variables can take and how probability is allocated among those pairs. Its support records the values that remain possible, including limiting values approached by points with positive probability. Comparing the joint support with the separate supports of the variables gives a useful first check for independence: if some pair of marginally possible values is excluded by the joint support, the variables cannot be independent.
This is a diagnostic, not a complete test. Independence requires the joint support to have the product shape, but that shape alone does not determine how probability is distributed within it. The probability calculations still matter.
This neighborhood definition matters for discrete distributions as well as continuous ones. For a finite discrete distribution, the support is exactly the set of values with positive probability. But for an infinite discrete distribution, the support may also contain accumulation points that have probability zero themselves. For example, if \(P(X=1/n)=2^{-n}\) for positive integers \(n\), then \(0\) is in the support: every neighborhood of \(0\) contains some \(1/n\) with positive probability, even though \(P(X=0)=0\).
For a continuous distribution, one often describes the support by the closure of the region on which its density is positive. The neighborhood definition is the underlying probability-based definition; it avoids dependence on how a density happens to be written at individual points.
The Product-Support Condition
Write \(\operatorname{supp}(X)\) for the support of \(X\) and \(\operatorname{supp}(X,Y)\) for the joint support. The product of the marginal supports is the set of all pairs whose first coordinate is in the support of \(X\) and whose second coordinate is in the support of \(Y\).
Here is why the theorem holds. Suppose first that \(x\) and \(y\) are in their respective marginal supports. Any open neighborhood of \((x,y)\) contains a sufficiently small product of neighborhoods \(U\times V\), where \(U\) surrounds \(x\) and \(V\) surrounds \(y\). By the support definitions, \(P(X\in U)>0\) and \(P(Y\in V)>0\). Independence gives \(P((X,Y)\in U\times V)=P(X\in U)P(Y\in V)>0\). Thus every neighborhood of \((x,y)\) has positive probability, so the pair is in the joint support.
For the other inclusion, suppose \(x\) is not in the support of \(X\). Some neighborhood \(U\) of \(x\) has \(P(X\in U)=0\). A sufficiently small neighborhood of any point \((x,y)\) lies inside the vertical strip whose first coordinate is in \(U\); that neighborhood has joint probability zero. Therefore \((x,y)\) is not in the joint support. The same argument applies if \(y\) is outside the support of \(Y\). Together, these arguments establish the product-support equality for independent variables.
The key implication for an exam problem is practical: identify the marginal supports, form their Cartesian product, and check whether the joint support includes all those pairs. A single excluded pair is enough to rule out independence.
Worked Example: A Finite Joint Table Restricted to the Diagonal
A hypothetical model records two binary indicators, \(X\) and \(Y\), for two features of a policy review. The only outcomes are \((0,0)\) with probability \(0.6\) and \((1,1)\) with probability \(0.4\). Find the marginal and joint supports and determine whether the variables are independent. No interest rate or mortality assumption is involved; the specified model is a finite discrete joint distribution with the probabilities given above.
Method. For this finite distribution, each support consists of the values or pairs with positive probability. The marginal probabilities are \(P(X=0)=0.6,\ P(X=1)=0.4\), and likewise \(P(Y=0)=0.6,\ P(Y=1)=0.4\). Hence \(\operatorname{supp}(X)=\operatorname{supp}(Y)=\{0,1\}\).
The product of the marginal supports is \(\{(0,0),(0,1),(1,0),(1,1)\}\), but the joint support contains only \((0,0)\) and \((1,1)\). The pairs \((0,1)\) and \((1,0)\) are missing, so the product-support condition fails and the variables are dependent.
As a probability check, independence would require \(P(X=0,Y=0)=P(X=0)P(Y=0)=0.6(0.6)=0.36\). The actual joint probability is \(0.6\), not \(0.36\). The support diagnosis and the probability calculation agree.
Common mistake avoided: Do not infer independence merely because both variables have the same two marginal values. The joint table determines which pairs occur. Here the diagonal shape rules independence out.
Worked Example: A Triangular Continuous Support
A hypothetical continuous model has joint density \(f_{X,Y}(x,y)=2\) for \(0<x<1,\ 0<y<1-x\), and \(0\) elsewhere. Determine the marginal supports and use geometry to assess independence. The density is the complete distribution specification; no interest or mortality assumption is involved.
Method. First check normalization. The region is a right triangle of area \(1/2\), so the integral of the constant density is \(2(1/2)=1\). The marginal density of \(X\), for \(0<x<1\), is \[ f_X(x)=\int_0^{1-x}2\,dy=2(1-x). \] It is positive throughout the interval \((0,1)\), and every neighborhood of either endpoint has positive probability. Thus \(\operatorname{supp}(X)=[0,1]\). By symmetry, \(f_Y(y)=2(1-y)\) for \(0<y<1\), so \(\operatorname{supp}(Y)=[0,1]\) as well.
The product of the marginal supports is the square \([0,1]\times[0,1]\). The joint support is the closed triangle \(\{(x,y):x\geq 0,\ y\geq 0,\ x+y\leq 1\}\). For example, \((0.8,0.8)\) belongs to the product square but is outside the triangle. Therefore the joint support is not the product of the marginal supports, and \(X\) and \(Y\) are dependent.
A numerical check gives the same conclusion. The event \(X\leq 0.5,\ Y\leq 0.5\) occupies a square of area \(0.25\) inside the triangular region, apart from a boundary point of probability zero. Its joint probability is \(2(0.25)=0.5\). Each marginal probability is \[ P(X\leq 0.5)=\int_0^{0.5}2(1-x)\,dx =[2x-x^2]_0^{0.5}=1-0.25=0.75, \] and \(P(Y\leq 0.5)=0.75\). If the variables were independent, the joint probability would be \(0.75(0.75)=0.5625\), which differs from \(0.5\).
Common mistake avoided: A triangular density region is not the product of its marginal ranges. The fact that both marginal supports span the same interval does not make every pair in the square jointly possible.
Why Matching Supports Do Not Prove Independence
A joint distribution can assign probability unevenly within a rectangular support while keeping that support equal to the product of its marginal supports. The support condition checks which regions are possible, not whether their probabilities factor into marginal probabilities. This distinction is why support geometry is a fast way to disprove independence but usually cannot establish it.
Worked Example: A Full-Square Support with Dependence
Let \(X,Y\) have density \(f_{X,Y}(x,y)=1+\frac12(2x-1)(2y-1)\) for \(0<x<1,\ 0<y<1\), and zero elsewhere. Determine the support and test independence using a rectangle probability. This is a continuous model with the displayed density and no interest or mortality assumption.
Method. Since each of \(2x-1\) and \(2y-1\) lies between \(-1\) and \(1\), the density is at least \(1-\frac12=\frac12>0\) throughout the open square. The density integrates to one because \[ \int_0^1\int_0^1 1\,dy\,dx=1 \] and the remaining term has integral zero: \(\int_0^1(2x-1)\,dx=0\). Thus the joint support is the closed square \([0,1]^2\).
Integrating over \(y\) gives \[ f_X(x)=\int_0^1\left[1+\frac12(2x-1)(2y-1)\right]dy=1, \quad 0<x<1, \] because \(\int_0^1(2y-1)\,dy=0\). Similarly, \(f_Y(y)=1\) on \((0,1)\). Both marginal supports are \([0,1]\), so their product is exactly the joint support.
Now calculate a rectangle probability: \[ \begin{aligned} P(X\leq 0.5,Y\leq 0.5) &=\int_0^{0.5}\int_0^{0.5} \left[1+\frac12(2x-1)(2y-1)\right]dy\,dx\\ &=0.25+\frac12(-0.25)(-0.25)\\ &=0.25+0.03125=0.28125=\frac{9}{32}. \end{aligned} \] The marginal probabilities are both \(0.5\), so independence would give \(0.5(0.5)=0.25\). Since \(9/32\neq 1/4\), the variables are dependent despite having the product support.
Common mistake avoided: Do not conclude independence from a rectangular support or from positive density throughout that rectangle. Independence is a factorization property of probabilities, not merely a statement about the shape of the support.
How to Use the Diagnostic Carefully
A reliable support check separates the geometry into two steps. First determine each marginal support; then compare the Cartesian product of those sets with the joint support. The product contains every combination of marginally possible values. Independence requires all those combinations to be in the joint support, not just the combinations that happen to have positive probability as individual points.
For a table, use the listed joint probabilities. For a density, identify its positive-probability region, remembering that boundaries can belong to the support through the neighborhood definition.
Use the marginal distributions, not just the visible shape of the joint region. For an infinite discrete distribution, include accumulation points whenever every neighborhood has positive probability.
If the joint support omits any pair in that product, independence is impossible. If the sets match, the support test is inconclusive.
For a suitable event, compare its joint probability with the product of the corresponding marginal probabilities. A mismatch proves dependence.
Several boundary and interpretation issues are worth keeping in view:
- Support is not just the set of positive-probability atoms in every discrete model. That shortcut works for finite distributions, but an infinite discrete distribution can have support points that carry no mass themselves.
- Boundary points can be in the support despite having probability zero. A continuous distribution may assign zero probability to each boundary point while every neighborhood of that point intersects a positive-density region.
- Marginal ranges can conceal restrictions. Knowing that \(X\) and \(Y\) each range from zero to one does not show that all pairs in the unit square can occur; a triangular or curved joint support may exclude many of them.
- Matching geometry is not a proof of independence. After supports match, probability factorization still needs to be checked if independence is at issue.
Check Your Understanding
For each problem, identify what the support comparison establishes; where the support matches, do not infer more than it warrants.
- A finite joint distribution assigns probabilities \(0.3,0.2,0.1,0.4\) to \((0,0),(0,1),(1,0),(1,1)\), respectively. Find the marginal supports and joint support. Does the support rule out independence? (Answer: Both marginal supports are \(\{0,1\}\), the joint support is the full four-pair product, so support geometry does not rule out independence.)
- A density is \(3\) on \(0<x<1,\ 0<y<1-x\), and zero elsewhere. Is it a normalized density, and what does its support imply about independence? (Answer: No; the triangular area is \(1/2\), so the integral is \(3/2\), not \(1\). As stated it is not a probability density, so no valid independence conclusion can be drawn without correcting the normalization.)
- Let \(P(X=1/n,Y=1/n)=2^{-n}\) for \(n=1,2,\ldots\), with no other outcomes. Describe the marginal supports and state whether support geometry rules out independence. (Answer: Both marginal supports are \(\{1,1/2,1/3,\ldots\}\cup\{0\}\); the joint support is \(\{(1/n,1/n):n\geq1\}\cup\{(0,0)\}\), not the product, so the variables are dependent.)
- A joint density is positive throughout the open rectangle \(0<x<2,\ 1<y<3\) and zero outside its closure, and the marginal supports are \([0,2]\) and \([1,3]\). What does the support comparison say? (Answer: The joint support matches \([0,2]\times[1,3]\), so support geometry alone is inconclusive about independence.)
- Suppose \(X,Y\) have uniform marginals on \([0,1]\), and \(P(X\leq 1/2,Y\leq 1/2)=0.30\). Their joint support is the full unit square. What conclusion follows from this event probability and the support? (Answer: Independence would require the probability to be \(0.5(0.5)=0.25\); since \(0.30\neq0.25\), the variables are dependent even though the support matches the product.)