What Makes a Discrete Joint Probability Function Valid?
When two discrete random variables are modeled together, a probability must be assigned to each possible ordered pair of values. The order matters: \((x,y)\) refers to \(X=x\) and \(Y=y\), not to an unordered pair. Before using a proposed joint model, check both that its probabilities form a legitimate probability distribution and that the stated rule assigns the intended probability to each pair.
A valid model must respect its support, assign no negative probabilities, and assign total probability one across all possible pairs. These checks apply whether the support contains a few points or countably infinitely many points. A probability formula that passes the total check can still misstate the model if it attaches a probability to the wrong ordered pair.
The Three Checks
A reliable verification keeps three questions separate. First, are the pairs listed in the support actually the allowed outcomes? Second, are all assigned probabilities nonnegative? Third, does summing the probabilities over every allowed pair give one? After those checks, compare the numerical assignments with the proposed rule: a normalized list can still have probabilities attached to the wrong pairs.
Write the allowed ordered pairs explicitly or describe them precisely. Interpret a formula only on that set, and assign probability zero outside it.
Show that the formula gives a value at least zero at every allowed pair. If it contains an unknown parameter, derive the parameter restrictions that ensure this condition.
Sum across all allowed pairs and verify the result is one. Then evaluate the rule at particular ordered pairs to ensure the probabilities have not been swapped or otherwise misassigned.
For a finite support, normalization is an ordinary finite sum. For an infinite countable support, it is a series; the series must converge and its sum must be one. In either case, a formula written only for supported pairs does not need to describe impossible pairs. Treating an impossible pair as if it were an additional outcome can produce an incorrect total.
Why the Sum Must Equal One
The normalization condition follows directly from what the probabilities represent. Each possible outcome of the pair is a singleton event \(\{X=x,Y=y\}\). Distinct ordered pairs give disjoint events, and the events over the full support cover every possible outcome of \((X,Y)\). Therefore their probabilities must add to the probability of the whole sample space, which is one.
Conversely, nonnegative assignments summing to one can define a probability distribution on a countable set: assign each supported singleton its stated probability, and assign an event consisting of several supported pairs the sum of their probabilities. The nonnegativity condition is essential; a total of one obtained by offsetting a negative entry with an excessively large positive entry is not a valid probability model.
Worked Example: Finding a Normalizing Constant on a Finite Support
Suppose \((X,Y)\) can take any pair in \(S=\{(0,0),(0,1),(1,0),(1,1)\}\), and a proposed rule is \(p(x,y)=c(1+2x+y)\) on \(S\), with probability zero outside \(S\). Find the value of \(c\) that makes the rule a valid joint probability mass function, and check the result. The method is to evaluate every supported pair and normalize the resulting weights.
Step 1: Evaluate the rule at each pair. At \((0,0)\), the weight is \(1+2(0)+0=1\), so \(p(0,0)=c\). At \((0,1)\), the weight is \(1+2(0)+1=2\), so \(p(0,1)=2c\). At \((1,0)\), the weight is \(1+2(1)+0=3\), so \(p(1,0)=3c\). At \((1,1)\), the weight is \(1+2(1)+1=4\), so \(p(1,1)=4c\). All four weights are positive.
Step 2: Normalize. The sum over the full support is \(c+2c+3c+4c=10c\). Requiring this to equal one gives \(10c=1\), hence \(c=1/10\).
Step 3: Check the assignments and total. The probabilities, in the same pair order, are \(1/10=0.1000\), \(2/10=0.2000\), \(3/10=0.3000\), and \(4/10=0.4000\). Their sum is \(0.1000+0.2000+0.3000+0.4000=1.0000\), and every value is nonnegative.
Check: Substituting \(c=1/10\) into the rule gives exactly the four listed probabilities. The rule is valid on the stated support. A common mistake is to normalize using only three of the four pairs; that would omit the probability \(4c\) at \((1,1)\) and would not normalize the distribution.
Nonnegativity and Parameter Restrictions
When a proposed rule includes an unknown parameter, validity may depend on its value. The appropriate procedure is to impose the nonnegativity condition on each supported pair and the normalization condition on the sum. Do not assume that a parameter is valid merely because the formula looks like a probability or because its entries add to one.
For example, if one supported pair has assigned probability \(\theta\), that entry requires \(\theta\geq 0\). If another has probability \(a-\theta\), it requires \(\theta\leq a\). Together these entries require \(0\leq\theta\leq a\), provided the other entries are already nonnegative and the total is one. A parameter value outside this range makes at least one probability negative.
Worked Example: Determining the Valid Parameter Range
Let the possible pairs be \(S=\{(0,0),(0,1),(1,0),(2,0)\}\). A proposed rule assigns \(p(0,0)=\theta\), \(p(0,1)=1/4\), \(p(1,0)=1/3\), and \(p(2,0)=5/12-\theta\). Determine all values of \(\theta\) for which these assignments form a valid joint probability mass function. The method is to check the total and then impose nonnegativity pair by pair.
Step 1: Check the total. Adding the four assignments gives \(\theta+1/4+1/3+5/12-\theta\). The \(\theta\) terms cancel, and the constants sum to \(3/12+4/12+5/12=12/12=1\). Thus the total is one for every \(\theta\).
Step 2: Check each probability for nonnegativity. The fixed entries \(1/4\) and \(1/3\) are positive. The entry at \((0,0)\) requires \(\theta\geq 0\). The entry at \((2,0)\) requires \(5/12-\theta\geq 0\), or \(\theta\leq 5/12\).
Step 3: Combine the conditions. The rule is valid exactly when \(0\leq\theta\leq5/12\). At either endpoint one supported pair has probability zero; that is permitted, although the exact support then omits that pair.
Check: At \(\theta=1/6\), for instance, the four entries are \(1/6\), \(1/4\), \(1/3\), and \(5/12-1/6=1/4\). Their sum is \(2/12+3/12+4/12+3/12=12/12=1\), and none is negative. A common mistake is to check only the total: the total is one even when \(\theta=-1/10\), but then \(p(0,0)<0\), so the assignment is invalid.
Correct Assignment to Ordered Pairs
Normalization does not by itself verify that a stated formula has been applied correctly. The coordinates have specific meanings: the first coordinate is the value of \(X\), and the second is the value of \(Y\). Swapping them can change the probability when the rule treats the coordinates differently. This distinction is especially important when a problem gives weights or a table as well as a formula.
A list with nonnegative entries that sum to one is a valid assignment on its own. But if the task is to implement a particular proposed rule, a list that swaps two entries does not represent that rule, even though the list remains normalized. State which question is being answered: “Does this list define some joint distribution?” or “Does this list correctly implement the specified joint model?”
Worked Example: Catching a Swap Between Coordinates
Let \(S=\{(0,0),(1,0),(0,1)\}\), and suppose the intended rule is \(p(x,y)=(1+x+2y)/6\) on \(S\). A draft assignment lists \(p(0,0)=1/6\), \(p(1,0)=3/6\), and \(p(0,1)=2/6\). Decide whether the draft correctly represents the stated model. The method is to evaluate the formula at each ordered pair, not merely to add the draft entries.
Step 1: Evaluate the intended formula. At \((0,0)\), it gives \((1+0+0)/6=1/6\). At \((1,0)\), it gives \((1+1+0)/6=2/6=1/3\). At \((0,1)\), it gives \((1+0+2)/6=3/6=1/2\).
Step 2: Compare with the draft. The draft assigns \(3/6\) to \((1,0)\) and \(2/6\) to \((0,1)\), reversing the two values required by the formula. Thus the draft does not correctly represent the specified model.
Step 3: Verify what normalization can and cannot establish. The draft entries are nonnegative and sum to \(1/6+3/6+2/6=6/6=1\). So, considered on its own, the draft is a valid probability assignment on these three pairs. However, it is not the assignment produced by the stated rule.
Check: The formula itself also gives a total of \(1/6+2/6+3/6=1\), so the intended model is valid. The common mistake is to stop after checking the draft total and conclude that it matches the rule. Correctness requires checking the probability attached to each ordered pair.
Countably Infinite Supports
For an infinite support, the same conditions apply, but normalization uses a convergent series. A useful strategy is to recognize a familiar series when possible, while still checking that the summation covers every supported pair exactly once. If a support is described by two indices, each pair must appear once in the double sum; an omitted row or duplicated pair changes the total.
Worked Example: Normalizing a Joint Rule on an Infinite Support
Suppose \(X\) takes values \(0,1,2,\ldots\), \(Y\) takes values \(0,1\), and every pair in \(S=\{(x,y):x=0,1,2,\ldots;\ y\in\{0,1\}\}\) is allowed. Let \(a_0=1\), \(a_1=2\), and propose \(p(x,y)=c(1/2)^x a_y\). Find \(c\), verify validity, and calculate \(p(2,1)\). The method is to sum the weights over both coordinates and use the geometric series.
Step 1: Check nonnegativity. For every supported pair, \(c(1/2)^x a_y\geq0\) if \(c\geq0\), because \((1/2)^x>0\) and \(a_0,a_1\) are positive.
Step 2: Sum over the support. The sum of the \(y\)-weights is \(a_0+a_1=1+2=3\). The geometric series over \(x\) is \(\sum_{x=0}^{\infty}(1/2)^x=1/(1-1/2)=2\). Therefore the total weight is \(3\cdot2=6\), and the total probability is \(6c\).
Step 3: Normalize. Requiring \(6c=1\) gives \(c=1/6\), which is nonnegative. Thus the rule is nonnegative throughout the support and sums to one.
Step 4: Evaluate the requested pair. At \((2,1)\), \(x=2\), \(y=1\), and \(a_1=2\). Hence \(p(2,1)=(1/6)(1/2)^2(2)=(1/6)(1/4)(2)=2/24=1/12=0.0833\) to four decimal places.
Check: The infinite geometric series converges, so the total calculation is legitimate; all pair probabilities are nonnegative and their total is one. A common mistake is to sum only over \(x\) and omit the two possible \(y\)-values, producing the wrong normalizing constant.
Common Validity Errors
Several shortcuts can lead to incorrect conclusions. A quick check is useful, but it should not replace checking the support, signs, total, and pair assignments separately.
- Checking only the sum. A total of one does not excuse a negative entry. Verify each supported pair or derive parameter restrictions that guarantee nonnegativity.
- Ignoring the support. A formula specified only for allowed pairs assigns no probability to pairs outside the support. Do not add impossible pairs to the normalization sum.
- Swapping coordinates. In general, \(p(x,y)\) need not equal \(p(y,x)\). Evaluate the rule using the stated order of \(X\) and \(Y\).
- Confusing model validity with model fidelity. A list may define a valid distribution while failing to match the formula or outcome labels it was meant to represent.
- Using a finite-sum argument for an infinite support. For infinitely many pairs, verify convergence and evaluate the full series, not just a finite initial portion.
These checks establish whether the proposed pair probabilities form the stated joint distribution. They do not by themselves determine the separate distributions of \(X\) or \(Y\), or whether \(X\) and \(Y\) are independent; those are distinct questions. A clear validity check ends with the support, nonnegativity, total, and any required confirmation of the ordered-pair assignments.
Check Your Understanding
For each problem, check the stated support and apply the validity conditions. Answers are included for self-checking.
- On \(S=\{(0,0),(0,1),(1,0)\}\), let \(p(x,y)=c(1+x+2y)\). Find \(c\) and \(p(1,0)\). (Answer: \(c=1/6\), \(p(1,0)=1/3\).)
- A distribution assigns probabilities \(0.15, 0.20, 0.25,\) and \(0.40\) to four distinct supported pairs and zero elsewhere. Is this a valid probability assignment? (Answer: Yes; all entries are nonnegative and their sum is \(1.0000\).)
- Let \(x=0,1,2,\ldots\), \(y\in\{0,1\}\), and \(p(x,y)=c(1/2)^x\) on every supported pair. Find \(c\) and \(p(3,1)\). (Answer: \(c=1/4\), \(p(3,1)=1/32=0.03125\).)
- On support \(\{(0,0),(0,1),(1,0),(1,1)\}\), suppose the assigned probabilities are \(\theta,1/5,1/2,\) and \(3/10-\theta\), respectively. Find the full range of \(\theta\) for which the assignments are valid. (Answer: \(0\leq\theta\leq3/10\).)
- On \(S=\{(0,0),(1,0),(0,1),(1,1)\}\), a proposed rule is \(p(x,y)=(x+y)/6\). Is it valid? If not, give a normalizing constant \(c\) for the corrected rule \(p(x,y)=c(x+y)\). (Answer: No; the proposed total is \(4/6=2/3\). The corrected constant is \(c=1/4\), with zero probability at \((0,0)\).)