Common Errors in Conditioning and Expectation

Three Checks Before Conditioning

Conditioning changes which probabilities or expectations are being calculated. Many errors arise from treating the conditioned-on information as if it were irrelevant, or from applying a rule without checking its assumptions. Before using a conditional probability or expectation, ask three questions: does the conditioning event have positive probability; am I using a conditional distribution or a marginal distribution; and is a factor being taken outside a conditional expectation actually known under the conditioning?

Throughout, random variables are defined on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). For an event \(B\in\mathcal{F}\) with \(\mathbb{P}(B)>0\), conditioning on \(B\) means using the probability measure \(\mathbb{P}(\,\cdot\mid B)\), where \(\mathbb{P}(C\mid B)=\mathbb{P}(C\cap B)/\mathbb{P}(B)\). For an integrable random variable \(X\), the corresponding conditional expectation is \(\mathbb{E}[X\mid B]=\mathbb{E}[X\mathbf{1}_B]/\mathbb{P}(B)\). The positive-probability requirement is essential: if \(\mathbb{P}(B)=0\), this ratio is undefined.

1. Conditioning on an Event of Probability Zero

A conditional probability given an event \(B\) is defined by a ratio whose denominator is \(\mathbb{P}(B)\). If that probability is zero, the elementary event-conditioning formula does not define a value. A common error is to insert a zero-probability event into the formula anyway, or to assume that there is one uniquely determined answer.

For conditioning on the value of a continuous random variable, conditional distributions can be represented by a family of probability distributions indexed by the conditioning value. Such a family is called a regular conditional distribution. Its defining integral identities determine its values only for almost every conditioning value under the distribution of the conditioning variable. A value at a particular point with probability zero can therefore depend on the chosen version.

Proof of non-uniqueness. A regular conditional distribution is required to satisfy its defining integral identities. The two families above differ only at \(x=0\), a set with \(X\)-probability zero. For each measurable \(A\), an integral with respect to the distribution of \(X\) is unchanged by changing an integrand on a set of measure zero. Thus both families satisfy the same defining identities while assigning different distributions at \(x=0\). \(\square\)

2. Conditional Distributions Are Not Marginals

The marginal distribution of \(Y\) describes \(Y\) without specifying another variable. The conditional distribution of \(Y\) given \(X=x\) describes \(Y\) after the value \(X=x\) is supplied. These distributions need not agree. To recover the marginal distribution from conditional distributions, average them over the distribution of \(X\); do not replace every conditional distribution by the marginal.

Proof. Since the \(A_i\) form a partition, \(\sum_{i=1}^{m}\mathbf{1}_{A_i}=1\) everywhere. Therefore \(X=\sum_{i=1}^{m}X\mathbf{1}_{A_i}\). Taking expectations and using the definition of conditioning on each positive-probability cell gives \(\mathbb{E}[X]=\sum_i\mathbb{E}[X\mathbf{1}_{A_i}]=\sum_i\mathbb{P}(A_i)\mathbb{E}[X\mid A_i]\). Taking \(X=\mathbf{1}_C\) proves the probability identity. If \(\mathbb{P}(A_i)=0\), then \(\mathbb{E}[|X|\mathbf{1}_{A_i}]=0\), so that cell contributes zero and its conditional expectation is not required. \(\square\)

3. Taking a Factor Outside a Conditional Expectation

The rule for moving a factor outside a conditional expectation depends on what is known under the conditioning. If \(H\) is measurable with respect to \(\mathcal{G}\), then \(H\) is known when conditioning on \(\mathcal{G}\). For integrable \(X\) and bounded \(\mathcal{G}\)-measurable \(H\), the standard taking-out-what-is-known property is \(\mathbb{E}[HX\mid\mathcal{G}]=H\mathbb{E}[X\mid\mathcal{G}]\) almost surely. Boundedness ensures \(HX\) is integrable. The measurability requirement is not optional: a factor that is not known under \(\mathcal{G}\) generally cannot be taken out.

Proof. Write \(H=\sum_{r=1}^{q}h_r\mathbf{1}_{C_r}\), where each \(C_r\in\mathcal{G}\) and the \(C_r\) form a finite partition. Set \(Z=\mathbb{E}[X\mid\mathcal{G}]\). The random variable \(HZ\) is \(\mathcal{G}\)-measurable. For every \(C\in\mathcal{G}\),

The middle equality follows from the defining property of \(Z\), because \(C\cap C_r\in\mathcal{G}\). Thus \(HZ\) satisfies the defining property of the conditional expectation of \(HX\), so the two are equal almost surely. \(\square\)

A Practical Error-Checking Routine

These errors can be caught by checking definitions before simplifying. For event conditioning, write down the denominator and verify it is positive. For conditioning on a random variable’s value, distinguish values with positive point probability from values in a continuous range: a conditional density or kernel is specified only up to almost-everywhere changes unless an additional convention selects a version. For a conditional distribution, identify which variable is conditioned on and calculate the relevant conditional probabilities rather than copying marginal probabilities. For a factor outside a conditional expectation, verify that it is measurable with respect to the conditioning sigma-field and check the integrability assumptions.

A final consistency check is to average conditional quantities over the conditioning information. On a finite partition, the weights are the probabilities of the partition cells and sum to 1. Conditional expectations can vary across cells even though their weighted average recovers the unconditional expectation. In contrast, taking a factor outside is a pointwise almost-sure identity that requires the factor to be known under the conditioning. These are different rules and should not be substituted for one another.