Conditional Probability

Suppose you roll a fair six-sided die. A friend tells you the outcome is even. How should this partial information change your assessment? Specifically, how should it change the chance that the outcome is a 2? Before hearing the clue, you would have said the probability is 1/6. Afterward, the number 2 is one of only three equally likely possibilities. So the answer should become 1/3. Conditional probability is the precise mathematical machinery for this process. It turns informal reasoning — updating beliefs in light of evidence — into a rigorous, computable quantity. This idea matters because almost no real inference happens in a vacuum. Medical tests are interpreted conditional on symptoms. Weather forecasts are conditional on atmospheric data. Machine-learning classifiers output probabilities conditional on observed features. Conditional probability is the bridge between raw uncertainty (encoded by the Axioms of Probability) and the everyday act of learning from data. Without it, probability theory could describe a single static state of ignorance. But it could never model the dynamic process of updating that ignorance. The big picture is that conditioning is nothing more — and nothing less — than restricting the sample space. When we learn that event $B$ has occurred, every outcome outside $B$ becomes impossible. So the new universe of possibilities is $B$ itself. The conditional probability of $A$ given $B$ measures what fraction of this smaller universe is occupied by $A$ . It turns out that this fraction is computed by renormalizing the probability measure so that $B$ has probability 1. This article gives the formal definition of $P(A\mid B)$ . It proves that conditioning on a fixed event $B$ with $P(B) > 0$ again produces a legitimate probability measure. This measure satisfies all three axioms. The article also works through examples that expose common pitfalls. It closes by connecting conditional probability to the ideas it directly enables: the Multiplication Rule and the Independence of Events.

Key takeaways

Conditional probability $P(A\mid B)$ measures the chance of $A$ once we know $B$ has occurred.
It is defined as a ratio, $P(A\mid B) = P(A\cap B)/P(B)$ , valid whenever $P(B) > 0$ .
Conditioning restricts the sample space to $B$ and renormalizes probabilities so that $P(B\mid B)=1$ .
The map $A \mapsto P(A\mid B)$ is itself a genuine probability measure, satisfying all the Axioms of Probability.
Conditional probability is not symmetric: $P(A\mid B) \neq P(B\mid A)$ in general — confusing the two is the "prosecutor's fallacy."
Key idea: conditioning on $B$ means recomputing probabilities inside the shrunken world where $B$ is certain.

Where this lives today

Conditional probability is the computational core of Bayesian statistics, where prior beliefs $P(\theta)$ are updated to posteriors $P(\theta \mid \text{data})$ via Bayes' Theorem, itself built directly from the definition given here. In medical diagnostics, the predictive value of a test — the probability a patient has a disease given a positive test result — is a conditional probability computed from the test's sensitivity, specificity, and disease prevalence. In machine learning, classifiers such as naive Bayes and logistic regression directly model or approximate $P(\text{label} \mid \text{features})$ . In cryptography and information theory, conditional entropy $H(X\mid Y)$ , built from conditional probability distributions, quantifies how much uncertainty about $X$ remains after observing $Y$ , underlying results like the security guarantees of one-time pads and the capacity of noisy channels (Shannon's coding theorems). In finance, conditional probabilities under a filtration of information drive martingale-based option pricing. In genetics, conditional probabilities compute the chance of inheriting a trait given known parental genotypes, and underlie recurrence-risk counseling. All of these rest, at bottom, on the same restrict-and-renormalize idea proved rigorously above, and all of them build toward the next natural questions: how to combine several conditionings via the Multiplication Rule, and when conditioning has no effect at all, which is precisely the definition of Independence of Events.

Ask the tutor: How does conditional probability generalize when we condition on a continuous random variable instead of a fixed event?

(\Omega, \mathcal{F}, P)

Intuition — restricting the sample space

The definition above is a formula, but the content of conditional probability is geometric. Picture the sample space $\Omega$ as a region whose total area represents probability 1, with events as sub-regions whose areas represent their probabilities. Learning that $B$ has occurred is like being told, "forget everything outside this region — the true outcome is definitely somewhere inside $B$ ." The new sample space is $B$ itself, and to make $P(B\mid B)=1$ hold we must rescale every area by the factor $1/P(B)$ . Under this rescaling, an event $A$ contributes only the part of it that overlaps $B$ , namely $A \cap B$ , and that overlap's area must be rescaled by the same factor $1/P(B)$ to remain a legitimate probability. This is exactly the ratio $P(A\cap B)/P(B)$ . Everything outside $B$ is simply erased from consideration — it does not get redistributed or reinterpreted, it is discarded because it is now known to be impossible. This picture immediately explains several facts that otherwise look like arbitrary algebra. It explains why $P(B \mid B) = 1$ (the whole new universe has full measure), why $P(\varnothing \mid B) = 0$ (the impossible event stays impossible), and why $P(A \mid B)$ can be much larger than $P(A)$ when $A$ and $B$ overlap heavily relative to $B$ 's size — conditioning "zooms in" and can make rare events look common if they are common within the zoomed-in region.

Figure

Restricting the sample space: conditioning on B zooms in and rescales

Central theorem

(\Omega, \mathcal{F}, P)

Q

A, B \in \mathcal{F}

Q(\cdot) = P(\cdot \mid B)

Worked examples

Example 1: Cards with unequal, non-symmetric events. A single card is drawn uniformly at random from a standard 52-card deck. Let

A = \{\text{the card is a King}\}, \qquad B = \{\text{the card is a face card (Jack, Queen, or King)}\}.

There are 4 Kings, 12 face cards total (4 Jacks, 4 Queens, 4 Kings), so $P(A) = 4/52 = 1/13$ and $P(B) = 12/52 = 3/13$ . Since every King is a face card, $A \cap B = A$ , so $P(A \cap B) = P(A) = 4/52$ . By definition,

P(A \mid B) = \frac{P(A \cap B)}{P(B)} = \frac{4/52}{12/52} = \frac{4}{12} = \frac{1}{3}.

This matches direct counting: among the 12 face cards, exactly 4 are Kings, so the "restricted sample space" $B$ makes $A$ occur with relative frequency $4/12 = 1/3$ — precisely illustrating the "restrict and renormalize" picture, with no symmetry between $A$ and $B$ 's sizes to hide any subtlety. Now compute the reverse conditional probability to emphasize the lack of symmetry:

P(B \mid A) = \frac{P(A \cap B)}{P(A)} = \frac{4/52}{4/52} = 1,

since every King is automatically a face card. So $P(A\mid B) = 1/3$ while $P(B \mid A) = 1$ — a stark numerical demonstration that $P(A\mid B) \neq P(B \mid A)$ in general.

Example 2: A biased die and a non-uniform event, verified two ways. Consider a biased six-sided die with outcomes $\Omega = \{1,2,3,4,5,6\}$ and probabilities

P(1) = 0.05,\; P(2) = 0.10,\; P(3) = 0.15,\; P(4) = 0.20,\; P(5) = 0.20,\; P(6) = 0.30.

(These sum to $0.05+0.10+0.15+0.20+0.20+0.30 = 1.00$ , confirming $P$ is a valid probability mass assignment consistent with Axiom 2.) Let $B = \{2,4,6\}$ (even outcomes) and $A = \{3,4,5,6\}$ (outcome at least 3). First,

P(B) = P(2)+P(4)+P(6) = 0.10+0.20+0.30 = 0.60.

Next, $A \cap B = \{4,6\}$ (the even numbers that are also at least 3), so

P(A \cap B) = P(4) + P(6) = 0.20 + 0.30 = 0.50.

Therefore

P(A \mid B) = \frac{P(A\cap B)}{P(B)} = \frac{0.50}{0.60} = \frac{5}{6} \approx 0.8333.

We can verify this using the restricted-sample-space picture directly, by computing $Q(\{k\}) = P(\{k\}\mid B)$ for each $k \in B$ and checking they form a valid probability distribution on $B$ :

Q(2) = \frac{0.10}{0.60} = \frac{1}{6}, \qquad Q(4) = \frac{0.20}{0.60} = \frac{1}{3}, \qquad Q(6) = \frac{0.30}{0.60} = \frac{1}{2}.

These sum to $\tfrac16+\tfrac13+\tfrac12 = \tfrac{1}{6}+\tfrac{2}{6}+\tfrac{3}{6} = 1$ , confirming $Q(B)=1$ as guaranteed by Central Theorem above. Now $A\cap B = \{4,6\}$ , so summing directly,

Q(A) = Q(4) + Q(6) = \frac{1}{3} + \frac{1}{2} = \frac{2}{6}+\frac{3}{6} = \frac{5}{6},

which matches the ratio computation exactly, cross-checking the arithmetic through two independent routes.

P(B) > 0

Scope, edge cases & common traps

Situation	What holds	What can go wrong
$P(B) > 0$	$P(\cdot \mid B)$ is a genuine probability measure with $P(B\mid B)=1$	None — the theory is fully rigorous
$P(B) = 0$	Ratio formula undefined	Students plug in $0/0$ ; must use conditional expectation / limiting arguments instead
$A \subseteq B$	$P(A\mid B) = P(A)/P(B) \ge P(A)$	Mistakenly assuming $P(A\mid B) \le P(A)$ always — false in general
$A, B$ mutually exclusive, $A\ne\varnothing$	$P(A\mid B) = 0$	Confusing "mutually exclusive" with "independent"
Reversing the conditioning	$P(A\mid B) \ne P(B\mid A)$ in general	The prosecutor's fallacy: equating $P(\text{evidence}\mid \text{innocent})$ with $P(\text{innocent}\mid\text{evidence})$

The mathematician behind it

Origins · 1 minAndrey Nikolaevich Kolmogorov and how this was discovered

Andrey Nikolaevich Kolmogorov

1903–1987 · Soviet Union, 20th century mathematics

Kolmogorov did not invent conditional reasoning about chance — that intuition long predates him, appearing informally in the work of Bayes, Laplace, and others in the 18th and 19th centuries. His decisive contribution was to place all of probability, including conditional probability, on an unimpeachably rigorous measure-theoretic foundation. In his 1933 monograph Grundbegriffe der Wahrscheinlichkeitsrechnung ("Foundations of the Theory of Probability"), Kolmogorov defined probability spaces via the now-standard axioms and showed that conditional probability, for events of positive probability, is simply the ratio $P(A\cap B)/P(B)$ — but he went further, using the Radon–Nikodym theorem to define conditional expectation with respect to an arbitrary sub- $\sigma$ -algebra, which subsumes and rigorously extends conditioning even to events of probability zero, resolving ambiguities like the one in the counterexample above.

History

- 1654 — Pascal and Fermat correspond on gambling problems, implicitly reasoning about probabilities updated by partial information.

- 1763 — Bayes' essay (published posthumously) formalizes inverse probability, an early conditional-probability calculation.

- 1812 — Laplace's Théorie analytique des probabilités develops conditional and inverse probability systematically.

- 1933 — Kolmogorov's Grundbegriffe axiomatizes probability spaces and defines

P(A\mid B) = P(A\cap B)/P(B)

rigorously.

- 1933–present — Conditional expectation via Radon–Nikodym becomes the modern tool for conditioning on zero-probability events and continuous random variables.

What is the intuitive meaning of P(A|B)?

It is the probability of $A$ once you know for certain that $B$ has occurred. Concretely, imagine repeating the underlying random experiment many times and discarding every trial in which $B$ fails to happen; $P(A\mid B)$ is the long-run fraction of the retained trials in which $A$ also happens. This matches the formal ratio $P(A\cap B)/P(B)$ : the numerator counts trials where both occur, the denominator counts trials where $B$ occurs, and the ratio is exactly that fraction. It is fundamentally a statement about a restricted universe of outcomes, not a claim about causation.

Why does the formula require dividing by P(B) rather than something else?

Dividing by $P(B)$ is exactly what forces $P(B\mid B) = 1$ , which is the defining feature of "restricting the sample space to $B$ ": once you know $B$ happened, $B$ itself must have new probability 1. Any other normalizing constant would violate Axiom 2 ( $P(\Omega)=1$ ) applied to the new measure, since $Q(\Omega)$ must equal 1 and $\Omega \cap B = B$ . The proof of the Central Theorem shows this normalization is also exactly what is needed to preserve countable additivity, so $P(B)$ is not an arbitrary choice — it is the unique constant making $Q$ a probability measure with $Q(B)=1$ .

Is P(A|B) the same as P(B|A)?

No, and conflating them is one of the most common statistical errors, sometimes called the prosecutor's fallacy. In the card example above, $P(\text{King}\mid\text{face card}) = 1/3$ but $P(\text{face card}\mid\text{King}) = 1$ — clearly different numbers. The correct relationship between the two is given by Bayes' Theorem, which states $P(A\mid B) = P(B\mid A)P(A)/P(B)$ ; the two conditional probabilities coincide only in the special case $P(A) = P(B)$ .

Can conditional probability ever be undefined even though B is a perfectly reasonable event?

Yes — whenever $P(B) = 0$ . This happens routinely with continuous random variables: for example, conditioning on "the temperature is exactly 20.0000...°C" involves an event of probability zero under a continuous distribution. The elementary ratio definition simply does not apply there; instead one must use conditional density functions or the measure-theoretic machinery of conditional expectation (built via the Radon–Nikodym theorem), and even then care is needed, since naive limiting procedures can give paradoxical, choice-dependent answers (the Borel–Kolmogorov paradox), as illustrated in the counterexample above.

Does conditioning ever make an event's probability decrease?

Yes, easily — conditioning can raise or lower a probability depending on how $A$ and $B$ relate. If $A$ and $B$ are disjoint, $P(A\mid B) = 0$ , a decrease from any positive $P(A)$ . If $A \supseteq B$ so that $B$ implies $A$ , then $P(A\mid B)=1$ , generally an increase. There is no universal direction; the only guarantee from the axioms is that $P(A\mid B)$ is itself a valid probability, i.e., a number in $[0,1]$ , and equals $1$ exactly when $B\setminus A$ has probability zero.

How does conditional probability relate to independence?

Two events $A$ and $B$ (with $P(B) > 0$ ) are defined to be independent precisely when conditioning on $B$ does not change the probability of $A$ , i.e., $P(A\mid B) = P(A)$ . Substituting the definition of conditional probability, this is equivalent to $P(A\cap B) = P(A)P(B)$ , which is the standard product-form definition of Independence of Events. So independence is the special, non-generic case where the "restrict and renormalize" operation happens to leave every probability of interest untouched.

What is the connection to the multiplication rule?

Rearranging the definition $P(A\mid B) = P(A\cap B)/P(B)$ algebraically (multiplying both sides by $P(B)$ , valid since $P(B)>0$ ) gives $P(A \cap B) = P(B)\,P(A\mid B)$ . This is exactly the statement of the Multiplication Rule, which extends by induction to chains of several events, $P(A_1\cap A_2\cap\cdots\cap A_n) = P(A_1)P(A_2\mid A_1)P(A_3\mid A_1\cap A_2)\cdots$ , and underlies tree-diagram calculations throughout applied probability.