The Multiplication Rule

Suppose a factory ships components through three inspection stations. Each station rejects a fraction of what it receives. You want the probability that a single component clears all three. You cannot simply multiply the three individual pass-rates unless the stations act independently. In reality, the second station's pass-rate depends on what already survived the first. How do you compute the probability of a sequence of dependent events without assuming independence at all? This is exactly the gap that the Multiplication Rule fills. It does so with no extra assumptions beyond the definition of Conditional Probability itself. The Multiplication Rule — also called the chain rule for probability — expresses the probability of an intersection of events. It does so as a product of successively conditioned probabilities. It is the algebraic engine that lets us build complicated joint probabilities. We start from simple, one-step-at-a-time conditional statements: first this happens, then that happens given the first, then a third thing happens given both previous ones, and so on. This "build sequentially" viewpoint mirrors how we naturally think about processes unfolding in time. Examples include draws from an urn without replacement, a sequence of coin flips, or a chain of genetic inheritance. This is precisely why the rule is indispensable in the study of Markov chains, sequential hypothesis testing, and stochastic processes generally. The big picture is that the rule is not an extra axiom bolted onto probability theory. It is a direct algebraic consequence of the definition of conditional probability. It is proved by nothing more than telescoping cancellation. Its power comes from generality: it holds for any finite collection of events, with no independence assumption. It reduces to the familiar "multiply the probabilities" shortcut exactly when independence is present. Understanding this rule carefully — including the precise condition under which it is even well-defined — is a prerequisite. It underlies the Law of Total Probability and the Bayes' Theorem, both of which decompose or invert exactly this kind of sequential product. This article gives the formal statement of the rule for two events and for $n$ events. It includes a complete inductive proof from the definition of conditional probability. It also offers a diagram illustrating the sequential "tree" picture. Two fully worked asymmetric examples follow: a without-replacement card problem and a manufacturing-defect chain. There is a counterexample showing why the positivity hypothesis cannot be dropped. Finally, it discusses where the rule surfaces across probability, statistics, and computer science.

Key takeaways

The Multiplication Rule writes $P(A\cap B) = P(A)P(B\mid A)$ , and generalizes to $n$ events by chaining conditional probabilities.
It is not a new axiom — it follows immediately from the definition $P(B\mid A) = P(A\cap B)/P(A)$ .
The $n$ -event version requires every partial intersection $P(A_1\cap\cdots\cap A_k)$ (for $k<n$ ) to be strictly positive, or the conditional probabilities are undefined.
The rule is exact for dependent events; independence is a special case where every conditional term collapses to an unconditional one.
It underlies sequential sampling without replacement, Markov chain path probabilities, and likelihood factorizations in statistics.
Key idea: to find the probability that a whole sequence of events all occur, multiply the probability of the first by the probability of each subsequent event conditional on all the ones before it.

Where this lives today

The Multiplication Rule is the algebraic backbone of far more than textbook urn problems. In the theory of Markov chains, the probability of an entire finite trajectory $X_0 = x_0, X_1 = x_1, \ldots, X_n = x_n$ is computed by chaining transition probabilities exactly via this rule; the Markov property then simplifies each conditional factor $P(X_k = x_k \mid X_0=x_0,\ldots,X_{k-1}=x_{k-1})$ down to $P(X_k=x_k\mid X_{k-1}=x_{k-1})$ , but the underlying decomposition into a product is the Multiplication Rule itself. In statistics, the likelihood function for a sequence of non-independent observations (such as an autoregressive time series) is written as a product of one-step-ahead conditional densities via the same chain-rule logic, which is the continuous analogue of this discrete rule. In Bayesian networks and probabilistic graphical models in computer science, the joint distribution over many variables is factored into a product of local conditional distributions according to a directed acyclic graph — a direct generalization of the Multiplication Rule to structured dependency graphs, underlying algorithms in machine learning and artificial intelligence. In cryptography, security proofs for protocols often bound the probability of a sequence of adversarial events (guessing a key, forging a signature, and so on) by chaining conditional probabilities across the stages of an attack, applying the same telescoping logic to obtain a union or product bound. In reliability engineering, the probability that a system survives a sequence of dependent stress tests (as in Worked Example 2) is computed exactly this way, forming the basis of sequential quality-control models. And pedagogically, the rule is the direct stepping stone to the Law of Total Probability, which reverses the roles by summing over a partition, and to Bayes' Theorem, which inverts the order of conditioning within a two-event application of this very rule.

Ask the tutor: Can you show me how the Multiplication Rule is used inside the derivation of Bayes' Theorem?

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

Intuition

Think of the event $A_1 \cap A_2 \cap \cdots \cap A_n$ as a path through a decision tree. At the root, the first branch splits according to whether $A_1$ occurs. Along the branch where $A_1$ occurs, a second split occurs according to whether $A_2$ occurs — but now the relevant probability is the conditional probability $P(A_2 \mid A_1)$ , because we are already standing inside the branch where $A_1$ happened. The probability of reaching the leaf corresponding to "all of $A_1, \ldots, A_n$ occur" is the product of the branch probabilities along the unique path from root to that leaf. This is exactly what the Multiplication Rule computes. The key conceptual point is that at each stage, the "sample space" effectively shrinks to the event that has already occurred, and the next probability is measured relative to that shrunken space. There is no new physics here, no new axiom — it is bookkeeping that follows directly from the ratio definition of conditional probability, applied one step at a time and telescoped.

Figure

A three-event probability tree illustrating the chain of conditioning

The Central Theorem

n

n

A_1, \ldots, A_n

Worked examples

Example 1 — Drawing cards without replacement (asymmetric, three-stage chain). A standard deck has 52 cards, 13 of which are hearts. Three cards are drawn sequentially without replacement, and not returned to the deck. Find the probability that the first card is a heart, the second card is a spade, and the third card is also a heart. Let $A_1$ = "first card is a heart," $A_2$ = "second card is a spade," $A_3$ = "third card is a heart." We want $P(A_1 \cap A_2 \cap A_3)$ . By the Multiplication Rule,

P(A_1 \cap A_2 \cap A_3) = P(A_1)\, P(A_2 \mid A_1)\, P(A_3 \mid A_1 \cap A_2).

$P(A_1) = \dfrac{13}{52}$ , since 13 of 52 cards are hearts.
$P(A_2 \mid A_1)$ : given the first card was a heart (and removed), 51 cards remain, of which 13 are spades (untouched, since spades weren't affected by removing a heart). So $P(A_2 \mid A_1) = \dfrac{13}{51}$ .
$P(A_3 \mid A_1 \cap A_2)$ : given the first card was a heart and the second was a spade, 50 cards remain. One heart has been removed, so 12 hearts remain among the 50. So $P(A_3 \mid A_1 \cap A_2) = \dfrac{12}{50}$ .

Therefore,

P(A_1 \cap A_2 \cap A_3) = \frac{13}{52}\cdot \frac{13}{51}\cdot\frac{12}{50}.

Compute step by step:

\frac{13}{52} = \frac{1}{4}, \qquad \frac{1}{4}\cdot\frac{13}{51} = \frac{13}{204}, \qquad \frac{13}{204}\cdot\frac{12}{50} = \frac{13 \cdot 12}{204 \cdot 50} = \frac{156}{10200}.

Simplify: $\gcd(156, 10200) = 12$ , giving $156/12 = 13$ and $10200/12 = 850$ . So

P(A_1\cap A_2\cap A_3) = \frac{13}{850} \approx 0.015294.

Note the asymmetry deliberately built in: the three events are of different suits ( $A_1, A_3$ hearts, $A_2$ spades), so the removal of a heart on draw 1 does not diminish the spade count on draw 2, but does diminish the heart count relevant to draw 3 — a subtlety the rule tracks automatically because each conditioning event is the actual accumulated history, not a symmetric placeholder.

Example 2 — A three-station manufacturing inspection chain. A component passes through three sequential inspection stations, $S_1, S_2, S_3$ . Let $A_i$ denote the event "the component passes station $i$ ." Historical data give:

$P(A_1) = 0.90$ (10% of components fail the first, coarsest inspection).
$P(A_2 \mid A_1) = 0.95$ (given it passed station 1, it passes the finer station 2 with probability 0.95).
$P(A_3 \mid A_1 \cap A_2) = 0.80$ (station 3 is the strictest test, applied only to components that already passed both earlier stations, and even among these survivors only 80% pass).

These are precisely the conditional inputs the Multiplication Rule requires — note they are not symmetric (0.90, 0.95, 0.80 are all different), which is realistic since later stations are typically stricter or test different failure modes with different base rates. We want the probability that a randomly selected component clears the entire inspection chain, $P(A_1 \cap A_2 \cap A_3)$ . By the theorem,

P(A_1 \cap A_2 \cap A_3) = P(A_1)\, P(A_2\mid A_1)\, P(A_3 \mid A_1\cap A_2) = 0.90 \times 0.95 \times 0.80.

Compute: $0.90 \times 0.95 = 0.855$ , and $0.855 \times 0.80 = 0.684$ . So $P(A_1\cap A_2 \cap A_3) = 0.684$ : about 68.4% of components clear all three stations. Note this is far below the naive (and wrong, if we don't check independence) guess of "roughly 90%," because each stage compounds a nontrivial rejection rate on the conditional survivors — exactly the sequential shrinking-sample-space effect the rule captures. As a sanity check, if a manager wrongly assumed independence and wrongly used the unconditional station pass-rates $P(A_1), P(A_2), P(A_3)$ estimated from the full population — say the marginal pass rate of station 3 in isolation happens to be $0.83$ rather than $0.80$ — multiplying $0.90 \times 0.95 \times 0.83 = 0.70965$ would silently give the wrong answer, because $P(A_3)$ (marginal) and $P(A_3\mid A_1\cap A_2)$ (the correct conditional term) need not coincide. This is the entire reason the Multiplication Rule insists on conditioning on the accumulated history rather than using marginal probabilities.

P(A_1 \cap \cdots \cap A_{n-1}) > 0

Scope, edge cases & common traps

Situation	What happens
$P(A_1 \cap \cdots \cap A_{n-1}) > 0$	Rule applies exactly as stated; every conditional factor well defined.
Some intersection $P(A_1\cap\cdots\cap A_k)=0$ for $k<n$	A later conditional probability is $0/0$ : undefined. Rule cannot be applied at that step.
Events mutually independent	Rule reduces to plain product $\prod P(A_i)$ (Independence of Events).
Order of conditioning events relabeled	The product's value is unchanged (it always equals $P(\bigcap A_i)$ ), but the individual factors change — conditioning on $A_1\cap A_2$ then $A_3$ gives different intermediate numbers than conditioning on $A_2\cap A_3$ then $A_1$ , even though the final product agrees.
Sampling without replacement	Canonical use case: each draw changes the composition of what remains, so successive conditional probabilities differ from the unconditional ones.
Sampling with replacement, independent trials	Conditional probabilities collapse to marginal ones; rule reduces to the independence corollary.

Common trapThe classic error

The most common mistake is using marginal (unconditional) probabilities $P(A_2), P(A_3), \ldots$ in place of the required conditional probabilities $P(A_2\mid A_1), P(A_3\mid A_1\cap A_2), \ldots$ — silently assuming independence without checking it. This is exactly the error illustrated in Worked Example 2's manager scenario: substituting $P(A_3) = 0.83$ (marginal) for $P(A_3\mid A_1\cap A_2)=0.80$ (the correct conditional) gives a numerically plausible but wrong answer, because nothing in the problem guaranteed the two coincide. Always verify which quantity is actually given before multiplying, and never chain marginal probabilities unless independence has been separately established.

The mathematician behind it

Origins · 1 minAndrey Nikolaevich Kolmogorov and how this was discovered

Andrey Nikolaevich Kolmogorov

1903–1987 · Soviet mathematician, architect of modern axiomatic probability theory

Kolmogorov's 1933 monograph Grundbegriffe der Wahrscheinlichkeitsrechnung placed conditional probability and its consequences, including the multiplication rule for sequential events, on a rigorous measure-theoretic foundation. Before Kolmogorov, probability calculations — including chained conditional arguments used implicitly since Bayes and Laplace — lacked a universally agreed axiomatic basis; Kolmogorov showed that conditional probability, and hence the Multiplication Rule, follows as a theorem once probability is defined as a measure satisfying countable additivity on a $\sigma$ -algebra of events.

History

- 1654 — Pascal and Fermat's correspondence on the problem of points implicitly uses sequential conditional reasoning to resolve a wager, an early precursor of chained probability arguments.

- 1763 — Thomas Bayes' posthumously published essay formalizes conditional probability reasoning that the Multiplication Rule makes precise and reversible.

- 1812 — Pierre-Simon Laplace's Théorie analytique des probabilités systematizes conditional and compound probability calculations, including sequential drawing-without-replacement problems.

- 1933 — Andrey Kolmogorov's axiomatization derives conditional probability, and consequently the Multiplication Rule, rigorously from measure theory.

- 1950s onward — The rule becomes the standard justification for factorizing joint distributions in Markov chains, hidden Markov models, and Bayesian networks.

Is the Multiplication Rule a new axiom of probability?

No. It is a direct algebraic consequence of the definition of conditional probability, $P(B\mid A) = P(A\cap B)/P(A)$ for $P(A)>0$ . Multiplying both sides by $P(A)$ and using that $P(A)>0$ lets you cancel the denominator, immediately giving $P(A\cap B) = P(A)P(B\mid A)$ . The $n$ -event version follows by induction, applying this two-event identity repeatedly and telescoping. No additional postulate about probability is needed beyond Kolmogorov's axioms and the ratio definition of conditioning.

How is the Multiplication Rule different from the Law of Total Probability?

The Multiplication Rule computes the probability of an intersection (all listed events happening together) by chaining conditional probabilities forward through a sequence. The Law of Total Probability, in contrast, computes the probability of a single event by summing its probability across a partition of the sample space, using the Multiplication Rule internally for each piece: $P(B) = \sum_i P(A_i)P(B\mid A_i)$ . So the Multiplication Rule is a tool used inside the Law of Total Probability, not a competing statement.

Does the order in which I condition on the events matter?

The final numerical value $P(A_1\cap\cdots\cap A_n)$ never depends on the order, since intersection is commutative and associative. However, the individual conditional factors in the product do depend on the order you choose to condition in — $P(A_2\mid A_1)$ is generally a different number from $P(A_1\mid A_2)$ , even though both orderings multiply out to the same final answer $P(A_1\cap A_2)$ . Choose whichever order makes the conditional probabilities easiest to state or estimate from context (e.g., chronological order in a sequential experiment).

Can the Multiplication Rule be applied to infinitely many events?

The rule as stated is for finite $n$ . Extending it to a countably infinite sequence $A_1, A_2, \ldots$ requires additional care: one considers $P(\bigcap_{i=1}^n A_i)$ for each finite $n$ and then takes a limit as $n\to\infty$ , using continuity of probability measures along decreasing sequences of events (a consequence of countable additivity). The resulting infinite product $\prod_{k=1}^\infty P(A_k\mid A_1\cap\cdots\cap A_{k-1})$ converges to $P(\bigcap_{i=1}^\infty A_i)$ , but this is a distinct, more advanced result built on top of the finite rule, not a direct restatement of it.

What if some of the events are guaranteed impossible partway through the chain?

If at some stage $P(A_1\cap\cdots\cap A_k)=0$ , then all subsequent conditional probabilities in the chain are of the form $0/0$ and are simply undefined; the Multiplication Rule cannot be mechanically applied past that point. In practice this usually signals that the sequence of events is already known to be impossible: since $A_1\cap\cdots\cap A_n \subseteq A_1\cap\cdots\cap A_k$ , if the latter has probability zero, so does the former, and you can directly conclude $P(A_1\cap\cdots\cap A_n)=0$ by monotonicity, without needing the conditional factorization at all.

How does this rule relate to sampling without replacement?

Sampling without replacement is the paradigm example where conditional probabilities genuinely differ from marginal ones at every stage, because removing an item changes the composition of what remains. This is exactly why the Multiplication Rule (rather than a naive product of unconditional probabilities) is needed: in Worked Example 1, the probability a third card is a heart depends on how many hearts remain, which depends on the first draw. Sampling with replacement, by contrast, keeps the population unchanged, so all conditional probabilities collapse to marginal ones and the independence corollary applies.

Can I use the Multiplication Rule if I only know unconditional probabilities of each event?

Not directly — the rule fundamentally requires the conditional probabilities $P(A_k\mid A_1\cap\cdots\cap A_{k-1})$ , not the marginal probabilities $P(A_k)$ . If you only have marginal probabilities and no independence assumption, you cannot recover the joint probability $P(\bigcap A_i)$ from them alone; additional information about dependence structure (correlations, a model, or conditional probabilities themselves) is required. Substituting marginals for conditionals without justification is the classic error described in this article's warning box.