Conditional Probability

Why we need it

In basic probability, you pick a sample space Ω (all possible outcomes) and assign probabilities to events (subsets of Ω). That’s great when you have no extra information.

But real reasoning almost always comes with information:

• •“Given that it rained, what’s the chance traffic is bad?”
• •“Given that the die roll is even, what’s the chance it’s a 6?”
• •“Given that a test is positive, what’s the chance the patient actually has the disease?”

That phrase “given that …” is exactly what conditional probability captures.

Intuition: shrinking the universe

Think of an event B as a filter. Once you learn B occurred, outcomes not in B are impossible, so you throw them away.

• •Original world: Ω
• •After learning B: your new world is B

Conditional probability asks: inside this new world B, how likely is A?

So P(A|B) is not “A and B” (that’s intersection). It’s “A measured within B.”

Formal definition

If P(B) > 0, the conditional probability of A given B is

P(A|B) = P(A ∩ B) / P(B)

Read it slowly:

• •A ∩ B: outcomes where both A and B happen
• •Divide by P(B): because we’re renormalizing probabilities so that the total probability of the new universe (B) becomes 1

A picture in words

Imagine 100 equally likely outcomes.

• •40 outcomes are in B
• •10 outcomes are in A ∩ B

Then:

• •P(B) = 40/100
• •P(A ∩ B) = 10/100
• •P(A|B) = (10/100) / (40/100) = 10/40 = 1/4

The key idea: you don’t compare A ∩ B to Ω anymore; you compare it to B.

The nonzero condition

The definition requires P(B) > 0.

Why? Because if P(B) = 0, then “given B happened” is conditioning on something that never happens (in your probability model). The fraction P(A ∩ B)/P(B) would divide by zero.

At difficulty level 2, the important takeaway is: only condition on events with positive probability (for basic discrete problems).

Why this matters

Most mistakes with conditional probability come from not fully committing to the new sample space. Learners often keep using the original denominator (Ω) instead of the conditioned denominator (B).

Sample-space rewrite rule

Once you know B happened:

• •Possible outcomes become: B
• •Your question becomes: “What fraction of B also lies in A?”

In equally likely discrete settings:

P(A|B) = |A ∩ B| / |B|

This is the same as the earlier formula, just using counts instead of probabilities.

Example: die roll with a condition

Let Ω = {1,2,3,4,5,6}.

• •A = “roll is 6” = {6}
• •B = “roll is even” = {2,4,6}

After conditioning on B, the new universe is {2,4,6}.

Within B:

• •Outcomes that satisfy A are just {6}

So:

P(A|B) = 1/3

Notice how different this is from P(A) = 1/6. The evidence “even” made 6 more plausible.

Two complementary views (and when each is useful)

Conditional probability can be approached in two equivalent ways:

Relationship to complements

If you are inside B, the complement of A becomes “not A, but still inside B.”

P(Aᶜ|B) = 1 − P(A|B)

This is often an easy way to compute a conditional probability when the direct event is awkward.

A common mental model: “re-normalize”

Conditioning on B does two things:

1) Deletes outcomes not in B

2) Scales the remaining probabilities so they sum to 1

If the outcomes in B were originally equally likely, they remain equally likely relative to each other after conditioning.

If outcomes in B were not equally likely, you can still condition, but you must keep their original weights and then re-normalize.

Mini-derivation: why the formula makes sense

Suppose we want a new probability measure P(·|B) that lives on the restricted universe B.

We want:

• •P(B|B) = 1
• •P(A|B) should be proportional to P(A ∩ B) (because only overlap with B matters)

So we set

P(A|B) = c · P(A ∩ B)

Choose c so that P(B|B) = 1:

1 = P(B|B) = c · P(B ∩ B) = c · P(B)

So c = 1 / P(B).

Therefore:

P(A|B) = P(A ∩ B) / P(B)

This is the cleanest justification for the definition: it’s the unique way to “renormalize” probabilities inside B.

Why we need a second mechanic

Sometimes you don’t want P(A|B). Instead, you want the probability that both events happen, P(A ∩ B). Conditional probability gives a direct bridge between these.

Starting from the definition:

P(A|B) = P(A ∩ B) / P(B)

Multiply both sides by P(B):

P(A ∩ B) = P(A|B) · P(B)

This is called the multiplication rule.

Two equivalent forms

Be careful with order: both are true (when denominators are nonzero).

P(A ∩ B) = P(A|B)P(B)

P(A ∩ B) = P(B|A)P(A)

They describe the same intersection, just conditioning in different directions.

Why this is powerful

The intersection P(A ∩ B) is often hard to estimate directly, but conditional probabilities are natural in real situations.

Example narrative:

• •P(B) = probability it rains
• •P(A|B) = probability of traffic given rain

Then P(A ∩ B) = probability it rains and traffic is bad.

Chaining more than two events

Conditional probability lets you build probabilities step-by-step. For three events A, B, C with appropriate nonzero probabilities:

P(A ∩ B ∩ C) = P(A|B ∩ C) · P(B|C) · P(C)

Interpretation: start from C, then within C consider B, then within (B ∩ C) consider A.

At this node’s level, you don’t need to memorize the general chain rule, but it’s useful to see that conditional probability is the building block for multi-step reasoning.

A note on symmetry (and why “given” is not symmetric)

Intersection is symmetric:

A ∩ B = B ∩ A

But conditional probability generally is not:

P(A|B) ≠ P(B|A)

Example:

• •A = “number is divisible by 2”
• •B = “number is divisible by 6” (in {1,…,12})

If a number is divisible by 6, it must be divisible by 2, so P(A|B) = 1.

But if a number is divisible by 2, it is not necessarily divisible by 6, so P(B|A) < 1.

This non-symmetry is the entire reason Bayes’ Theorem is interesting later: it provides a way to relate the two directions.

Translating English to events

A lot of conditional probability skill is language parsing.

• •“Probability of A given B” → P(A|B)
• •“Probability of A and B” → P(A ∩ B)
• •“Probability of A or B” → P(A ∪ B)
• •“Only look at cases where B happened” → restrict sample space to B

A practical technique: rewrite the question as

“Among outcomes where B is true, what fraction have A true?”

Diagnostic tests (setup for Bayes)

Consider medical testing language:

• •D = person has disease
• •T = test is positive

Two different conditionals:

• •P(T|D): sensitivity (test detects disease when disease is present)
• •P(D|T): probability of disease given a positive test (what you actually care about)

These are not the same. Conditional probability makes that distinction precise.

Even before Bayes’ Theorem, you can see the structure:

• •P(D|T) depends on how common the disease is (base rate P(D))
• •P(T|D) is about the test’s behavior

This node equips you to keep the symbols straight so Bayes later feels like algebra, not magic.

Independence as a special case

Independence will be unlocked soon. Conditional probability is the quickest way to express it.

A and B are independent exactly when

P(A|B) = P(A)

(assuming P(B) > 0)

Interpretation: learning B doesn’t change your belief about A.

Equivalently:

P(A ∩ B) = P(A)P(B)

But conceptually, the conditional form is often more intuitive: no update.

Markov chains preview

A Markov chain is about transitions like

P(Xₜ₊₁ = j | Xₜ = i)

That is literally conditional probability: the next state given the current state.

So this node is foundational: without comfort reading and manipulating P(·|·), transition matrices and “memoryless” properties will feel opaque.

Quick checklist for solving conditional probability problems

1) Identify events A and B clearly.

2) Confirm P(B) > 0 (in discrete problems, B must have at least one outcome).

3) Decide approach:

• •Restrict-and-count if equally likely
• •Use P(A|B) = P(A ∩ B)/P(B) if probabilities are given

4) Be explicit about the denominator: after conditioning, your denominator is B.

5) Sanity-check: 0 ≤ P(A|B) ≤ 1, and if A ⊂ B then P(A|B) = P(A)/P(B) and should be ≤ 1.

That last point is a great self-check: if you compute something bigger than 1, your denominator or event interpretation is wrong.