Independence and mutual exclusivity: Visual walkthrough

This page rebuilds the central fact of Independence and mutual exclusivity from nothing: why two events that "can't happen together" are the exact opposite of two events that "don't affect each other." We do it with pictures of a single square — the picture of all possible outcomes.

We only assume you know: an event is a collection of possible outcomes, and its probability is a number between $0$ and $1$ telling you how "likely" it is. Everything else we build. If you want the ground layer, see 1.3.01-Basic-probability-concepts and 1.3.02-Conditional-probability.

Step 1 — Draw the world as a square of area 1

WHAT. We picture every possible outcome of an experiment as points filling a square whose total area is exactly $1$ .

WHY a square of area 1? Because "probability" behaves like area: the whole square (everything that could happen) must have probability $1$ , and the area of a region is the probability of that region. Area is the friendliest picture for probability — it adds up the way probabilities add up, and it lets us literally see overlap.

PICTURE. Look at the figure. The full board is the square. An event $A$ is a shaded blob inside it. The fraction of the square that blob covers is $P (A)$ — the probability of $A$ .

Figure — Independence and mutual exclusivity

Step 2 — The overlap is where BOTH happen

WHAT. We draw a second event $B$ as a second blob, and mark the region where the two blobs cross each other.

WHY this region matters. A point in the crossing region is an outcome that belongs to $A$ and to $B$ at the same time. That crossing is called the intersection, written $A \cap B$ (read " $A$ and $B$ "). Its area is the probability that both happen at once.

PICTURE. The pink lens in the middle is $A \cap B$ . Everything hinges on how big this lens is — the whole rest of the page is a story about its size.

Step 3 — Case one: push the blobs apart (mutually exclusive)

WHAT. We slide $A$ and $B$ until they no longer touch. The lens shrinks to nothing.

WHY do this? This models events that compete: rolling a $2$ versus rolling a $5$ , "spam" versus "not spam." If one happens, the other simply cannot. In the picture that means zero overlap.

PICTURE. No pink lens survives. The two blobs are separate islands.

Before we add the areas, one new symbol. The region covered by $A$ or $B$ (all points in at least one blob — either blob alone, plus their overlap if any) is called the union, written $A \cup B$ (read " $A$ or $B$ "). It is the total shaded footprint on the board.

Because in Step 3 the overlap is gone, the area of " $A$ or $B$ " is just the two areas stacked with nothing double-counted:

Step 4 — What "knowing B happened" does to the picture

WHAT. Before comparing to independence, we build one tool: conditional probability. When we learn " $B$ happened," we throw away every outcome outside $B$ and treat $B$ as our new, smaller world.

WHY a new tool, and why this one? "Does $A$ affect $B$ ?" is a question about information. The only honest way to ask it is: if I restrict my attention to the world where $B$ is true, how much of that world is also $A$ ? That ratio is exactly $P (A ∣ B)$ (read " $A$ given $B$ "). We need it because independence is literally the statement "this restriction changes nothing."

PICTURE. The board dims everywhere except inside $B$ . Now $B$ is our whole universe. The bright slice inside it is the part that is also $A$ — its fraction of $B$ is $P (A ∣ B)$ .

Step 5 — Case two: overlap that "just fits" (independent)

WHAT. Now we ask: what size must the lens be so that learning $B$ tells you nothing about $A$ ? That is, so that $P (A ∣ B) = P (A)$ : the fraction of $B$ that is $A$ equals the fraction of the whole square that is $A$ .

WHY this is the definition. "Tells you nothing" is precisely "the proportion of $A$ is the same whether you look at all of the square or only at $B$ ." Independence is a statement about proportions matching, not about blobs being apart.

PICTURE. $A$ cuts across $B$ so that it grabs the same fraction of $B$ as it does of the whole board. Think of $A$ as a vertical stripe and $B$ as a horizontal stripe: their crossing is a rectangle whose area is width $\times$ height.

Now turn " $P (A ∣ B) = P (A)$ " into an equation about areas. Start from Step 4:

$demand: unchanged P (A) = P (A ∣ B) = \frac{P ( A \cap B )}{P ( B )}$

Multiply both sides by $P (B)$ (WHAT: clear the denominator; WHY: to isolate the overlap area):

Step 6 — The punchline: these two cases are enemies

WHAT. We overlay Step 3 (disjoint) and Step 5 (independent) and compare the required lens size.

WHY. People think "mutually exclusive" and "independent" are cousins. The picture proves they are opposites whenever both events are actually possible.

PICTURE. Left board: overlap forced to $0$ . Right board: overlap forced to $P (A) P (B)$ , a positive rectangle. They cannot both be true unless one blob has zero area.

Suppose $P (A) > 0$ and $P (B) > 0$ (both events genuinely possible). Then:

$independence wants this P (A) \cdot P (B) > 0 but exclusivity wants this P (A \cap B) = 0$

A positive number cannot equal zero. So the independence equation fails. Concretely:

$P (A ∣ B) = \frac{P ( A \cap B )}{P ( B )} = \frac{0}{P ( B )} = 0 \neq = P (A)$

Knowing $B$ happened crushes $A$ 's probability to $0$ — maximum information, the exact opposite of "no information."

Step 7 — The degenerate edges (never skip these)

WHAT. We check the corner cases where the two ideas do touch — both the empty blob (area $0$ ) and the full blob (area $1$ ).

WHY. A derivation you can trust must survive its own boundary. Here the boundary is an event that is either impossible ( $P = 0$ ) or certain ( $P = 1$ ).

PICTURE. Left: a blob shrunk to a single dot has area $0$ . Right: a blob swollen to fill the whole board has area $1$ . In both extremes disjointness and independence quietly agree.

If $P (A) = 0$ (impossible event): the overlap is $0$ (so disjoint holds) and $P (A) P (B) = 0$ (so independence holds). Both properties are true at once — but this is a trivial, degenerate case, not a counterexample to Step 6.
If $P (B) = 0$ : the conditional $P (A ∣ B)$ is undefined — you cannot stand inside a world of area zero. The area-form $P (A \cap B) = P (A) P (B)$ still works as the general definition, which is why we prefer it.
If $P (A) = 1$ (certain event): $A$ fills the whole square, so every point of $B$ already lies inside $A$ . Then the overlap is all of $B$ : $P (A \cap B) = P (B)$ . Check independence: $P (A) P (B) = 1 \cdot P (B) = P (B) = P (A \cap B)$ ✓ — a certain event is independent of everything, since learning $B$ cannot change a probability that is already stuck at $1$ .
If $P (A) = 1$ and we want exclusivity: it collapses too — $P (A \cap B) = P (B) = 0$ forces $P (B) = 0$ , so the only way a certain event is disjoint from $B$ is if $B$ is impossible.
The full union formula always holds and reduces correctly: $P (A \cup B) = P (A) + P (B) - = 0 if disjoint P (A \cap B)$ When disjoint, the last term vanishes and we recover Step 3.

Step 8 — Worked check: bag of balls & two dice

WHAT. We drive the pictures onto the parent note's numbers.

Bag: 3 red, 2 blue, one draw. $P (R) = \frac{3}{5}$ , $P (B) = \frac{2}{5}$ .

Overlap: cannot draw red and blue $\Rightarrow P (R \cap B) = 0$ (Step 3 → disjoint).
Independence test: $P (R) P (B) = \frac{3}{5} \cdot \frac{2}{5} = \frac{6}{25} \neq = 0$ → not independent (Step 6 confirmed).

Two dice. $A =$ "first is 4", $B =$ "second is 4", $C =$ "sum is 8".

$P (A) = \frac{1}{6}$ , $P (B) = \frac{1}{6}$ , $P (A \cap B) = \frac{1}{36}$ , and $\frac{1}{6} \cdot \frac{1}{6} = \frac{1}{36}$ → independent (Step 5 rectangle fits perfectly).
$P (C) = \frac{5}{36}$ , $P (A \cap C) = \frac{1}{36}$ , but $P (A) P (C) = \frac{1}{6} \cdot \frac{5}{36} = \frac{5}{216} \neq = \frac{1}{36}$ → not independent.

The one-picture summary

One number rules everything: the overlap area $P (A \cap B)$ .

Push it to $0$ → mutually exclusive.
Set it to width $\times$ height, $P (A) P (B)$ → independent.
These meet only when a blob has zero area, or degenerates to the whole board (Step 7). Otherwise they are opposite ends of a ruler running from maximum dependence to zero dependence.

This single dial is exactly the machinery 1.3.03-Bayes-theorem leans on, and its conditional version — features that stop informing each other once you fix the class — is what makes the 3.1.01-Naive-Bayes-classifier and 5.2.03-Probabilistic-graphical-models tractable.