Probability has been favourable outcomes over total for two chapters — count the ones you want, count them all, divide — and that dividing only works while every outcome carries equal weight. So load a coin until heads lands 7 times in every 10. The count cannot move: heads is still one face out of two, so counting insists on 1/2. The real chance of heads is 0.7. The counting machine we built is wonderful, and it is blind to weight. So this chapter builds the language that fixes it. It starts, of all places, with set theory. That will feel dry for about ninety seconds, and then turn out to be the exact vocabulary probability was missing. We'll pin down the sample space Ω, the list of everything that can happen, which for three coin flips holds 8 entries. We'll find that an event is nothing more than a subset of Ω, and that the words you already use — or, and, not — are precise operations on those sets. Then comes the upgrade the whole chapter turns on. We stop counting outcomes and start measuring them, giving weight directly to whole regions of Ω. That one move, from a count to a measure, is what lets probability handle a loaded die, a lopsided coin, a spinner that can stop anywhere. And once we've written down the three short rules a measure must obey, we'll watch nearly every formula you'll ever use fall out of them for free. No magic — just the right words, said carefully enough that the machine has no room left to be vague.
01Outcomes, and the sets that name them
Flip a coin three times and the list of outcomes is short enough to fit on one line: HHH, HHT, HTH, HTT, THH, THT, TTH, TTT, all eight of them. That complete list has a name and a symbol. It is the sample space, written Ω — the Greek capital omega. Ω is not a formula or a mystery. It is the set of every outcome the experiment can produce. One actual run of the experiment, say HTH, is a single element of that set, and we write one element with a little omega, ω. So Ω holds all 8 outcomes, and ω is one of them. Build Ω below. Set the number of flips and watch the whole space of possibilities list itself out: 2 flips gives 4 outcomes, 3 flips gives 8, 4 flips gives 16. Then click any single outcome to see it light up as one ω sitting inside the larger Ω.
ω = HH ∈ Ω · |Ω| = 4
Ω is the whole list — click any tile to pick one ω from it.
Fig. 1. Set n and watch Ω auto-enumerate — every one of its 2ⁿ outcomes, written out as a tile inside the bracket. Click any tile and it locks in as one ω, a single element sitting inside the set, and the readout updates to show exactly which one. That's why Ω isn't mysterious: it's just the complete, writable list of what can happen, and every ω is one point in it.
So Ω is the whole labelled world of what-can-happen, and each ω is one point inside it — one of the 8 points, for three flips. That is the entire idea, and every later page of this chapter is built on it. Now comes the move that makes set theory suddenly worth the trouble. It is the hinge the whole language swings on.
A probability question in English is already a set, written in words. Ask "what's the chance the first flip is heads?" and that sentence is quietly selecting some outcomes out of Ω and ignoring the rest. "First flip is heads" points at exactly four of the eight outcomes — HHH, HHT, HTH, HTT — and at nothing else. That selected collection is an event, and an event is nothing grander than a subset of the sample space. The sentence and the subset are the same object. Pick different sentences below and watch the matching outcomes light up. The words on the left literally are the set on the right.
the sentence is the subset it names — translate the words, and the outcomes light up.
"first flip is heads" → 4/8 = 0.500
Fig. 2. Each button is an English sentence; behind it sits a fixed rule for testing any outcome. Click one and watch the rule sweep across all 8 outcomes of Ω — the ones it keeps light up green, and the count assembles live into count / |Ω|. Notice "first flip is heads" and "more heads than tails" both light up 4 cells — the same probability, 0.5 — but not the same 4 outcomes. The fraction only ever measures the subset; the sentence is what decided which subset that was.
The dividing is the easy half. A claim in words becomes a subset of Ω, and its probability, in this equally-likely world, is just how many outcomes are in the subset divided by how many are in Ω. For "first flip is heads" that is 4 divided by 8, so 1/2. Count the subset, count the space, divide. The hard half — the real skill, and the one the rest of probability runs on — is turning fuzzy language into an exact set. And the moment you have sets, you can do things to them.
02Three moves: OR, AND, NOT
Real questions combine events, and the joining words are the operations. "First flip heads or second flip tails." "Red and face card." "Not a six." Each of those small words is a genuine operation on subsets, with a name and a picture, and there are only three of them. Let's meet all three on one workbench. Below, Ω is the eight three-flip outcomes. A is "first flip is heads," which holds 4 of them, and B is "second flip is tails," which holds another 4 — each drawn as a loop around its outcomes. Toggle between the three combinations and watch which outcomes get selected, and how the probability changes as you do.
∪ union = or · ∩ intersection = and · ᶜ complement = not
A∪B: in A OR B — 6 of 8
Same 8 dots, same 2 loops — each word just picks a different region. That's why or, and, not aren't vague: each is one drawable operation.
Fig. 3. The Venn workbench. Eight three-flip outcomes sit as dots; loop A holds every outcome whose first flip is Heads, loop B every outcome whose second flip is Tails. Toggle A∪B (or), A∩B (and), Aᶜ (not) — each word lights a different region of the same picture, and the probability is just how many of the 8 dots it caught.
Three operations, three pictures — and the differences between them are exactly where beginners lose probabilities, so let's earn each one properly. First, union, written A ∪ B and read "A or B." Union collects everything in either set: an outcome qualifies if it is in A, or in B, or in both. It is the big umbrella, the permissive one, always at least as large as the pieces you started with. Now the trap, and it is worth walking straight into so it never catches you. In everyday speech "or" usually means one or the other, not both. In set language "or" always includes both. Guess first, then reveal — how many of the eight outcomes satisfy "first heads or second tails"?
Flip 3 coins (H/T). Each dot is one of the 8 equally-likely outcomes, like HTT. Guess how many satisfy “first = H OR second = T”.
Pick a guess, then hit Reveal.
The trap: everyday “or” means pick one. Set “or” (∪) means either — or both. That’s why OR is the bigger set, not the overlap.
Fig. 4. Every dot is one of the 8 equally-likely outcomes of 3 coin flips. Circle A holds every outcome whose first flip is heads; circle B holds every outcome whose second flip is tails. Guess how many satisfy “A OR B,” then hit Reveal: the everyday instinct is to point at the tiny overlap (2, the AND) — but set “or” keeps everything in either circle, so the true answer is 6. OR is the big umbrella, not the overlap.
Six of the eight, not two. If you pictured the tiny overlap instead, you read or as and, and that single slip is one of the most common ways a probability comes out wrong. OR is the big set. That tiny overlap has its own name, and it is union's opposite twin: the intersection, written A ∩ B and read "A and B." Here an outcome qualifies only if it sits in both sets at once, which for first-heads and second-tails leaves just 2 of the 8: HTH and HTT. So intersection is the strict, demanding operation, and it can only ever make the set smaller. Stack conditions below and watch the surviving set shrink with every "and" you add.
start: all 8 outcomes possible
AND is a hurdle every survivor must clear — it can only shrink the group, never grow it.
Fig. 5. Eight equally-likely outcomes of 3 coin flips start alive (green). Click 1 to demand first = heads — half the group dies. Click 2 to also demand second = tails — half of what's left dies. Click 3 to also demand third = heads, and one lone outcome survives: 8 → 4 → 2 → 1. Every added and is one more hurdle the same outcome must clear at once, so the surviving group can only shrink — it never has anywhere to grow.
So "and" narrows and "or" widens. Every extra and is another hurdle an outcome must clear. Every extra or is another door it may enter. The third move is the simplest and quietly the most useful. The complement of A, written Aᶜ, is everything in Ω that is not in A. If A is "first flip heads," then Aᶜ is "first flip tails." A holds 4 outcomes, Aᶜ holds the other 4, and 4 + 4 = 8 is all of Ω. They never overlap, and together they tile the whole space with nothing left over. Peel A off Ω below and watch what's left become its complement.
not peeled yet — A sits inside Ω
Aha: A and Aᶜ never overlap and never leave a gap — so the instant you know P(A), you already know P(Aᶜ) = 1 − P(A) for free.
Fig. 6. Drag the gold tab (or the slider) to peel A — the outcomes with exactly two heads, HHT, HTH, THH — straight up and out of Ω. Those three tiles lift free and turn solid blue; the five left behind need no extra work to name — they turn solid red and becomeAᶜ. Watch the bar underneath as you drag: the blue slice and the red slice grow together, never overlapping, and however far you've peeled, they always sum to exactly that much. Peel all the way and P(A) + P(Aᶜ) lands on exactly 1.000 — not 0.9, not 1.1 — because A and its complement tile Ω with nothing left over and nothing double-counted.
You now own all three moves: ∪ for or, ∩ for and, and the complement for not. Every set so far, though, arrived pre-drawn for you. The real skill this chapter promised was turning words into the right subset yourself. So do it with your own hands. Here are two plain-English clauses, with nothing shaded yet. Tag each of the outcomes by eye, then build their or, their and, and their not. Then check whether the sets you drew match the truth.
tag all 8 the way you read the sentences
tag all 8, then reveal
A wrong tile flashes — that's the exact spot your reading of the words slipped.
Your sets vs. the true sets
Areveal to check
Breveal to check
A∪Breveal to check
A∩Breveal to check
Aᶜreveal to check
Every number above came only from the 8 tiles you tagged — nothing was pre-loaded. Change which tiles you'd call "in A" and every one of ∪, ∩, ᶜ would recompute to match. The hard part of probability was never this arithmetic; it was turning "at least two heads" into the right four tiles, and you just did that by hand.
Fig. 7. Two plain sentences over the same 8 outcomes — A = “at least two heads,” B = “the last flip is tails” — and an empty grid. Click each tile to tag it —, A, B, or A∩B, then hit reveal & grade. The true split is A: 4 of 8, B: 4 of 8, A∪B: 7 of 8, A∩B: 1 of 8 (only HHT), Aᶜ: 4 of 8 — and every one of those five numbers was built only from the tiles you tagged. A wrong tile flashes once so the miss registers before it's explained: the arithmetic was never the hard part; reading the sentence into the right four tiles was.
A and its complement exactly fill Ω and never touch: 4 outcomes in A, the other 4 in Aᶜ, none in both. It is a fact so plain it is easy to miss, and so powerful that we will build an entire chapter on it. For now, hold the picture. A here, Aᶜ there, nothing outside the two. Three operations, and you have the full grammar. Time to draw it in a way that makes the probabilities readable by eye.
03The picture that's exactly right
Here is a picture that is exactly to scale, where every region's area literally equals its probability. The Venn loops we have been drawing are cartoons by comparison: good for intuition, loose about sizes. Take two coin flips. There are four equally-likely outcomes, so cut Ω into four equal tiles, each one worth 1/4. Let A = "first flip heads" be the left column, which is 2 tiles, and let B = "second flip heads" be the bottom row, another 2 tiles. Now every combination is a region you can count in tiles. Read them straight off below.
click to toggle each region
A · P = 1/2
Equal-weight tiles ⇒ AREA already IS probability — that's why the cartoon Venn can finally be trusted here, no formula needed.
Fig. 8.Click A, then B, then ¬. Two fair flips cut Ω into four equal tiles — each one worth exactly ¼ because HH, HT, TH, TT are equally likely. A is the whole left column, B is the whole bottom row; turn both on and the ∪ OR pill shows their union (3 tiles, ¾) while ∩ AND isolates just the overlap tile, A∩B (¼). Hit ¬ NOT from any state to flip to the complement — Aᶜ is the two tiles A left behind, still ½. Nothing here is computed by formula: you are just counting lit tiles out of four, because at equal weight area IS probability.
Every answer here is a tile count. A ∩ B is the single corner where the column meets the row, one tile out of four, so P(A∩B)=1/4. The complement of A is the whole right column, two tiles, so P(Aᶜ)=1/2. The union A ∪ B covers three of the four tiles, so P(A∪B)=3/4. No formulas anywhere — you are just seeing areas. And this picture reveals the thing that trips people up when they only have the algebra: what actually happens when you compute a union.
Try to get P(A∪B) the lazy way, by just adding P(A) and P(B), and you get 1/2 + 1/2 = 1: the union looks certain. The tiles say the union is only 3/4. The corner tile where the column and the row overlap lives in both A and B, so adding the two areas counts that one tile twice. Subtract the overlap back out once and the sum lands on 3/4. Watch the double-count happen and then get corrected, region by region.
two sets A and B — they share 2 tiles
Adding A then B pays for the shared column twice — so 0.60 overshoots the truth. The minus sign isn’t a new rule: it’s the one refund you owe on what you double-charged.
Fig. 9. Two events drawn on a 20-tile sample space: A (six blue tiles) and B (six coral tiles), overlapping in the shared column A∩B. Step through it: add P(A)=0.30, then add P(B)=0.30 — and the running total climbs to 0.60, overshooting the gold line. Look at the shared column: it got shaded twice, so its 0.10 was charged twice. Subtracting P(A∩B) once removes exactly that overshoot and lands on the truth, 0.50. The reader-facing symbols: ∪ means “A or B” (the whole shaded region), ∩ means “A and B” (only the shared part). That’s the whole reason the minus sign is there and isn’t optional — you owe exactly one refund on what you paid for twice.
That subtraction is the whole of inclusion–exclusion: P(A∪B) = P(A) + P(B) − P(A∩B). On our four tiles it reads 3/4 = 1/2 + 1/2 − 1/4. It is not a rule to memorize. It is the visible, unavoidable correction for having counted the shared region one time too many. And because our sample space is small and finite, we do not have to take it on faith. We can grind out every outcome and check that the identity holds on the nose.
▸ for ω in Ω: classify(ω)
click run to enumerate Ω
A∩B gets counted once inside |A| and once inside |B| — subtracting it once removes exactly that double count.
Fig. 10. Hit run enumeration and all eight three-flip outcomes classify themselves one by one: blue if only in A (first flip H), cyan if only in B (second flip T), gold if in both, grey if in neither. Straight counting gives |A|=4, |B|=4, |A∩B|=2 — feed those into |A|+|B|−|A∩B| and you get 6; count the coloured tiles in the union by hand and you also get 6, to the last outcome. That's why the picture can be trusted: the tiles, the counted outcomes, and the formula are three views of one identity that checks out exactly.
The enumeration agrees to the last outcome. That is the point of a small example: the picture, the count, and the formula are three views of one true thing. But all three views have quietly leaned on the four tiles being equal, each worth 1/4. It is time to kick that assumption out and watch what breaks, and what survives.
04From a count to a measure
Probability is the area of A divided by the area of Ω — throw darts uniformly at a board, and P(A) is simply how often a dart lands inside A. The professor slipped that line in almost as an aside, and it is the most important thing in the chapter. On a fair board of 4 equal tiles, a dart lands in a 2-tile region half the time, which is the 1/2 you already trust. Play with that picture first, on a fair board, and confirm it gives back exactly the count-and-divide answers.
no darts yet — fire away
Every cell is the same area, so counting darts in A over darts total is the same count-over-total answer — just reached by throwing instead of listing.
Fig. 11. The board is Ω — every possible outcome — cut into 20 cells of exactly equal area, 7 of them gold and called A. Fire darts (+1 / +10 / +100): each one lands uniformly, and hits in A ⁄ darts settles onto area(A) ⁄ area(Ω) = 0.350 — the same count-over-total answer as "7 of 20," reached by throwing instead of listing. Because every cell carries equal weight, the area model can't help but agree with counting — which is exactly why it still works once the pieces stop being equal-sized cells and become a genuinely uneven board.
On a fair board the area model and the counting ratio agree perfectly: same number, two roads. Now do the thing counting cannot survive. Make the coin biased, weighted so heads lands 7 times in 10, and the two roads split. The naive count can only see faces — one heads out of two — so it stays frozen at 1/2 however loaded the coin is. The area picture simply lets the heads region grow until it covers 0.7 of the board, so P(heads) tracks the real chance. 0.7 against 0.5, from the same coin. Drag the bias and watch the count lie while the measure tells the truth.
fair coin — both roads land on 0.50
Counting can only ever say 1 of 2. It owns no slot for “heavier.” Real coins do — so we upgrade a count of faces into a measure of area.
Fig. 12. Bias the coin and watch the two roads split. The count road (top) is frozen at ½ forever — “heads” is always 1 of 2 faces, so its halves never move. The weight road (bottom) resizes with the real bias: the dashed gold line marks where counting insists the boundary sits, the white line marks where the weight actually puts it, and the orange slice between them is the chance counting is physically blind to. Counting has no slot for “heavier” — which is exactly why probability upgrades a bare count of outcomes into a measure of area.
That is why we upgrade. The counting ratio was never exactly wrong. It was a special case, valid only when every outcome weighs the same, and it has no way to say that heads weighs 0.7 while tails weighs 0.3. The area picture has room for weight built in, and it keeps working when the counting picture goes blind. So we promote probability from a count to a measure: a rule that hands every subset of Ω a number between 0 and 1, saying how much weight that region carries. Below, drop weight onto the individual outcomes however you like. Pile it unevenly if you want, and P(A) becomes the total weight sitting inside A. Then flatten the weights to equal and watch the measure collapse back into plain count-over-total.
drag any white dot — weight piles there, every other outcome shrinks to keep the total at 1
piled — P(A) tracks the weight
that's why Ch 1–2 were never wrong, just narrow: counting is the equal-weight corner of one measure.
Fig. 13. Six outcomes share the sample space Ω; the green band, three outcomes wide, is event A. Drag any white dot up and weight piles onto that outcome while every other outcome shrinks to compensate — a measure can never hand out more weight than the whole, so the pile always sums to exactly 1. P(A) is just the weight caught inside the green band. Hit make weights equal and it snaps to |A|/|Ω| = 3/6 — the plain count-and-divide you already knew, now revealed as one special case of the general rule.
When the weights are all equal, the measure is favourable-over-total — set all 8 outcomes to weight 1/8, and P(A) with 4 outcomes inside comes back as 4/8, exactly 1/2. So the whole of Chapters 1 and 2 is the equal-weight corner of this one general idea. We did not throw counting away. We found the bigger room it had always been sitting inside. A measure is the real object, and counting was its fair-world shadow. But a measure cannot be just any rule that spits out numbers. To deserve the name "probability" it has to obey a few laws. Remarkably few.
05The constitution: three axioms
Everything — every distribution, every theorem, the entire rest of this book — is downstream of exactly three rules. Think of them as the constitution of probability: short, non-negotiable, generative. One: P(Ω) = 1, since something happens. The whole space carries all the weight there is, so our 8 three-flip outcomes must have weights summing to 1. Two: for any event A, P(A) ≥ 0. No negative chances, ever. Three: if A and B are disjoint, meaning no outcome is in common, then P(A∪B) = P(A) + P(B). Separate pieces just add: 1/4 for one tile plus 1/4 for another tile is 1/2. That is the entire founding document. Read each clause below.
holds for every valid measure
three short lines, yet they force every coin, die, and dataset into the same shape — the entire founding document.
Fig. 14. The whole founding document, in one card. Click I / II / III to see each clause's tiny picture: Ω (the sample space — "everything that could happen") always fills completely, no outcome ever gets a negative share, and disjoint pieces just add. Now hit break it: a fake "measure" tries each rule anyway — a gap left over, a negative share, or overlapping sets double-counted — and every time, the picture visibly fails. Three short non-negotiable lines are all it takes to pin down every probability that will ever be computed.
Axiom three slipped in a condition you may have skimmed: the pieces must be disjoint. Why should a quiet word like disjoint matter to a plain plus? Feel it by breaking it. Slide two events from apart to overlapping, and the naive P(A)+P(B) drifts above the true P(A∪B) by exactly the part they now share. On our tiles that overshoot was 1 against 3/4, an error of 1/4.
Predict first: before you slide, guess where the true-union bar will land — then slide and watch it split away from the frozen naive sum.
disjoint: the '+' is exact
Neither block changes size — so the naive sum never moves. Only the shared tiles, counted twice, open the gap.
Fig. 15. Slide B across A and predict where the true-union bar will fall. At overlap 0.00 both bars meet at 1.00 — the one case axiom 3 licenses, where A and B share no tile and P(A∪B) = P(A)+P(B) exactly. Now slide: the naive sum stays frozen (neither block changed size) while the true union drops, and the gold gap between them is precisely the overlap you'd be double-counting. Switch to the over-stuffed pair and the naive sum freezes at 1.20 — a “probability” past the red ceiling, which is impossible — while the true union never climbs above 1.00. That gold gap is the price of the ‘+’: it is exact only when the pieces are disjoint, which is exactly why the axiom insists on it.
The rules you would have assumed were separate facts turn out to be theorems hiding inside those three lines. Three lines look like far too little to build a subject on, and that is exactly the wonder of it. Take the complement rule. You already believe P(Aᶜ) = 1 − P(A) is obvious. But watch it fall out. A and its complement Aᶜ are disjoint, and together they make all of Ω. So by rule three their probabilities add, and by rule one that sum is 1, meaning P(A) + P(Aᶜ) = 1. Rearrange, and there it is: if P(A) is 1/2 then P(Aᶜ) is 1/2, and if P(A) is 0.7 then P(Aᶜ) is 0.3. Step through the derivation.
step 1 of 4
try another A — the rule doesn't care
A and Aᶜ tile Ω — no gap, no overlap
That's why the complement rule feels obvious: it's axiom 3 and axiom 1 standing side by side — nothing new was added.
Fig. 16. Step through the proof: A ⊔ Aᶜ = Ω — the event and its complement tile the whole sample space, no gap and no overlap. Hit next: because they don't overlap, axiom 3 says their probabilities simply add. Hit next again: together they are Ω, and axiom 1 says P(Ω) = 1. One more next and it's just algebra — subtract P(A) from both sides and P(Aᶜ) = 1 − P(A) falls out. Try A=0.2 or A=0.8: the bar's split changes, the derivation above it doesn't — because it never depended on which A you picked. That's the whole trick: the "complement rule" isn't a fact someone added, it's axiom 3 and axiom 1 stood next to each other.
Nothing new entered that derivation — just axiom three followed by axiom one. The complement rule was never an extra law. It was those two axioms, stood next to each other. The same thing happens with monotonicity, the plain-sense claim that a bigger event cannot be less likely. If A sits inside B, then B is A plus the leftover part of B. Those two pieces are disjoint, and the leftover has probability ≥ 0 by axiom two. So P(B) is P(A) plus something non-negative, and cannot be smaller. One tile inside a two-tile row means 1/4 can never beat 1/2. Watch it drop out.
Here is the trap that catches almost everyone. You have leaned on disjoint all section, and it simply means A and B share no outcome. But disjoint does not mean the two events have nothing to do with each other. Two events that cannot both happen are the most tightly linked of all. Take "first flip heads" and "first flip tails", which share 0 outcomes. Learn that one occurred, and the other is instantly ruled out, its probability dropping from 1/2 to 0. Feel that difference now, well before a later chapter gives it a name.
Same 12 equally-likely outcomes, two setups. A = the left block (dashed). B = the green cells. Both start at P(B) = 1/3. Guess first: when A just happens, in which panel does B's chance change?
Guess, then press the button.
Disjoint is maximum influence: A happening screams B did not. The panel that holds steady at 1/3 is the real “no effect” case — a later chapter names it independent.
Fig. 17. Two panels over one 12-outcome Ω, both reading P(B) = 1/3. Guess where B's chance moves, then press A just happened: the everyday instinct picks the overlap panel, but it’s the disjoint one that collapses to given A → P(B) = 0, while the tuned overlap holds flat at 1/3. Disjoint isn’t “no effect” — it’s the strongest effect: A happening proves B did not. The panel that doesn’t budge is the different idea a later chapter will call independent.
Drag the gold dot up to grow A inside B — or use the slider. Watch the green ring (the leftover) shrink to feed it.
0.16 ≤ 0.64 — always holds
the ring is real area — axiom 2 forbids negative probability, so it can never owe back more than B has.
Fig. 18.A and B are events — collections of outcomes — and A ⊆ B ("A is a subset of B") means every outcome in A is also in B: A sits entirely inside B. Drag the gold dot (or the slider) to grow A — the boundary is clamped so A can never spill past B, exactly because it's contained. The green ring is the leftover, B minus A: the outcomes in B but not in A. Since probability P(·) can never be negative (axiom 2), that ring's probability can't dip below zero either — so the meter's identity P(B) = P(A) + leftover means P(B) can only sit at or above P(A). Drag all the way out: leftover shrinks to 0.00 exactly when A fills B, the one case where P(A) = P(B). That's the whole proof of monotonicity: a bigger event isn't "probably" at least as likely — it's forced to be, because the extra area is real, non-negative weight.
The inclusion–exclusion rule we read off the tiles earlier — P(A∪B) = P(A)+P(B)−P(A∩B) — is the same story once more. Carve the union into disjoint slices, add them with axiom three, and the double-counted overlap has to come back out. On the four tiles those slices are 1/4, 1/4 and 1/4, giving 3/4. The picture told us that the rule is true. The axioms tell us why it must be. Re-derive it from the constitution below.
one region — but which pieces make it?
Nothing new is drawn after step 1 — every later step only re-shades this same picture.
Fig. 19. The same union, re-derived instead of read off. Step through: cut A∪B into three pieces that share no outcome, add them with axiom 3 (disjoint pieces just add), then regroup — and the formula the tiles showed, P(A∪B) = P(A)+P(B)−P(A∩B), drops out on its own.
Three axioms sit at the root of the whole subject, and a spreading tree of consequences grows out of them: complement, monotonicity, inclusion–exclusion, and eventually every distribution in the book. That is the shape of probability, and it is worth saying out loud. Each branch is a theorem you could re-derive rather than a rule you have to store, so the whole subject asks you to remember 3 things. That is the payoff of the dry set-theory detour. It bought us a foundation small enough to hold in one hand and strong enough to carry everything.
06You are here
The same three axioms work on sets far stranger than coin flips. Our sample spaces were finite lists you could enumerate — 8 outcomes for three flips. But those same rules also assign probabilities to the space of all natural images, to a fractal, to the Lorenz attractor coiling through three dimensions. Asking "how big is this set" for objects like those is a deep field called measure theory, and what we just did is quite literally its first page. So before we place this chapter on the map, take one glance over the edge at the horizon.
count → area: same axioms, finite Ω
Discrete counting is just the Ω = finite-tiles case — page one of a machine that scales to measure images, fractals, even chaos, as long as it keeps μ(Ω)=1, μ(A)≥0, and disjoint pieces add.
Fig. 20. Four wildly different Ω — a coin, the space of every possible image, a fractal, the phase space of a chaotic system — click through them. Each caption names what Ω is here and what “measuring” means for it, but the locked panel below never changes: every domain still owes you μ(Ω)=1, μ(A)≥0, and disjoint pieces adding. That's the whole point — the finite tiles you've been counting all chapter aren't a special case of probability, they're the smallest instance of the same measure that weighs images, fractals, and chaos.
You do not need any of that machinery yet. But it is worth knowing that the modest thing you learned here is the seed of a large tree. "Probability is a measure on a sample space" is a sentence a mathematician means literally, every word of it. Now let's find our rung on the family tree of the course.
Hover — or tap — the two earlier rungs to see exactly what each one handed forward to today's measure.
hover / tap 01 or 02 — trace the climb
Fig. 21. The spine so far, with today lit gold: chapter 1 handed you the ratio, chapter 2 handed you the machine that counts it, and this chapter promoted both into a measure on Ω. Hover — or tap — 01 and 02 to watch exactly what each one fed forward. Nothing got replaced: the ratio still works, the counting engine still runs the sums — they were the equal-weight corner of a bigger room all along.
There it is, lit. We climbed from the counting ratio up to a measure on a sample space. It is the same favourable fraction, finally freed from the equally-likely assumption that had been quietly holding it — which is why a coin at 0.7 heads no longer breaks anything. Everything above us on the tree, from conditioning to the Bayes flip to the whole family of distributions, is a measure being asked sharper and sharper questions. And one of the little tools we derived along the way is about to do something spectacular.
Look again at the complement rule, P(Aᶜ) = 1 − P(A). It reads like bookkeeping. But watch what it becomes when the event you want is a tangle — a mess of overlapping cases that would be a nightmare to count head-on. "At least one" is exactly that kind of tangle. Its complement, "none," is often a single clean thing. And when the servers fail independently of each other, that clean thing is one probability multiplied n times. Say 3 servers each fail 10% of the time: all three stay up with probability 0.9 × 0.9 × 0.9 = 0.729, so at least one fails 27.1% of the time. We'll earn that multiplication properly in the next chapter. Get a first taste below: flip the hard question into its easy opposite.
15 overlapping cases
Flip the question: “none” is one clean product; “at least one” is every overlap you'd have to untangle by hand. Same fact — pick the easy side.
Fig. 22. Click at least one and the circles sprawl — each ring is one server's "it timed out," and to count their union by hand you'd need 2ⁿ−1 overlapping cases (already 255 at n = 8). Click none instead and the same n servers collapse into one tidy bar, split into n identical factors of 0.9 chained by multiplication — valid because the servers fail independently — no overlaps to untangle, ever. Whichever button you're on, the readout on the right never actually counts the mess: it always resolves through P(at least one) = 1 − P(none). That's why one boring line of bookkeeping is secretly a power tool — when the event you want is a tangle, its opposite is often a single tidy thing.
The tell is in the wording, not the numbers. The servers showed the flip paying off once, on a problem already chosen for you, but an expert reaches for the flip unprompted. So here is a short menu of questions, some worth flipping and some not. Predict which ones the clean opposite 'none' cracks open. Then reveal the case-count each way and see the trigger for yourself.
Each card is a worded question. Toggle it to count it straight or flip it to its opposite — then Reveal the separate cases each path makes you add. The shorter bar is the path you want.
Sort all four, then Reveal.
The trigger is the phrase, not the sum. “At least one” has a single tidy opposite — none — so flip it. But “exactly two” and “sum 7” have no clean opposite, so flipping just trades one tangle for another.
Fig. 23. Four worded questions, one menu. Toggle each card to count it straight or flip it to the opposite, then hit Reveal: the bars show how many separate cases each path forces you to add. For the two “at least one” questions the direct path is a tangle (4 and 10 cases) while the flip collapses to a single clean none (1 case) — flip those. But “exactly 2 heads” and “sum to 7” are already one clean case going direct (1), and flipping explodes them into 5 and 10 — so count those straight. The trigger is the phrase “at least one,” whose one tidy opposite is “none”; recognising which phrase has a clean complement is the whole skill.
One subtraction turned a horrible count into an easy one: P(at least one) = 1 − P(none). That is the whole idea the next chapter is built on, and it leads somewhere genuinely shocking. A room needs only 23 people before two of them very probably share a birthday — the chance passes 50% at exactly 23. The complement rule, one line we got for free from three axioms, is what cracks it open. Turn the page.