02Sequences, Series & the Constant e

Chapter 1 ended on a promise and a question. The promise was that "how many?" is answered by pairing and never by looking, because you pair the pile with the ruler {1, 2, …, n} and read off where you stopped. The question is the one that ruler quietly raises, which is that it never stops. So what happens when the pairing runs on forever, and you are adding up infinitely many shrinking pieces? Most tellings answer that with a dot-dot-dot and a formula, and we are going to do something else instead. Here's the plan. We'll build a sequence as Chapter 1's function with the counting numbers as its domain — no new object, just an old one pointed somewhere new. Then we'll drag a band around a number until you've used the definition of a limit before you've ever seen its Greek letter. Then I'll do the thing nobody admits to: I'll change what the plus sign means, in front of you, on purpose. Everything after that is harvest. You'll get a/(1−r) along with the exact conditions under which it's allowed to exist, and you'll get e from both ends and build the bridge between them yourself, out of Chapter 1's C(n,k). And you'll walk out with one discipline that separates a quant from a formula-user: prove the thing exists, then do algebra to it. Skip that order and one day you'll prove that 1 + 2 + 4 + 8 + … = −1 and not flinch.

This chapter looks like a maintenance stop on the way to the interesting material. It is not. The limit you build in the next twenty minutes is the load-bearing beam under the integral in Chapter 3, under Poisson in Chapter 12, under the Central Limit Theorem in Chapter 13, and under dW² = dt in Chapter 26. Every equals sign in the rest of the book is signed by this chapter.

So let's go and get it. We start exactly where Chapter 1 stopped.

01A sequence is a rule, not a list

Fold a sheet of paper in half, and then fold it again. Its thickness doubles every time, so after n folds it is 2^n sheets thick. Nothing exotic there — you just did it in your head.

Now a second one, and it's pure Chapter 1. Roll a die n times and ask what fraction of all the possible outcome-strings contain no 6 at all. Each roll independently has 6 faces, so the multiplication principle gives 6^n total strings, and the strings that dodge the 6 have 5 choices per roll, so there are 5^n of them. Favourable over total is (5/6)^n, and both of those tools were already yours.

And a third, the simplest of the lot: 1/2^n — half a cake, then half of what's left, then half of that.

Three tables. Here's the sentence that matters, and almost every telling skips straight past it. Every one of those three is a function whose domain is Chapter 1's counting numbers ℕ = {1, 2, 3, …}. The subscript in a_n is not decoration and it is not multiplication, it is just the input, written low. We call the whole object a sequence. No new machinery entered the room — we took Chapter 1's function and pointed its domain at the counting numbers.

Why do I care so much about that? Because of what it lets you ask. Type any n you like into the machines below — 7, 40, 200 — and the rule answers instantly. Nobody wrote 200 things down first.

That is the whole point of the last five minutes. If a sequence were a list, then "what happens far out?" would be a question about nothing, because there'd be no last item to go and look at. But a rule you can query at any n makes "far out" a perfectly legitimate place to point. That's the rung, and everything in this chapter stands on it.

02What "settles down" actually means

Look at 1/2^n running: 0.5, 0.25, 0.125, 0.0625. Your instinct says it's heading for zero, and mine does too. But I want to know what that sentence actually claims, because as written it claims nothing checkable.

Here's the trouble, and it's one word doing two jobs. "Approaches" is a hedge. You are handed a process that never finishes, and then shown an equals sign. Most people quietly decide the equals sign is a teacher's white lie. It isn't, and we're going to fix that permanently.

There's a second trap, sharper than the first. n → ∞ reads like an instruction to plug infinity in, and you cannot, because ∞ is not in ℕ. There is no a_∞, so whatever a limit is, it is not a value of the sequence. Anyone who thinks the limit is "the last term" is stuck forever, because there is no last term.

So let's build the only honest test available. Draw the candidate number L on a line, and draw a band around it of half-width g. Then ask one question: is there a point from which the sequence is inside the band and never leaves? Drag g thinner and find out.

Drag it thinner and the entry point N slides right. Thinner again, and it slides further out still. But it always exists, and that's the claim. Notice what you never once consulted: a_∞. Only the tail — everything from some N on — was ever asked a question.

That thing you were dragging has a name: the band's half-width is what every textbook calls ε, and the entry point is N. That is the entire ε–N definition of a limit, and you used it before you saw it. Written out: a_n → L means that for every positive band-width you name, however thin, there is some N from which the whole tail sits inside the band. When such an L exists we say the sequence converges, and when none does, it diverges.

Now do 1/2^n → 0 for real, and it takes one line. Name any band-width you like — 0.001, say — and halving from 1 gets under 0.001 by the 10th term, since 1/1024 < 0.001. Name 0.000001 and I need 20 halvings instead. Whatever you name, enough halvings get under it, and once under it the sequence stays under it. So the tail is in the band forever, and the limit is 0.

Read that back and notice how strange it is. L is a number. The sequence never has to arrive, because arriving was never part of the deal.

Now, does every sequence have one of these? Try the band test on something that refuses to cooperate.

(−1)^n is +1, −1, +1, −1, forever. Put your band on +1 and half the terms are a full 2 away, put it on −1 and the other half are, and put it in the middle at 0 and every single term is 1 away. There is no L, because it gets close to +1 infinitely often and close to −1 infinitely often, and lands on neither. So a destination has to be one number, and the sequence has to hold it for good. That's what "converges" was claiming all along.

03★ The infinite sum — the keystone

Now the most important twenty minutes in the chapter, and possibly in the first half of this course. I want to write down 1 + 1/2 + 1/4 + 1/8 + … and ask what it equals.

And I have to stop before I even start, because there is a problem in that line that nobody ever flags. The plus sign has never been defined for infinitely many terms. Not once, in your entire education. Addition is a thing you do to two numbers, and by repetition, to any finite pile of them, and then someone writes a "…" on the end and moves on, and the sleight of hand is never named.

Two opposite failures grow from that one missed inch. Some readers reject the equals sign: "you can never finish, so it can't EQUAL 2." Other readers trust it far too much: "it's just addition, so I can reorder and regroup at will." That second one is the seed of every horror later in analysis. Opposite errors, one unspoken gap.

So here's what I'm about to do, said out loud first. I am about to change what the plus sign means. Watch me do it, because nobody usually admits to it.

Build the partial sums. S_1 = 1, then S_2 = 1 + 1/2 = 1.5, then S_3 = 1.75, then S_4 = 1.875. Every one of those is a perfectly legal finite addition, so nothing illegal has happened. And S_n is a sequence — a function on ℕ, exactly the object we built in Section 1 — so point Section 2 at it. We define the infinite sum to be the destination of the partial sums:

Σ a_k := lim S_n

That's it, and that's the whole move. No new arithmetic was invented, and no addition was performed forever. The Σ with an ∞ on top is not an instruction to do something infinitely many times, it is shorthand for "take the limit of the partial sums."

Now spend the definition. What does 1 + 1/2 + 1/4 + 1/8 + … equal? If it isn't 2, tell me what it is, and I'll knock your answer out. Commit before you press anything.

Most people say "just under 2" or "it never gets there." Here's why every candidate dies, using nothing but your own partial sums S_n = 2 − 1/2^{n−1}, which you can check by hand: 1, 1.5, 1.75, 1.875.

Suppose the answer is some L below 2. Then the gap g = 2 − L is a positive number, and Section 2 says halving drives 1/2^{n−1} under any positive number you name. So by that point the partial sums have already sailed past L, and they're still rising. The sequence didn't land on L, it left L behind. Dead.

Suppose instead the answer is some L above 2. Every single S_n sits below 2, so the sequence never gets within (L − 2) of L, and it fails L's band forever. Dead.

Every candidate you can name dies except exactly one. 2 is the only number left standing. So the sum does not "approach" 2 and it does not "get close to" 2 — it IS 2. The equals sign was never a lie, it was a claim about which number survives. And look at what you never needed: the sum never had to finish.

One more thing falls out of this, and you've been carrying it around since you were about nine years old. What is 0.999…?

0.999… = 9/10 + 9/100 + 9/1000 + …. That's a geometric series, and it's the same machine you just built. The partial sums are 0.9, 0.99, 0.999, and the gap to 1 is 1/10^n, which gets under any band you name. So 0.999… = 1 — not approximately, not "for practical purposes." Equal. Nobody ever justified that notation to you, and you believed it anyway. Now you can prove it yourself.

Hold onto what just happened, because the rest of this course leans on it. An infinite series is a number, not a journey, and a number is something you are allowed to do algebra to.

04Shrinking terms prove nothing

We now have a definition: a series converges when its partial sums have a destination. That leaves exactly one honest question: does every series with shrinking terms have a destination?

Test the smallest natural candidate anyone would write down: 1 + 1/2 + 1/3 + 1/4 + …. It's called the harmonic series, and its terms clearly go to zero. Commit to an answer before you read on. Does it settle, or does it grow forever?

Nearly everyone says settles. It's the most confident wrong belief in this entire topic, and the eye corroborates it — plot the partial sums and they look flat on any screen you own. But watch what a 14th-century monk did to it with no machinery at all.

Group the terms in doubling blocks. 1/3 + 1/4 is bigger than 1/4 + 1/4 = 1/2. The next block, 1/5 + 1/6 + 1/7 + 1/8, beats 4 × 1/8 = 1/2, and the next one beats 8 × 1/16 = 1/2. Every block clears 1/2, and there are infinitely many blocks, so you can pass any number you name. Want to beat 100? Take 200 blocks. There is no ceiling, so there is no destination, and the harmonic series diverges.

And now the sting, which is delicious and also completely honest. It really does run away, but so slowly that no computer will ever show it to you, because to pass 100 you need more terms than there are atoms in the observable universe. The eye was lying and the block argument was not. And notice which one of them you were about to trust.

So the belief is dead: "the terms shrink to nothing, so the sum must settle" is false. Convergence is about the total, not the size of the terms. The terms have to shrink fast enough, and we now have a real example of "not fast enough" to point at.

Which raises the complementary question. If a rising sum can run away, what could possibly stop it? Only a ceiling. So is a ceiling enough?

Look at the two rising sequences side by side. The harmonic has no ceiling, so it runs away, while 1 + 1/2 + 1/4 + … has a ceiling at 2, so it lands. Same shape, opposite fate. The difference is the ceiling and nothing else.

Here's the rule, and it's the one that changes everything downstream. Partial sums of positive terms only ever go up, and a rising sequence has exactly one way to fail: run away. So if it rises and there's a ceiling it never passes, it must land, and that's the ceiling rule. The mirror is also true: falling with a floor means it lands.

Why can't it wobble forever instead? Because rising means it can never go back below where it has already been, so the room it's allowed to live in only ever shrinks. It has exactly one place left to be.

And where does it land? Drag the ceiling line downward in that figure and watch the landing spot rise to meet it, because it lands on the lowest ceiling that still holds. That object has a formal name — the supremum — and you've just watched it, which is more than the name will ever give you.

Now an honest flag, because I don't simplify in silence. This rule is true because the real number line has no holes. That property is called completeness, and I am stating it, not proving it, because it's an axiom of what the real numbers are. I'm telling you plainly so you don't spend an hour hunting for the proof you think you missed.

Take a second to see what you just bought. You can now prove a limit exists before you have the faintest idea what it equals. Books assert "this converges" and "e exists" with no reason at all, and readers quietly conclude that existence is something teachers just say. It isn't, and we'll cash this in twice: once in about five minutes, and once on e itself.

05The same step, forever

Look back at Section 1, where two of those three sequences were built the identical way. 1/2^n multiplies by 1/2 every step, and (5/6)^n multiplies by 5/6 every step. Same construction, different number. That family deserves a name.

But the name is the boring half, so feel the property first, on the die. After each roll, whatever fraction of outcome-strings was still clean, you keep 5/6 of it. Not "you subtract a bit" — you keep a fixed proportion of what's left. So from any point onward, the rest of the story is the whole story, shrunk.

That's a geometric sequence: a, ar, ar², …, with a constant ratio r. Ratio, not difference. A constant difference gives you the arithmetic family, a completely different animal. And the property is worth more than the formula: self-similarity, meaning the tail is a scaled copy of the whole.

Now turn the one knob and watch six personalities fall out of a single number.

At r = 0.5 it collapses to nothing, and at r = 0.9 it collapses slowly but it collapses. At r = 1 it just sits there, at r = 1.02 it creeps up and then explodes, and at r = 1.5 it explodes immediately. At r = −0.5 it flips sign every step while shrinking. You can see the frontier at |r| < 1 long before any formula needs it, so remember that you saw it. When the condition arrives in the next section it should feel like a memory, not a rule.

Let's prove the frontier instead of pointing at it. The claim is r^n → 0 when |r| < 1, and every book says "and of course it does." It is the silent brick under a/(1−r), under e, under Poisson, and under every discount factor in this course, and nobody ever argues it.

First, a prediction, because your gut is wrong here in an interesting way. Take r = 0.9: how many steps to get under 0.001? Most people say thousands, or honestly suspect it never really gets there. The answer is 66, because every step deletes 10% of what's left — a fixed slice of a shrinking thing — so nothing survives.

Now the argument, and watch the order carefully. |r|^n is falling and it's floored at 0, so the ceiling rule's mirror hands us a landing place L ≥ 0. We don't know what L is, but we know it exists, and that's enough to start doing algebra.

Write |r|^{n+1} = |r| · |r|^n. Read the left side: it's the same sequence shifted one seat, so it has the same destination, L. Read the right side: it's the sequence scaled by |r|, so its destination is |r|·L. Both sides are the same sequence, so L = |r|L, and therefore L(1 − |r|) = 0. Since |r| < 1, the bracket isn't zero, so L = 0.

Two small tools got used there and I'm naming them rather than sneaking them past you. Shifting a sequence doesn't move its destination, because a shifted sequence has the same tail. Scaling a sequence scales its destination, because the band scales along with it. Both fall straight out of the band picture in Section 2, which is why they cost nothing.

And now the inch that matters more than the result. Writing "let L be the limit" and then building an equation about L feels illegal, because you're using the answer before you have it. It is legal here, but only because the ceiling rule already proved an L exists. Existence first, algebra second. That order is exactly where careless work manufactures nonsense with a straight face, and we'll watch it happen on purpose in about four minutes.

06The harvest, and the warranty

Time to sum the thing. Take the finite sum first, because nothing infinite needs to happen yet: S_n = a + ar + … + ar^{n−1}.

The standard telling multiplies by r and subtracts, and it looks like a rabbit out of a hat, because you can verify every line and still have no idea why any human would try it. So let's make it inevitable instead of clever. Print the two lists stacked and just look.

There it is. rS_n = ar + ar² + … + ar^n is the same list slid one seat right, and nobody "thought of" multiplying by r. Self-similarity IS the shift — that's what Section 5's property means. So subtract the slid copy away, and every term that appears in both lists cancels itself, and only the two ends survive: S_n − rS_n = a − ar^n. Factor and divide:

S_n = a(1 − r^n) / (1 − r)

Exact. Finite. No limits involved. The trick evaporated into the property. And notice the denominator, because I want you to build a reflex here: that formula dies at r = 1. Look at what happens — you'd be dividing by zero — and the answer there is honestly just na, since every term is a. Whenever you meet a denominator, ask when it dies. It costs you two seconds and saves you an afternoon.

Now take the limit, and every brick is already on the wall. Section 3 says the infinite sum IS lim S_n, and Section 5 says r^n → 0 when |r| < 1. So the r^n term vanishes and we're left with the formula that powers half the "hard" tricks in this course:

S = a / (1 − r), and this holds when |r| < 1.

Check it against the number you already own. 1 + 1/2 + 1/4 + … has a = 1 and r = 1/2, so S = 1/(1 − 1/2) = 2. Same 2 you watched every other candidate lose to.

Now let's break it on purpose, because a formula whose warranty you can't state is a formula that will eventually embarrass you. Put r = 2 in and read it out with a straight face.

1 + 2 + 4 + 8 + … = 1/(1 − 2) = −1. A sum of positive whole numbers is minus one. Let the outrage land for a second, then see what actually happened. The formula never made a claim about r = 2. r^n → 0 was the only reason the −ar^n/(1−r) piece disappeared, and at r = 2 that piece explodes instead of dying. The formula didn't lie — we drove it outside its warranty, and it did what machines do when you do that.

So |r| < 1 is not decoration on the formula. It is the load-bearing reason the formula exists, and that's the difference between memorising it and owning it.

And now the form actually worth keeping, which isn't the fraction at all. Look at the series again with self-similarity in mind. The whole thing is the first term, plus r times the whole thing again.

S = a + rS. One line of algebra: S(1 − r) = a, so S = a/(1−r). Same answer, and this time you didn't need shift-and-subtract at all, because you just wrote down what self-similarity says.

But it's legal only because Section 3 made S a number and Sections 4 and 5 proved it exists. Watch what that same line does at r = 2, where S = 1 + 2S gives S = −1, cheerfully. It isn't false there, it's meaningless — there is no number S to do algebra to. Existence first, algebra second. That habit is what this chapter is really selling.

Keep S = a + rS somewhere you can reach it. It is the direct ancestor of a perpetuity's price C/r in Chapter 25, and the ancestor of the expected-flips state machine E = 1 + …E in Chapter 14. It's every "the rest of the game looks like the start of the game" argument in the course. Whenever you spot self-similarity in the wild, this is the one line it turns into.

One more series before we leave, and it's here to prove a point about method rather than family. Take 1/2 + 1/6 + 1/12 + …, which is Σ 1/(k(k+1)). That's not geometric, and watch us sum it exactly anyway.

Don't start from 1/(k(k+1)), because then the partial fraction looks like a second rabbit. Start from the difference instead: add up d_k = 1/k − 1/(k+1) for k = 1 to n and watch the middle annihilate. The −1/2 kills the +1/2, and the −1/3 kills the +1/3. All that's left standing is S_n = 1 − 1/(n+1), which goes to 1.

And now do the one line of algebra: 1/k − 1/(k+1) = 1/(k(k+1)). The ugly series nobody asked for turns out to be the pretty one you just summed. Nothing came out of a hat, because it was there the whole time, written the other way round. That collapse is called telescoping.

It was the same recipe as everything else: build S_n, look at it, take the limit. Different collapse, identical method, and what you own now is the recipe, not a zoo of tricks.

07Compounding, and the number it lands on

Section 5 gave us repeated multiplication by a fixed ratio. What if the ratio itself depends on n, shrinking as the steps multiply? A smaller step, but many more of them.

Put it in money, since this is a quant course. One dollar, 100% interest, one year: compound once and you have $2. Compound twice — 50% at the half-year, 50% at the end — and you have $2.25, because the second half-year earns interest on the first half's interest. Compound n times and you hold (1 + 1/n)^n.

Now predict, before you look at anything: compound infinitely often. Does the bank go bust?

Quarterly gives $2.4414, monthly gives $2.6130, daily gives $2.7146, and every second gives $2.71828. It doesn't blow up, and it doesn't sit at 1.

Both of those outcomes matter, because your gut had two answers and they contradict each other. Answer (a): 1 + 1/n → 1, and 1 to any power is 1, so the whole thing goes to 1. Answer (b): you're multiplying more and more numbers, every one of them bigger than 1, so it goes to ∞. Both arguments are honest. Both are wrong, and the table kills them both at once.

This is the 1^∞ indeterminate form, and now you can see exactly why it's indeterminate. Two forces are fighting, because the base falls toward 1 while the exponent runs to ∞. The rates decide the winner, and staring at the symbols cannot tell you the rates. That's the most valuable thing this step teaches, and it's worth more than the number it produces.

So it's headed somewhere between 2 and 3. Or is it headed anywhere? Nothing we own yet answers that, and we can't even name the number, because we have no ceiling for it.

To open (1 + 1/n)^n we need a tool for a bracket raised to the n-th power. We have one, except Chapter 1 never named it, so let's earn it in ninety seconds — and it's Chapter 1's counting read out loud.

Multiply out (1+t)(1+t)(1+t) and don't collect terms. You pick one item from each bracket, so the multiplication principle gives 8 words: 111, 11t, 1t1, t11, 1tt, t1t, tt1, ttt. Now ask the only question left, which is how many of those words have exactly two t's. That is verbatim Chapter 1's question: choose 2 of the 3 brackets to donate a t. That's C(3,2) = 3, so count them: 1tt, t1t, tt1. Three.

So the coefficient IS the count, and there's no theorem here, just Chapter 1 read aloud: (1+t)^n = Σ C(n,k) t^k. That's the binomial theorem, and C(n,k) is the number of ways to choose which k brackets donated the t. It's a counting statement, not a formula to memorise.

Now put it on the compounding sequence, and two harvests fall out of one line.

Expand: (1 + 1/n)^n = Σ_{k=0}^{n} C(n,k)/n^k. Rewrite term k and it becomes (1/k!) · (1)(1 − 1/n)(1 − 2/n)…(1 − (k−1)/n). Drag n in that figure and watch the brackets fatten toward 1 while 1/k! sits still, because you are watching 1/k! get uncovered. And you can see why Chapter 1's factorial had to be there: each t came from a different bracket, so the orderings divide out.

Harvest one — it exists. Every bracket is under 1, so term k is smaller than 1/k!. And k! ≥ 2^{k−1} for every k ≥ 1, since a factorial multiplies by 2, 3, 4, … while 2^{k−1} only ever multiplies by 2. So the whole sum is under 1 + Σ_{k≥1} 1/2^{k−1} = 1 + 2 = 3. Look at what just pinned it: a geometric series, and the one you built in Section 6 is literally the ceiling on e. Meanwhile each bracket grows with n and there are more terms, so the sequence rises, and rising plus a ceiling means the ceiling rule applies. It lands. Call the landing place e.

That's the moment this chapter stops being two topics. The series machinery didn't just precede e — it's what makes e exist.

Harvest two — where it lands. Each bracket goes to 1, so term k goes to 1/k!. That gives e = Σ_{k≥0} 1/k! = 1 + 1 + 1/2 + 1/6 + 1/24 + … = 2.718281828…

So e has two faces and they are one number. The slow face is the one compounding lands on, and it's hopeless for computing — (1+1/n)^n at n = 1000 only gets you 2.7169. The fast face is Σ 1/k!, and it nails 2.71828 in six terms. Books hand you both and say "these agree," and readers take it on faith, but you just built the bridge yourself, out of Chapter 1's C(n,k). So if you ever forget it, you can rebuild it at a whiteboard from nothing.

Two honest flags, because I won't simplify in silence. First: letting each bracket go to 1 inside an infinite sum swaps a limit past a sum. That move is not always legal, and it's almost always done quietly, but here the ceiling buys the licence. I'm telling you because this is exactly the kind of step that gets waved through.

Second: nothing in this chapter earns the word "natural" for e. You'll hear e called the natural base and you should notice that nobody said why. That debt gets paid in Chapter 5, where e^x turns out to be the one function equal to its own derivative. Here, e is simply the number compounding lands on, and that's all we've earned, so that's all I'll claim.

08e^x, and the 0.3679 that follows you

One question left, and it pays for a lot of the course. What is e^x?

Not "e multiplied by itself x times." Say that phrase for x = −1 and listen to it. Multiply e by itself minus one times? It's meaningless, and the same goes for x = 0.5 and x = π. This is the same disease as the infinite plus sign in Section 3: exponentiation past whole numbers was never defined for you, and every textbook proceeds as though it had been.

So don't build e^x out of e, build it out of the process that made e. Money at rate x, compounded continuously for one unit of time, is (1 + x/n)^n, so turn the same binomial crank and you get e^x := lim (1 + x/n)^n = Σ x^k/k!. That's the definition. e^x is a process, not repeated multiplication.

And now x = −1 isn't weird at all — it's an account shrinking at 100% a year, continuously. For that one you don't even need the crank. Just flip the fraction.

Watch: (1 − 1/n)^n = ((n−1)/n)^n = 1 / (n/(n−1))^n, and n/(n−1) = 1 + 1/(n−1). So the denominator is (1 + 1/(n−1))^n, which we split as (1 + 1/(n−1))^{n−1} · (1 + 1/(n−1)). The first factor is Section 7's sequence at index m = n−1, so it goes to e, and the second goes to 1. So the whole thing goes to 1/e = 0.367879…. Pure algebra on a fraction, until e walked in wearing a different index.

That step leaned on one more limit law, and the band picture hands it over free: if a_n → L and L ≠ 0, then 1/a_n → 1/L. Cross-check against the other face: Σ (−1)^k/k! = 1 − 1 + 1/2 − 1/6 + 1/24 − … agrees to every digit you care to check.

0.3679 is worth owning outright, and you don't need a scrap of probability machinery to earn it. This is pure Chapter 1 counting.

Throw n balls into n boxes. Each ball has n choices, so the multiplication principle gives n^n arrangements. Now count the arrangements that miss box 1: each ball has n−1 choices, so there are (n−1)^n. Favourable over total is ((n−1)/n)^n = (1 − 1/n)^n, which is exactly the sequence you just did.

Predict first: with as many balls as boxes, surely a given box is nearly certain to get hit?

36.8% of all arrangements leave it empty, and the number barely twitches from n = 10 all the way to n = ∞. Sit with that for a second, because it's genuinely surprising and it's also just (1−1/n)^n wearing a costume.

Here's what you just bought. That is the ancestor of "nothing arrived" in Chapter 12's Poisson distribution, where it shows up as e^{−λ}, and it's the ancestor of the derangement 1/e in Chapter 15. You own it before you've met either name, and you got it with counting alone.

The whole chapter fits in one picture.

You can sum an infinite geometric series, and far more importantly, you can say exactly when you're allowed to, because you know precisely what killed the −ar^n/(1−r) term and what happens when it refuses to die. You can spot self-similarity in the wild and turn it into one line, S = a + rS. You know that shrinking terms prove nothing, which will keep you honest about tails in Chapters 29 and 31. And you own e from both ends, plus 0.3679, built out of nothing but Chapter 1.

But the thing I actually want you to carry is smaller than any of that. A limit is a number, not a journey. That single re-definition is the licence behind every equals sign in the rest of this book — the integral in Chapter 3, Poisson as a binomial limit in Chapter 12, the CLT in Chapter 13, dW² = dt in Chapter 26. And underneath it sits the discipline: prove the thing exists, then do algebra to it. Get that order backwards and one day you'll prove that 1 + 2 + 4 + 8 + … = −1 and not flinch.

There's a bridge already under your feet. We spent this chapter adding up infinitely many shrinking pieces and taking a limit. Now look at a curve and slice the area under it into infinitely many thin rectangles. Every rectangle is a legal finite area, and their sum is a partial sum. And the area under the curve is the limit of those partial sums — which is a definition you already own, pointed at a new picture. That's the integral. Its inverse is the derivative, and the two of them are Chapter 3.