Keep adding terms and one of two things happens: the running total creeps toward a finite number, or it wanders off without bound. This chapter is about telling the two apart — and why shrinking terms alone don't decide it.
A series is what you get when you try to add infinitely many numbers at once: a₁ + a₂ + a₃ + ⋯, never stopping. That's a strange thing to ask a calculator for — you can't literally perform infinitely many additions. So the whole subject is really about giving that "⋯" an honest meaning.
The trick is one you already know from Chapter 2: don't add them all at once. Add the first few, watch where the running total is heading, and let "the sum" be that destination. Everything in this chapter — geometric sums, the surprising harmonic series, the tests, even eˣ written as an infinite polynomial — is a variation on that single move.
You can't add infinitely many terms in one go. But you can add the first one, then the first two, then the first three — and keep a running total as you march down the list. That running total is the whole idea. Give it a name: the nth partial sum,
S₁ = a₁, S₂ = a₁ + a₂, S₃ = a₁ + a₂ + a₃, … Sₙ = a₁ + ⋯ + aₙ
Notice what just happened. A pile of infinitely many terms became a plain sequence S₁, S₂, S₃, … — exactly the kind of object Chapter 2 taught us to chase with ε and N. The "infinite sum" is defined as wherever this sequence of running totals goes. If a programmer built it, they'd call scanLeft and hand you back the stream of prefixes.
Below is that stream, made visible. Each dot is a partial sum Sₙ plotted against how many terms you've added. Pick a series and watch the running total either home in on a line — or keep climbing.
That reframing is the engine of the chapter, so let's say it plainly and then use it everywhere.
Here is the definition, and it's almost anticlimactic after Chapter 2:
The series Σ aₙ converges exactly when its sequence
of partial sums S₁, S₂, S₃, … converges to a finite limit S.
Then we write Σ aₙ = S. Otherwise the series diverges.
So there is nothing genuinely new to define. "Does this infinite sum have a value?" is the very same question as "does this sequence of running totals have a limit?" — and you already have the machinery. The sum S isn't reached by adding forever; it's the destination the finite sums approach, the same way √2 was the destination of 1, 1.4, 1.41, … back in Chapter 1.
Because it's a sequence in disguise, you can compute it like one. A partial sum is just a left fold over a prefix; the whole stream of partial sums is a scanLeft. Building it lazily lets you peek at S₂₀ without committing to the infinite tail.
The lazy stream is the perfect mental model: the terms are an infinite list, and scanLeft hands you the running totals one at a time. Convergence is just the question of whether that stream of totals settles down. Time to look at the two most important cases — one that always settles, one that famously doesn't.
Start with 1 and keep multiplying by the same factor r each step: 1 + r + r² + r³ + ⋯. This is the friendliest infinite sum there is, and it has a clean closed form. The classic picture: ½ + ¼ + ⅛ + 1⁄16 + ⋯ = 1 — cut a chocolate bar in half, then the leftover in half, forever; the pieces reassemble into exactly one whole bar.
Why? There's a one-line algebra trick. Let Sₙ = 1 + r + ⋯ + rⁿ⁻¹. Multiply by r and subtract:
Sₙ − r·Sₙ = 1 − rⁿ ⟹ Sₙ = (1 − rⁿ) / (1 − r)
Now let n → ∞. If |r| < 1, the term rⁿ shrinks to zero (each step scales it down), so Sₙ homes in on a finite limit:
1 + r + r² + r³ + ⋯ = 1 / (1 − r) (when |r| < 1)
Plug in r = ½: the sum is 1/(1 − ½) = 2, i.e. 1 + ½ + ¼ + ⋯ = 2 (drop the leading 1 and you're back to the chocolate bar equalling 1). But if |r| ≥ 1 the terms don't shrink — rⁿ holds steady or blows up — and the partial sums run off to infinity. Set the walker above to Geometric and drag the r slider: watch the target line 1/(1−r) move, and Sₙ chase it. Push r toward 1 and the target flees upward; the sum still converges, just ever more slowly.
|r| < 1 ⟹ converges to a₁/(1−r); |r| ≥ 1 ⟹ diverges. That single number decides everything — no other test needed.Here's the trap everyone falls into once. Consider 1 + ½ + ⅓ + ¼ + ⅕ + ⋯ — the harmonic series. The terms 1/n get smaller and smaller, sliding all the way down to 0. Surely a sum of things that vanish must settle somewhere finite?
It doesn't. The harmonic series diverges — its partial sums crawl upward without bound, past 10, past 100, past any ceiling you name. They just do it excruciatingly slowly (you need over 10⁴³ terms to pass 100). Flip the walker to Harmonic and watch: no target line, no settling — just a slow, relentless climb labelled diverges — slowly.
The proof is a gorgeous grouping trick. Bracket the terms in runs whose lengths double:
1 + ½ + (⅓ + ¼) + (⅕ + ⅙ + ⅐ + ⅛) + ⋯
Each bracket is at least ½: for instance ⅓ + ¼ > ¼ + ¼ = ½, and ⅕ + ⅙ + ⅐ + ⅛ > 4 × ⅛ = ½. So the sum keeps picking up another ½, again and again, forever. Adding ½ infinitely many times can only run off to infinity.
0" is a hurdle a convergent series must clear — but clearing it guarantees nothing. The harmonic series clears it and diverges anyway. Whether the terms shrink fast enough is the real question, and it's exactly what the convergence tests measure.Most series don't hand you a tidy formula like the geometric one. So instead of computing the limit, we ask a cheaper question: does a finite limit exist? Three workhorse tests answer it.
The nth-term test (a quick rejection). If the terms aₙ do not shrink to 0, the series has no hope — it diverges, full stop. This is the harmonic warning read backwards: term-goes-to-zero is required, so failing it is an instant disqualification. But heed the one-way arrow: passing this test (terms → 0) proves nothing, as the harmonic series just showed.
The comparison test (lean on a series you know). If 0 ≤ aₙ ≤ bₙ and the bigger series Σ bₙ converges, then the smaller Σ aₙ — trapped underneath — must converge too. And the reverse: if a smaller series diverges, anything larger diverges with it. This is how the harmonic series condemns its cousins: Σ 1/√n has terms bigger than 1/n, so it diverges too.
The ratio test (measure the shrink rate). Look at how each term compares to the one before: L = lim |aₙ₊₁ / aₙ|. This is the "am I secretly geometric?" test.
The ratio test is a programmer's favourite because it's local and mechanical: compute one ratio, take a limit, read off the verdict. It's tailor-made for series with factorials and powers — which is precisely where we're headed next.
Here's where series stop being a curiosity and become a superpower. Instead of summing fixed numbers, let the terms carry a variable x: c₀ + c₁x + c₂x² + c₃x³ + ⋯. That's a power series — an "infinite polynomial." For each x you plug in, it's an ordinary numeric series that either converges (giving a value) or doesn't.
The headline example is the exponential:
eˣ = 1 + x + x²/2! + x³/3! + ⋯ = Σₙ₌₀ xⁿ / n!
Run the ratio test on it: the ratio of consecutive terms is |x| / (n+1), which marches to 0 for every x. So this series converges everywhere — the factorial in the denominator crushes any power of x. The function eˣ literally is this sum; the same trick writes sin, cos, and log as infinite polynomials too. This is how a computer actually evaluates them — truncate the series, sum a handful of terms, stop when the tail is smaller than your tolerance.
The general recipe — read off the coefficients of a function's power series from its derivatives — is the Taylor series, and it's a course of its own. We're only planting the flag here: infinite series are the bridge from "adding numbers" to "representing functions."
1 − 1 + 1 − 1 + ⋯, mostly getting away with it through sheer genius. It couldn't last. In the 1820s Augustin-Louis Cauchy insisted a series only means something once its partial sums converge, and Niels Henrik Abel warned that divergent series were "an invention of the devil." Their reckoning — pinning "sum" to the limit of partial sums — is the definition we used all chapter, and the opening act of the rigor crisis that Chapter 7 tells in full.One last twist, and it's a subtle one. Take the harmonic series and flip the sign of every other term:
1 − ½ + ⅓ − ¼ + ⅕ − ⅙ + ⋯ = ln 2 ≈ 0.693
This alternating harmonic series converges — set the walker to Alternating harmonic and watch the partial sums zig-zag inward, closing on ln 2. But look what changed. The terms 1/n are the same as before; only the alternating signs are new. Strip the signs away — sum 1 + ½ + ⅓ + ⋯ — and you're back to the diverging harmonic series.
So this convergence is fragile: it survives only because positive and negative terms keep cancelling. We call that conditional convergence — the series converges, but the series of absolute values does not. Contrast a series that converges even after you make every term positive (like the geometric Σ (½)ⁿ): that's absolute convergence, the sturdy kind.
Why care? Absolutely convergent series behave like finite sums — you can reorder the terms freely and the total is unchanged. Conditionally convergent ones are treacherous: Riemann proved you can rearrange the alternating harmonic series to sum to any number you like — 0, 42, even ∞ — just by choosing the order you pick positive and negative terms. The cancellation is load-bearing, and shuffling it demolishes the answer. For a programmer: absolute convergence is order-independent and safe to parallelize; conditional convergence is a sum where evaluation order silently changes the result.
Σ|aₙ| converges — is robust: reorder at will, the sum holds. Conditional — Σ aₙ converges but Σ|aₙ| diverges — leans entirely on cancellation, and rearranging can send it anywhere.