Series: Adding Infinitely Many Numbers (and Getting a Finite Answer)

21.1 The Problem of Infinite Addition

Here is a sum that should be impossible:

$$\frac{1}{2} + \frac{1}{4} + \frac{1}{8} + \frac{1}{16} + \cdots = 1.$$

The left side has no last term. You can never stop adding. And yet the right side claims a finite, exact answer. How can adding infinitely many positive numbers produce anything but infinity?

The resolution is the central idea of this chapter, and it is more subtle than it first appears. Addition, as you learned it, is a binary operation: it takes two numbers and returns one. Repeat it and you can add any finite list. But "add infinitely many things" is not a meaningful instruction — there is no last step to reach. So before we can write the equation above, we have to define what an infinite sum even means. The definition we choose is the only one that makes the calculus of the previous chapter pay off.

The Key Insight. An infinite series is not an act of addition you perform; it is a limit you take. You add finitely many terms — getting a partial sum — then ask what those partial sums approach as you include more and more. The series "equals" that limit, or it has no value at all. Every question about infinite addition becomes a question about a sequence, and you already know how sequences behave (Chapter 20).

An infinite series is a formal sum of the terms of a sequence $\{a_n\}$:

$$\sum_{n=1}^\infty a_n = a_1 + a_2 + a_3 + \cdots.$$

This chapter introduces the workhorses: the geometric series (by far the most useful), the harmonic series (the most instructive divergent example), telescoping series, the divergence test, and the algebraic properties of convergent series. The fuller battery of convergence tests — comparison, ratio, root, integral, and alternating — is the business of Chapter 22, and the specialization to power and Taylor series is Chapter 23. Here we build the foundation those chapters stand on.

This is also where calculus finally pays off a debt. Zeno's paradox (the Chapter 3 case study) insisted that motion is impossible because it requires completing infinitely many sub-journeys. We can now say precisely why Zeno was wrong: an infinite sum of shrinking positive terms can be finite.

21.2 The Partial Sum Definition

Fix a sequence $\{a_n\}$. Its $N$-th partial sum is the ordinary finite sum of the first $N$ terms:

$$S_N = \sum_{n=1}^N a_n = a_1 + a_2 + \cdots + a_N.$$

Each $S_N$ is a perfectly legitimate finite addition. As $N$ increases, the partial sums form a new sequence $\{S_N\}$ — and now we are back on the familiar ground of Chapter 20. We define the value of the infinite series to be the limit of that sequence:

$$\sum_{n=1}^\infty a_n := \lim_{N \to \infty} S_N.$$

If this limit exists as a finite number $S$, we say the series converges to $S$. If the limit fails to exist — because the partial sums march off to $\pm\infty$, or oscillate without settling — the series diverges.

The Key Insight. Convergence of the series IS convergence of the sequence of partial sums. There is no separate theory. Whenever you are stuck on a series, retreat to $\{S_N\}$ and ask what the sequence does. Every theorem about sequence limits — limit laws, the squeeze theorem, the monotone convergence theorem — transfers instantly to series.

Geometric Intuition. Picture each term $a_n$ as a horizontal step you take along the number line: forward when $a_n > 0$, backward when $a_n < 0$. The partial sum $S_N$ is where you are standing after $N$ steps. The series converges if and only if your position settles toward a single point — even though you keep taking (ever smaller) steps forever. The harmonic series will show that "smaller and smaller steps" is not enough to keep you from wandering off to infinity.

Three quick cases. The cleanest way to absorb the definition is to compute partial sums directly.

A convergent series. Take $a_n = (1/2)^n$:

$$S_1 = \tfrac12, \quad S_2 = \tfrac34, \quad S_3 = \tfrac78, \quad \ldots, \quad S_N = 1 - \frac{1}{2^N}.$$

(The closed form $S_N = 1 - 2^{-N}$ follows from the geometric formula we derive in §21.3, but you can also verify it by induction.) Since $2^{-N} \to 0$, we get $S_N \to 1$. The series converges to $1$ — the impossible sum from §21.1, now rigorous.

A series diverging to infinity. Take $a_n = 1$ for every $n$. Then $S_N = N$, and $N \to \infty$. Diverges.

A series diverging by oscillation. Take $a_n = (-1)^{n+1}$, so the terms are $1, -1, 1, -1, \ldots$. The partial sums are $S_1 = 1$, $S_2 = 0$, $S_3 = 1$, $S_4 = 0, \ldots$ — they bounce between $1$ and $0$ forever and never settle. No limit, so the series diverges. Note that it does not run off to infinity; "divergence" only means "the partial sums fail to converge."

Check Your Understanding. The partial sums of a series are $S_N = \dfrac{3N}{N+1}$. Does the series converge, and if so, to what? What is its third term $a_3$?

Answer

$\lim_{N\to\infty} \frac{3N}{N+1} = 3$, so the series converges to $3$. To recover a single term, undo the accumulation: $a_3 = S_3 - S_2 = \frac{9}{4} - \frac{6}{3} = \frac{9}{4} - 2 = \frac14$. In general $a_n = S_n - S_{n-1}$ — the term sequence is the "difference" of the partial-sum sequence, the discrete echo of differentiating an accumulation function (Chapter 14).

21.3 Geometric Series — The Most Important Series

A geometric series is one whose terms form a geometric sequence: each term is a fixed multiple of the one before. It has the form

$$\sum_{n=0}^\infty a r^n = a + ar + ar^2 + ar^3 + \cdots,$$

where $a \neq 0$ is the first term and $r$ is the common ratio. Geometric series deserve their own special place because so many other series — repeating decimals, perpetuities, loan balances, probabilities — are geometric series in disguise, and because, uniquely, we can write down their exact sum.

Deriving the sum

The derivation is a beautiful one-line trick. Start from the partial sum and multiply it by $r$:

$$S_N = a + ar + ar^2 + \cdots + ar^N, \qquad r S_N = ar + ar^2 + \cdots + ar^N + ar^{N+1}.$$

Subtract the second from the first. Every interior term cancels — this is a telescoping in disguise — leaving only the endpoints:

$$S_N - rS_N = a - ar^{N+1} \quad\Longrightarrow\quad S_N(1 - r) = a\big(1 - r^{N+1}\big).$$

Provided $r \neq 1$, divide to get the closed form for the partial sum:

$$\boxed{\,S_N = a\,\frac{1 - r^{N+1}}{1 - r}\,}.$$

Now take the limit. The only piece that depends on $N$ is $r^{N+1}$, and from Chapter 20 we know exactly how the powers of $r$ behave.

• If $|r| < 1$: then $r^{N+1} \to 0$, so $S_N \to \dfrac{a}{1 - r}$. The series converges, and

$$\boxed{\,\sum_{n=0}^\infty a r^n = \frac{a}{1 - r} \quad (|r| < 1).\,}$$

• If $|r| > 1$: then $|r^{N+1}| \to \infty$, the partial sums blow up, and the series diverges.
• If $r = 1$: the series is $a + a + a + \cdots$, with $S_N = (N+1)a \to \pm\infty$. Diverges.
• If $r = -1$: the terms are $a, -a, a, -a, \ldots$ and the partial sums oscillate between $a$ and $0$. Diverges.

So a geometric series converges exactly when $|r| < 1$, and then to first term over (one minus ratio).

Geometric Intuition. Why $a/(1-r)$? Look at the bar-stacking picture for $1 + \tfrac12 + \tfrac14 + \cdots$. Lay down a bar of length $1$, then a bar half as long, then half that, marching toward a wall. Each bar covers half of the remaining gap to the wall at position $2$, so the gap never quite closes but shrinks to nothing — the total length is exactly $2 = 1/(1 - \tfrac12)$. The factor $1/(1-r)$ is precisely the "amplification" that summing the whole infinite tail provides over the first term.

Common Pitfall. Many students memorize $\sum r^n = \dfrac{1}{1-r}$ and apply it blindly, but that formula assumes the series starts at $n = 0$ (so the first term is $r^0 = 1$). When a series starts at $n = 1$, the first term is $r$, not $1$, and the sum is $\dfrac{r}{1-r}$. Safest practice: identify the actual first term $a$ and the ratio $r$, then use $\dfrac{a}{1-r}$. Always read off $a$ as the literal first term of the series as written, never as a coefficient ripped from the formula.

Worked examples, graduated

Example 21.3.1 (textbook case). $\displaystyle\sum_{n=0}^\infty \left(\tfrac13\right)^n$. Here $a = 1$, $r = \tfrac13$, and $|r| < 1$, so

$$\sum_{n=0}^\infty \left(\tfrac13\right)^n = \frac{1}{1 - \tfrac13} = \frac{1}{\,2/3\,} = \frac{3}{2}.$$

Example 21.3.2 (negative ratio, shifted start). $\displaystyle\sum_{n=1}^\infty 2\left(-\tfrac14\right)^n$. The constant $2$ factors out (linearity, §21.6). The series inside starts at $n=1$, so its first term is $\left(-\tfrac14\right)^1 = -\tfrac14$, and the ratio is $-\tfrac14$:

$$2 \sum_{n=1}^\infty \left(-\tfrac14\right)^n = 2 \cdot \frac{-\tfrac14}{\,1 - (-\tfrac14)\,} = 2 \cdot \frac{-\tfrac14}{\,5/4\,} = 2 \cdot \left(-\tfrac15\right) = -\frac{2}{5}.$$

The negative ratio means the partial sums approach $-\tfrac25$ from alternating sides, not monotonically — a useful sanity check when you graph them.

Example 21.3.3 (a repeating decimal). What rational number is $0.\overline{3} = 0.3333\ldots$? Write it as the sum of its place values:

$$0.\overline{3} = \frac{3}{10} + \frac{3}{100} + \frac{3}{1000} + \cdots = \sum_{n=1}^\infty \frac{3}{10^n}.$$

This is geometric with first term $a = \tfrac{3}{10}$ and ratio $r = \tfrac{1}{10}$:

$$0.\overline{3} = \frac{3/10}{1 - 1/10} = \frac{3/10}{9/10} = \frac{1}{3}.$$

So the disputed equality "$0.\overline{3} = \tfrac13$," which troubles so many students, is just the geometric series formula. The same machinery turns any repeating decimal into a fraction: $0.\overline{142857} = \dfrac{142857}{999999} = \dfrac17$. (The denominator is all nines because the ratio is $10^{-k}$, where $k$ is the length of the repeating block, and $1 - 10^{-k} = (10^k - 1)/10^k$.)

Real-World Application — Present value of a perpetuity (economics). A perpetuity is a financial instrument that pays a fixed amount $\$C$ every year, forever — British "consols" and some endowment payouts work this way. Money arriving in the future is worth less today, so a payment $C$ arriving $n$ years from now is *discounted* to $C/(1+r)^n$ at annual interest rate $r$. The fair price today is the sum of all discounted future payments: $$PV = \sum_{n=1}^\infty \frac{C}{(1+r)^n} = \frac{C/(1+r)}{1 - 1/(1+r)} = \frac{C}{r}.$$ A perpetuity paying \$1 per year at a 5% discount rate is worth exactly $1/0.05 = \$20$ today. This single formula — a geometric series in disguise — underlies bond pricing, the dividend-discount model for valuing stocks, and the "cap rate" rule of thumb in real estate.

21.4 Telescoping Series

A geometric series collapses because of its multiplicative structure. A telescoping series collapses for a different reason: its terms are built to cancel in pairs, like the sections of a collapsing spyglass.

The model example is

$$\sum_{n=1}^\infty \left(\frac{1}{n} - \frac{1}{n+1}\right).$$

Write out the partial sum and watch the cancellation:

$$S_N = \left(1 - \tfrac12\right) + \left(\tfrac12 - \tfrac13\right) + \left(\tfrac13 - \tfrac14\right) + \cdots + \left(\tfrac1N - \tfrac1{N+1}\right).$$

Every negative piece is annihilated by the positive piece of the next term. Only the very first and very last survivors remain:

$$S_N = 1 - \frac{1}{N+1}.$$

As $N \to \infty$, $\dfrac{1}{N+1} \to 0$, so the series converges to $1$.

The general pattern

If you can express the term as a difference $a_n = b_n - b_{n+1}$ for some auxiliary sequence $\{b_n\}$, then the sum telescopes:

$$S_N = (b_1 - b_2) + (b_2 - b_3) + \cdots + (b_N - b_{N+1}) = b_1 - b_{N+1}.$$

If $b_{N+1} \to L$, the series converges to $b_1 - L$. The whole game is recognizing the difference structure — and the usual tool for revealing it is partial fractions (Chapter 16).

Example 21.4.1. $\displaystyle\sum_{n=1}^\infty \frac{1}{n(n+1)}$. Partial fractions give $\dfrac{1}{n(n+1)} = \dfrac{1}{n} - \dfrac{1}{n+1}$, so this is exactly the model series. $S_N = 1 - \dfrac{1}{N+1} \to 1$. Converges to $1$.

Example 21.4.2. $\displaystyle\sum_{n=1}^\infty \frac{1}{n^2 + 3n + 2}$. Factor the denominator: $n^2 + 3n + 2 = (n+1)(n+2)$. Partial fractions give $\dfrac{1}{(n+1)(n+2)} = \dfrac{1}{n+1} - \dfrac{1}{n+2}$. Telescoping, $b_n = \dfrac{1}{n+1}$, so

$$S_N = b_1 - b_{N+1} = \frac{1}{2} - \frac{1}{N+2} \to \frac{1}{2}.$$

The series converges to $\tfrac12$. Notice the sum is not $1$ this time — the leading survivor is $b_1 = \tfrac12$ because the indexing is shifted.

Common Pitfall. Telescoping does not always leave just two survivors. If the term is $b_n - b_{n+2}$, two strands of cancellation run in parallel and four terms survive: $S_N = b_1 + b_2 - b_{N+1} - b_{N+2}$. Always write out the first three or four terms and the last two before declaring what cancels — do not assume "everything in the middle vanishes" without checking the offset.

21.5 The Harmonic Series — When Shrinking Terms Aren't Enough

The geometric series converges because its terms shrink geometrically (each a fixed fraction of the last). It is tempting to conclude that any series whose terms shrink to zero must converge. The harmonic series is the famous, decisive counterexample:

$$\sum_{n=1}^\infty \frac{1}{n} = 1 + \frac{1}{2} + \frac{1}{3} + \frac{1}{4} + \frac{1}{5} + \cdots.$$

Its terms $\tfrac1n$ march obediently to zero. And yet the harmonic series diverges — its partial sums grow without bound, all the way to infinity. This is one of the genuine surprises of mathematics, and the proof is short enough to commit to memory.

The grouping proof (Oresme, c. 1350)

Group the terms into blocks whose lengths double: $\{1\}$, $\{\tfrac12\}$, $\{\tfrac13, \tfrac14\}$, $\{\tfrac15, \ldots, \tfrac18\}$, $\{\tfrac19, \ldots, \tfrac1{16}\}$, and so on. The trick is to replace every term in a block by the smallest term in that block — which only makes the sum smaller — and watch each block clear the same hurdle:

$$\underbrace{\frac13 + \frac14}_{> \frac14 + \frac14 = \frac12} ,\qquad \underbrace{\frac15 + \frac16 + \frac17 + \frac18}_{> 4 \cdot \frac18 = \frac12} ,\qquad \underbrace{\frac19 + \cdots + \frac1{16}}_{> 8 \cdot \frac1{16} = \frac12} ,\qquad \ldots$$

Each block (the one ending at $\tfrac{1}{2^k}$) contains $2^{k-1}$ terms, each at least $\tfrac{1}{2^k}$, so each block sums to more than $2^{k-1} \cdot \tfrac{1}{2^k} = \tfrac12$. There are infinitely many such blocks, so the partial sums exceed $1 + \tfrac12 \cdot k$ — they grow past any bound. The series diverges.

Geometric Intuition. Compare the harmonic series to the area under $y = 1/x$. The rectangles of width $1$ and height $1/n$ erected on the intervals $[1,2], [2,3], \ldots$ each sit above the curve, so $\sum_{n=1}^N \tfrac1n > \int_1^{N+1} \tfrac{1}{x}\,dx = \ln(N+1)$. Since $\ln(N+1) \to \infty$, the harmonic partial sums must too. This is the integral test in embryo — the rigorous version arrives in Chapter 22 — and it even tells you the rate: the partial sums grow like $\ln N$.

How slowly it diverges

The harmonic series diverges, but only barely. A precise estimate (provable with the integral comparison above) is

$$S_N = \sum_{n=1}^N \frac{1}{n} \approx \ln N + \gamma, \qquad \gamma \approx 0.5772,$$

where $\gamma$ is the Euler–Mascheroni constant. Growth like $\ln N$ is astonishingly slow. To push the partial sum past $100$ you would need roughly $N \approx e^{100 - \gamma} \approx 1.5 \times 10^{43}$ terms — vastly more than the number of atoms in the observable universe. If you summed a billion terms per second since the Big Bang, you would not be close. And yet, given forever, the sum exceeds every finite target. The harmonic series is the picture of "diverges, but you would never guess it from a numerical experiment."

Historical Note. The medieval scholar Nicole Oresme (c. 1320–1382), Bishop of Lisieux, gave the grouping proof above around 1350 — three centuries before calculus existed. It was promptly forgotten and rediscovered by the Bernoulli brothers in the 1680s; Jacob Bernoulli published it, generously crediting his late brother Johann. The constant $\gamma$ was studied by Euler in 1734 and refined by Mascheroni in 1790. Remarkably, no one to this day knows whether $\gamma$ is even rational — it is one of the oldest open questions in mathematics.

Check Your Understanding. The terms of the series $\sum \dfrac{1}{2n}$ also go to zero. Does this series converge or diverge? (Hint: factor.)

Answer

Diverges. Factor out the constant: $\sum_{n=1}^\infty \frac{1}{2n} = \frac12 \sum_{n=1}^\infty \frac{1}{n}$, which is exactly half the harmonic series. Multiplying a divergent series by a nonzero constant cannot rescue it — by the constant-multiple property (§21.6), if the sum converged we could double it to make the harmonic series converge, a contradiction. "Half of infinity" is still infinity.

21.6 The Divergence Test — and Its Crucial Warning

The harmonic series teaches a hard lesson about what shrinking terms can and cannot guarantee. Let us state both directions carefully, because this is the most misremembered result in the chapter.

The Divergence Test ($n$-th term test). If $\displaystyle\lim_{n\to\infty} a_n \neq 0$ (or the limit fails to exist), then $\displaystyle\sum_{n=1}^\infty a_n$ diverges.

This is the cheapest convergence check there is: just look at the terms. If they do not die out to zero, the series cannot possibly converge. The reasoning is the contrapositive of a clean fact:

Necessary condition for convergence. If $\displaystyle\sum a_n$ converges, then $a_n \to 0$.

Why: if the series converges to $S$, then both $S_N \to S$ and $S_{N-1} \to S$. Subtracting, $a_N = S_N - S_{N-1} \to S - S = 0$. So the terms of any convergent series are forced to zero. Flip this around and you get the divergence test.

Warning. The converse is false, and this is the single most important caution in the chapter. $a_n \to 0$ does NOT imply that $\sum a_n$ converges. The harmonic series is the standing counterexample: $\tfrac1n \to 0$, yet $\sum \tfrac1n = \infty$. So the divergence test is a one-way tool. It can only ever prove divergence. If the terms go to zero, the test says nothing — the series might converge (like $\sum 1/n^2$) or diverge (like $\sum 1/n$), and you must reach for a stronger test from Chapter 22. Never write "$a_n \to 0$, therefore the series converges." That single sentence is the most common error in all of series theory.

Example 21.6.1. $\displaystyle\sum_{n=1}^\infty \frac{n}{n+1}$. The terms approach $\dfrac{n}{n+1} \to 1 \neq 0$, so the series diverges by the divergence test. (Intuitively, you are adding numbers that are eventually all close to $1$ — of course the total blows up.)

Example 21.6.2. $\displaystyle\sum_{n=1}^\infty (-1)^n$. The terms $-1, 1, -1, 1, \ldots$ have no limit, so the limit is certainly not $0$, and the series diverges.

Example 21.6.3 (the test is inconclusive). $\displaystyle\sum_{n=1}^\infty \frac{n^2}{n^3 + 1}$. The terms behave like $\dfrac{n^2}{n^3} = \dfrac1n \to 0$. The divergence test is silent — it neither confirms nor denies convergence. (In fact this series diverges, by comparison with the harmonic series — but that requires the comparison test of Chapter 22. The divergence test alone cannot settle it.)

Real-World Application — Convergence diagnostics in numerical software (data science / engineering). Iterative algorithms — solving linear systems, training a model with gradient descent, simulating a PDE — generate a sequence of correction terms, and the total correction is an infinite series. Numerical libraries cannot test the true tail, so they monitor the most recent terms: if the per-step update fails to shrink toward zero, the divergence test guarantees the process is not converging, and the solver halts with a "did not converge" flag. The converse trap bites here too: a shrinking update is necessary but not sufficient, which is exactly why robust solvers also bound the cumulative change, not just the latest step.

21.7 Properties of Convergent Series

Convergent series behave like the finite sums they are limits of — but only for the operations that survive a limit. The safe ones come straight from the limit laws for sequences (Chapter 20) applied to the partial sums.

Suppose $\sum a_n = A$ and $\sum b_n = B$ both converge, and $c$ is a constant. Then:

• Linearity (sum/difference): $\displaystyle\sum (a_n \pm b_n) = \sum a_n \pm \sum b_n = A \pm B$. (The partial sums add: $\sum_{n\le N}(a_n+b_n) = S_N^{(a)} + S_N^{(b)} \to A + B$.)
• Constant multiple: $\displaystyle\sum c\,a_n = c \sum a_n = cA$.
• Finitely many terms don't affect convergence: adding, deleting, or altering any finite number of terms cannot change whether a series converges (though it changes the sum). Convergence is a statement about the tail.

A useful structural fact for non-negative series sits underneath all of this: a series with $a_n \geq 0$ converges if and only if its partial sums are bounded above. The reason is the Monotone Convergence Theorem from Chapter 20 — with non-negative terms, $\{S_N\}$ is increasing, and an increasing sequence converges exactly when it is bounded. This is precisely how we will justify the comparison test in Chapter 22.

Warning — what does NOT carry over. Two tempting "operations" fail for infinite series. (1) You cannot multiply two series term-by-term: $(\sum a_n)(\sum b_n) \neq \sum a_n b_n$ in general; the correct product is the more elaborate Cauchy product $\sum c_n$ with $c_n = \sum_{k=0}^n a_k b_{n-k}$, and even that requires absolute convergence to be valid. (2) You cannot freely rearrange the terms of a series unless it converges absolutely (a notion defined in Chapter 22). For a merely conditionally convergent series — like $1 - \tfrac12 + \tfrac13 - \tfrac14 + \cdots = \ln 2$ — the Riemann Rearrangement Theorem says the terms can be reordered to sum to any real number you like, or even to diverge. Order matters at infinity in a way it never does for finite sums. We return to these subtleties in Chapter 22; for now, only add, scale, and trust the tail.

Math Major Sidebar — The Cauchy criterion. Behind every convergence statement in this chapter is the completeness of $\mathbb{R}$. The Cauchy criterion for series states: $\sum a_n$ converges if and only if for every $\varepsilon > 0$ there is an $N$ such that $\left|\sum_{n=m+1}^{k} a_n\right| < \varepsilon$ for all $k > m \geq N$. In words, every sufficiently late block of terms contributes arbitrarily little. This is just the statement that the partial sums $\{S_N\}$ form a Cauchy sequence, which in the complete space $\mathbb{R}$ is equivalent to convergence (Chapter 20's Bolzano–Weierstrass circle of ideas). Setting $k = m+1$ recovers the necessary condition $a_n \to 0$ as the simplest special case — but the full criterion is genuinely stronger, and it is what fails for the harmonic series: the blocks $\frac{1}{m+1} + \cdots + \frac{1}{2m}$ stay larger than $\tfrac12$ no matter how large $m$ is.

21.8 The p-Series: Where the Boundary Lies

The harmonic series $\sum 1/n$ sits exactly on the knife-edge between convergence and divergence. To see the boundary clearly, compare powers:

$$\sum_{n=1}^\infty \frac{1}{n^p} \qquad (\text{a } \textbf{p-series}).$$

The governing rule — which we will prove with the integral test in Chapter 22 — is strikingly simple:

$$\boxed{\ \sum_{n=1}^\infty \frac{1}{n^p} \ \text{converges if and only if } p > 1.\ }$$

• $p = 1$ is the harmonic series: diverges (just barely).
• $p = 2$ gives $\sum \dfrac{1}{n^2} = \dfrac{\pi^2}{6} \approx 1.6449$ — the celebrated Basel problem, solved by Euler in 1734. (The appearance of $\pi$ in a sum of reciprocal squares is one of the most beautiful surprises in mathematics; we derive it via Taylor and Fourier methods in Chapter 24.)
• $p = 4$ gives $\sum \dfrac{1}{n^4} = \dfrac{\pi^4}{90}$, and every even integer $p$ has a similar closed form built from $\pi$.
• $p = 3$ gives Apéry's constant $\zeta(3) \approx 1.20206$ — known to be irrational (Apéry, 1979) but with no known closed form.
• $p = 1.001$ converges (just barely on the convergent side); $p = 0.999$ diverges.

The function $\zeta(s) = \sum_{n=1}^\infty n^{-s}$ that collects these values is the Riemann zeta function, whose deeper behavior connects to the distribution of prime numbers and the most famous open problem in mathematics, the Riemann hypothesis. For this chapter, the one thing to remember is the boundary: $p > 1$ converges, $p \leq 1$ diverges, and the harmonic series is the borderline case that fails.

Check Your Understanding. Without summing, decide whether each converges: (a) $\sum \dfrac{1}{n^{3/2}}$, (b) $\sum \dfrac{1}{\sqrt{n}}$, (c) $\sum \dfrac{5}{n^2}$.

Answer

(a) $p = \tfrac32 > 1$, converges. (b) $\sqrt n = n^{1/2}$, so $p = \tfrac12 \leq 1$, diverges. (c) $\sum \frac{5}{n^2} = 5\sum \frac{1}{n^2}$ is a constant times a $p = 2$ series, so it converges (to $5 \cdot \pi^2/6$). The constant multiple cannot change convergence (§21.7).

21.9 Computation: Watching Series Converge and Diverge

Python lets you see the partial-sum sequence and check your hand answers — but it also reveals the limits of numerical experiments, which is itself a lesson. Following the textbook's three-tier pattern (analytic, by hand, machine check), we first verify a convergent sum, then watch the harmonic series fail to settle.

import numpy as np

# Tier 3 check of three hand results: geometric, telescoping, and the Basel sum.
def partial_sum(term, N: int) -> float:
    """Sum term(n) for n = 1..N."""
    return float(np.sum([term(n) for n in range(1, N + 1)]))

# Geometric: sum_{n=0}^inf (1/3)^n = 3/2  (Example 21.3.1; shift index to start at 0)
geo = 1.0 + partial_sum(lambda n: (1/3)**n, 30)
print(f"geometric  -> {geo:.6f}  (exact 1.500000)")     # 1.500000

# Telescoping: sum 1/(n(n+1)) = 1  (Example 21.4.1)
tele = partial_sum(lambda n: 1/(n*(n+1)), 10_000)
print(f"telescope  -> {tele:.6f}  (exact 1.000000)")     # 0.999900 -> 1

# Basel p-series: sum 1/n^2 = pi^2/6 (converges, but SLOWLY)
basel = partial_sum(lambda n: 1/n**2, 1000)
print(f"basel(1e3) -> {basel:.6f}  (exact {np.pi**2/6:.6f})")  # 1.643935 vs 1.644934

The geometric and telescoping sums lock onto their exact values almost immediately. The Basel sum is correct but slow: even after a thousand terms the error is about $10^{-3}$, because the tail of $\sum 1/n^2$ decays only like $1/N$. Now watch divergence, which looks deceptively tame:

import numpy as np

# Harmonic partial sums grow like ln N + gamma -- visibly, painfully slowly.
gamma = 0.5772156649
for N in [10, 1_000, 100_000, 10_000_000]:
    S = float(np.sum(1.0 / np.arange(1, N + 1)))
    print(f"N={N:>10}:  S_N = {S:8.4f}   ln N + gamma = {np.log(N)+gamma:8.4f}")
# N=        10:  S_N =   2.9290   ln N + gamma =   2.8799
# N=      1000:  S_N =   7.4855   ln N + gamma =   7.4848
# N= 100000:    S_N =  12.0901   ln N + gamma =  12.0901
# N=  10000000: S_N =  16.6953   ln N + gamma =  16.6953

Computational Note. The harmonic output is a cautionary tale: after ten million terms the partial sum has crawled only to about $16.7$. A naive observer running this experiment might conclude the series "converges to something near $17$" — and be completely wrong. No finite computation can prove divergence here; only the grouping proof of §21.5 can. This is why we insist on rigorous tests rather than numerical eyeballing. Note also a floating-point subtlety: summing $1/n$ from largest $n$ to smallest is more accurate than the natural order, because adding tiny terms last avoids swamping them — a real concern when conditional convergence makes order matter.

21.10 Applications: Finite Answers from Infinite Processes

Series are not an abstraction invented to torment students; they are the natural language for any process that repeats forever with diminishing effect. Three classic settings, spanning physics and economics, show the geometric series doing real work.

Zeno's paradox, finally resolved

Zeno argued that to walk across a room you must first cover half the distance, then half of what remains, then half of that, ad infinitum — infinitely many tasks, so (he claimed) you can never arrive. As a series, the total distance is

$$\frac12 + \frac14 + \frac18 + \cdots = \sum_{n=1}^\infty \frac{1}{2^n} = \frac{1/2}{1 - 1/2} = 1.$$

The infinitely many sub-distances sum to a finite total: you cross the whole room. The same holds for time if your speed is constant — infinitely many shrinking time intervals sum to a finite total time. Calculus does not deny Zeno's infinitely many steps; it shows their sum is finite, which is exactly what Zeno overlooked.

The bouncing ball

A ball is dropped from height $h$ and, on each bounce, rebounds to a fraction $r$ of its previous height ($0 < r < 1$). What is the total vertical distance it travels before coming to rest? It falls $h$, then rises $rh$ and falls $rh$, then rises $r^2 h$ and falls $r^2 h$, and so on:

$$d = h + 2rh + 2r^2 h + 2r^3 h + \cdots = h + 2rh\sum_{n=0}^\infty r^n = h + 2rh \cdot \frac{1}{1-r} = h\cdot\frac{1+r}{1-r}.$$

For a ball dropped from $h = 10$ m with rebound fraction $r = 0.8$, the total distance is $d = 10 \cdot \dfrac{1.8}{0.2} = 90$ m. The ball bounces infinitely many times yet travels a finite total distance — Zeno's paradox made physical. (Even more strikingly, it also comes to rest in finite total time, computable as another convergent series, which is why a real bouncing ball actually stops.)

Compounding and the perpetuity, by another route

The perpetuity value $PV = C/r$ of §21.3 can be reached through calculus as well as through a series, and the agreement is reassuring. Model a fund earning continuous interest at rate $r$ while paying out continuously at rate $W$ per year:

$$\frac{dA}{dt} = rA - W.$$

The fund holds steady (neither growing nor shrinking) when $dA/dt = 0$, i.e. when $A = W/r$. That steady-state balance — the principal whose interest exactly funds the payout — is the same $C/r$ the geometric series produced for the discrete perpetuity. Discrete summation and continuous differential equation tell one story: a perpetual payout of $W$ is "worth" the principal $W/r$. This is the recurring theme of the book in miniature — the same mathematical structure appears across every quantitative field, and the discrete (series) and continuous (calculus) pictures are two views of one truth.

Add to Your Modeling Portfolio. Add an infinite-horizon accumulation to your model — a quantity built from a process that repeats forever with geometric decay — and compute its finite total with the geometric-series formula $a/(1-r)$. Biology: model a drug taken on a fixed schedule; if a fraction $r$ of each dose remains when the next is taken, the long-run steady-state body load is $D/(1-r)$ (dose over one-minus-retention). Identify $r$ and the plateau concentration for your scenario. Economics: value a perpetuity or a growing-dividend stock; compute $PV = \sum C(1+g)^{n}/(1+r)^{n} = C/(r-g)$ for $g < r$, and report what discount rate your model implies. Physics: compute the total distance and total time for a bouncing ball (or a damped oscillation losing a fixed fraction of energy per cycle), showing both are finite despite infinitely many cycles. Data Science: model an iterative process whose per-step correction shrinks by a factor $r$ each round (a geometric error decay); sum the corrections to predict the total convergence "budget," and check it against a numerical run.

21.11 Summary and the Road Ahead

You have turned the impossible-sounding act of "adding infinitely many numbers" into the rigorous, computable act of taking the limit of partial sums. The essential results:

• Definition. $\displaystyle\sum a_n = \lim_{N\to\infty} S_N$, where $S_N = a_1 + \cdots + a_N$. Series convergence is convergence of the partial-sum sequence.
• Geometric series (the most important): $\displaystyle\sum_{n=0}^\infty ar^n = \dfrac{a}{1-r}$ when $|r| < 1$; diverges when $|r| \geq 1$. First term over one-minus-ratio.
• Telescoping series: if $a_n = b_n - b_{n+1}$, then $S_N = b_1 - b_{N+1}$; the sum is $b_1 - \lim b_{N+1}$.
• Harmonic series: $\sum 1/n$ diverges (grouping proof), even though $1/n \to 0$. It grows like $\ln N$ — divergence so slow no computer could detect it.
• p-series: $\sum 1/n^p$ converges $\iff p > 1$. The harmonic case $p=1$ is the boundary.
• Divergence test: if $a_n \not\to 0$, the series diverges. But $a_n \to 0$ does NOT imply convergence — the harmonic series is the eternal counterexample.
• Properties: convergent series are linear (add and scale freely); they do not multiply term-by-term, and only absolutely convergent series may be safely rearranged.

Chapter 22 builds the full convergence toolkit — comparison, limit comparison, ratio, root, integral, and alternating series tests — and the crucial distinction between absolute and conditional convergence that we kept deferring here. Chapter 23 lets the terms depend on a variable $x$, producing power series and Taylor series, which represent functions like $e^x$, $\sin x$, and the normal-curve integral as infinite polynomials. Chapter 24 cashes in those representations, deriving the Basel sum $\pi^2/6$ and Euler's identity $e^{i\pi} + 1 = 0$. The humble geometric series you mastered here is the seed of all of it.

Reflection

The deepest idea in this chapter is also the simplest: an infinite sum is a limit, not an act of addition. Once you grant that, the paradoxes dissolve. Zeno's runner arrives because $\tfrac12 + \tfrac14 + \cdots = 1$. The bouncing ball stops because its total path is finite. And the harmonic series — whose terms shrink to nothing, yet whose sum is infinite — stands as a permanent warning that intuition about infinity must be checked against proof. You now hold the two facts that organize everything to come: a geometric series with $|r| < 1$ converges to $a/(1-r)$, and shrinking terms are necessary but never sufficient. Keep them close. Then turn the page, and learn how to decide any series.

Continue to: Exercises · Quiz · Case Study 1 · Case Study 2 · Key Takeaways · Further Reading