Chapter VIII · The convergence of infinite series and infinite integrals

The series ∑n−s\sum n^{-s}

§181 30 min

The series 1 + 2^(−s) + 3^(−s) + … converges exactly when s > 1, and it becomes the standard yardstick for every series whose terms behave like a power of n.

Why this matters

The geometric series decides series whose terms shrink geometrically. Most series that arise from algebra shrink like a power of nn: 3n+1n3+5\dfrac{3n+1}{n^3+5} behaves like 3/n23/n^2, 1n2+1\dfrac{1}{\sqrt{n^2+1}} like 1/n1/n. To decide those we need to know exactly which powers give convergent series. §181 answers this with the integral test, and the answer is the single most useful fact in the chapter: ∑n−s\sum n^{-s} converges if s>1s > 1 and diverges if s≤1s \le 1. Combined with the comparison theorem D, it decides at a glance every series whose terms are asymptotic to a power of nn.

The idea

For s>0s > 0 the terms n−sn^{-s} decrease, so the integral test applies with ϕ(x)=x−s\phi(x) = x^{-s}. The question becomes: does the area under y=x−sy = x^{-s} from 11 to ξ\xi stay bounded as ξ→∞\xi \to \infty? The integral is easy to compute, and the answer turns on whether ξ1−s\xi^{1-s} tends to 00 or to ∞\infty, that is, on whether s>1s > 1 or s<1s < 1. The borderline s=1s = 1 is the harmonic series, which diverges.

But knowing that a series converges is not the same as being able to sum it in practice. Near the borderline the convergence is extremely slow.

Explore ∑n−s\sum n^{-s} for every ss

Try this. Start at s=1.5s = 1.5, then move ss slowly down to 11 and below. Then set s=1.1s = 1.1 and compare the partial sum at N=105N = 10^5 with the true sum. Finally try s=3s = 3.

∑n=1N1n1.5\sum_{n=1}^{N} \frac{1}{n^{1.5}} for NN up to 10510^5
N (log scale)S_N
  • partial sums S_N
  • 1 + ∫₁ᴺ x⁻ˢ dx, always above S_N
  • Hardy’s bound s/(s − 1)
  • the sum ζ(s)
S100 000=2.60605S_{100\,000} = 2.60605. The series converges, to 2.612382.61238, which is below s/(s−1)=3s/(s-1) = 3. The part still missing after NN terms is about N1−s/(s−1)N^{1-s}/(s-1), so an error below 10−310^{-3} needs roughly N≈106.6N \approx 10^{6.6} terms.
What to notice

For s>1s > 1 the partial sums level off under the ochre curve, below the dashed bound s/(s−1)s/(s-1). For s≤1s \le 1 they keep rising. At s=1.1s = 1.1 the picture looks almost the same as at s=1s = 1: after 10510^5 terms the partial sum is still far below the sum ≈10.58\approx 10.58, and an accuracy of 0.0010.001 would need about 104010^{40} terms. Convergence and practical summability are different things.

Precise statements

At this point in the book xsx^s is defined only for rational ss (irrational powers come in Ch. IX), so Hardy states the result for rational ss. The proof is word for word the same for every real ss once powers are available.

Theorem The series ∑n−s\sum n^{-s} Hardy: §181

The series

1−s+2−s+3−s+⋯+n−s+⋯1^{-s} + 2^{-s} + 3^{-s} + \cdots + n^{-s} + \cdots

converges if s>1s > 1 and diverges if s≤1s \le 1. When s>1s > 1 its sum is less than s/(s−1)s/(s-1).

Together with Theorem D (§173) this gives a test that covers almost every series with “algebraic” terms.

Corollary Comparison with ∑n−s\sum n^{-s}

If vn>0v_n > 0 and nsvnn^s v_n tends to a limit l≠0l \ne 0 (that is, vn∼l n−sv_n \sim l\,n^{-s}), then ∑vn\sum v_n converges if s>1s > 1 and diverges if s≤1s \le 1.

For example, 3n+1n3+5∼3n−2\dfrac{3n+1}{n^3+5} \sim 3n^{-2}, so ∑3n+1n3+5\sum \dfrac{3n+1}{n^3+5} converges; 1n2+1∼n−1\dfrac{1}{\sqrt{n^2+1}} \sim n^{-1}, so ∑1n2+1\sum \dfrac{1}{\sqrt{n^2+1}} diverges. More generally (Examples LXXI, 3), ∑ns/(nt+a)\sum n^s/(n^t + a) with a,t>0a, t > 0 converges exactly when t>1+st > 1 + s.

Proofs

Proof The series ∑n−s\sum n^{-s}
Strategy

Dispose of s≤0s \le 0 (the terms do not tend to 00). For s>0s > 0 apply the integral test to ϕ(x)=x−s\phi(x) = x^{-s}, which is positive, continuous and decreasing. Compute Φ(ξ)=∫1ξx−s dx\Phi(\xi) = \int_1^\xi x^{-s}\,dx and see whether it has a limit. The case s=1s = 1 needs a separate argument, since its integral is a logarithm.

s≤0s \le 0: divergence

Then n−s≥1n^{-s} \ge 1 for every nn, so the terms do not tend to 00 and the series diverges.

The integral for s>0s > 0, s≠1s \ne 1

An antiderivative of x−sx^{-s} is x1−s/(1−s)x^{1-s}/(1-s), so

Φ(ξ)=∫1ξdxxs=ξ1−s−11−s.\Phi(\xi) = \int_1^\xi \frac{dx}{x^s} = \frac{\xi^{1-s} - 1}{1 - s}.

s>1s > 1: convergence, with the bound

Now 1−s<01 - s < 0, so ξ1−s→0\xi^{1-s} \to 0 and Φ(ξ)→1s−1=l\Phi(\xi) \to \dfrac{1}{s-1} = l. By the integral test the series converges, and its sum is less than ϕ(1)+l=1+1s−1=ss−1\phi(1) + l = 1 + \dfrac{1}{s-1} = \dfrac{s}{s-1}.

0<s<10 < s < 1: divergence

Now 1−s>01 - s > 0, so ξ1−s→∞\xi^{1-s} \to \infty and Φ(ξ)→∞\Phi(\xi) \to \infty. The series diverges. (So, more simply, does comparison with ∑1/n\sum 1/n, since n−s≥n−1n^{-s} \ge n^{-1}.)

s=1s = 1 without logarithms

Split ∫12ndx/x\int_1^{2^n} dx/x at 2,4,…,2n−12, 4, \dots, 2^{n-1}. The substitution x=2rux = 2^r u turns ∫2r2r+1dx/x\int_{2^r}^{2^{r+1}} dx/x into ∫12du/u\int_1^2 du/u, the same positive number cc for every rr. So Φ(2n)=nc→∞\Phi(2^n) = nc \to \infty; since Φ\Phi increases, Φ(ξ)→∞\Phi(\xi) \to \infty, and the harmonic series diverges. (This is the integral version of the grouping argument of §77.)

Proof Comparison with ∑n−s\sum n^{-s}
Strategy

Theorem D with un=n−su_n = n^{-s}.

Apply Theorem D

vn/un=nsvn→l≠0v_n/u_n = n^s v_n \to l \ne 0, so ∑vn\sum v_n and ∑n−s\sum n^{-s} converge or diverge together.

Read off the answer

By the theorem, ∑n−s\sum n^{-s} converges exactly when s>1s > 1.

The picture of the integral test shows where the bound s/(s−1)s/(s-1) comes from, and how good it is.

Explore Where s/(s−1)s/(s-1) comes from

Try this. Set s=2s = 2 and slide the slivers left. Then try s=1.2s = 1.2 and s=3s = 3.

ϕ(x)=x−2\phi(x) = x^{-2}: rectangles of height ϕ(ν)\phi(\nu) against the area under the curve
xy
  • rectangles: total area ϕ(1)+⋯+ϕ(n)\phi(1) + \cdots + \phi(n); shaded: ∫1n+1ϕ\int_1^{n+1}\phi
  • the excess slivers
Slivers
ϕ(1)+⋯+ϕ(10)=1.5498\phi(1) + \cdots + \phi(10) = 1.5498, ∫111ϕ=0.90909\int_1^{11}\phi = 0.90909; the slivers have total area 0.64070.6407, which is less than ϕ(1)=1\phi(1) = 1 however large nn is. So the sums stay bounded exactly when the integrals do: here ∫1∞ϕ=1\int_1^\infty \phi = 1, the series converges (to 1.64491.6449), and its sum lies between 11 and ϕ(1)+1=2\phi(1) + 1 = 2.
What to notice

The rectangles exceed the area under the curve, ∫1∞x−s dx=1/(s−1)\int_1^\infty x^{-s}\,dx = 1/(s-1), by the slivers, which fit in the column of height ϕ(1)=1\phi(1) = 1. Hence the sum lies between 1/(s−1)1/(s-1) and 1+1/(s−1)=s/(s−1)1 + 1/(s-1) = s/(s-1). For s=2s = 2 these are 11 and 22, and the true sum is π2/6≈1.645\pi^2/6 \approx 1.645.

Worked examples and exercises

Exercise ●○○ Hardy: Examples LXXI, 2

Show that ∑n−2\sum n^{-2}, ∑n−3/2\sum n^{-3/2} and ∑n−11/10\sum n^{-11/10} converge, with sums at most 22, 33 and 1111 respectively, and that ∑n−1/2\sum n^{-1/2} and ∑n−10/11\sum n^{-10/11} diverge.

Solution

The exponents 22, 32\tfrac32, 1110\tfrac{11}{10} exceed 11, so the series converge, with sums below s/(s−1)s/(s-1): 2/1=22/1 = 2, 32/12=3\tfrac32/\tfrac12 = 3, 1110/110=11\tfrac{11}{10}/\tfrac1{10} = 11. The exponents 12\tfrac12 and 1011\tfrac{10}{11} are at most 11, so those series diverge. (The bounds are crude near s=1s = 1: ∑n−11/10≈10.58\sum n^{-11/10} \approx 10.58.)

Exercise ●○○ Hardy: Examples LXXI, 3

Let a>0a > 0 and t>0t > 0. Show that ∑nsnt+a\sum \dfrac{n^s}{n^t + a} converges if t>1+st > 1 + s and diverges if t≤1+st \le 1 + s.

Hint
Divide the term by ns−tn^{s-t}.

Solution

ns/(nt+a)ns−t=ntnt+a=11+an−t→1\dfrac{n^s/(n^t + a)}{n^{s-t}} = \dfrac{n^t}{n^t + a} = \dfrac{1}{1 + a n^{-t}} \to 1, since n−t→0n^{-t} \to 0 for t>0t > 0. By the Corollary the series behaves like ∑n−(t−s)\sum n^{-(t-s)}, which converges exactly when t−s>1t - s > 1, that is t>1+st > 1 + s. (Hardy’s examples tacitly have t>0t > 0; if t<0t < 0 the term behaves like ns/an^s/a instead.)

Exercise ●●○ Hardy: Examples LXXI, 5

Show that, for every positive integer mm,

1m2+1(m+1)2+1(m+2)2+⋯<m+1m2.\frac{1}{m^2} + \frac{1}{(m+1)^2} + \frac{1}{(m+2)^2} + \cdots < \frac{m+1}{m^2}.
Hint
Apply the bound of the integral test to the series that starts at n=mn = m.
Solution

The integral test, started at x=mx = m instead of x=1x = 1, says that ∑n≥mϕ(n)<ϕ(m)+∫m∞ϕ\sum_{n \ge m}\phi(n) < \phi(m) + \int_m^\infty \phi. With ϕ(x)=x−2\phi(x) = x^{-2}: ϕ(m)=1/m2\phi(m) = 1/m^2 and ∫m∞x−2 dx=1/m\int_m^\infty x^{-2}\,dx = 1/m. So the tail is less than 1m2+1m=m+1m2\dfrac{1}{m^2} + \dfrac1m = \dfrac{m+1}{m^2}. The same picture gives the lower bound ∫m∞x−2 dx=1/m\int_m^\infty x^{-2}\,dx = 1/m: the tail after m−1m - 1 terms is about 1/m1/m, which is why ∑1/n2\sum 1/n^2 needs about 10710^7 terms for 77 correct decimals.

Exercise ●●○ Hardy: Examples LXXI, 6

Show that ∑n=1∞1n2+1<12+14π\displaystyle\sum_{n=1}^\infty \frac{1}{n^2 + 1} < \frac12 + \frac14\pi.

Hint
Use the integral test with ϕ(x)=1/(1+x2)\phi(x) = 1/(1 + x^2) and ∫dx/(1+x2)=arctan⁡x\int dx/(1+x^2) = \arctan x.

Solution

ϕ(x)=1/(1+x2)\phi(x) = 1/(1+x^2) is positive, continuous and decreasing for x≥1x \ge 1. ∫1ξϕ=arctan⁡ξ−14π→12π−14π=14π=l\int_1^\xi \phi = \arctan\xi - \tfrac14\pi \to \tfrac12\pi - \tfrac14\pi = \tfrac14\pi = l. By the integral test the series converges and its sum is less than ϕ(1)+l=12+14π≈1.285\phi(1) + l = \tfrac12 + \tfrac14\pi \approx 1.285. (The sum is about 1.0771.077; the lower bound from the same picture is 14π≈0.785\tfrac14\pi \approx 0.785.) Choose 1/(1+x2)1/(1+x^2) in the widget above to see the picture.

Exercise ●●○ Hardy: Examples LXXI, 9

If ϕ(n)→l>1\phi(n) \to l > 1, show that ∑n−ϕ(n)\sum n^{-\phi(n)} converges; if ϕ(n)→l<1\phi(n) \to l < 1, show that it diverges.

Hint
Pick ss strictly between ll and 11.

Solution

If l>1l > 1, choose ss with 1<s<l1 < s < l. For large nn, ϕ(n)>s\phi(n) > s, so n−ϕ(n)<n−sn^{-\phi(n)} < n^{-s} (because n≥1n \ge 1), and ∑n−s\sum n^{-s} converges. If l<1l < 1, choose ss with l<s<1l < s < 1; for large nn, ϕ(n)<s\phi(n) < s and n−ϕ(n)>n−sn^{-\phi(n)} > n^{-s}, with ∑n−s\sum n^{-s} divergent. The case l=1l = 1 is genuinely undecided: ϕ(n)=1\phi(n) = 1 diverges, ϕ(n)=1+2log⁡log⁡n/log⁡n\phi(n) = 1 + 2\log\log n/\log n converges (Ch. IX).

Exercise ●●● Hardy: Examples LXXI, 1

Without integrating, show that ∫1ξx−s dx→∞\int_1^\xi x^{-s}\,dx \to \infty as ξ→∞\xi \to \infty when s<1s < 1.

Hint
Imitate the case s=1s = 1: split at the powers of 22 and substitute x=2rux = 2^r u.

Solution

For s<1s < 1, substituting x=2rux = 2^r u gives

∫2r2r+1dxxs=2r(1−s)∫12duus=2r(1−s)c,c>0.\int_{2^r}^{2^{r+1}} \frac{dx}{x^s} = 2^{r(1-s)} \int_1^2 \frac{du}{u^s} = 2^{r(1-s)} c, \qquad c > 0 .

Since 1−s>01 - s > 0, 2r(1−s)≥12^{r(1-s)} \ge 1, so each of these pieces is at least cc. Therefore ∫12nx−s dx≥nc→∞\int_1^{2^n} x^{-s}\,dx \ge nc \to \infty, and since the integral increases with ξ\xi it tends to ∞\infty.

Check yourself

Check yourself True or false?

∑1/n1.0001\sum 1/n^{1.0001} converges.

Why

True: s=1.0001>1s = 1.0001 > 1. The sum is below s/(s−1)=10 001s/(s-1) = 10\,001 (it is about 10 000.5810\,000.58), and in practice the series is hopeless to sum by adding terms. Convergence says nothing about speed.

Check yourself True or false?

∑n−s\sum n^{-s} converges for every s>0s > 0, because then n−s→0n^{-s} \to 0.

Why

False. Terms tending to 00 are necessary, not sufficient. For 0<s≤10 < s \le 1 the series diverges, for example ∑1/n\sum 1/\sqrt n. The dividing line is s=1s = 1, not s=0s = 0.

Check yourself Choose one

Which of these series converges?

Why

Find the power of nn that each term behaves like. 1/n(n+1)∼n−11/\sqrt{n(n+1)} \sim n^{-1}: diverges. n/(n2−1)∼n−3/2\sqrt n/(n^2 - 1) \sim n^{-3/2}: converges. n/(n2+1)∼n−1n/(n^2+1) \sim n^{-1}: diverges. 1/(n1/2+n1/3)∼n−1/21/(n^{1/2} + n^{1/3}) \sim n^{-1/2}: diverges (the larger power in the denominator dominates).

Check yourself Enter a number

Hardy’s bound for the sum of ∑n−3/2\sum n^{-3/2} is s/(s−1)s/(s-1). What is it?

Why

32/12=3\tfrac32 / \tfrac12 = 3. The true sum is ζ(32)≈2.612\zeta(\tfrac32) \approx 2.612.

Check yourself Enter a number

The tail ∑n>N1/n2\sum_{n > N} 1/n^2 is about 1/N1/N. Roughly how many terms of ∑1/n2\sum 1/n^2 are needed for an error below 0.0010.001?

Why

1/N<0.0011/N < 0.001 requires N>1000N > 1000. Each extra correct decimal costs ten times as many terms, which is Hardy’s footnote: some ten million terms for seven decimals.

Hardy’s language

Much the most important application of the integral test is to the series 1⁻ˢ + 2⁻ˢ + 3⁻ˢ + … + n⁻ˢ + …, where s is any rational number.

— G. H. Hardy, §181

Summary

Summary
  • ∑n−s\sum n^{-s} converges for s>1s > 1 (with sum below s/(s−1)s/(s-1)) and diverges for s≤1s \le 1.
  • Proof: integral test with ∫1ξx−sdx=(ξ1−s−1)/(1−s)\int_1^\xi x^{-s}dx = (\xi^{1-s} - 1)/(1-s); the harmonic case s=1s = 1 by doubling intervals.
  • If vn∼l n−sv_n \sim l\,n^{-s} with l≠0l \ne 0, ∑vn\sum v_n behaves like ∑n−s\sum n^{-s}: find the dominant power and compare it with 11.
  • Near s=1s = 1 convergence is extremely slow: the tail after NN terms is about N1−s/(s−1)N^{1-s}/(s-1).