Chapter VIII · The convergence of infinite series and infinite integrals

Finer tests: Abel's theorem and the integral test

§§178–180 40 min

Why comparison with the geometric series is too crude, the necessary condition nun→0n u_n \to 0 for decreasing terms, and Maclaurin's integral test.

Why this matters

The root and ratio tests settle ∑n2/2n\sum n^2/2^n and ∑1/n!\sum 1/n! in a line, but they are silent on ∑1/n2\sum 1/n^2, ∑1/n3/2\sum 1/n^{3/2}, ∑1/(nlog⁡n)\sum 1/(n \log n): on every series whose terms shrink like a power of nn or more slowly. Those are exactly the series that matter most in analysis. §§178–180 explain why the old tests fail there and introduce two new tools: a necessary condition for convergence (Abel’s theorem), and a test that compares a series with an integral (Maclaurin’s integral test). The integral test settles ∑n−s\sum n^{-s} in the next lesson.

The idea

The geometric series is too good a yardstick

Comparison with ∑rn\sum r^n can only prove convergence of series whose terms eventually shrink geometrically. That is extremely fast. Hardy’s example: (2/3)n(2/3)^n at first looks much larger than n−12n^{-12}, but far enough out the geometric sequence is smaller by an astronomical factor.

Explore Geometric decay always wins

Try this. Start with Hardy’s pair r=2/3r = 2/3, k=12k = 12 and find where the curves cross. Then push rr up towards 0.990.99 and kk up to 3030.

rn against n−kr^n \text{ against } n^{-k}, on a logarithmic scale
nlog₁₀ of the term
  • log⁡10rn\log_{10} r^n
  • log⁡10n−k\log_{10} n^{-k}
  • from here on rn<n−kr^n < n^{-k}
At n=296n = 296: rn≈10−52.12r^n \approx 10^{−52.12} and n−k≈10−29.66n^{-k} \approx 10^{−29.66}, so rnr^n is smaller by a factor of about 1022.510^{22.5}. From n=148n = 148 on, rn<n−kr^n < n^{-k} for good: the straight line always wins against the logarithmic curve.
What to notice

On a logarithmic scale rnr^n is a straight line falling at a fixed slope, and n−kn^{-k} is a curve whose slope flattens out like −k/n-k/n. However gentle the line and however steep the curve at first, the line wins. So a series like ∑1/n2\sum 1/n^2, whose terms are not eventually below any rnr^n, can never be handled by a test built on the geometric series.

We need a slower yardstick. The series ∑n−s\sum n^{-s} will be it, but first we must decide for which ss it converges, and for that the geometric series is useless. The new tests assume something the old ones did not: that the terms decrease steadily. This is true of nearly every series that occurs naturally.

A necessary condition: nun→0n u_n \to 0

For a convergent series the terms must tend to 00. If the terms also decrease, much more is true: they must tend to 00 faster than 1/n1/n. The idea is that a decreasing sequence cannot hide its size: the nn terms un+1,…,u2nu_{n+1}, \dots, u_{2n} are all at least u2nu_{2n}, so their sum is at least nu2nn u_{2n}, and that sum is a piece of the tail of a convergent series, which tends to 00.

Explore Does nunn u_n tend to 00?

Try this. Look at nunn u_n for each of the four series, with 300 or 1000 terms. Note in the read-out which ones decrease and which converge.

un=1n2u_n = \frac1{n^2}
nn·uₙ
  • n unn\,u_n
Terms decreasing: yes. Series converges: yes. nun→0n u_n \to 0: yes. (At n=100n = 100: nun=0.01n u_n = 0.01, partial sum 1.6351.635.) Convergent and decreasing, so Abel’s theorem applies: n·uₙ = 1/n → 0, as promised.
What to notice

For 1/n21/n^2 (decreasing, convergent) nun→0n u_n \to 0, as the theorem says. For 1/n1/n it does not, so ∑1/n\sum 1/n cannot converge. The third series converges although nun=1n u_n = 1 infinitely often, because its terms are not decreasing. The fourth decreases and has nun→0n u_n \to 0, yet diverges: the condition is necessary, not sufficient.

Comparing a sum with an area

If ϕ\phi is a decreasing function, the sum ϕ(1)+ϕ(2)+⋯+ϕ(n)\phi(1) + \phi(2) + \cdots + \phi(n) is the area of nn rectangles of width 11 standing over the graph, and ∫1n+1ϕ(x) dx\int_1^{n+1} \phi(x)\,dx is the area under it. The difference consists of thin curved slivers, one per rectangle. Slide them all to the left, into the strip 0≤x≤10 \le x \le 1: because ϕ\phi decreases, they stack one above the other without overlapping, inside a column of height ϕ(1)\phi(1). So the sum and the integral never differ by more than ϕ(1)\phi(1), and one is bounded exactly when the other is.

Explore The integral test as a picture

Try this. With ϕ(x)=1/x\phi(x) = 1/x, increase nn and switch between “in place” and “slide them left”. Then change ss to 22, and try 1/(1+x2)1/(1+x^2).

ϕ(x)=1x\phi(x) = \frac1x: 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(8)=2.7179\phi(1) + \cdots + \phi(8) = 2.7179, ∫19ϕ=2.1972\int_1^{9}\phi = 2.1972; the slivers have total area 0.52060.5206, 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 ∫1ξϕ→∞\int_1^\xi \phi \to \infty as ξ→∞\xi \to \infty, so the series diverges.
What to notice

The slivers always fit inside the dashed column, so their total area stays below ϕ(1)\phi(1). For 1/x1/x both the sum and the integral grow without bound (like log⁡n\log n); for x−2x^{-2} both stay bounded. The sum of 1/(1+n2)1/(1+n^2) is trapped between π/4\pi/4 and 12+π/4\tfrac12 + \pi/4.

Precise statements

“Decreasing” here means never increasing: un+1≤unu_{n+1} \le u_n for every nn.

Theorem Abel’s (or Pringsheim’s) theorem Hardy: §179

If ∑un\sum u_n is a convergent series of positive, decreasing terms, then nun→0n u_n \to 0.

Two warnings, both shown in the widget. Without “decreasing” the conclusion can fail (Examples LXX, 2). And the converse is false: nun→0n u_n \to 0 with decreasing terms does not imply convergence (Examples LXX, 3).

Theorem Maclaurin’s (or Cauchy’s) integral test Hardy: §180

Let ϕ(x)\phi(x) be positive, continuous and decreasing for x≥1x \ge 1, and let

Φ(ξ)=∫1ξϕ(x) dx.\Phi(\xi) = \int_1^{\xi} \phi(x)\,dx .

Then the series ϕ(1)+ϕ(2)+ϕ(3)+⋯\phi(1) + \phi(2) + \phi(3) + \cdots converges if Φ(ξ)\Phi(\xi) tends to a limit ll as ξ→∞\xi \to \infty, and diverges if Φ(ξ)→∞\Phi(\xi) \to \infty. In the first case its sum is less than ϕ(1)+l\phi(1) + l.

Since ϕ>0\phi > 0, the function Φ\Phi increases, so these are the only two possibilities. The proof gives more than the theorem.

Corollary Sum minus integral Hardy: §180

Under the same hypotheses,

∑ν=1n−1ϕ(ν)−∫1nϕ(x) dx\sum_{\nu=1}^{n-1} \phi(\nu) - \int_1^n \phi(x)\,dx

increases with nn and tends to a limit between 00 and ϕ(1)\phi(1).

For ϕ(x)=1/x\phi(x) = 1/x this limit is Euler’s constant γ=0.5772…\gamma = 0.5772\ldots (Examples XC, 1, in Ch. IX).

Proofs

Proof Abel’s theorem
Strategy

Bound nu2nn u_{2n} by a block of the tail of the series, which tends to 00. That handles even indices; odd indices are squeezed between even ones because the terms decrease.

A block of the tail tends to 00

The series converges, so its tails un+1+un+2+⋯u_{n+1} + u_{n+2} + \cdots tend to 00. The block un+1+⋯+u2nu_{n+1} + \cdots + u_{2n} is part of such a tail (positive terms), so it tends to 00 too.

The block is at least nu2nn u_{2n}

It has nn terms, each at least u2nu_{2n} because the terms decrease. So 0≤nu2n≤un+1+⋯+u2n→00 \le n u_{2n} \le u_{n+1} + \cdots + u_{2n} \to 0. Hence 2n u2n=2(nu2n)→02n\,u_{2n} = 2(n u_{2n}) \to 0: the claim holds along even indices.

Odd indices

Since u2n+1≤u2nu_{2n+1} \le u_{2n},

0≤(2n+1) u2n+1≤2n+12n⋅2n u2n→1⋅0=0.0 \le (2n+1)\,u_{2n+1} \le \frac{2n+1}{2n}\cdot 2n\,u_{2n} \to 1 \cdot 0 = 0 .

Conclude

Both the even and the odd terms of the sequence nunn u_n tend to 00, so nun→0n u_n \to 0.

Proof The integral test
Strategy

Measure the gap between each term ϕ(ν−1)\phi(\nu - 1) and the integral over [ν−1,ν][\nu - 1, \nu]. These gaps are positive and add up to at most ϕ(1)\phi(1) (the slivers in the column), so “sum minus integral” converges. Then the sum and the integral are bounded together.

Squeeze ϕ\phi on each unit interval

Because ϕ\phi decreases, ϕ(ν−1)≥ϕ(x)≥ϕ(ν)\phi(\nu - 1) \ge \phi(x) \ge \phi(\nu) for ν−1≤x≤ν\nu - 1 \le x \le \nu.

Define the slivers

Let

vν=ϕ(ν−1)−∫ν−1νϕ(x) dx=∫ν−1ν(ϕ(ν−1)−ϕ(x)) dx.v_\nu = \phi(\nu - 1) - \int_{\nu-1}^{\nu} \phi(x)\,dx = \int_{\nu-1}^{\nu} \big(\phi(\nu-1) - \phi(x)\big)\,dx .

The integrand lies between 00 and ϕ(ν−1)−ϕ(ν)\phi(\nu - 1) - \phi(\nu), so 0≤vν≤ϕ(ν−1)−ϕ(ν)0 \le v_\nu \le \phi(\nu - 1) - \phi(\nu). (This is the area of the sliver over [ν−1,ν][\nu-1, \nu], and the bound says it fits in the horizontal strip between heights ϕ(ν)\phi(\nu) and ϕ(ν−1)\phi(\nu - 1).)

The slivers add up to at most ϕ(1)\phi(1)

The upper bounds telescope: v2+⋯+vn≤ϕ(1)−ϕ(n)≤ϕ(1)v_2 + \cdots + v_n \le \phi(1) - \phi(n) \le \phi(1). So ∑vν\sum v_\nu is a series of positive terms with bounded partial sums, and it converges, to a sum between 00 and ϕ(1)\phi(1).

Rewrite the partial sums

Adding the definitions, v2+⋯+vn=∑ν=1n−1ϕ(ν)−Φ(n)v_2 + \cdots + v_n = \sum_{\nu=1}^{n-1}\phi(\nu) - \Phi(n). This is the Corollary: it increases and tends to a limit in [0,ϕ(1)][0, \phi(1)].

Conclude

Write ∑1n−1ϕ(ν)=Φ(n)+(v2+⋯+vn)\sum_{1}^{n-1}\phi(\nu) = \Phi(n) + (v_2 + \cdots + v_n). The second part converges, so the partial sums of the series converge exactly when Φ(n)\Phi(n) does. Since Φ\Phi increases, Φ(n)\Phi(n) has a limit exactly when Φ(ξ)\Phi(\xi) has one as ξ→∞\xi \to \infty through all real values, and otherwise Φ(ξ)→∞\Phi(\xi) \to \infty. In the convergent case the sum is l+lim⁡(v2+⋯+vn)≤l+ϕ(1)l + \lim(v_2 + \cdots + v_n) \le l + \phi(1). (Hardy notes that the inequality is strict, since the vνv_\nu cannot all reach their upper bounds.)

Worked examples and exercises

Exercise ●○○ Hardy: Examples LXX, 1

Use Abel’s theorem to show that ∑1/n\sum 1/n and ∑1/(an+b)\sum 1/(an + b) diverge, where a>0a > 0 and an+b>0an + b > 0 for all nn considered.

Solution

Both series have positive decreasing terms. For un=1/nu_n = 1/n, nun=1n u_n = 1, which does not tend to 00. For un=1/(an+b)u_n = 1/(an+b), nun=n/(an+b)→1/a≠0n u_n = n/(an + b) \to 1/a \ne 0. If either series converged, Abel’s theorem would force nun→0n u_n \to 0. So both diverge.

Exercise ●●○ Hardy: Examples LXX, 2

Let un=1/nu_n = 1/n when nn is a perfect square and un=1/n2u_n = 1/n^2 otherwise. Show that ∑un\sum u_n converges but nunn u_n does not tend to 00. Why is Abel’s theorem not contradicted?

Hint
Split the series into the terms at squares and the others. Each part can be compared with ∑1/k2\sum 1/k^2.

Solution

The terms at the squares n=k2n = k^2 are 1/k21/k^2, and these form the convergent series ∑1/k2\sum 1/k^2. The remaining terms are some of the terms of ∑1/n2\sum 1/n^2, so their partial sums are bounded by ∑1/n2\sum 1/n^2. Every partial sum of ∑un\sum u_n is at most the sum of the two bounds, so ∑un\sum u_n converges (Theorem B). But nun=1n u_n = 1 whenever nn is a square, which happens for arbitrarily large nn, so nun↛0n u_n \not\to 0. There is no contradiction because the terms do not decrease: u4=14>u3=19u_4 = \tfrac14 > u_3 = \tfrac19.

Exercise ●●● Hardy: Examples LXX, 3

Let u1=1u_1 = 1 and un=1/(nv)u_n = 1/(nv) when 2v−2<n≤2v−12^{v-2} < n \le 2^{v-1} (v=2,3,…v = 2, 3, \dots). Show that the terms decrease and nun→0n u_n \to 0, but ∑un\sum u_n diverges.

Hint
Group the terms into the blocks 2v−2<n≤2v−12^{v-2} < n \le 2^{v-1}. How many terms does a block have, and how small can each be?

Solution

Decreasing. Inside a block 1/(nv)1/(nv) decreases with nn; passing to the next block the factor 1/v1/v drops to 1/(v+1)1/(v+1) and nn increases, so the terms keep decreasing.

nun→0n u_n \to 0. nun=1/vn u_n = 1/v, and v→∞v \to \infty as n→∞n \to \infty (we have v−1≥log⁡2nv - 1 \ge \log_2 n).

Divergence. The block for v≥2v \ge 2 has 2v−22^{v-2} terms, each at least 1/(2v−1v)1/(2^{v-1} v). So its sum is at least 2v−2/(2v−1v)=1/(2v)2^{v-2}/(2^{v-1}v) = 1/(2v). The partial sums up to n=2V−1n = 2^{V-1} are therefore at least 12(12+13+⋯+1V)\tfrac12(\tfrac12 + \tfrac13 + \cdots + \tfrac1V), which tends to ∞\infty with VV because ∑1/v\sum 1/v diverges. Hence the converse of Abel’s theorem is false: nun→0n u_n \to 0 can happen arbitrarily slowly.

Exercise ●●○ Hardy: Examples LXXI, 8

Show that 2n−2<1+12+⋯+1n<2n−12\sqrt n - 2 < 1 + \dfrac{1}{\sqrt2} + \cdots + \dfrac{1}{\sqrt n} < 2\sqrt n - 1 for n≥2n \ge 2.

Hint
Use the rectangles of the integral test on both sides of the curve y=1/xy = 1/\sqrt x.

Solution

Let ϕ(x)=x−1/2\phi(x) = x^{-1/2}, decreasing, with ∫abϕ=2b−2a\int_a^b \phi = 2\sqrt b - 2\sqrt a.

Lower bound. On [ν,ν+1][\nu, \nu+1], ϕ(x)≤ϕ(ν)\phi(x) \le \phi(\nu), with strict inequality for x>νx > \nu; integrating, ∫νν+1ϕ<ϕ(ν)\int_\nu^{\nu+1}\phi < \phi(\nu). Summing for ν=1,…,n\nu = 1, \dots, n: ∑1nϕ(ν)>∫1n+1ϕ=2n+1−2>2n−2\sum_1^n \phi(\nu) > \int_1^{n+1}\phi = 2\sqrt{n+1} - 2 > 2\sqrt n - 2.

Upper bound. On [ν−1,ν][\nu - 1, \nu], ϕ(x)≥ϕ(ν)\phi(x) \ge \phi(\nu), strictly for x<νx < \nu, so ϕ(ν)<∫ν−1νϕ\phi(\nu) < \int_{\nu-1}^\nu \phi. Summing for ν=2,…,n\nu = 2, \dots, n and adding ϕ(1)=1\phi(1) = 1: ∑1nϕ(ν)<1+∫1nϕ=2n−1\sum_1^n \phi(\nu) < 1 + \int_1^n\phi = 2\sqrt n - 1.

Exercise ●●○ Hardy: Examples XC, 1 (from §180)

Show that 1+12+⋯+1n−log⁡n1 + \tfrac12 + \cdots + \tfrac1n - \log n tends to a limit γ\gamma with 0≤γ≤10 \le \gamma \le 1. (You may use ∫1ndx/x=log⁡n\int_1^n dx/x = \log n.)

Hint
Apply the Corollary with ϕ(x)=1/x\phi(x) = 1/x, and account for the last term 1/n1/n.

Solution

By the Corollary with ϕ(x)=1/x\phi(x) = 1/x, dn=∑ν=1n−11ν−log⁡nd_n = \sum_{\nu=1}^{n-1}\frac1\nu - \log n increases and tends to a limit γ\gamma with 0≤γ≤ϕ(1)=10 \le \gamma \le \phi(1) = 1. Our sequence is dn+1/nd_n + 1/n, and 1/n→01/n \to 0, so it has the same limit. (In fact γ=0.5772…\gamma = 0.5772\ldots; it is Euler’s constant.) This also shows how fast ∑1/n\sum 1/n diverges: its partial sums are log⁡n+γ+o(1)\log n + \gamma + o(1).

Check yourself

Check yourself True or false?

If ∑un\sum u_n is a convergent series of positive terms, then nun→0n u_n \to 0.

Why

False without “decreasing”: the series with un=1/nu_n = 1/n at squares and 1/n21/n^2 elsewhere converges, yet nun=1n u_n = 1 at every square. Abel’s theorem uses the decrease to say that u2nu_{2n} is the smallest of the block un+1,…,u2nu_{n+1}, \dots, u_{2n}.

Check yourself True or false?

If unu_n decreases and nun→0n u_n \to 0, then ∑un\sum u_n converges.

Why

False. ∑1/(nlog⁡n)\sum 1/(n \log n) has decreasing terms with nun=1/log⁡n→0n u_n = 1/\log n \to 0, and it diverges (integral test, with ∫dx/(xlog⁡x)=log⁡log⁡x\int dx/(x\log x) = \log\log x). Hardy’s Examples LXX, 3 is an example that avoids logarithms.

Check yourself Choose one

Which series can the integral test be applied to directly, as stated?

Why

The test needs ϕ\phi positive and decreasing. sin⁡2x/x2\sin^2 x/x^2 is not decreasing (it vanishes at multiples of π\pi); compare it with 1/n21/n^2 instead. (−1)n/n(-1)^n/n is not positive. x2−xx2^{-x} increases at first (up to x=1/log⁡2≈1.44x = 1/\log 2 \approx 1.44); the test could be applied from x=2x = 2 on, but the ratio test is far simpler.

Check yourself Enter a number

For ϕ(x)=1/x2\phi(x) = 1/x^2, the integral test bounds ∑n≥11/n2\sum_{n\ge1} 1/n^2 above by ϕ(1)+l\phi(1) + l. What is this bound?

Why

ϕ(1)=1\phi(1) = 1 and l=∫1∞x−2 dx=1l = \int_1^\infty x^{-2}\,dx = 1, so the bound is 22. The true sum is π2/6≈1.645\pi^2/6 \approx 1.645, and the lower bound from the same picture is l=1l = 1.

Check yourself True or false?

For every rr with 0<r<10 < r < 1 and every k>0k > 0, we have rn<n−kr^n < n^{-k} for all sufficiently large nn.

Why

True: nkrn→0n^k r^n \to 0 (Examples XXVII, 7), so nkrn<1n^k r^n < 1 eventually. Even r=0.99r = 0.99 and k=100k = 100 only postpone the moment (to beyond n=100 000n = 100\,000).

Hardy’s language

The geometric series … is not only convergent but very rapidly convergent. The tests derived from comparison with it are therefore naturally very crude, and much more delicate tests are often wanted.

— G. H. Hardy, §178

Summary

Summary
  • Tests built on the geometric series only detect geometric decay; they fail for terms like n−sn^{-s}.
  • Abel’s theorem: positive, decreasing, convergent implies nun→0n u_n \to 0. Both hypotheses matter, and the converse is false.
  • Integral test: for positive decreasing continuous ϕ\phi, ∑ϕ(n)\sum \phi(n) converges exactly when ∫1ξϕ\int_1^\xi \phi tends to a limit ll; the sum then lies between ll and ϕ(1)+l\phi(1) + l.
  • Sum minus integral converges, to a limit between 00 and ϕ(1)\phi(1): the slivers stack inside one column.