Chapter VIII · The convergence of infinite series and infinite integrals

Series of positive terms and the comparison test

§§171–174 35 min

Why a series of positive terms can only converge or diverge to infinity, how to compare one such series with another, and Cauchy's root test.

Why this matters

Given an infinite series, the first question is whether it converges at all. In Chapter IV Hardy answered it for the geometric series and a few relatives. §§171–174 start the systematic theory, with series whose terms are all positive. They are the easiest case, and the most important one: the question for a series with terms of both signs, or complex terms, is usually settled by looking at the positive series of their absolute values (§191). Everything in this lesson rests on one fact about increasing sequences and one idea, comparison.

The idea

Write sn=u0+u1+⋯+uns_n = u_0 + u_1 + \cdots + u_n for the partial sums. If every un≥0u_n \ge 0, each new term can only push sns_n up or leave it where it is. An increasing sequence has only two possible fates (§69): either it is bounded, and then it converges, or it is not, and then it tends to +∞+\infty. Oscillation, the third possibility in Chapter IV, cannot happen.

Explore Two positive series that look alike

Try this. Compare the partial sums of ∑1/n\sum 1/n (blue) and ∑1/n2\sum 1/n^2 (ochre). Switch to 300 and then 1000 terms.

∑n=1∞1n\sum_{n=1}^{\infty} \frac1n compared with ∑n=1∞1n2\sum_{n=1}^{\infty} \frac{1}{n^2}
nsₙ
  • sₙ for 1n\frac1n
  • sₙ for 1n2\frac{1}{n^2}
  • the sum s
Plot
After 100 terms s100=5.187377518s_{100} = 5.187377518. The series diverges. The comparison series converges to 1.6449341.
What to notice

Both sequences of partial sums only climb; neither can turn back. The ochre one levels off below 1.6451.645, so it is bounded and converges. The blue one keeps climbing, slowly (about log⁡n\log n), but with no ceiling, so it diverges. The terms tend to 00 in both cases, so that alone decides nothing.

So to show that a positive series converges, it is enough to find some ceiling for its partial sums. The cheapest way to find a ceiling is to sit the series under another one whose sum we know. If vn≤unv_n \le u_n for every nn and ∑un=s\sum u_n = s, then every partial sum of ∑vn\sum v_n is at most the corresponding partial sum of ∑un\sum u_n, hence at most ss. That is the comparison test. Its most useful yardstick is the geometric series ∑rn\sum r^n with 0≤r<10 \le r < 1, which converges to 1/(1−r)1/(1-r).

How do we tell whether vn≤rnv_n \le r^n? Take nn-th roots: vn≤rnv_n \le r^n is the same as vn1/n≤rv_n^{1/n} \le r. This gives Cauchy’s root test.

Explore Comparison and the root test on a gallery of series

Try this. Start with n2/2nn^2/2^n under each test: compare it with (3/4)n(3/4)^n, then look at vn1/nv_n^{1/n}. Then choose 1/n21/n^2 and 1/n1/n under the root test, and (n/(n+1))n(n/(n+1))^n.

∑n=1∞vn,vn=n22n\sum_{n=1}^{\infty} v_n,\quad v_n = \frac{n^2}{2^n}
nvₙ/uₙ
  • vn/unv_n/u_n
  • the limit
Test
At n=50n = 50: vn/un=3.9208×10−6v_n / u_n = 3.9208 × 10⁻⁶. Converges. The ratio vn/un=n2(2/3)nv_n/u_n = n^2(2/3)^n tends to 0, so it is bounded: vn≤K(3/4)nv_n \le K(3/4)^n. The geometric series converges, so by Theorem C so does ∑vn\sum v_n.
What to notice

For n2/2nn^2/2^n the roots settle near 12\tfrac12, comfortably below the red line at 11, so the series converges. For 1/n21/n^2 and 1/n1/n the roots creep up towards 11 from below: the test is silent, although one series converges and the other diverges. For (n/(n+1))n(n/(n+1))^n the roots also approach 11, yet the terms tend to 1/e1/e, so the series diverges for a much simpler reason.

Precise statements

From now on “positive” includes zero, as it does in Hardy, and ∑un\sum u_n means u0+u1+u2+⋯u_0 + u_1 + u_2 + \cdots (or u1+u2+⋯u_1 + u_2 + \cdots when that is more natural). Hardy also writes

∑mnuν=um+um+1+⋯+un.\sum_{m}^{n} u_\nu = u_m + u_{m+1} + \cdots + u_n .

The four basic facts, all proved in Chapter IV (§77), are restated in §173.

Theorem No oscillation Hardy: A, §173

A series of positive terms either converges or diverges to +∞+\infty. It cannot oscillate.

Theorem Bounded partial sums Hardy: B, §173

A series of positive terms ∑un\sum u_n converges if and only if there is a number KK such that u0+u1+⋯+un<Ku_0 + u_1 + \cdots + u_n < K for every nn.

Theorem The comparison theorem Hardy: C, §173

Let ∑un\sum u_n and ∑vn\sum v_n be series of positive terms and KK a constant.

  1. If ∑un\sum u_n converges and vn≤Kunv_n \le K u_n for every nn, then ∑vn\sum v_n converges, and ∑vn≤K∑un\sum v_n \le K \sum u_n.
  2. If ∑un\sum u_n diverges and vn≥Kunv_n \ge K u_n for every nn, with K>0K > 0, then ∑vn\sum v_n diverges.

The conclusions about convergence and divergence remain true if the inequality holds only for n≥n0n \ge n_0. The inequality between the sums need not.

Corollary Comparison by limits Hardy: D, §173

If un>0u_n > 0, vn>0v_n > 0 and un/vnu_n / v_n tends to a limit other than zero as n→∞n \to \infty, then ∑un\sum u_n and ∑vn\sum v_n both converge or both diverge.

Comparing with the geometric series gives the first two tests (§174).

Theorem Cauchy’s test Hardy: §174, 1–3

Let ∑vn\sum v_n be a series of positive terms.

  1. If vn≤Krnv_n \le K r^n for all sufficiently large nn, where r<1r < 1, the series converges.
  2. (Cauchy’s test.) If vn1/n≤rv_n^{1/n} \le r for all sufficiently large nn, where r<1r < 1, the series converges.
  3. If vn1/n≥1v_n^{1/n} \ge 1 for infinitely many values of nn, the series diverges.

Notice what (2) asks: not merely vn1/n<1v_n^{1/n} < 1, but vn1/n≤rv_n^{1/n} \le r for a fixed rr below 11. The roots must stay away from 11. For 1/n21/n^2 every root is less than 11, but the roots approach 11, and the test says nothing.

Proofs

Proof Theorems A and B
Strategy

The partial sums of a positive series form an increasing sequence, and Chapter IV classified increasing sequences completely: bounded ones converge, unbounded ones tend to ∞\infty.

The partial sums increase

sn−sn−1=un≥0s_{n} - s_{n-1} = u_n \ge 0, so s0≤s1≤s2≤⋯s_0 \le s_1 \le s_2 \le \cdots.

If they are bounded, they converge

If sn<Ks_n < K for all nn, then by §69 the increasing sequence sns_n tends to a limit, at most KK. So the series converges.

If they are not bounded, they tend to ∞\infty

If no KK bounds them, then for any Δ\Delta some sNs_N exceeds Δ\Delta, and since the sequence increases, sn≥sN>Δs_n \ge s_N > \Delta for every n≥Nn \ge N. That is exactly sn→∞s_n \to \infty.

Conclude

These two cases exhaust the possibilities, which proves A. And convergence happens exactly in the first case, which proves B.

Proof The comparison theorem C
Strategy

Find a ceiling for the partial sums of ∑vn\sum v_n using the ceiling we already have for ∑un\sum u_n, then apply Theorem B. For divergence, run the same inequality the other way.

Partial sums inherit the inequality

If vν≤Kuνv_\nu \le K u_\nu for every ν\nu, adding up gives tn=v0+⋯+vn≤K(u0+⋯+un)=Ksnt_n = v_0 + \cdots + v_n \le K(u_0 + \cdots + u_n) = K s_n.

A ceiling for tnt_n

Since uν≥0u_\nu \ge 0, sns_n increases to its limit ss, so sn≤ss_n \le s. Hence tn≤Kst_n \le Ks for every nn. By Theorem B, ∑vn\sum v_n converges, and letting n→∞n \to \infty in tn≤Kst_n \le Ks gives ∑vn≤K∑un\sum v_n \le K\sum u_n.

Divergence

If instead vν≥Kuνv_\nu \ge K u_\nu with K>0K > 0 and ∑un\sum u_n diverges, then tn≥Ksnt_n \ge K s_n and sn→∞s_n \to \infty, so tn→∞t_n \to \infty.

Only large nn matter

If the inequality holds only for n≥n0n \ge n_0, apply the argument to the series that begin at n0n_0. Removing the finitely many earlier terms changes every partial sum by the same fixed amount, so it changes neither convergence nor divergence (§77). It does change the sums, so ∑vn≤K∑un\sum v_n \le K\sum u_n can fail: the discarded terms of ∑vn\sum v_n may be enormous.

Proof Comparison by limits, D
Strategy

A non-zero limit pins the ratio between two positive constants for large nn, which gives comparison inequalities in both directions.

Squeeze the ratio

Let un/vn→lu_n/v_n \to l. Since the terms are positive, l≥0l \ge 0, and by hypothesis l>0l > 0. Taking δ=12l\delta = \tfrac12 l in the definition of a limit, 12l<un/vn<32l\tfrac12 l < u_n/v_n < \tfrac32 l for all n≥n0n \ge n_0.

Two comparisons

So vn<(2/l) unv_n < (2/l)\,u_n and vn>(2/3l) unv_n > (2/3l)\,u_n for n≥n0n \ge n_0. If ∑un\sum u_n converges, the first inequality and Theorem C give convergence of ∑vn\sum v_n. If ∑un\sum u_n diverges, the second gives divergence.

Why l≠0l \ne 0 is needed

If l=0l = 0 only the first inequality survives: vnv_n may be much larger than unu_n. For un=1/n2u_n = 1/n^2 and vn=1/nv_n = 1/n the ratio tends to 00, one series converges and the other diverges.

Proof Cauchy’s test
Strategy

Translate the condition on vn1/nv_n^{1/n} into a comparison with a geometric series.

Convergence (parts 1 and 2)

If vn≤Krnv_n \le K r^n for n≥n0n \ge n_0 with 0≤r<10 \le r < 1, Theorem C applies with the convergent series ∑rn\sum r^n. If vn1/n≤rv_n^{1/n} \le r, then raising both (positive) sides to the power nn gives vn≤rnv_n \le r^n, which is part 1 with K=1K = 1.

Divergence (part 3)

If vn1/n≥1v_n^{1/n} \ge 1 then vn≥1v_n \ge 1. If this happens for infinitely many nn, the terms do not tend to 00. But the terms of a convergent series must tend to 00 (§77), since vn=sn−sn−1→s−s=0v_n = s_n - s_{n-1} \to s - s = 0. So the series diverges.

Worked examples and exercises

Exercise ●○○ Hardy: Examples LXVIII, 5

Show that ∑n−p\sum n^{-p} converges for every integer p≥2p \ge 2, starting from the convergent telescoping series ∑1/(n(n+1))\sum 1/(n(n+1)).

Hint
The ratio of 1/n21/n^2 to 1/(n(n+1))1/(n(n+1)) is (n+1)/n(n+1)/n. What does Theorem D say?
Hint
For p>2p > 2, compare n−pn^{-p} with n−2n^{-2} directly.

Solution

First, 1n(n+1)=1n−1n+1\dfrac{1}{n(n+1)} = \dfrac1n - \dfrac1{n+1}, so ∑1N1n(n+1)=1−1N+1→1\sum_1^N \dfrac{1}{n(n+1)} = 1 - \dfrac{1}{N+1} \to 1: the series converges.

For p=2p = 2: 1/n21/(n(n+1))=n+1n→1≠0\dfrac{1/n^2}{1/(n(n+1))} = \dfrac{n+1}{n} \to 1 \ne 0, so by Theorem D ∑1/n2\sum 1/n^2 converges too. (Even more simply, 1/n2≤2/(n(n+1))1/n^2 \le 2/(n(n+1)) since n+1≤2nn+1 \le 2n.)

For p>2p > 2: n−p≤n−2n^{-p} \le n^{-2} for n≥1n \ge 1, so Theorem C gives convergence. Hardy’s own route uses ∑1/(n(n+1)⋯(n+p−1))\sum 1/(n(n+1)\cdots(n+p-1)), which also telescopes, and n(n+1)⋯(n+p−1)∼npn(n+1)\cdots(n+p-1) \sim n^p. Divergence for p=1p = 1 was proved in §77; for p≤0p \le 0 the terms do not tend to 00.

Exercise ●○○ Hardy: Examples LXVIII, 7

Let m1<m2<m3<⋯m_1 < m_2 < m_3 < \cdots be positive integers. Show that ∑rmn\sum r^{m_n} converges if 0≤r<10 \le r < 1 and diverges if r≥1r \ge 1. For example, 1+r+r4+r9+⋯1 + r + r^4 + r^9 + \cdots.

Hint
How large must mnm_n be, given that the mm‘s are distinct positive integers in increasing order?

Solution

By induction mn≥nm_n \ge n: m1≥1m_1 \ge 1, and mn+1>mn≥nm_{n+1} > m_n \ge n forces mn+1≥n+1m_{n+1} \ge n + 1. If 0≤r<10 \le r < 1 then rmn≤rnr^{m_n} \le r^n, and ∑rn\sum r^n converges, so by Theorem C so does ∑rmn\sum r^{m_n}. If r≥1r \ge 1 every term is at least 11, so the terms do not tend to 00 and the series diverges. The example is the case mn=(n−1)2m_n = (n-1)^2, shifted to start at n=1n = 1 with m1=0m_1 = 0; the first term r0=1r^0 = 1 does not affect convergence.

Exercise ●○○ Hardy: Examples LXVIII, 12

If ∑un\sum u_n is a convergent series of positive terms, show that ∑un2\sum u_n^2 and ∑un/(1+un)\sum u_n/(1 + u_n) converge.

Hint
The terms of a convergent series tend to 00, so eventually un≤1u_n \le 1.

Solution

Since un→0u_n \to 0, there is an n0n_0 with un≤1u_n \le 1 for n≥n0n \ge n_0; then un2=un⋅un≤unu_n^2 = u_n \cdot u_n \le u_n. By Theorem C (for large nn) ∑un2\sum u_n^2 converges. For the second series no such care is needed: 1+un≥11 + u_n \ge 1, so un/(1+un)≤unu_n/(1+u_n) \le u_n for every nn.

Exercise ●●○ Hardy: Examples LXVIII, 13

If un≥0u_n \ge 0 and ∑un2\sum u_n^2 converges, show that ∑un/n\sum u_n/n converges.

Hint
Use 2ab≤a2+b22ab \le a^2 + b^2 with a=una = u_n and b=1/nb = 1/n.

Solution

Since (un−1/n)2≥0(u_n - 1/n)^2 \ge 0, we get 2un/n≤un2+1/n22u_n/n \le u_n^2 + 1/n^2. Both ∑un2\sum u_n^2 and ∑1/n2\sum 1/n^2 converge, so their sum does, and by Theorem C (with K=12K = \tfrac12) so does ∑un/n\sum u_n/n. Note that we never needed to know how large unu_n is compared with 1/n1/n: the inequality covers both cases at once.

Exercise ●●○ Hardy: Examples LXVIII, 15

Prove by contradiction that ∑1/n\sum 1/n diverges: suppose it converges to SS and split it into its odd and even terms.

Hint
The even terms 12+14+16+⋯\tfrac12 + \tfrac14 + \tfrac16 + \cdots are half of the whole series.
Hint
Compare the odd terms 1,13,15,…1, \tfrac13, \tfrac15, \dots with the even terms 12,14,16,…\tfrac12, \tfrac14, \tfrac16, \dots one by one.

Solution

Suppose ∑1/n=S\sum 1/n = S. The odd-term series O=1+13+15+⋯O = 1 + \tfrac13 + \tfrac15 + \cdots and the even-term series E=12+14+⋯E = \tfrac12 + \tfrac14 + \cdots converge by Theorem C, since their terms are terms of the original series and their partial sums are at most SS. Adding partial sums, O+E=SO + E = S. Also E=12∑1/n=12SE = \tfrac12 \sum 1/n = \tfrac12 S, so O=12S=EO = \tfrac12 S = E.

But O−E=(1−12)+(13−14)+(15−16)+⋯O - E = (1 - \tfrac12) + (\tfrac13 - \tfrac14) + (\tfrac15 - \tfrac16) + \cdots, a series of strictly positive terms whose first term is already 12\tfrac12. So O−E≥12>0O - E \ge \tfrac12 > 0, a contradiction. Hence ∑1/n\sum 1/n diverges.

Check yourself

Check yourself True or false?

A series of positive terms whose terms tend to 00 must converge.

Why

False. un→0u_n \to 0 is necessary for convergence, never sufficient: ∑1/n\sum 1/n and ∑1/n\sum 1/\sqrt n diverge although their terms tend to 00. What Theorem A guarantees is only that such a series cannot oscillate.

Check yourself True or false?

If vn≤unv_n \le u_n for all n≥100n \ge 100 and ∑un=s\sum u_n = s, then ∑vn\sum v_n converges and ∑vn≤s\sum v_n \le s.

Why

False, although only half of it fails. ∑vn\sum v_n does converge, since only large nn matter for convergence. But the first 99 terms of ∑vn\sum v_n are not controlled at all; take v1=106v_1 = 10^6 and vn=unv_n = u_n afterwards. The inequality between the sums needs vn≤unv_n \le u_n for every nn.

Check yourself Choose one

Suppose vn>0v_n > 0 and vn1/n<1v_n^{1/n} < 1 for every nn. What follows?

Why

Cauchy’s test needs vn1/n≤rv_n^{1/n} \le r for a fixed r<1r < 1. Both vn=1/(n+1)v_n = 1/(n+1) and vn=1/(n+1)2v_n = 1/(n+1)^2 have vn1/n<1v_n^{1/n} < 1 for every n≥1n \ge 1 and vn→0v_n \to 0, but the first diverges and the second converges. Their roots tend to 11, so no fixed r<1r < 1 works.

Check yourself Choose one

Does ∑n=1∞3n+1n3+5\displaystyle\sum_{n=1}^{\infty} \frac{3n + 1}{n^3 + 5} converge?

Why

Theorem D needs a limit other than zero. Against 1/n21/n^2 the ratio is (3n3+n2)/(n3+5)→3(3n^3 + n^2)/(n^3 + 5) \to 3, so the series behaves like ∑1/n2\sum 1/n^2 and converges. The second option uses a limit 00, from which nothing follows. The root test would fail here (the roots tend to 11).

Check yourself Enter a number

For vn=(n2n+1)nv_n = \left(\dfrac{n}{2n+1}\right)^n, what is lim⁡vn1/n\lim v_n^{1/n}? (So Cauchy’s test gives convergence.)

Why

vn1/n=n/(2n+1)=1/(2+1/n)→12v_n^{1/n} = n/(2n+1) = 1/(2 + 1/n) \to \tfrac12. Indeed vn1/n<12v_n^{1/n} < \tfrac12 for every nn, so we may take r=12r = \tfrac12 in Cauchy’s test.

Hardy’s language

The theory of the convergence of series is comparatively simple when all the terms of the series considered are positive.

— G. H. Hardy, §172

Summary

Summary
  • The partial sums of a positive series increase, so the series converges exactly when they are bounded; otherwise it diverges to ∞\infty. It never oscillates.
  • Comparison: a positive series below a convergent one converges; one above a divergent one diverges. Only large nn matter for convergence, but not for the size of the sum.
  • If un/vnu_n/v_n tends to a limit other than zero, the two series behave alike. A limit zero proves nothing.
  • Cauchy’s (root) test: vn1/n≤r<1v_n^{1/n} \le r < 1 for large nn gives convergence. If the roots tend to 11, the test is silent.
  • Always check first whether the terms tend to 00.