Chapter VIII · The convergence of infinite series and infinite integrals

Series of positive and negative terms

§190 20 min

Why a few negative terms change nothing, why infinitely many sign changes change everything, and what tools survive.

Why this matters

Every test in the first part of this chapter (comparison, the ratio and root tests, the integral test, condensation) was proved for series whose terms are all positive. Most series that matter are not like that: 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots, ∑(sin⁡n)/n\sum (\sin n)/n, the power series of the next lessons at negative or complex zz. §190 asks a simple question: when do the old tests still apply, and what new difficulty appears when they do not? The answer organises the rest of the chapter.

The idea

A series of positive terms has increasing partial sums. An increasing sequence has only two possible fates: it is bounded and converges, or it is unbounded and diverges to +∞+\infty. It can never oscillate. That is why the tests of §§173–183 could concentrate on one question, “bounded or not?”.

Now allow negative terms. If only finitely many terms are negative, nothing changes. After the last negative term, the partial sums increase again, and a finite number of terms at the start can only shift the sum by a fixed amount. The same is true if the terms are eventually all negative: then the partial sums eventually decrease.

The real change comes when the signs keep changing for ever. Then the partial sums go up and down without end, and a third fate becomes possible: oscillation.

Explore When do the signs stop changing?

Try this. Start with ∑(n2−10)/n4\sum (n^2 - 10)/n^4 and look at the terms, then the partial sums. Then step through the menu: 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots, ∑(sin⁡n)/n\sum (\sin n)/n, 1−1+1−⋯1 - 1 + 1 - \cdots, ∑sin⁡n\sum \sin n and 1−2+3−⋯1 - 2 + 3 - \cdots.

∑n=1∞n2−10n4\sum_{n=1}^{\infty} \dfrac{n^2-10}{n^4}
nsₙ
  • sₙ after adding a positive term
  • sₙ after adding a negative term
  • from here on, one sign only
Show
The last change of sign is at n=4n = 4. From there on every term is positive, so the partial sums only increase: they must converge or diverge, and cannot oscillate. Here s40=−9.20294s_{40} = -9.20294; the series converges (its tail is a positive series comparable with Σ1/n²).
What to notice

For the first series the first three terms are negative and all the rest positive: in the green region the partial sums only climb, so they converge or diverge, and comparison with ∑1/n2\sum 1/n^2 decides which. For every other series the signs never settle. Two of those still converge, 1−1+1−⋯1 - 1 + 1 - \cdots and ∑sin⁡n\sum \sin n oscillate between bounds, and 1−2+3−⋯1 - 2 + 3 - \cdots swings wider and wider.

This is the content of Hardy’s remark that every condition imposed on unu_n in the earlier tests “may be violated for a finite number of terms”. The tests are about the tail of a series, and only the sign pattern of the tail matters.

Precise statements

Throughout, “positive” includes zero, as it does in Hardy (his footnote to §172). Recall that a series ∑un\sum u_n converges to ss if its partial sums sn=u1+⋯+uns_n = u_1 + \cdots + u_n tend to ss; it diverges if sn→+∞s_n \to +\infty or sn→−∞s_n \to -\infty; and it oscillates otherwise, finitely if sns_n stays bounded and infinitely if it does not (§77).

Theorem Series whose terms are eventually of one sign Hardy: §190

Suppose that un≥0u_n \ge 0 for all n≥n0n \ge n_0. Then ∑un\sum u_n either converges or diverges to +∞+\infty; it cannot oscillate. It converges if and only if the series of positive terms ∑n≥n0un\sum_{n \ge n_0} u_n converges, so every test for series of positive terms may be applied to that tail. The same holds with “un≤0u_n \le 0” and “−∞-\infty”, applying the tests to ∑(−un)\sum (-u_n).

When the signs change infinitely often, the old tests do not apply, but two general facts from Chapter IV still do.

Theorem What survives for every series Hardy: §77 (7), §84

(i) If ∑un\sum u_n converges, then un→0u_n \to 0. (ii) (Cauchy’s principle) ∑un\sum u_n converges if and only if for every δ>0\delta > 0 there is an n0n_0 such that

∣un1+1+un1+2+⋯+un2∣<δwhenever n2>n1≥n0.|u_{n_1 + 1} + u_{n_1 + 2} + \cdots + u_{n_2}| < \delta \qquad \text{whenever } n_2 > n_1 \ge n_0 .

Part (i) only rules series out: ∑(sin⁡n)\sum (\sin n) cannot converge because sin⁡n↛0\sin n \not\to 0. Part (ii) is a complete criterion, but using it means estimating sums in which positive and negative terms cancel, and that is exactly the hard part. The rest of this chapter supplies tools for it: absolute convergence (§191), the alternating series test (§195) and the tests of Abel and Dirichlet (§196).

The same distinction applies to infinite integrals. The tests of §185 for ∫a∞ϕ(x) dx\int_a^\infty \phi(x)\,dx need ϕ(x)≥0\phi(x) \ge 0 only for xx beyond some x0x_0, and for an integrand that is unbounded near x=ax = a they need a fixed sign only on some interval (a,a+δ)(a, a + \delta). The difficult cases are integrands that change sign infinitely often, such as

∫1∞sin⁡xxs dxor∫aAsin⁡(1x−a)dxx−a,\int_1^\infty \frac{\sin x}{x^s}\,dx \qquad\text{or}\qquad \int_a^A \sin\Big(\frac{1}{x - a}\Big)\frac{dx}{x - a},

where the first changes sign as x→∞x \to \infty and the second as x→ax \to a. We return to these in §203.

Proofs

Proof Eventually one sign: no oscillation
Strategy

After n0n_0 the partial sums are monotone. A monotone sequence either converges or tends to infinity (§69). The finitely many earlier terms add a fixed number, which changes neither behaviour.

Split off the first terms

Let C=u1+⋯+un0−1C = u_1 + \cdots + u_{n_0 - 1}, a fixed number (zero if n0=1n_0 = 1). For n≥n0n \ge n_0 write sn=C+tns_n = C + t_n, where tn=un0+⋯+unt_n = u_{n_0} + \cdots + u_n.

The tail sums are increasing

For n≥n0n \ge n_0 we have tn+1−tn=un+1≥0t_{n+1} - t_n = u_{n+1} \ge 0, so tnt_n increases (in Hardy’s wide sense: it never decreases).

Two cases

If tnt_n is bounded above it tends to a limit tt, and then sn→C+ts_n \to C + t. If it is not bounded above, it tends to +∞+\infty (an increasing sequence that exceeds every bound once exceeds it for ever after), and so does sn=C+tns_n = C + t_n.

Conclude

In neither case can sns_n oscillate, and ∑un\sum u_n converges exactly when the series of positive terms ∑n≥n0un\sum_{n \ge n_0} u_n does. For un≤0u_n \le 0 apply this to ∑(−un)\sum(-u_n).

Notice what the proof used: only monotonicity of the tail. Once infinitely many terms have each sign, the partial sums are no longer monotone on any tail, and the argument has nothing to hold on to.

Explore Oscillation, finite and infinite

Try this. Look at the partial sums of 1−1+1−⋯1 - 1 + 1 - \cdots (here r=−1r = -1 in the geometric series). Then move rr slightly below −1-1, and slightly above it.

∑n=0∞rn\sum_{n=0}^{\infty} r^{n}
nsₙ
  • sₙ for rnr^{n}
Plot
After 30 terms s29=0s_{29} = 0. The series oscillates.
What to notice

At r=−1r = -1 the sums jump between 11 and 00 for ever: bounded, but with no limit (finite oscillation). For r<−1r < -1 the swings grow without bound (infinite oscillation). For −1<r<0-1 < r < 0 the signs still alternate for ever, yet the series converges: persistent sign changes allow oscillation but do not force it.

Worked examples and exercises

Exercise A few negative terms ●○○ Hardy: §190

Show that ∑n=1∞n2−10n4\displaystyle\sum_{n=1}^\infty \frac{n^2 - 10}{n^4} converges.

Hint
For which nn is n2−10>0n^2 - 10 > 0?
Hint
For those nn, compare n2−10n4\dfrac{n^2 - 10}{n^4} with 1n2\dfrac{1}{n^2}.

Solution

n2−10>0n^2 - 10 > 0 exactly when n≥4n \ge 4, so the terms are positive from n0=4n_0 = 4 on, and by the theorem we may test the tail ∑n≥4\sum_{n \ge 4} as a series of positive terms. For n≥4n \ge 4, 0<(n2−10)/n4<n2/n4=1/n20 < (n^2 - 10)/n^4 < n^2/n^4 = 1/n^2, and ∑1/n2\sum 1/n^2 converges (§181). By comparison the tail converges, and adding the three negative terms −9-9, −616-\tfrac{6}{16}, −181-\tfrac{1}{81} does not affect convergence.

Exercise Two integrals ●○○ Hardy: §190

Show that ∫1∞3x−7(x+1)3 dx\displaystyle\int_1^\infty \frac{3x - 7}{(x+1)^3}\,dx and ∫011−2xx dx\displaystyle\int_0^1 \frac{1 - 2x}{\sqrt x}\,dx converge, and find their values.

Hint
Where does each integrand change sign? Only the behaviour near ∞\infty (first) and near 00 (second) can cause trouble.
Hint
Put t=x+1t = x + 1 in the first. Split the second as x−1/2−2x1/2x^{-1/2} - 2x^{1/2}.

Solution

First. The integrand is positive for x>7/3x > 7/3, so convergence is decided by a positive integrand, and 0<(3x−7)/(x+1)3<3/(x+1)20 < (3x - 7)/(x+1)^3 < 3/(x+1)^2 there, whose integral to ∞\infty converges. To find the value, put t=x+1t = x + 1: the integrand is (3t−10)/t3=3t−2−10t−3(3t - 10)/t^3 = 3t^{-2} - 10t^{-3}, and

∫2T(3t−2−10t−3) dt=[−3t−1+5t−2]2T→32−54=14.\int_2^T (3t^{-2} - 10t^{-3})\,dt = \Big[-3t^{-1} + 5t^{-2}\Big]_2^T \to \tfrac32 - \tfrac54 = \tfrac14 .

Second. The integrand is unbounded only near x=0x = 0, where it is positive (for x<12x < \tfrac12), and it is less than x−1/2x^{-1/2}, whose integral over (0,1)(0, 1) converges. Its value is ∫ϵ1(x−1/2−2x1/2) dx=[2x1/2−43x3/2]ϵ1→2−43=23\int_\epsilon^1 (x^{-1/2} - 2x^{1/2})\,dx = \big[2x^{1/2} - \tfrac43 x^{3/2}\big]_\epsilon^1 \to 2 - \tfrac43 = \tfrac23.

Exercise Which kind of oscillation? ●○○ Hardy: §77, §190

Find the partial sums of 1−1+1−⋯1 - 1 + 1 - \cdots and of 1−2+3−4+⋯1 - 2 + 3 - 4 + \cdots, and say how each series behaves.

Hint
Treat even and odd nn separately.

Solution

For 1−1+1−⋯1 - 1 + 1 - \cdots: sn=1s_n = 1 for odd nn and 00 for even nn. The sums are bounded but have no limit: finite oscillation.

For 1−2+3−⋯1 - 2 + 3 - \cdots: pairing terms, s2m=(1−2)+⋯+(2m−1−2m)=−ms_{2m} = (1 - 2) + \cdots + (2m - 1 - 2m) = -m, and s2m+1=−m+(2m+1)=m+1s_{2m+1} = -m + (2m + 1) = m + 1. The sums are unbounded in both directions: infinite oscillation. In both cases un↛0u_n \not\to 0, which already rules out convergence.

Exercise Terms tending to zero are not enough ●●○ Hardy: §190

Show that 1+12−13+14+15−16+⋯1 + \tfrac12 - \tfrac13 + \tfrac14 + \tfrac15 - \tfrac16 + \cdots (two plus signs, then one minus, repeating) diverges, although its terms tend to 00 and change sign infinitely often.

Hint
Group the terms in threes. What is the sign and size of each group?
Hint
If the series converged, its partial sums s3ms_{3m} would converge too.

Solution

The kk-th group of three is

gk=13k−2+13k−1−13k>13k−2>13k,g_k = \frac{1}{3k-2} + \frac{1}{3k-1} - \frac{1}{3k} > \frac{1}{3k - 2} > \frac{1}{3k},

since 13k−1>13k\tfrac{1}{3k-1} > \tfrac{1}{3k}. So s3m=g1+⋯+gm>13(1+12+⋯+1m)→∞s_{3m} = g_1 + \cdots + g_m > \tfrac13(1 + \tfrac12 + \cdots + \tfrac1m) \to \infty. If ∑un\sum u_n converged, the subsequence s3ms_{3m} of its partial sums would have the same finite limit. So the series does not converge. In fact it diverges to +∞+\infty, since s3m+1s_{3m+1} and s3m+2s_{3m+2} exceed s3ms_{3m}.

Check yourself

Check yourself True or false?

A series with only finitely many negative terms cannot oscillate.

Why

True. After the last negative term the partial sums increase, so they converge or tend to +∞+\infty. Finitely many terms only shift the sums by a constant.

Check yourself True or false?

If un→0u_n \to 0 and the signs of unu_n change infinitely often, then ∑un\sum u_n converges.

Why

False. The series 1+12−13+14+15−16+⋯1 + \tfrac12 - \tfrac13 + \tfrac14 + \tfrac15 - \tfrac16 + \cdots in the last exercise has both properties and diverges. Sign changes make cancellation possible; they do not guarantee enough of it.

Check yourself Choose one

To decide whether ∑(n2−10)/n4\sum (n^2 - 10)/n^4 converges, which statement is right?

Why

The terms are positive from n=4n = 4 on, so the tail is a series of positive terms and all the earlier tests apply to it. A single change of sign cannot cause oscillation. The bound ∣un∣≤1/n4|u_n| \le 1/n^4 is false: for large nn the terms are close to 1/n21/n^2.

Check yourself Enter a number

For 1−1+1−1+⋯1 - 1 + 1 - 1 + \cdots, what is the partial sum s100s_{100} of the first 100100 terms?

Why

The terms pair off as (1−1)+(1−1)+⋯(1 - 1) + (1 - 1) + \cdots, so every even partial sum is 00, and every odd one is 11. The series oscillates finitely.

Hardy’s language

For now we have to consider the possibility of oscillation as well as of convergence or divergence.

— G. H. Hardy, §190

Summary

Summary
  • Every test for series of positive terms applies to a series whose terms are of one sign from some point on: finitely many exceptions never matter.
  • Such a series converges or diverges; it cannot oscillate, because its tail sums are monotone.
  • When signs change infinitely often, oscillation (finite or infinite) becomes possible, and convergence depends on cancellation.
  • What always remains: un→0u_n \to 0 is necessary, and Cauchy’s principle is necessary and sufficient. The next lessons build practical tests on top of them.