Chapter VIII · The convergence of infinite series and infinite integrals

Conditionally convergent series

§§193–194 35 min

Series that converge only through cancellation; why both their halves diverge, why comparison tests fail for them, and how rearranging their terms can produce any sum at all.

Why this matters

Some series converge although the series of their moduli does not. The standard example is 1−12+13−14+⋯1 - \tfrac12 + \tfrac13 - \tfrac14 + \cdots, which converges (to log⁡2\log 2, as Chapter IX will show), while 1+12+13+⋯1 + \tfrac12 + \tfrac13 + \cdots diverges. §193 and §194 explain what is special about such series: their convergence rests entirely on cancellation, so the familiar tools stop working. You cannot compare them with other series, you cannot pick out some of their terms, and, most strikingly, you cannot reorder them without risking a different sum.

The idea

Recall from the previous lesson that every series is the difference of its positive part and its negative part, ∑un=∑vn−∑wn\sum u_n = \sum v_n - \sum w_n. For 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots the positive part is 1+13+15+⋯1 + \tfrac13 + \tfrac15 + \cdots and the negative part is 12+14+16+⋯\tfrac12 + \tfrac14 + \tfrac16 + \cdots. Both diverge. The series converges only because the order of the terms makes the two infinite supplies cancel each other step by step.

Now imagine you are free to choose the order. You have an unlimited supply of positive terms and an unlimited supply of negative terms, both shrinking to zero. To steer the partial sums towards any number TT: add positive terms until you are above TT, then negative terms until you are below it, and repeat. Because both supplies are unlimited, you never get stuck. Because the terms shrink, the overshoots shrink.

Explore Hit any target

Try this. With “chase a target” selected, set the target to 1.51.5, then to −1-1, then to 33, and watch the partial sums of the rearranged series. Try the other two series in the menu as well, and finally build a rearrangement yourself with “by hand”.

Rearranging 1−12+13−14+⋯1 - \tfrac12 + \tfrac13 - \tfrac14 + \cdots
terms usedpartial sum
  • after a positive term
  • after a negative term
  • original sum 0.69315
  • your target
Order
After 300 terms (250 positive, 50 negative) the partial sum is 1.492881.49288. Both kinds of term keep being used, and the sums cross the target ever more closely, because the terms themselves tend to 0.
What to notice

For 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots and 1−12+13−⋯1 - \tfrac{1}{\sqrt2} + \tfrac{1}{\sqrt3} - \cdots the partial sums zigzag around any target, closer and closer, and both kinds of term keep being used: every term is used exactly once, so this is a genuine rearrangement with sum TT. For the absolutely convergent series 1−122+⋯1 - \tfrac{1}{2^2} + \cdots the same rule gets stuck, because each half has a finite total.

Precise statements

Definition Conditional convergence Hardy: §193

A series ∑un\sum u_n is conditionally convergent if it converges but ∑∣un∣\sum |u_n| diverges.

Since ∑∣un∣\sum |u_n| is a series of positive terms, “diverges” here means ∑∣un∣=+∞\sum |u_n| = +\infty. Every convergent real series is therefore either absolutely or conditionally convergent.

Theorem Both halves diverge Hardy: §193

If ∑un\sum u_n is conditionally convergent, then its positive part ∑vn\sum v_n and its negative part ∑wn\sum w_n both diverge to +∞+\infty. Its sum is the limit of VN−WNV_N - W_N, a difference of two quantities that both tend to infinity.

The consequences that Hardy draws in §193 and §194:

  • Selections fail. A series made of some of the terms need not converge (the positive terms alone diverge).
  • Comparison fails. Knowing ∣vn∣≤∣un∣|v_n| \le |u_n| and that ∑un\sum u_n converges conditionally tells us nothing about ∑vn\sum v_n. For example 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots converges, while 12+13+14+⋯\tfrac12 + \tfrac13 + \tfrac14 + \cdots diverges, although each of its terms is numerically smaller. The proof of comparison for positive series needs the terms of the larger series to be positive, and the conclusion ∣∑v∣≤∑∣v∣≤∑∣u∣|\sum v| \le \sum |v| \le \sum |u| only helps when ∑∣un∣\sum |u_n| is finite. Any test for conditional convergence must therefore look at how the terms cancel, not just at their sizes.
  • Rearrangements fail. The proof of §192 used the convergence of ∑vn\sum v_n and ∑wn\sum w_n separately, and now we know both diverge. The theorem itself fails too, as Riemann showed.
Theorem Riemann’s rearrangement theorem

If ∑un\sum u_n is a conditionally convergent real series and TT is any real number, the terms of ∑un\sum u_n can be rearranged to give a series that converges to TT. They can also be rearranged so that the new series diverges to +∞+\infty, or to −∞-\infty.

Hardy only mentions this theorem (referring to Bromwich for the proof), but the proof is just the game above made precise.

Proofs

Proof Both halves diverge
Strategy

Rule out the other two possibilities: both halves converging, and exactly one converging.

They cannot both converge

If ∑vn\sum v_n and ∑wn\sum w_n both converged, then so would ∑(vn+wn)=∑∣un∣\sum (v_n + w_n) = \sum |u_n|, contradicting conditional convergence.

They cannot converge one at a time

Suppose ∑wn\sum w_n converges to WW while ∑vn\sum v_n diverges. Then sN=VN−WN≥VN−Ws_N = V_N - W_N \ge V_N - W, and VN→+∞V_N \to +\infty, so sN→+∞s_N \to +\infty, contradicting the convergence of ∑un\sum u_n. The case with vv and ww exchanged is the same with signs reversed.

Conclude

Both are series of positive terms, so “not convergent” means “divergent to +∞+\infty”.

Proof Riemann’s rearrangement theorem (sum TT)
Strategy

List the positive terms p1,p2,…p_1, p_2, \dots and the sizes of the negative terms q1,q2,…q_1, q_2, \dots in their original order. Both lists tend to 00 (because un→0u_n \to 0), and both have infinite sums. Build the new series greedily: positive terms while the partial sum is at most TT, negative terms while it exceeds TT.

Each stage ends

Suppose at some stage the partial sum is ≤T\le T and we start adding the next unused positive terms. The unused pkp_k still have an infinite sum (removing finitely many terms from a divergent positive series leaves it divergent), so after finitely many of them the partial sum exceeds TT. Similarly, a stage of negative terms ends after finitely many terms.

Every term is used

Since each stage ends, the stages alternate for ever, so infinitely many positive and infinitely many negative terms are taken, always in their listed order. Hence every pkp_k and every qkq_k is used exactly once. (Terms equal to 00, if any, can be slipped in one at a time after each stage.) The new series is a rearrangement of ∑un\sum u_n.

The overshoots tend to zero

When a positive stage ends, the partial sum has just crossed TT upwards with the term pkp_k, so it lies in (T,T+pk](T, T + p_k]. During the following negative stage it decreases but stays above TT until the last step, which leaves it in [T−qj,T][T - q_j, T]. So at every moment after the stage ending with pkp_k (or qjq_j), the distance from TT is at most the size of the term that ended the latest stage.

Conclude

The terms that end the stages are pkp_k and qjq_j with k,j→∞k, j \to \infty, and pk,qj→0p_k, q_j \to 0. Hence the partial sums of the rearranged series tend to TT.

To diverge to +∞+\infty instead, aim at a moving target: take positive terms until the sum exceeds 11, then one negative term, then positive terms until the sum exceeds 22, then one negative term, and so on (exercise 3 below).

Rearrangements need not be clever to change the sum. Hardy’s example takes two positive terms of 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots for every negative one: 1+13−12+15+17−14+⋯1 + \tfrac13 - \tfrac12 + \tfrac15 + \tfrac17 - \tfrac14 + \cdots.

Explore Two up, one down

Try this. Choose “fixed pattern” with 22 positive and 11 negative term per block, and look at 5000 terms. Then try 11 and 22, and 44 and 11. Finally switch the series to 1−12+13−⋯1 - \tfrac{1}{\sqrt 2} + \tfrac{1}{\sqrt 3} - \cdots with the pattern 22 and 11.

Rearranging 1−12+13−14+⋯1 - \tfrac12 + \tfrac13 - \tfrac14 + \cdots
terms usedpartial sum
  • after a positive term
  • after a negative term
  • original sum 0.69315
  • limit of this pattern
Order
After 300 terms (200 positive, 100 negative) the partial sum is 1.037231.03723. Every term is used eventually, so this is a genuine rearrangement. Its sum is 1.039721.03972 =log⁡2+12log⁡(2/1)= \log 2 + \tfrac12\log(2/1) (a classical formula; Hardy, Ex. LXXVIII. 7, treats the pattern 2 : 1), not the original sum.
What to notice

For the alternating harmonic series the pattern pp positive, qq negative converges to log⁡2+12log⁡(p/q)\log 2 + \tfrac12 \log(p/q) (a classical formula, quoted here; the exercise below proves the case 22 and 11): about 1.0401.040 for 22 and 11, which is 32log⁡2\tfrac32 \log 2, and about 0.347=12log⁡20.347 = \tfrac12\log 2 for 11 and 22. For the square-root series the pattern 22 and 11 does not converge at all: its partial sums climb to infinity.

Worked examples and exercises

Exercise Two positive terms to one negative ●●● Hardy: Examples LXXVIII, 7

Let s=1−12+13−⋯s = 1 - \tfrac12 + \tfrac13 - \cdots. Show that the rearrangement 1+13−12+15+17−14+⋯1 + \tfrac13 - \tfrac12 + \tfrac15 + \tfrac17 - \tfrac14 + \cdots converges to s+12log⁡2s + \tfrac12 \log 2, which equals 32log⁡2\tfrac32\log 2.

Hint
Write Hm=1+12+⋯+1mH_m = 1 + \tfrac12 + \cdots + \tfrac1m. Express the sum of the first nn odd reciprocals, and of the first nn even reciprocals, in terms of HH.
Hint
Use Hm−log⁡m→γH_m - \log m \to \gamma (Euler’s constant), which follows from the integral test of §180.

Solution

The sum of the first nn even reciprocals is 12+14+⋯+12n=12Hn\tfrac12 + \tfrac14 + \cdots + \tfrac1{2n} = \tfrac12 H_n, and so the sum of the first nn odd reciprocals is H2n−12HnH_{2n} - \tfrac12 H_n. The first 3n3n terms of the new series contain the first 2n2n positive terms and the first nn negative ones:

t3n=(H4n−12H2n)−12Hn.t_{3n} = \big(H_{4n} - \tfrac12 H_{2n}\big) - \tfrac12 H_n .

Write Hm=log⁡m+γ+ϵmH_m = \log m + \gamma + \epsilon_m with ϵm→0\epsilon_m \to 0. Then

t3n=log⁡4n−12log⁡2n−12log⁡n+(γ−12γ−12γ)+o(1)=log⁡4−12log⁡2+o(1)→32log⁡2.t_{3n} = \log 4n - \tfrac12 \log 2n - \tfrac12\log n + (\gamma - \tfrac12\gamma - \tfrac12\gamma) + o(1) = \log 4 - \tfrac12 \log 2 + o(1) \to \tfrac32 \log 2 .

Also t3n+1t_{3n+1} and t3n+2t_{3n+2} differ from t3nt_{3n} by at most two terms, which tend to 00, so the whole series converges to 32log⁡2\tfrac32\log 2. By the same method s=lim⁡(H2n−Hn)=log⁡2s = \lim (H_{2n} - H_n) = \log 2, so the new sum is s+12log⁡2s + \tfrac12\log 2. (Hardy reaches the same answer by writing t3n−s2nt_{3n} - s_{2n} as a sum close to 12∫12dx/x\tfrac12\int_1^2 dx/x.)

Exercise A rearrangement that diverges ●●○ Hardy: Examples LXXVIII, 8

Show that 1+13−12+15+17−14+⋯1 + \tfrac{1}{\sqrt3} - \tfrac{1}{\sqrt2} + \tfrac{1}{\sqrt5} + \tfrac{1}{\sqrt7} - \tfrac{1}{\sqrt4} + \cdots diverges to +∞+\infty, although it is a rearrangement of the convergent series 1−12+13−⋯1 - \tfrac{1}{\sqrt2} + \tfrac{1}{\sqrt3} - \cdots.

Hint
Look at the kk-th block of three terms, 14k−3+14k−1−12k\tfrac{1}{\sqrt{4k-3}} + \tfrac{1}{\sqrt{4k-1}} - \tfrac{1}{\sqrt{2k}}.
Hint
Each of the two positive terms exceeds 14k\tfrac{1}{\sqrt{4k}}.

Solution

The kk-th block satisfies

14k−3+14k−1−12k>24k−12k=(1−12)1k.\frac{1}{\sqrt{4k-3}} + \frac{1}{\sqrt{4k-1}} - \frac{1}{\sqrt{2k}} > \frac{2}{\sqrt{4k}} - \frac{1}{\sqrt{2k}} = \Big(1 - \frac{1}{\sqrt2}\Big)\frac{1}{\sqrt k} .

Since ∑1/k\sum 1/\sqrt k diverges, the partial sums t3nt_{3n} tend to +∞+\infty. The partial sums t3n+1t_{3n+1} and t3n+2t_{3n+2} exceed t3nt_{3n} (they add positive terms), so the whole sequence of partial sums tends to +∞+\infty.

Exercise Rearranging to infinity ●●○ Hardy: §193

Describe a rearrangement of 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots that diverges to +∞+\infty, and prove that it does.

Hint
Use a moving target 1,2,3,…1, 2, 3, \dots and only one negative term per stage.

Solution

Take positive terms (in order) until the partial sum exceeds 11, then the next negative term; then positive terms until it exceeds 22, then the next negative term; and so on. Each stage ends because the unused positive terms 12k−1\tfrac{1}{2k-1} have an infinite sum. Every term is used: there are infinitely many stages, each using at least one positive and exactly one negative term. After the mm-th stage the sum exceeds m−12m - \tfrac12 (it exceeded mm before the single negative term, which is at most 12\tfrac12), and during stage m+1m+1 it only increases. So from the end of stage mm on, every partial sum exceeds m−12m - \tfrac12, and the partial sums tend to +∞+\infty.

Exercise Why there is no comparison test ●○○ Hardy: §194

Give a convergent series ∑un\sum u_n and a series ∑vn\sum v_n with ∣vn∣≤∣un∣|v_n| \le |u_n| for every nn, such that ∑vn\sum v_n diverges. Explain which step of the comparison test breaks.

Hint
Keep the sizes, change the signs.

Solution

Take un=(−1)n−1/nu_n = (-1)^{n-1}/n and vn=1/nv_n = 1/n: then ∣vn∣=∣un∣|v_n| = |u_n|, ∑un\sum u_n converges, and ∑vn\sum v_n diverges. (Hardy’s version uses vn=1/(n+1)v_n = 1/(n+1), which is even strictly smaller.) The comparison test bounds the partial sums of ∑∣vn∣\sum |v_n| by those of ∑∣un∣\sum |u_n|; that is useful only when ∑∣un∣\sum |u_n| is finite, that is, when ∑un\sum u_n converges absolutely. The sizes of the terms of a conditionally convergent series are too large to bound anything.

Explore Two diverging halves, one converging difference

Try this. Choose ∑(sin⁡n)/n\sum (\sin n)/n, then ∑(−1)n−1/n\sum (-1)^{n-1}/\sqrt n, and increase NN.

∑n=1∞sin⁡nn\sum_{n=1}^{\infty} \dfrac{\sin n}{n}
Npartial sums
  • positive part VN=∑n≤NvnV_N = \sum_{n \le N} v_n
  • negative part WN=∑n≤NwnW_N = \sum_{n \le N} w_n
  • the series sN=VN−WNs_N = V_N - W_N
  • its sum
At N=200N = 200: VN=2.52656V_N = 2.52656, WN=1.46015W_N = 1.46015, so ∑n≤N∣un∣=VN+WN=3.9867\sum_{n\le N}|u_n| = V_N + W_N = 3.9867 and sN=VN−WN=1.06641s_N = V_N - W_N = 1.06641. Both parts keep growing without bound (each like 1πlog⁡N\tfrac1\pi \log N), yet their difference settles down: the series is only conditionally convergent.
What to notice

In each case the blue and ochre curves (the positive and negative parts) rise without bound, while the purple curve, their difference, settles: the theorem “both halves diverge” in a picture.

Check yourself

Check yourself Choose one

∑un\sum u_n converges but ∑∣un∣\sum |u_n| diverges. What can you say about the series of positive terms ∑vn\sum v_n?

Why

If both halves converged, ∑∣un∣\sum|u_n| would converge; if only one did, the partial sums of ∑un\sum u_n would tend to ±∞\pm\infty. A series of positive terms cannot oscillate.

Check yourself True or false?

If ∑un\sum u_n converges and ∣vn∣≤∣un∣|v_n| \le |u_n| for every nn, then ∑vn\sum v_n converges.

Why

False unless ∑un\sum u_n converges absolutely. Take un=(−1)n−1/nu_n = (-1)^{n-1}/n and vn=1/(n+1)v_n = 1/(n+1).

Check yourself Enter a number

To four decimal places, what is the sum of 1+13−12+15+17−14+⋯1 + \tfrac13 - \tfrac12 + \tfrac15 + \tfrac17 - \tfrac14 + \cdots?

Why

32log⁡2≈1.0397\tfrac32 \log 2 \approx 1.0397, half as much again as the sum log⁡2≈0.6931\log 2 \approx 0.6931 of the original order.

Check yourself True or false?

The terms of 1−12+13−⋯1 - \tfrac12 + \tfrac13 - \cdots can be rearranged so that the new series oscillates between 00 and 11.

Why

True. Steer the partial sums alternately above 11 and below 00, using the greedy method with targets 1,0,1,0,…1, 0, 1, 0, \dots. Both supplies of terms are inexhaustible, so every stage ends and every term is used; the lim sup is 11 and the lim inf is 00.

Hardy’s language

There are no comparison tests for convergence of conditionally convergent series.

— G. H. Hardy, §194

Summary

Summary
  • A conditionally convergent series converges, but its series of moduli diverges.
  • Its positive part and negative part both diverge to +∞+\infty: the sum exists only through cancellation in the given order.
  • So comparison tests, selections and rearrangements all fail for it.
  • Riemann: a conditionally convergent series can be rearranged to converge to any number, or to diverge to ±∞\pm\infty; for example 1+13−12+⋯1 + \tfrac13 - \tfrac12 + \cdots sums to 32log⁡2\tfrac32 \log 2.