Chapter VIII · The convergence of infinite series and infinite integrals

Dirichlet's theorem: rearranging a positive series

§176 20 min

The sum of a convergent series of positive terms does not depend on the order of its terms, and the two-line reason why.

Why this matters

Addition of finitely many numbers is commutative: the order does not matter. Is the same true of the infinitely many terms of a series? A series is a limit of partial sums, and changing the order changes every partial sum, so there is something to prove. §176 proves that for series of positive terms the answer is yes. The answer for series with terms of both signs is a dramatic no (§193), and the contrast is one of the main themes of this chapter. The next lesson uses Dirichlet’s theorem at once, to multiply two series.

The idea

Take a convergent series of positive terms u1+u2+u3+⋯u_1 + u_2 + u_3 + \cdots with sum ss, and write the same terms in another order: v1+v2+v3+⋯v_1 + v_2 + v_3 + \cdots. Every uu must appear exactly once among the vv’s, and every vv is one of the uu’s.

Look at the first nn terms of the new series. They are finitely many of the uu’s, so they all have indices at most some number MM. Since all terms are positive, their sum is at most the sum of all of u1,…,uMu_1, \dots, u_M, which is at most ss. So the new partial sums never exceed ss. That is half the proof, and the widget shows it.

Explore Taking the terms in a new order

Try this. With 1/n21/n^2 and the reversed-block order, move nn and watch the purple dot (the new partial sum tnt_n) and the hollow dot (the old partial sum sMs_M). Then try the other orders, and finally the divergent series 1/n1/n.

∑un, un=1n2\sum u_n,\ u_n = \frac{1}{n^2}, terms taken in a new order
number of termspartial sums
  • new order: tnt_n
  • original order: sns_n
  • the sum s
  • your nn, and sMs_M (hollow)
The first 12 terms of the new series are uiu_i for i=1,3,2,7,6,5,4,15,14,13,12,11i = 1, 3, 2, 7, 6, 5, 4, 15, 14, 13, 12, 11. The largest index among them is M=15M = 15, so they are some of u1,…,u15u_1, \dots, u_{15} and t12=1.54247≤s15=1.58044t_{12} = 1.54247 \le s_{15} = 1.58044 ≤s=1.64493\le s = 1.64493. Conversely, all of u1,…,u12u_1, \dots, u_{12} have appeared in the new order by term 15, so t15≥s12t_{15} \ge s_{12}.
What to notice

The new partial sums sometimes lag behind the old ones and never overtake the dashed line: each tnt_n is at most sMs_M, which is at most ss. And they always catch up, because after enough new terms every one of u1,…,unu_1, \dots, u_n has appeared. For 1/n1/n both sequences grow without bound in every order.

The other half is the same argument with the two series exchanged: the original series is a rearrangement of the new one. So each sum is at most the other, and they are equal.

Precise statements

A rearrangement of the series ∑un\sum u_n is a series ∑vn\sum v_n whose terms are the same terms in another order: there is a one-to-one correspondence between the positions, so that every umu_m is exactly one vnv_n and every vnv_n is exactly one umu_m. No term may be left out and none repeated. (Putting “all the even-numbered terms first, then the odd-numbered ones” is not a rearrangement in this sense: the odd-numbered terms would never be reached.)

Theorem Dirichlet’s theorem Hardy: §176

If ∑un\sum u_n is a convergent series of positive terms with sum ss, then every rearrangement ∑vn\sum v_n of it converges, and its sum is ss.

Corollary Divergence is also preserved

If a series of positive terms diverges, every rearrangement of it diverges.

The corollary follows from the theorem: if some rearrangement converged, the original series, being a rearrangement of it, would converge too.

Proofs

Proof Dirichlet’s theorem
Strategy

Use Theorem B: a positive series converges exactly when its partial sums are bounded. Show that ss bounds the partial sums of the new series, so it converges to some t≤st \le s. Then swap the roles of the two series to get s≤ts \le t.

Any finite selection of uu’s adds up to at most ss

Take finitely many distinct terms ui1,…,uinu_{i_1}, \dots, u_{i_n} and let MM be the largest of the indices. Since all terms are positive, adding the missing terms with indices up to MM can only increase the total: ui1+⋯+uin≤u1+⋯+uM=sM≤su_{i_1} + \cdots + u_{i_n} \le u_1 + \cdots + u_M = s_M \le s.

So the new partial sums are bounded by ss

The partial sum tn=v1+⋯+vnt_n = v_1 + \cdots + v_n is such a finite selection, since the vv’s are distinct terms of the original series. Hence tn≤st_n \le s for every nn.

The new series converges, with sum t≤st \le s

By Theorem B (§173), ∑vn\sum v_n converges. Its sum tt is the limit of numbers tn≤st_n \le s, so t≤st \le s.

Exchange the roles

The original series is a rearrangement of the new one: every uu is some vv. The argument of the first two steps, with uu and vv exchanged, shows sn≤ts_n \le t for every nn, hence s≤ts \le t.

Conclude

t≤st \le s and s≤ts \le t, so t=st = s.

Where did we use positivity? In the first step only: “adding the missing terms can only increase the total”. With negative terms allowed, a finite selection of terms can add up to more than any partial sum, and the argument collapses. It is not merely the proof that fails: for the series 1−12+13−14+⋯1 - \tfrac12 + \tfrac13 - \tfrac14 + \cdots a suitable rearrangement converges to any number we like (§193). What survives is the case in which the series of absolute values converges (§192).

Worked examples and exercises

Exercise ●●○ Hardy: Examples LXVIII, 14

Show that

1+132+152+⋯=34(1+122+132+⋯ ),1 + \frac{1}{3^2} + \frac{1}{5^2} + \cdots = \frac34\left(1 + \frac1{2^2} + \frac1{3^2} + \cdots\right),

and that the series obtained from ∑1/n2\sum 1/n^2 by omitting every term whose index is a multiple of 44 has sum 1516∑1/n2\tfrac{15}{16}\sum 1/n^2.

Hint
Write S=∑1/n2S = \sum 1/n^2 as the sum of the odd-indexed terms and the even-indexed terms. Why is that allowed?
Hint
The even-indexed terms are 1(2k)2=14⋅1k2\dfrac{1}{(2k)^2} = \dfrac14\cdot\dfrac{1}{k^2}.

Solution

Let S=∑1/n2S = \sum 1/n^2, OO the sum over odd nn and EE the sum over even nn. Both OO and EE converge, by comparison with SS. The series OO and EE written alternately (1,122,132,…1, \tfrac1{2^2}, \tfrac1{3^2}, \dots) is just ∑1/n2\sum 1/n^2 again, and the partial sums of the original series are sums of partial sums of OO and EE; letting n→∞n \to \infty gives S=O+ES = O + E. (Equally, write the terms in the order “all of OO’s first kk terms, then EE’s first kk terms” blockwise: Dirichlet’s theorem guarantees the sum is unchanged.)

Now E=∑k14k2=14SE = \sum_k \dfrac{1}{4k^2} = \tfrac14 S, so O=S−14S=34SO = S - \tfrac14 S = \tfrac34 S.

In the same way the terms with index divisible by 44 are ∑k116k2=116S\sum_k \dfrac{1}{16k^2} = \tfrac{1}{16}S, so the remaining terms add up to 1516S\tfrac{15}{16}S.

Exercise Brackets in a positive series ●○○

Let un≥0u_n \ge 0. Show that if the bracketed series (u1+u2)+(u3+u4)+⋯(u_1 + u_2) + (u_3 + u_4) + \cdots converges, then ∑un\sum u_n converges to the same sum. Why does this fail for 1−1+1−1+⋯1 - 1 + 1 - 1 + \cdots?

Hint
The partial sums of the bracketed series are the even partial sums s2,s4,…s_2, s_4, \dots of the original. What do you know about the odd ones?

Solution

Let σ\sigma be the sum of the bracketed series; its partial sums are s2,s4,s6,…s_2, s_4, s_6, \dots, so s2k→σs_{2k} \to \sigma. Since the terms are positive, sns_n increases, and s2k≤s2k+1≤s2k+2s_{2k} \le s_{2k+1} \le s_{2k+2}. So the odd partial sums are squeezed between even ones and also tend to σ\sigma. Hence sn→σs_n \to \sigma.

For 1−1+1−⋯1 - 1 + 1 - \cdots the bracketed series (1−1)+(1−1)+⋯(1-1) + (1-1) + \cdots converges to 00, but the partial sums alternate 1,0,1,0,…1, 0, 1, 0, \dots. The squeeze used s2k≤s2k+1≤s2k+2s_{2k} \le s_{2k+1} \le s_{2k+2}, which needs positive terms. Hardy removes brackets in exactly this way in the next section.

Exercise A double sum in any order ●●○

The numbers 2−m3−n2^{-m}3^{-n} with m,n=0,1,2,…m, n = 0, 1, 2, \dots are written as a single sequence in some order (every pair (m,n)(m,n) used exactly once). Show that the resulting series converges to 33, whatever the order.

Hint
By Dirichlet’s theorem it is enough to find one order in which the sum is easy.
Hint
Take the terms row by row in growing squares: first m,n≤0m, n \le 0, then the new terms with max⁡(m,n)=1\max(m, n) = 1, and so on.

Solution

In the “growing squares” order, the sum of all terms with m,n≤km, n \le k is

(∑m=0k2−m)(∑n=0k3−n)→2⋅32=3.\Big(\sum_{m=0}^k 2^{-m}\Big)\Big(\sum_{n=0}^k 3^{-n}\Big) \to 2 \cdot \tfrac32 = 3 .

These are some of the partial sums of the series in that order; since the terms are positive, the partial sums increase, and an increasing sequence with a subsequence tending to 33 tends to 33. So this arrangement has sum 33, and by Dirichlet’s theorem every other arrangement does too. This is the idea of the next lesson.

Exercise Not a rearrangement ●○○

Explain why the sequence “u2,u4,u6,…u_2, u_4, u_6, \dots, followed by u1,u3,u5,…u_1, u_3, u_5, \dots” does not define a rearrangement of ∑un\sum u_n, and what its “partial sums” actually are.

Solution

A series is indexed by 1,2,3,…1, 2, 3, \dots: every term must stand at some finite position. Here the term u1u_1 would have to come after infinitely many others, so it has no position nn. The partial sums tnt_n of the sequence that is defined, u2+u4+⋯+u2nu_2 + u_4 + \cdots + u_{2n}, only ever see the even-indexed terms; for un=1/n2u_n = 1/n^2 they tend to 14⋅π26\tfrac14 \cdot \tfrac{\pi^2}{6}, not to π26\tfrac{\pi^2}{6}. Dirichlet’s theorem does not apply because no term may be left out.

Check yourself

Check yourself True or false?

The sum of every convergent series is unchanged by rearranging its terms.

Why

False. Dirichlet’s theorem needs positive terms (more generally, absolute convergence, §192). The series 1−12+13−14+⋯1 - \tfrac12 + \tfrac13 - \tfrac14 + \cdots converges to log⁡2\log 2, but it can be rearranged to converge to any other number, or to diverge (§193).

Check yourself True or false?

If ∑un\sum u_n is a convergent series of positive terms with sum ss, then every partial sum of every rearrangement is at most ss.

Why

True: this is the first half of the proof. A partial sum of a rearrangement is a finite selection of the uu’s, and a finite selection of positive terms is at most some sMs_M, hence at most ss.

Check yourself Choose one

In the widget, the new partial sums tnt_n sometimes lie below the old partial sums sns_n. Why does this not contradict t=st = s?

Why

For each fixed mm, the terms u1,…,umu_1, \dots, u_m all appear among the first PP terms of the new series for some PP, and then tn≥smt_n \ge s_m for all n≥Pn \ge P. So the tnt_n may lag, but they catch up with every sms_m and never pass ss. The theorem claims equality of the limits, not of the partial sums.

Check yourself Enter a number

Given that ∑n=1∞1/n2=π2/6\sum_{n=1}^\infty 1/n^2 = \pi^2/6, what is 1+132+152+⋯1 + \tfrac{1}{3^2} + \tfrac{1}{5^2} + \cdots? (Give a decimal.)

Why

By the first exercise it is 34⋅π26=π28≈1.2337\tfrac34 \cdot \tfrac{\pi^2}{6} = \tfrac{\pi^2}{8} \approx 1.2337.

Hardy’s language

Of course no terms must be omitted: every u must come somewhere among the v’s, and vice versa.

— G. H. Hardy, §176

Summary

Summary
  • A rearrangement uses every term exactly once, each at a finite position.
  • Dirichlet’s theorem: a convergent positive series converges to the same sum in every order; a divergent one diverges in every order.
  • Proof: any finite set of terms adds up to at most ss, so the new series converges with t≤st \le s; by symmetry s≤ts \le t.
  • Positivity is used essentially; for series with infinitely many terms of each sign the theorem can fail.