Skip to content
Topos Uranos

Absolute convergence, rearrangements and products of series

Summary

A finite sum does not depend on the order of its summands; a series, in general, does. This lesson studies two operations with series that finite sums carry out without risk: changing the order of the terms and multiplying two series. A rearrangement takes the terms in another order, given by a bijection of the natural numbers. Dirichlet’s theorem proves that an absolutely convergent series can be rearranged without changing its sum. By contrast, in a conditionally convergent series the series of the positive terms and that of the negative terms diverge, and the Riemann rearrangement theorem proves, with an explicit algorithm, that a suitable rearrangement converges to any number given in advance, or diverges to ∞\infty or to −∞-\infty. The lesson also proves that the alternating harmonic series has sum ln⁡2\ln 2, and that one of its rearrangements has half that sum. The Cauchy product of two series gathers the products of their terms whose indices add up to the same number, as when multiplying polynomials; Mertens’s theorem proves that its sum is the product of the sums if one of the series converges absolutely and the other converges, and an example shows that with two conditionally convergent series the product can diverge.

Learning objectives

  1. Prove that rearranging an absolutely convergent series does not change its sum, and that in a conditionally convergent series the series of the positive terms and that of the negative terms diverge.
  2. Prove the Riemann rearrangement theorem with its algorithm, and compute the sum of specific rearrangements of the alternating harmonic series.
  3. Prove Mertens’s theorem on the Cauchy product and its case for two absolutely convergent series, and recognize when the product of two convergent series diverges.
  4. Compute sums of series as Cauchy products of known series.

Rearrangements

By generalized commutativity, a fact admitted in the lesson on prime numbers and the fundamental theorem of arithmetic, a sum of finitely many summands does not change if their order is altered. The sum of a series is a limit, and each partial sum only sees the first terms: changing the order changes the partial sums, and nothing ensures that their limit is the same. The lesson on numerical series and their convergence announced that for some convergent series rearranging does change the sum, and the lesson on alternating series and conditional convergence, that this does not happen with an absolutely convergent series. The conventions of those lessons are kept: partial sums SnS_n, sum SS and, for a second series, TnT_n and TT.

DefinitionRearrangement of a series

Let ∑n=1∞an\sum_{n=1}^{\infty} a_n be a series, and let σ\sigma be a bijection from N\mathbb{N} to N\mathbb{N}. The series ∑n=1∞aσ(n)\sum_{n=1}^{\infty} a_{\sigma(n)}, whose partial sums are Tn=∑k=1naσ(k)T_n = \sum_{k=1}^{n} a_{\sigma(k)}, is called a rearrangement of ∑n=1∞an\sum_{n=1}^{\infty} a_n. Its nnth term is the term with index σ(n)\sigma(n) of the given series: each term appears exactly once, at the place mm such that σ(m)\sigma(m) is its index.

If σ\sigma is a bijection, for each natural number kk there is a unique natural number mm with σ(m)=k\sigma(m) = k; that mm is the place occupied in the rearrangement by the term aka_k. For sequences, changing the order does not change the limit: only finitely many terms are far from the limit, and once all of them have been placed only nearby terms remain.

LemmaRearranging a sequence does not change its limit

Let (xn)(x_n) be a sequence with xn→Lx_n \to L, with LL real, and σ\sigma a bijection from N\mathbb{N} to N\mathbb{N}. Then xσ(n)→Lx_{\sigma(n)} \to L. The terms x1,…,xNx_1, \ldots, x_N occupy finite places in the rearrangement; past the largest of them, all the terms have index greater than NN.

Proof

  1. σ(mk)=k,mk≤M(1≤k≤N)\dato{lug}{\sigma(m_k) = k, \quad m_k \leq M \quad (1 \leq k \leq N)}

    Given ε>0\varepsilon > 0, let NN be a natural number such that ∣xk−L∣<ε\abs{x_k - L} < \varepsilon for every k≥Nk \geq N. For each kk with 1≤k≤N1 \leq k \leq N, let mkm_k be the natural number with σ(mk)=k\sigma(m_k) = k, and let MM be the largest of m1,…,mNm_1, \ldots, m_N, which exists because they are finitely many (the larger of two numbers, applied N−1N - 1 times).

  2. n>M→σ(n)>N→∣xσ(n)−L∣<εn > M \rightarrow \sigma(n) > N \rightarrow \resaltar{\abs{x_{\sigma(n)} - L} < \varepsilon}

    Let n>Mn > M. If we had σ(n)=k\sigma(n) = k with k≤Nk \leq N, the injectivity of σ\sigma would give n=mk≤Mn = m_k \leq M, contrary to the assumption. Therefore, σ(n)>N\sigma(n) > N, and MM depends only on ε\varepsilon and on σ\sigma.

With series, the same argument works if the error left by the distant terms is controlled by the series of absolute values.

TheoremDirichlet’s theorem on rearrangements

If the series ∑n=1∞an\sum_{n=1}^{\infty} a_n converges absolutely and its sum is SS, every rearrangement of it converges absolutely, and its sum is SS:

∑n=1∞aσ(n)=∑n=1∞an\sum_{n=1}^{\infty} a_{\sigma(n)} = \sum_{n=1}^{\infty} a_n

When nn is large, the partial sum TnT_n contains all the terms a1,…,aNa_1, \ldots, a_N, and the others, with indices greater than NN, add up in absolute value to less than the tail of the series of absolute values.

Proof

  1. ∑k=1n∣aσ(k)∣≤∑k=1m∣ak∣≤P\sum_{k=1}^{n} \abs{a_{\sigma(k)}} \leq \sum_{k=1}^{m} \abs{a_k} \leq \resaltar{P}

    Let P=∑n=1∞∣an∣P = \sum_{n=1}^{\infty} \abs{a_n}. By the monotone convergence theorem, from the lesson on monotonic sequences and the number e, PP is the supremum of the increasing partial sums ∑k=1m∣ak∣\sum_{k=1}^{m} \abs{a_k}. Let nn be a natural number, and let mm be the largest of σ(1),…,σ(n)\sigma(1), \ldots, \sigma(n). Those nn indices are distinct and lie between 11 and mm; by generalized commutativity, ∑k=1m∣ak∣\sum_{k=1}^{m} \abs{a_k} is the sum of the ∣aσ(k)∣\abs{a_{\sigma(k)}} with k≤nk \leq n plus that of the remaining ∣aj∣\abs{a_j}, which is not negative. The partial sums of the series ∑n=1∞∣aσ(n)∣\sum_{n=1}^{\infty} \abs{a_{\sigma(n)}} increase and are bounded by PP, and the same theorem gives its convergence.

  2. m>N→∑k=N+1m∣ak∣<ε2,∣S−SN∣≤ε2\dato{cola}{m > N \rightarrow \sum_{k=N + 1}^{m} \abs{a_k} < \frac{\varepsilon}{2}, \qquad \abs{S - S_N} \leq \frac{\varepsilon}{2}}

    Given ε>0\varepsilon > 0, the Cauchy criterion for series, from the lesson on numerical series and their convergence, applied to the series of absolute values with ε2\frac{\varepsilon}{2}, gives a natural number NN such that ∑k=n+1m∣ak∣<ε2\sum_{k=n + 1}^{m} \abs{a_k} < \frac{\varepsilon}{2} if m>n≥Nm > n \geq N. With n=Nn = N, the triangle inequality theorem, from the lesson on the real line, extended by induction to any number of summands, gives ∣Sm−SN∣<ε2\abs{S_m - S_N} < \frac{\varepsilon}{2} for m>Nm > N, and the preservation of inequalities, from the theorem stating that the limit theorems hold for sequences, from the lesson on sequences and their limit, as m→∞m \to \infty, gives the bound on S−SNS - S_N.

  3. n≥M→∣Tn−SN∣≤∑k=N+1m∣ak∣<ε2n \geq M \rightarrow \abs{T_n - S_N} \leq \sum_{k=N + 1}^{m} \abs{a_k} < \resaltar{\frac{\varepsilon}{2}}

    As in the lemma above, let MM be the largest of the places m1,…,mNm_1, \ldots, m_N of the terms a1,…,aNa_1, \ldots, a_N. If n≥Mn \geq M, the indices σ(1),…,σ(n)\sigma(1), \ldots, \sigma(n) contain 1,…,N1, \ldots, N; let mm be the largest of them. By generalized commutativity, Tn−SNT_n - S_N is the sum of the aja_j with jj among σ(1),…,σ(n)\sigma(1), \ldots, \sigma(n) and with j>Nj > N, which is 00 if m=Nm = N; if m>Nm > N, those jj are distinct and lie between N+1N + 1 and mm: the extended triangle inequality bounds ∣Tn−SN∣\abs{T_n - S_N} by the sum of their absolute values, which does not exceed that of all the ∣ak∣\abs{a_k} with N+1≤k≤mN + 1 \leq k \leq m, because the summands that are added are not negative.

  4. n≥M→∣Tn−S∣≤∣Tn−SN∣+∣SN−S∣<ε2+ε2=εn \geq M \rightarrow \abs{T_n - S} \leq \abs{T_n - S_N} + \abs{S_N - S} < \frac{\varepsilon}{2} + \frac{\varepsilon}{2} = \resaltar{\varepsilon}

    For n≥Mn \geq M, the triangle inequality combines the two bounds. The natural number MM depends only on ε\varepsilon and on σ\sigma; therefore, Tn→ST_n \to S.

The property has no analogue for improper integrals: an integral has no order of summands to alter. A series with nonnegative terms that converges converges absolutely, and the theorem applies to it; if it diverges, every rearrangement of it also diverges to ∞\infty, as the proposed problem “The rearrangements of a series with nonnegative terms” asks to prove.