math.wiki
Analysis › Sequences and series › Series convergence, proved
Sign in

Series convergence, proved

Calculus decides whether a series converges with a handful of tests, some resting on facts taken on trust. What do the tests really rest on, and may the terms of a convergent series be added in any order?

Where you are 5 in Sequences and series

Why this exists

Convergence tests gathers the calculus tests: the The nth term test for divergence, comparison (The direct and limit comparison tests), the ratio and root tests, the alternating series test and absolute convergence. Their proofs took on trust that a nondecreasing bounded sequence converges, and left both rearrangement theorems unproved. A series is its sequence of partial sums, and this course now has the sequence theorems that answer every such question: the Cauchy criterion of Cauchy sequences for series whose sum cannot be named, The monotone convergence theorem for nonnegative terms, nested intervals for alternating signs.

Definition

Let \{a_n\} be a sequence of real numbers. The series \sum_{n=1}^{\infty} a_n has partial sums S_n = \sum_{k=1}^{n} a_k (Infinite series and partial sums). It converges to S, written \sum_{n=1}^{\infty} a_n = S, if S_n \to S; otherwise it diverges.

Blocks and tails. For m > n, S_m - S_n = \sum_{k=n+1}^{m} a_k is the block of terms from n + 1 to m. For a convergent series, the tail after n is t_n = S - S_n = \sum_{k=n+1}^{\infty} a_k.

Absolute and conditional. The series converges absolutely if \sum \lvert a_n \rvert converges, conditionally if it converges but not absolutely.

Rearrangement. For a bijection \sigma of the positive integers onto themselves, \sum_{n=1}^{\infty} a_{\sigma(n)} is a rearrangement: every term used once, in a new order.

Grouping. For integers 0 = n_0 < n_1 < n_2 < \cdots, the series \sum_{k=1}^{\infty} G_k with G_k = \sum_{i=n_{k-1}+1}^{n_k} a_i is a grouping: parentheses inserted, order kept.

Intuition

A series converges exactly when its late blocks are small, every late block, however long. Small terms are only the blocks of length one, which is why a_n \to 0 is necessary and not sufficient: for the harmonic series the block from n + 1 to 2n has n terms, each at least \displaystyle \frac{1}{2n}, so it is at least \displaystyle \frac{1}{2} for every n.

Order is a separate question. An absolutely convergent series has a far tail of small total size, and any reordering eventually contains all the early terms, so the sum cannot move. A conditionally convergent series has positive terms adding to \infty and negative terms adding to -\infty, and the order decides how fast each supply is spent. Figure 1 shows one such pair of orders.

10
20
30
0.5
1
m
\ln 2
\tfrac{3}{2}\ln 2
Figure 1The same moves in two orders (Example 4). Dots: the offsets S_m in the listed order, settling at \ln 2. Squares: Andre’s offsets R_m, two rightward moves then one leftward, settling at \displaystyle \frac{3}{2}\ln 2.

Results

Theorem 1 (Cauchy criterion for series). \sum a_n converges if and only if for every \varepsilon > 0 there is N such that \left\lvert \sum_{k=n+1}^{m} a_k \right\rvert < \varepsilon for all m > n > N.

Theorem 2 (terms, tails, late starts). If \sum a_n = S, then a_n \to 0 and t_n \to 0. For every integer K \ge 0, \sum_{n=1}^{\infty} a_n converges if and only if \sum_{n=K+1}^{\infty} a_n does, and then S = S_K + \sum_{n=K+1}^{\infty} a_n.

Theorem 3 (nonnegative terms, comparison). If a_n \ge 0 for every n, then \sum a_n converges if and only if \{S_n\} is bounded above, and then \sum a_n = \sup\{S_n : n \ge 1\}. If 0 \le a_n \le b_n for every n > M and \sum b_n converges, then \sum a_n converges; if \sum a_n diverges, so does \sum b_n.

Theorem 4 (absolute convergence). If \sum \lvert a_n \rvert converges, then \sum a_n converges and \left\lvert \sum_{n=1}^{\infty} a_n \right\rvert \le \sum_{n=1}^{\infty} \lvert a_n \rvert.

Theorem 5 (root test with the limit superior). Let \rho_n = \lvert a_n \rvert^{1/n}. (a) If \{\rho_n\} is bounded and \limsup_{n \to \infty} \rho_n < 1, then \sum a_n converges absolutely. (b) If \{\rho_n\} is unbounded, or \limsup_{n \to \infty} \rho_n > 1, then \sum a_n diverges.

Theorem 6 (alternating series test, by nested intervals). Let \{b_n\} be nonincreasing with b_n \to 0. Then \sum_{n=1}^{\infty} (-1)^{n+1} b_n converges to a number S lying in every interval I_k = [S_{2k}, S_{2k-1}], and \lvert S - S_n \rvert \le b_{n+1} for every n \ge 1.

Theorem 7 (grouping). If \sum a_n = S, every grouping of it converges to S.

Theorem 8 (rearranging an absolutely convergent series). If \sum a_n converges absolutely to S, every rearrangement converges absolutely, to S.

Theorem 9 (Riemann rearrangement theorem). If \sum a_n converges conditionally, then for every real number L some rearrangement converges to L.

The ratio test’s proof in The ratio and root tests is complete once Theorem 3 is in hand.

Why it is true

Theorem 1. Proof. S_m - S_n is the block from n + 1 to m. By Theorem 4 of Cauchy sequences, \{S_n\} converges if and only if it is Cauchy. Since \lvert S_m - S_n \rvert = \lvert S_n - S_m \rvert, and it is 0 when m = n, the Cauchy condition only needs pairs with m > n. \square

Theorem 2. Proof. Given \varepsilon > 0, take N from Theorem 1; the block with m = n + 1 gives \lvert a_{n+1} \rvert < \varepsilon for n > N. Next, t_n = S - S_n \to S - S = 0. For n > K, S_n = S_K + \sum_{k=K+1}^{n} a_k: the partial sums of the late start differ from those of the series by the constant S_K, so one sequence converges if and only if the other does, by the sum law and the tails theorem (Theorem 3 of Limits of sequences, epsilon and N). \square

Theorem 3. Proof. If a_n \ge 0, then S_{n+1} = S_n + a_{n+1} \ge S_n, and Theorem 1 of The monotone convergence theorem gives convergence, to the supremum, exactly when \{S_n\} is bounded above. For comparison, by Theorem 2 we may start both series at M + 1. The partial sums of \sum_{n > M} b_n are nondecreasing, so each is at most their limit B. Then

\sum_{k=M+1}^{n} a_k \le \sum_{k=M+1}^{n} b_k \le B

for every n > M, and the first part applies. The second statement is the contrapositive of the first. \square

Theorem 4. Proof. Let \varepsilon > 0. Theorem 1 for \sum \lvert a_n \rvert gives N with \sum_{k=n+1}^{m} \lvert a_k \rvert < \varepsilon for m > n > N, and the Triangle inequality for finite sums gives

\left\lvert \sum_{k=n+1}^{m} a_k \right\rvert \le \sum_{k=n+1}^{m} \lvert a_k \rvert < \varepsilon

so Theorem 1 gives convergence. With A = \sum \lvert a_n \rvert, \lvert S_n \rvert \le \sum_{k=1}^{n} \lvert a_k \rvert \le A by Theorem 3. Since \bigl\lvert \lvert S_n \rvert - \lvert S \rvert \bigr\rvert \le \lvert S_n - S \rvert, \lvert S_n \rvert \to \lvert S \rvert, and the order limit theorem (Theorem 8 of Limits of sequences, epsilon and N) gives \lvert S \rvert \le A. \square

Theorem 5. Proof. (a) Let L = \limsup \rho_n < 1 and choose r with L < r < 1. The suprema s_n = \sup\{\rho_k : k \ge n\} decrease to L (the definition in Subsequences and Bolzano-Weierstrass), so some s_M < r, and then \lvert a_k \rvert < r^k for every k \ge M. The geometric series \sum r^k converges, so Theorem 3 makes \sum_{k \ge M} \lvert a_k \rvert converge, and Theorem 2 adds the first terms. (b) If \{\rho_n\} is unbounded, \rho_n > 1 for infinitely many n. If it is bounded with \limsup \rho_n = L > 1, Theorem 7 of Subsequences and Bolzano-Weierstrass gives a subsequence \rho_{n_k} \to L, so \rho_{n_k} > 1 for all large k (sign persistence). Either way \lvert a_n \rvert > 1 for infinitely many n, so a_n \not\to 0 and Theorem 2 gives divergence. \square

Theorem 6. Proof. Grouping in pairs,

\begin{aligned} S_{2k+1} &= S_{2k-1} - (b_{2k} - b_{2k+1}) \le S_{2k-1} \\ S_{2k+2} &= S_{2k} + (b_{2k+1} - b_{2k+2}) \ge S_{2k} \end{aligned}

and S_{2k-1} - S_{2k} = b_{2k} \ge 0. So the I_k are nested closed bounded intervals with lengths b_{2k} \to 0. The nested interval property (Theorem 3 of The monotone convergence theorem) gives a point S in every I_k, and only one, since two such points are within b_{2k} of each other for every k. If n \ge 2k - 1, then S_n is an endpoint of I_j for some j \ge k, and I_j \subseteq I_k, so \lvert S_n - S \rvert \le b_{2k}. Given \varepsilon > 0, choose k with b_{2k} < \varepsilon: every n \ge 2k - 1 has \lvert S_n - S \rvert < \varepsilon, so S_n \to S.

For the bound, if n = 2k - 1, then S_{2k} \le S \le S_{2k-1} gives 0 \le S_n - S \le b_{2k} = b_{n+1}. If n = 2k, then S \in I_{k+1} gives S_{2k} \le S_{2k+2} \le S \le S_{2k+1}, so 0 \le S - S_n \le b_{2k+1} = b_{n+1}. \square

Theorem 7. Proof. The kth partial sum of the grouping is S_{n_k}, and n_1 < n_2 < \cdots, so these form a subsequence of \{S_n\}, which converges to S by Theorem 2 of Subsequences and Bolzano-Weierstrass. \square

Theorem 8. Proof. Let A = \sum \lvert a_n \rvert and T_m = \sum_{j=1}^{m} a_{\sigma(j)}. The indices \sigma(1), \dots, \sigma(m) are distinct, so \sum_{j=1}^{m} \lvert a_{\sigma(j)} \rvert \le A, and Theorem 3 gives absolute convergence. Now let \varepsilon > 0. By Theorem 2 for \sum \lvert a_n \rvert, choose M with \displaystyle \sum_{n > M} \lvert a_n \rvert < \frac{\varepsilon}{2}. Let K = \max\{\sigma^{-1}(1), \dots, \sigma^{-1}(M)\}, so for m \ge K the indices \sigma(1), \dots, \sigma(m) include 1, \dots, M. Then T_m - S_M is a sum of a_n over finitely many distinct n > M, so \displaystyle \lvert T_m - S_M \rvert \le \sum_{n > M} \lvert a_n \rvert < \frac{\varepsilon}{2}. Theorem 4 applied to the tail gives \displaystyle \lvert S - S_M \rvert \le \sum_{n > M} \lvert a_n \rvert < \frac{\varepsilon}{2}. So \lvert T_m - S \rvert < \varepsilon for every m \ge K. \square

Theorem 9. Proof. List the terms with a_n \ge 0 in order as u_1, u_2, \dots, and the terms with a_n < 0 in order as -v_1, -v_2, \dots. Since the series converges conditionally, its positive and negative parts both diverge (Absolute and conditional convergence), so both lists are infinite and every tail of either list has unbounded partial sums. Both u_j \to 0 and v_j \to 0, as subsequences of a_n \to 0.

Let T_0 = 0. At step m, if T_{m-1} \le L, add the next unused u (a u-step); otherwise subtract the next unused v (a v-step). Each kind of step occurs infinitely often: were there a last u-step, the later steps would subtract a tail of the v’s, the sum would fall to L or below, and the next step would be a u-step; symmetrically for v-steps. So every term is used exactly once, and the result is a rearrangement. Let c_j be the term added at step j.

Suppose T_m > L, and let j \le m be the last u-step. Then T_j = T_{j-1} + c_j \le L + c_j, and steps j + 1, \dots, m only subtract, so 0 < T_m - L \le c_j. Suppose T_m \le L, and let j \le m be the last v-step. Then T_j = T_{j-1} + c_j > L + c_j, and later steps only add, so 0 \le L - T_m < \lvert c_j \rvert. Given \varepsilon > 0, only finitely many u_i and v_i are at least \varepsilon; let J be a step after all of them are used. Once m is past a u-step and a v-step that both come after J, the step j above comes after J, and \lvert T_m - L \rvert \le \lvert c_j \rvert < \varepsilon. \square

Worked examples

Example 1

A cutoff from blocks alone

Use Theorem 1 to prove that \displaystyle \sum_{k=1}^{\infty} \frac{1}{k^3} converges, find the least N that the block bound certifies for \varepsilon = 10^{-4}, and bound S - S_n.

Solution

For k \ge 2, k^3 > k^3 - k = (k - 1)k(k + 1), and partial fractions give

\frac{1}{(k - 1)k(k + 1)} = \frac{1}{2}\left(\frac{1}{(k - 1)k} - \frac{1}{k(k + 1)}\right)

For m > n \ge 1 the right side telescopes over the block:

\begin{aligned} \sum_{k=n+1}^{m} \frac{1}{k^3} &< \frac{1}{2}\left(\frac{1}{n(n + 1)} - \frac{1}{m(m + 1)}\right) \\ &< \frac{1}{2n(n + 1)} \end{aligned}

For n > N this is at most \displaystyle \frac{1}{2(N + 1)(N + 2)}, which is below any \varepsilon > 0 once N is large, so the series converges by Theorem 1. For \varepsilon = 10^{-4} we need (N + 1)(N + 2) \ge 5000; 70 \cdot 71 = 4970 falls short and 71 \cdot 72 = 5112 suffices, so N = 70.

Fix n and let m \to \infty in the block bound: the order limit theorem gives \displaystyle 0 < S - S_n \le \frac{1}{2n(n + 1)}. For n = 10, S_{10} \approx 1.197532 and \displaystyle S - S_{10} \le \frac{1}{220} \approx 0.004545; in fact S \approx 1.202057, so the bound is nearly sharp.

Example 2

Nested intervals around a sum

Apply Theorem 6 to \displaystyle \sum_{n=1}^{\infty} \frac{(-1)^{n+1}}{2n + 1} = \frac{1}{3} - \frac{1}{5} + \frac{1}{7} - \cdots. Find I_1 and I_2, check that they nest, find the least k with I_k shorter than 10^{-3}, and check the intervals against the value \displaystyle 1 - \frac{\pi}{4}, taken from the Leibniz series \displaystyle 1 - \frac{1}{3} + \frac{1}{5} - \cdots = \frac{\pi}{4}.

Solution

Here \displaystyle b_n = \frac{1}{2n + 1} decreases to 0. The partial sums are \displaystyle S_1 = \frac{1}{3}, \displaystyle S_2 = \frac{2}{15}, \displaystyle S_3 = \frac{29}{105}, \displaystyle S_4 = \frac{52}{315}, so

I_1 = \left[\frac{2}{15}, \frac{1}{3}\right], \qquad I_2 = \left[\frac{52}{315}, \frac{29}{105}\right]

As \displaystyle \frac{2}{15} = \frac{42}{315} and \displaystyle \frac{29}{105} = \frac{87}{315} < \frac{1}{3}, I_2 \subseteq I_1. The length of I_k is \displaystyle b_{2k} = \frac{1}{4k + 1}, below 10^{-3} exactly when 4k + 1 > 1000, so k = 250: the sum lies between S_{500} and S_{499}. Removing the first term of the Leibniz series (Theorem 2) gives \displaystyle S = 1 - \frac{\pi}{4} \approx 0.214602, which lies in I_1 \approx [0.133333, 0.333333] and in I_2 \approx [0.165079, 0.276190].

Example 3

Decreasing terms of a convergent series

Let a_n \ge 0 be nonincreasing with \sum a_n convergent. Prove that n a_n \to 0.

Solution

Proof. Let \varepsilon > 0. By Theorem 1 there is N with \displaystyle \sum_{k=n+1}^{m} a_k < \frac{\varepsilon}{3} for m > n > N. For n > N, the block from n + 1 to 2n has n terms, each at least a_{2n} since the terms do not increase, so \displaystyle n a_{2n} < \frac{\varepsilon}{3}. Hence 2n a_{2n} < \varepsilon, and (2n + 1)a_{2n+1} \le (2n + 1)a_{2n} \le 3n a_{2n} < \varepsilon. Every index k > 2N + 1 is 2n or 2n + 1 with n > N, so k a_k < \varepsilon. \square

Example 4

Mei’s gantry: same moves, new order

Mei programs the gantry stage of a lab microscope, which slides a camera along a rail. Move n is \displaystyle \frac{1}{n} millimeters, rightward for odd n and leftward for even n, so after n moves the offset is \displaystyle S_n = \sum_{k=1}^{n} \frac{(-1)^{k+1}}{k}. Andre runs the same moves two rightward, then one leftward, each kind in its own order: \displaystyle 1 + \frac{1}{3} - \frac{1}{2} + \frac{1}{5} + \frac{1}{7} - \frac{1}{4} + \cdots. Prove that S_n \to \ln 2 and that Andre’s offsets \displaystyle R_m \to \frac{3}{2}\ln 2.

Solution

Write \displaystyle H_n = \sum_{k=1}^{n} \frac{1}{k} and d_n = H_n - \ln n. First, d_n converges. Since \displaystyle \frac{1}{x} > \frac{1}{n+1} on (n, n + 1),

d_n - d_{n+1} = \int_n^{n+1} \frac{dx}{x} - \frac{1}{n + 1} > 0

and \displaystyle H_n > \int_1^{n+1} \frac{dx}{x} = \ln(n + 1) > \ln n, so d_n > 0. By monotone convergence d_n \to \gamma for some number \gamma, and for each fixed positive integer c, d_{cn} \to \gamma too, as a subsequence.

Mei’s order. Removing the even terms twice, \displaystyle S_{2k} = H_{2k} - 2 \cdot \frac{1}{2}H_k = H_{2k} - H_k, so

S_{2k} = d_{2k} - d_k + \ln 2 \to \ln 2

and \displaystyle S_{2k+1} = S_{2k} + \frac{1}{2k + 1} \to \ln 2. Even and odd indices together give S_n \to \ln 2.

Andre’s order. After k blocks of three he has used the odd reciprocals up to \displaystyle \frac{1}{4k - 1}, which sum to \displaystyle H_{4k} - \frac{1}{2}H_{2k}, and the even ones up to \displaystyle \frac{1}{2k}, which sum to \displaystyle \frac{1}{2}H_k. The logarithms \displaystyle \ln 4k - \frac{1}{2}\ln 2k - \frac{1}{2}\ln k combine to \displaystyle \frac{3}{2}\ln 2, so

\begin{aligned} R_{3k} &= H_{4k} - \tfrac{1}{2}H_{2k} - \tfrac{1}{2}H_k \\ &= d_{4k} - \tfrac{1}{2}d_{2k} - \tfrac{1}{2}d_k \\ &\quad + \tfrac{3}{2}\ln 2 \end{aligned}

The d terms tend to \displaystyle \gamma - \frac{\gamma}{2} - \frac{\gamma}{2} = 0, so \displaystyle R_{3k} \to \frac{3}{2}\ln 2 \approx 1.0397. Also \displaystyle R_{3k+1} = R_{3k} + \frac{1}{4k + 1} and \displaystyle R_{3k+2} = R_{3k+1} + \frac{1}{4k + 3}, so all three residues of m modulo 3 give the same limit, and \displaystyle R_m \to \frac{3}{2}\ln 2.

m3303003000
R_m0.83331.01521.03721.0395

Andre’s camera stops about 0.3466 millimeters right of Mei’s (Figure 1), which Theorem 8 rules out for an absolutely convergent series.

Example 5

Two groupings, two sums

Let \displaystyle a_n = (-1)^{n+1}\frac{n}{n + 1}. Show that the grouping (a_1 + a_2) + (a_3 + a_4) + \cdots converges to \ln 2 - 1 and the grouping a_1 + (a_2 + a_3) + (a_4 + a_5) + \cdots converges to \ln 2, and conclude that \sum a_n diverges.

Solution

For k \ge 1, using (2k - 1)(2k + 1) - (2k)^2 = -1 and (2k + 1)^2 - 2k(2k + 2) = 1,

\begin{aligned} a_{2k-1} + a_{2k} &= \frac{2k - 1}{2k} - \frac{2k}{2k + 1} \\ &= \frac{1}{2k + 1} - \frac{1}{2k} \\ a_{2k} + a_{2k+1} &= \frac{2k + 1}{2k + 2} - \frac{2k}{2k + 1} \\ &= \frac{1}{2k + 1} - \frac{1}{2k + 2} \end{aligned}

With S_n the alternating harmonic partial sums of Example 4, the first grouping has partial sums \displaystyle \sum_{k=1}^{K}\left(\frac{1}{2k + 1} - \frac{1}{2k}\right) = S_{2K+1} - 1, which tend to \ln 2 - 1. The second starts with \displaystyle a_1 = \frac{1}{2}, and its partial sums \displaystyle \frac{1}{2} + \sum_{k=1}^{K}\left(\frac{1}{2k + 1} - \frac{1}{2k + 2}\right) = S_{2K+2} tend to \ln 2. If \sum a_n converged, Theorem 7 would give both groupings the same sum; they differ by 1, so the series diverges.

Boundaries

  • Completeness carries Theorem 1. The partial sums of \displaystyle \sum_{k=0}^{\infty} \frac{1}{k!} are rational and Cauchy, but their limit e is irrational (stated without proof), so the criterion fails in \mathbb{Q}.
  • Example 3 needs monotone terms. With \displaystyle a_n = \frac{1}{n} at perfect squares and \displaystyle \frac{1}{n^2} elsewhere, \displaystyle \sum a_n \le 2\sum \frac{1}{n^2} converges, yet n a_n = 1 at every square.
  • Grouping needs convergence first (Example 5). With a_n \to 0 and groups of bounded length the converse holds (Exercise 11).
  • Rearrangement needs absolute convergence (Example 4, Theorem 9), unless every index moves by a bounded amount (Exercise 13).

Where this goes next

Power series, proved applies Theorem 5 to \sum c_n (x - a)^n, which converges absolutely when \lvert x - a \rvert \limsup \lvert c_n \rvert^{1/n} < 1 and diverges when this exceeds 1: a formula for the radius of convergence. Pointwise and uniform convergence runs the Cauchy criterion uniformly in x for series of functions.

Exercises

The Cauchy criterion

  1. For \displaystyle \sum_{k=1}^{\infty} \frac{1}{4k^2 - 1}, find the least N with \displaystyle \left\lvert \sum_{k=n+1}^{m} \frac{1}{4k^2 - 1} \right\rvert < \frac{1}{1000} for all m > n > N, and a pair m > n that defeats N - 1.
  2. Show that \displaystyle \sum_{k=n+1}^{2n} \frac{1}{2k - 1} > \frac{1}{4} for every n \ge 1, conclude that \displaystyle \sum \frac{1}{2k - 1} diverges, and find \displaystyle \lim_{n \to \infty} \sum_{k=n+1}^{2n} \frac{1}{2k - 1}.
  3. Show that every block of \displaystyle \sum_{k=1}^{\infty} \frac{\cos k}{2^k} from n + 1 on is less than 2^{-n} in absolute value. Find the least N this bound certifies for \varepsilon = 10^{-6}, and bound \lvert S \rvert.

Nested intervals and the root test

  1. For \displaystyle \sum_{n=1}^{\infty} \frac{(-1)^{n+1}}{n(n + 1)}, find I_1 and I_2, find the sum using Example 4, and find the least n for which Theorem 6 guarantees \lvert S - S_n \rvert \le 10^{-3}.
  2. Let \displaystyle a_n = \frac{n^5 \left(2 + (-1)^n\right)^n}{4^n}. Find \limsup \lvert a_n \rvert^{1/n} and \liminf \lvert a_n \rvert^{1/n}, decide convergence, and show that the ratios \displaystyle \frac{a_{n+1}}{a_n} have no limit, finite or infinite.

Rearrangements

  1. Given \displaystyle \sum_{n=1}^{\infty} \frac{1}{n^2} = \frac{\pi^2}{6}, find the sum of the rearrangement \displaystyle 1 - \frac{1}{4} - \frac{1}{16} + \frac{1}{9} - \frac{1}{36} - \frac{1}{64} + \cdots of \displaystyle \sum \frac{(-1)^{n+1}}{n^2} (one positive term, then two negative), and the sums of its positive and of its negative terms.
  2. Andre runs Mei’s moves one rightward, then two leftward: \displaystyle 1 - \frac{1}{2} - \frac{1}{4} + \frac{1}{3} - \frac{1}{6} - \frac{1}{8} + \cdots. Where does the camera settle?
  3. With p rightward moves, then q leftward, repeated, show that the camera settles at \displaystyle \ln 2 + \frac{1}{2}\ln\frac{p}{q}. Which pattern, with the fewest moves per block, settles at 2\ln 2, and which at 0?

Proofs

  1. (Cauchy condensation) Let a_n \ge 0 be nonincreasing. Prove that \sum_{n=1}^{\infty} a_n converges if and only if \sum_{k=0}^{\infty} 2^k a_{2^k} converges.
  2. Use Exercise 9 to prove that \displaystyle \sum_{n=2}^{\infty} \frac{1}{n \ln n} diverges and \displaystyle \sum_{n=2}^{\infty} \frac{1}{n (\ln n)^2} converges. Conclude that the converse of Example 3 is false.
  3. Let a_n \to 0, and let a grouping of \sum a_n have every group of at most d terms. Prove that if the grouping converges to S, then \sum a_n = S.
  4. Prove that a conditionally convergent series has a rearrangement whose partial sums tend to \infty.
  5. Let \sum a_n = S, and let \sigma be a bijection with \lvert \sigma(n) - n \rvert \le d for every n. Prove that \sum a_{\sigma(n)} = S.
  6. Prove that if a_n \ge 0 and \sum a_n converges, then \displaystyle \sum \frac{\sqrt{a_n}}{n} converges.
Answers
  1. N = 249, defeated for N = 248 by n = 249, m = 124{,}750. Since \displaystyle \frac{1}{4k^2 - 1} = \frac{1}{2}\left(\frac{1}{2k - 1} - \frac{1}{2k + 1}\right), the block is \displaystyle \frac{1}{2}\left(\frac{1}{2n + 1} - \frac{1}{2m + 1}\right) < \frac{1}{2(2n + 1)} \le \frac{1}{2(2N + 3)} for n > N, at most \displaystyle \frac{1}{1000} when 2N + 3 \ge 500. For the pair, \displaystyle \frac{1}{2}\left(\frac{1}{499} - \frac{1}{249{,}501}\right) > \frac{1}{1000}.
  2. The limit is \displaystyle \frac{1}{2}\ln 2. The block has n terms, each at least \displaystyle \frac{1}{4n - 1} > \frac{1}{4n}. With \displaystyle \varepsilon = \frac{1}{4}, the pair n = N + 1, m = 2n defeats every N, so Theorem 1 gives divergence. As in Example 4, the block is \displaystyle H_{4n} - \frac{3}{2}H_{2n} + \frac{1}{2}H_n, which is \displaystyle d_{4n} - \frac{3}{2}d_{2n} + \frac{1}{2}d_n + \frac{1}{2}\ln 2 and tends to \displaystyle \frac{1}{2}\ln 2.
  3. N = 19 and \lvert S \rvert \le 1. The block is at most \sum_{k=n+1}^{m} 2^{-k} = 2^{-n} - 2^{-m} < 2^{-n} \le 2^{-(N+1)} for n > N; 2^{19} = 524{,}288 < 10^6 \le 2^{20}. Theorem 4 gives \lvert S \rvert \le \sum 2^{-k} = 1.
  4. \displaystyle I_1 = \left[\frac{1}{3}, \frac{1}{2}\right], \displaystyle I_2 = \left[\frac{11}{30}, \frac{5}{12}\right], S = 2\ln 2 - 1 \approx 0.386294, and n = 31. Since \displaystyle \frac{1}{n(n + 1)} = \frac{1}{n} - \frac{1}{n + 1}, the series is \displaystyle \sum \frac{(-1)^{n+1}}{n} - \sum \frac{(-1)^{n+1}}{n + 1}, that is \ln 2 - (1 - \ln 2). We need \displaystyle \frac{1}{(n + 1)(n + 2)} \le 10^{-3}: 31 \cdot 32 = 992, 32 \cdot 33 = 1056.
  5. \displaystyle \limsup = \frac{3}{4} and \displaystyle \liminf = \frac{1}{4}, so the series converges absolutely by Theorem 5. \displaystyle \lvert a_n \rvert^{1/n} = n^{5/n}\frac{2 + (-1)^n}{4} and n^{1/n} \to 1 (The ratio and root tests), so even indices give \displaystyle \frac{3}{4} and odd ones \displaystyle \frac{1}{4}. For even n, \displaystyle \frac{a_{n+1}}{a_n} = \left(\frac{n + 1}{n}\right)^5 \frac{1}{4 \cdot 3^n} \to 0; for odd n it is \displaystyle \left(\frac{n + 1}{n}\right)^5 \frac{3^{n+1}}{4} \to \infty.
  6. The sum is \displaystyle \frac{\pi^2}{12}; positive terms \displaystyle \frac{\pi^2}{8}, negative terms \displaystyle -\frac{\pi^2}{24}. The even terms give \displaystyle \sum \frac{1}{4j^2} = \frac{\pi^2}{24}, the odd ones \displaystyle \frac{\pi^2}{6} - \frac{\pi^2}{24} = \frac{\pi^2}{8}. By Theorem 8 the order does not matter.
  7. At \displaystyle \frac{1}{2}\ln 2 \approx 0.3466 millimeters. After k blocks, \displaystyle R_{3k} = \left(H_{2k} - \frac{1}{2}H_k\right) - \frac{1}{2}H_{2k}, which is \displaystyle \frac{1}{2}(H_{2k} - H_k) and tends to \displaystyle \frac{1}{2}\ln 2; the two partial sums inside a block differ from R_{3k} by terms tending to 0.
  8. Four rightward then one leftward gives 2\ln 2; one rightward then four leftward gives 0. After k blocks the sum is \displaystyle H_{2pk} - \frac{1}{2}H_{pk} - \frac{1}{2}H_{qk}; its logarithms \displaystyle \ln 2pk - \frac{1}{2}\ln pk - \frac{1}{2}\ln qk give \displaystyle \ln 2 + \frac{1}{2}\ln\frac{p}{q}, its d terms tend to 0, and within a block the sums move by at most p + q terms tending to 0. The two targets need \displaystyle \frac{p}{q} = 4 and \displaystyle \frac{p}{q} = \frac{1}{4}.
  9. Proof. Let T_k = \sum_{j=0}^{k} 2^j a_{2^j}. The block of indices from 2^j to 2^{j+1} - 1 has 2^j terms, each at most a_{2^j}, so S_n \le T_k whenever n < 2^{k+1}. The block from 2^{j-1} + 1 to 2^j has 2^{j-1} terms, each at least a_{2^j}, so \displaystyle S_{2^k} \ge a_1 + \sum_{j=1}^{k} 2^{j-1}a_{2^j} \ge \frac{1}{2}T_k. Both sequences are nondecreasing, so one is bounded if and only if the other is, and Theorem 3 finishes. \square
  10. Proof. Put a_1 = a_2; the terms are positive and nonincreasing, and by Theorem 2 the first term does not matter. For k \ge 1, \displaystyle 2^k a_{2^k} = \frac{1}{k\ln 2} for the first series, a multiple of the harmonic series, so it diverges, and \displaystyle \frac{1}{k^2 (\ln 2)^2} for the second, a convergent p-series. For the first, \displaystyle n a_n = \frac{1}{\ln n} \to 0 with nonincreasing terms, yet the series diverges. \square
  11. Proof. Let \varepsilon > 0. Take K with \displaystyle \lvert S_{n_k} - S \rvert < \frac{\varepsilon}{2} for k \ge K, and M with \displaystyle \lvert a_i \rvert < \frac{\varepsilon}{2d} for i > M. Let n > \max\{n_K, M + d\} and take k with n_k \le n < n_{k+1}; then k \ge K, and since n - n_k < d, S_n - S_{n_k} is a sum of fewer than d terms with indices above n - d > M. So \displaystyle \lvert S_n - S \rvert < \frac{\varepsilon}{2} + d \cdot \frac{\varepsilon}{2d} = \varepsilon. \square
  12. Proof. With u_j and v_j as in the proof of Theorem 9, at stage s = 1, 2, \dots add unused u’s until the sum exceeds s, then subtract one unused v. Each stage ends, as the unused u’s have unbounded partial sums, and every term is used, as each stage uses one v and the targets outgrow any finite set of u’s. Take s_0 with v_j < 1 for j \ge s_0. For s \ge s_0, after stage s the sum is above s - 1 and stays there: each later stage adds until it passes the next target, then subtracts less than 1. So the partial sums tend to \infty. \square
  13. Proof. Let T_m = \sum_{j=1}^{m} a_{\sigma(j)}, A = \{1, \dots, m\} and B = \sigma(A). If j \le m, then \sigma(j) \le m + d; if i \le m - d, then \sigma^{-1}(i) \le i + d \le m, so i \in B. Hence A and B differ only in indices from m - d + 1 to m + d, at most 2d of them, and \lvert T_m - S_m \rvert \le 2d \max\{\lvert a_i \rvert : m - d < i \le m + d\}, which tends to 0 because a_n \to 0. So T_m \to S. \square
  14. Proof. From \displaystyle \left(\sqrt{a_n} - \frac{1}{n}\right)^2 \ge 0, \displaystyle \frac{\sqrt{a_n}}{n} \le \frac{1}{2}\left(a_n + \frac{1}{n^2}\right). The right side is half the sum of two convergent series, so Theorem 3 gives convergence. \square
math.wiki · CC-BY-SA-4.0