Subsequences and Bolzano-Weierstrass
A bounded sequence need not converge, but some part of it always does. Why does every bounded sequence of real numbers have a convergent subsequence, and what do the limits of its subsequences say about the sequence itself?
Where you are 5 in Sequences and series
Why this exists
A bounded sequence can diverge: Limits of sequences, epsilon and N proved that \displaystyle \cos \frac{2n\pi}{3} diverges, yet along the multiples of 3 it is constantly 1. The Bolzano-Weierstrass theorem says this always happens: every bounded sequence of real numbers has a convergent subsequence. It is the form of completeness that analysis uses most. The subsequences’ limits also measure a divergent sequence: they lie between its limit inferior and limit superior.
Definition
Subsequence. Let \{a_n\} be a sequence and let n_1 < n_2 < n_3 < \cdots be positive integers. The sequence \{a_{n_k}\}_{k=1}^{\infty} is a subsequence of \{a_n\}: infinitely many of its terms, in their original order. For example, \displaystyle \left\{\frac{1}{k^2}\right\} is the subsequence of \displaystyle \left\{\frac{1}{n}\right\} with n_k = k^2, and the K-tail of Limits of sequences, epsilon and N is the one with n_k = k + K.
Subsequential limit. A real number L is a subsequential limit of \{a_n\} if some subsequence converges to L.
Peak. An index m is a peak of \{a_n\} if a_m \ge a_n for every n > m: no later term is higher.
Limit superior and limit inferior. Let \{a_n\} be bounded. For each n \ge 1 let
the least upper bound and greatest lower bound of the tail from n; they need not be terms. The limit superior and limit inferior are
Theorem 7 shows that both limits exist. In running text we write \limsup a_n and \liminf a_n.
Intuition
Two pictures give two proofs.
Look to the right. An index is a peak when, looking right from its dot, nothing is higher (Figure 1). Infinitely many peaks form a nonincreasing subsequence. Past the last of finitely many, every term has a higher later term, and stepping from term to higher term builds an increasing one. Monotone convergence then finishes.
Halve the interval. If every term lies in [-M, M], one half of the interval holds terms for infinitely many indices; keep it and halve again. The kept intervals shrink to a point, and one term from each, at increasing indices, gives a subsequence squeezed onto it.
Results
The negation, tails and order limit theorems are those of Limits of sequences, epsilon and N; monotone convergence (Theorem 1) and nested intervals (Theorem 3) are those of The monotone convergence theorem; the squeeze theorem is in Convergence of sequences. Throughout, \{a_n\} is a sequence of real numbers.
Theorem 1 (indices grow). If n_1 < n_2 < \cdots are positive integers, then n_k \ge k for every k \ge 1.
Theorem 2 (subsequences of a convergent sequence). If a_n \to L, then every subsequence of \{a_n\} converges to L. Consequently, if two subsequences converge to different limits, or some subsequence diverges, then \{a_n\} diverges.
Theorem 3 (subsequential limits). L is a subsequential limit of \{a_n\} if and only if, for every \varepsilon > 0, \lvert a_n - L \rvert < \varepsilon for infinitely many n.
Theorem 4 (monotone subsequence). Every sequence of real numbers has a monotone subsequence.
Theorem 5 (Bolzano-Weierstrass). Every bounded sequence of real numbers has a convergent subsequence.
Theorem 6 (one subsequential limit). Let \{a_n\} be bounded. If every convergent subsequence of \{a_n\} has limit L, then a_n \to L.
Theorem 7 (limit superior and inferior). Let \{a_n\} be bounded, with s_n and t_n as in the definition.
(a) \{s_n\} is nonincreasing, \{t_n\} is nondecreasing, both converge, and \liminf a_n \le \limsup a_n.
(b) \limsup a_n and \liminf a_n are subsequential limits, and every subsequential limit L has \liminf a_n \le L \le \limsup a_n: they are the largest and the smallest.
(c) \{a_n\} converges if and only if \liminf a_n = \limsup a_n, and then its limit is their common value.
Part (b) gives a third proof of Theorem 5; The pigeonhole principle has a finite version of Theorem 4.
Why it is true
Theorem 1. Proof. By induction on k. Since n_1 is a positive integer, n_1 \ge 1. If n_k \ge k, then n_{k+1} > n_k \ge k, and n_{k+1} is an integer, so n_{k+1} \ge k + 1. \square
Theorem 2. Proof. Given \varepsilon > 0, take N with \lvert a_n - L \rvert < \varepsilon for every n > N. If k > N, then n_k \ge k > N by Theorem 1, so \lvert a_{n_k} - L \rvert < \varepsilon. The consequence follows: a convergent \{a_n\} would force one limit on every subsequence. \square
Theorem 3. Proof. If a_{n_k} \to L and \varepsilon > 0, there is K with \lvert a_{n_k} - L \rvert < \varepsilon for every k > K, and these n_k are infinitely many different indices.
Conversely, suppose the condition holds. Choose n_1 with \lvert a_{n_1} - L \rvert < 1. Having n_k, the set of n with \displaystyle \lvert a_n - L \rvert < \frac{1}{k+1} is infinite, so it has a member greater than n_k; let n_{k+1} be the least. Then n_1 < n_2 < \cdots and \displaystyle \lvert a_{n_k} - L \rvert < \frac{1}{k}. Given \varepsilon > 0, the The Archimedean property and density of the rationals gives an integer \displaystyle K > \frac{1}{\varepsilon}, and for k > K, \displaystyle \lvert a_{n_k} - L \rvert < \frac{1}{k} < \varepsilon. \square
Theorem 4. Proof. Infinitely many peaks. List them as m_1 < m_2 < \cdots. Since m_k is a peak and m_{k+1} > m_k, a_{m_k} \ge a_{m_{k+1}}: a nonincreasing subsequence.
Finitely many peaks. Let n_1 be larger than every peak (any n_1 \ge 1 if there is none). Having n_k larger than every peak, n_k is not a peak, so some n_{k+1} > n_k has a_{n_{k+1}} > a_{n_k}, and n_{k+1} is again larger than every peak. By induction this gives an increasing subsequence. \square
Theorem 5. First proof. Theorem 4 gives a monotone subsequence. Its terms are terms of \{a_n\}, so it is bounded, and monotone convergence makes it converge. \square
Second proof (bisection). Take M > 0 with \lvert a_n \rvert \le M for every n, and let I_1 = [-M, M]. Suppose I_k = [c_k, d_k] contains a_n for infinitely many n, and let \displaystyle m = \frac{c_k + d_k}{2}. Each such n has a_n \in [c_k, m] or a_n \in [m, d_k], and if both sets of indices were finite, so would be their union. Let I_{k+1} = [c_k, m] if it holds a_n for infinitely many n, and I_{k+1} = [m, d_k] otherwise. The intervals are nested with lengths \displaystyle \frac{2M}{2^{k-1}} \le \frac{2M}{k}, so the nested interval theorem gives one point L in every I_k.
Let n_1 = 1 and, having n_k, let n_{k+1} > n_k be an index with a_{n_{k+1}} \in I_{k+1}; infinitely many indices qualify. Both a_{n_k} and L lie in I_k, so \displaystyle \lvert a_{n_k} - L \rvert \le \frac{2M}{k}, and the squeeze theorem gives a_{n_k} \to L. \square
Theorem 6. Proof. Suppose a_n \not\to L. By the negation theorem there is \varepsilon > 0 with \lvert a_n - L \rvert \ge \varepsilon for infinitely many n; list them as m_1 < m_2 < \cdots. The bounded subsequence \{a_{m_j}\} has, by Theorem 5, a convergent subsequence \{a_{m_{j_i}}\}_{i=1}^{\infty}. Since j_1 < j_2 < \cdots and the m_j increase, m_{j_1} < m_{j_2} < \cdots: a subsequence of a subsequence is a subsequence of \{a_n\}. By hypothesis its limit is L, so \lvert a_{m_{j_i}} - L \rvert < \varepsilon for large i, against the choice of the m_j. \square
Theorem 7. Proof. (a) Take M with \lvert a_n \rvert \le M. Each tail is nonempty and lies in [-M, M], so -M \le t_n \le s_n \le M. Tails shrink as n grows, so s_{n+1} \le s_n and t_{n+1} \ge t_n. Monotone convergence gives both limits, and the order limit theorem gives \liminf a_n \le \limsup a_n.
(b) Write S = \limsup a_n. By monotone convergence S = \inf\{s_n : n \ge 1\}, so s_n \ge S for every n. Let \varepsilon > 0 and N \ge 1. Since s_n \to S, some p > N has s_p < S + \varepsilon. By the approximation property of the supremum (Least upper bounds and completeness), some k \ge p has a_k > s_p - \varepsilon \ge S - \varepsilon, and a_k \le s_p < S + \varepsilon. So \lvert a_k - S \rvert < \varepsilon for some k > N, for every N, and Theorem 3 makes S a subsequential limit.
If a_{n_k} \to L, then t_{n_k} \le a_{n_k} \le s_{n_k}. The outer sequences are subsequences of \{t_n\} and \{s_n\}, so by Theorem 2 they tend to \liminf a_n and S, and the order limit theorem gives \liminf a_n \le L \le S. The statements about \liminf a_n follow by applying these to \{-a_n\}, whose tail suprema are -t_n.
(c) If a_n \to L, then \limsup a_n and \liminf a_n are subsequential limits by (b), so both equal L by Theorem 2. If they are equal, t_n \le a_n \le s_n and the squeeze theorem gives convergence to the common value. \square
Worked examples
Two subsequential limits
Let \displaystyle a_n = (-1)^{n(n+1)/2}\left(1 + \frac{2}{n}\right). Find s_n, t_n, \limsup a_n, \liminf a_n and every subsequential limit, and decide whether \{a_n\} converges.
SolutionThe exponent \displaystyle \frac{n(n+1)}{2} is even exactly when n \equiv 0 or 3 \pmod 4, so the signs run -, -, +, + and repeat: the terms begin \displaystyle -3, -2, \frac{5}{3}, \frac{3}{2}, -\frac{7}{5}, -\frac{4}{3}, \frac{9}{7}, \frac{5}{4}.
The positive terms \displaystyle 1 + \frac{2}{m} decrease as m grows, so the largest term of a tail is its first positive one: \displaystyle s_n = 1 + \frac{2}{m} for the least m \ge n with m \equiv 0 or 3 \pmod 4. Thus \displaystyle s_1 = s_2 = s_3 = \frac{5}{3}, \displaystyle s_4 = \frac{3}{2} and \displaystyle s_5 = s_6 = s_7 = \frac{9}{7}. Likewise \displaystyle t_n = -\left(1 + \frac{2}{m}\right) for the least m \ge n with m \equiv 1 or 2 \pmod 4: t_1 = -3, t_2 = -2 and \displaystyle t_3 = t_4 = t_5 = -\frac{7}{5}.
Since m \ge n, \displaystyle 1 < s_n \le 1 + \frac{2}{n} and \displaystyle -1 - \frac{2}{n} \le t_n < -1, so the squeeze theorem gives \limsup a_n = 1 and \liminf a_n = -1. These differ, so \{a_n\} diverges by Theorem 7(c).
There are no other subsequential limits. Let L \ne \pm 1 and \displaystyle \varepsilon = \frac{1}{2}\min\{\lvert L - 1 \rvert, \lvert L + 1 \rvert\}. Each a_n is within \displaystyle \frac{2}{n} of its sign \sigma_n = \pm 1, so for \displaystyle n > \frac{2}{\varepsilon} the triangle inequality gives
By Theorem 3 the subsequential limits are exactly 1 and -1.
Peaks and the subsequence they build
For each sequence, find the peaks and the monotone subsequence that the proof of Theorem 4 builds, taking the least index allowed at each step: (a) \displaystyle a_n = \frac{2 + (-1)^n}{n}; (b) \displaystyle b_n = (-1)^n \frac{n}{n+1}.
Solution(a) The terms are \displaystyle \frac{3}{n} at even n and \displaystyle \frac{1}{n} at odd n. An even n is a peak, since every later term is \displaystyle \frac{3}{m} or \displaystyle \frac{1}{m} with m > n, below \displaystyle \frac{3}{n}. An odd n is not, since \displaystyle \frac{3}{n+1} > \frac{1}{n} (as 3n > n + 1). The proof takes all the peaks: \displaystyle a_{2k} = \frac{3}{2k}, decreasing to 0 (Figure 1).
(b) The terms are \displaystyle -\frac{1}{2}, \frac{2}{3}, -\frac{3}{4}, \frac{4}{5}, \ldots. No index is a peak: an odd index is followed by a positive term, and an even n has \displaystyle b_{n+2} = \frac{n+2}{n+3} > \frac{n}{n+1} = b_n. So the proof starts at n_1 = 1 and each time takes the least later index with a strictly larger term: \displaystyle b_2 = \frac{2}{3} > -\frac{1}{2} gives n_2 = 2, the odd terms are negative, so n_3 = 4, and in general n_k = 2k - 2 for k \ge 2. The subsequence \displaystyle -\frac{1}{2}, \frac{2}{3}, \frac{4}{5}, \frac{6}{7}, \ldots increases to 1.
The odd terms of (b) form another monotone subsequence, decreasing to -1.
A zoetrope under a strobe light
Naomi spins a zoetrope drum at a steady \displaystyle \sqrt{2} turns per second, and Arjun lights it with a strobe that flashes once a second, starting at time 0 with a marked slit on the start line. At the nth flash the slit appears
of a turn past the line. Show, before computing anything, that some subsequence of \{r_n\} converges. Then use the pairs (p, q) = (1, 1), (7, 5), (41, 29), \ldots and (3, 2), (17, 12), (99, 70), \ldots, each obtained from the one before by (p, q) \mapsto (3p + 4q, 2p + 3q), to find a subsequence converging to 0 and one converging to 1.
SolutionEvery r_n lies in [0, 1), so Theorem 5 gives a convergent subsequence. It names no limit; the pairs do. Expanding both squares shows (3p + 4q)^2 - 2(2p + 3q)^2 = p^2 - 2q^2, so the map keeps p^2 - 2q^2 fixed, and it makes q strictly larger, so each family of flashes q indexes a subsequence.
The first family has p^2 - 2q^2 = -1, so
which lies in (0, 1). So \displaystyle \lfloor q\sqrt{2} \rfloor = p and \displaystyle r_q = \frac{1}{q\sqrt{2} + p} < \frac{1}{q}. At the flashes q = 1, 5, 29, 169 the slit is 0.4142, 0.0711, 0.0122, 0.0021 of a turn past the line. The kth of these q is at least k (Theorem 1), so \displaystyle r_q < \frac{1}{k} along the subsequence, which tends to 0.
The second family has p^2 - 2q^2 = 1, so \displaystyle p - q\sqrt{2} = \frac{1}{p + q\sqrt{2}} lies in (0, 1), \displaystyle \lfloor q\sqrt{2} \rfloor = p - 1, and \displaystyle r_q = 1 - \frac{1}{p + q\sqrt{2}}. At q = 2, 12, 70 the slit is 0.8284, 0.9706, 0.9949 of a turn past the line, and this subsequence tends to 1 (Figure 2).
So \{r_n\} diverges (Theorem 2), and its subsequential limit 1 is no value of r_n; on the drum, 0 and 1 are the start line, approached from its two sides. In fact every number in [0, 1] is a subsequential limit, because \displaystyle \sqrt{2} is irrational; that is proved in number theory, not here.
A recursion that zigzags to the square root of 2
Let x_1 = 1 and \displaystyle x_{n+1} = 1 + \frac{1}{1 + x_n}. Prove that \displaystyle x_n \to \sqrt{2}.
SolutionProof. By induction every x_n is positive. The terms \displaystyle 1, \frac{3}{2}, \frac{7}{5}, \frac{17}{12}, \frac{41}{29}, \ldots are the fractions \displaystyle \frac{p}{q} of Example 3 and alternate around \displaystyle \sqrt{2}, so \{x_n\} is not monotone; its odd- and even-indexed subsequences are. Two steps of the recursion give x_{n+2} = h(x_n), where
For x, y > 0, \displaystyle h(x) - h(y) = \frac{x - y}{(2x + 3)(2y + 3)}, so h is increasing. Also \displaystyle h(\sqrt{2}) = \sqrt{2}, since \displaystyle (3\sqrt{2} + 4)(3 - 2\sqrt{2}) = \sqrt{2} and \displaystyle (2\sqrt{2} + 3)(3 - 2\sqrt{2}) = 1, and
Odd terms. \displaystyle x_1 = 1 < \sqrt{2}, and if \displaystyle 0 < x_n < \sqrt{2}, then \displaystyle x_{n+2} = h(x_n) < h(\sqrt{2}) = \sqrt{2} and x_{n+2} - x_n = h(x_n) - x_n > 0. So \{x_{2k-1}\} increases and is bounded above by \displaystyle \sqrt{2}, and by monotone convergence it has a limit A \ge 1. Even terms. \displaystyle x_2 = \frac{3}{2} > \sqrt{2}, and the same argument with the inequalities reversed shows that \{x_{2k}\} decreases and is bounded below by \displaystyle \sqrt{2}, with a limit \displaystyle B \ge \sqrt{2}.
In x_{2k+1} = h(x_{2k-1}) the left side tends to A (tails) and the right side to h(A) (sum, product and quotient laws), so A = h(A), 2 - A^2 = 0, and \displaystyle A = \sqrt{2} since A > 0. Likewise \displaystyle B = \sqrt{2}.
The two subsequences use every index. Given \varepsilon > 0, there is K with \displaystyle \lvert x_{2k-1} - \sqrt{2} \rvert < \varepsilon and \displaystyle \lvert x_{2k} - \sqrt{2} \rvert < \varepsilon for every k > K. Every n > 2K is 2k - 1 or 2k with k > K, so \displaystyle \lvert x_n - \sqrt{2} \rvert < \varepsilon. \square
Boundaries
- Bounded is needed. The sequence a_n = n has no convergent subsequence: a_{n_k} = n_k \ge k, so every subsequence is unbounded. An unbounded sequence may still have one (Exercise 5).
- Completeness is needed. The fractions x_n of Example 4 are rational and lie in \displaystyle \left[1, \frac{3}{2}\right], but every subsequence converges in \mathbb{R} to \displaystyle \sqrt{2} (Theorem 2), so none converges to a rational number. Theorem 5 fails in \mathbb{Q}, and Exercise 11 shows that it is one more form of the least-upper-bound property.
- The limit can leave the set. In Example 3 every term lies in [0, 1) and a subsequence tends to 1. A set in which every sequence has a subsequence converging to a point of the set is called compact; Open, closed and compact sets in R proves that in \mathbb{R} these are the closed bounded sets.
- Existence, not construction. Whether a half holds infinitely many terms is not decided by any finite list of terms, so Theorem 5 promises a limit without naming it.
- Higher dimensions. The theorem holds in \mathbb{R}^2 (Exercise 14) and in \mathbb{R}^n. In infinite-dimensional spaces it fails, a fact stated without proof that belongs to a later analysis course.
Where this goes next
Cauchy sequences proves that a Cauchy sequence converges: Theorem 5 gives a convergent subsequence, and the closeness of the terms pulls the rest along. Open, closed and compact sets in R proves the Heine-Borel theorem on \mathbb{R} from Theorem 5, and Extreme and intermediate value theorems, proved uses it to show that a continuous function on [a, b] attains a maximum.
Exercises
Subsequential limits
- Let \displaystyle a_n = \frac{(-1)^n n + 2}{n + 1}. (a) Find the limits of \{a_{2k}\} and \{a_{2k-1}\}, and conclude that \{a_n\} diverges. (b) Find s_1, t_1, \limsup a_n and \liminf a_n.
- Find every subsequential limit of \displaystyle c_n = \cos\frac{n\pi}{4} + \frac{1}{n}, and find \limsup c_n and \liminf c_n.
- Let \displaystyle d_n = \left(1 + \frac{1}{n}\right)^{(-1)^n n}. Show that s_n = e and \displaystyle t_n = \frac{1}{e} for every n, and find \limsup d_n and \liminf d_n.
- Find s_n and t_n for \displaystyle g_n = \frac{(-1)^n}{n}, and check that \limsup g_n = \liminf g_n.
- Let \displaystyle f_n = \frac{(1 - (-1)^n)\,n^2 + 2}{n}. Show that \{f_n\} is unbounded, that 0 is its only subsequential limit, and that it diverges. Which hypothesis of Theorem 6 fails?
Peaks and monotone subsequences
- Find every peak of \displaystyle a_n = \frac{n^2}{2^n} and the monotone subsequence that the proof of Theorem 4 builds.
- Find every peak of \displaystyle b_n = \frac{(n - 3)^2}{n^2 + 1} and the increasing subsequence that the proof of Theorem 4 builds when each index is the least allowed. What is its limit?
Bisection and explicit subsequences
- Run the bisection proof of Theorem 5 on \displaystyle h_n = \sin\frac{n\pi}{6} with M = 1, through I_5. Then run it again, choosing the right half whenever it holds h_n for infinitely many n. Which subsequential limit does each run find?
- Arjun speeds the zoetrope to \displaystyle \sqrt{3} turns per second, so the slit appears at \displaystyle r_n = n\sqrt{3} - \lfloor n\sqrt{3} \rfloor. The map (p, q) \mapsto (2p + 3q, p + 2q) keeps p^2 - 3q^2 fixed. Use the pairs (1, 1), (5, 3), (19, 11), \ldots to find a subsequence of \{r_n\} tending to 0, and give r_q to four decimal places at q = 1, 3, 11.
Proofs
- Prove that a_n \to L if and only if every subsequence of \{a_n\} has a further subsequence converging to L.
- Prove that a monotone sequence with a convergent subsequence converges. Then prove that an ordered field in which every bounded sequence has a convergent subsequence has the least-upper-bound property.
- Let \{a_n\} be bounded and S = \limsup a_n. Prove that for every \varepsilon > 0, a_n < S + \varepsilon for every n from some index on, and a_n > S - \varepsilon for infinitely many n. Prove conversely that a number with both properties is \limsup a_n.
- Prove that \limsup (a_n + b_n) \le \limsup a_n + \limsup b_n for bounded \{a_n\} and \{b_n\}, and give an example where the inequality is strict.
- Prove that every bounded sequence of points (x_n, y_n) in \mathbb{R}^2 has a subsequence along which both x_{n_k} and y_{n_k} converge.
Answers
- (a) \displaystyle a_{2k} = 1 + \frac{1}{2k + 1} \to 1 and \displaystyle a_{2k-1} = -1 + \frac{3}{2k} \to -1, different limits, so \{a_n\} diverges by Theorem 2. (b) \displaystyle s_1 = \frac{4}{3}, t_1 = -1, \limsup a_n = 1, \liminf a_n = -1. Both subsequences decrease, the even-indexed terms exceed 1 and the odd-indexed ones are at most \displaystyle \frac{1}{2}, so s_n is the first even-indexed term from n on and \displaystyle 1 < s_n \le 1 + \frac{1}{n+1}. The odd-indexed terms decrease to -1 without reaching it, so t_n = -1 for every n.
- The subsequential limits are \displaystyle 1, \frac{\sqrt{2}}{2}, 0, -\frac{\sqrt{2}}{2}, -1, with \limsup c_n = 1 and \liminf c_n = -1. Along each class n \equiv j \pmod 8 the cosine is the constant \displaystyle \cos\frac{j\pi}{4}, so c_n tends to it. For any other L, let \varepsilon be half its distance to the nearest of the five values; for \displaystyle n > \frac{1}{\varepsilon}, c_n is within \displaystyle \frac{1}{n} < \varepsilon of one of them, so \lvert c_n - L \rvert > \varepsilon, and Theorem 3 excludes L.
- \limsup d_n = e and \displaystyle \liminf d_n = \frac{1}{e}. The even-indexed terms \displaystyle \left(1 + \frac{1}{n}\right)^n increase to e and stay in \displaystyle \left[\frac{9}{4}, e\right) (The monotone convergence theorem, Theorem 5). The odd-indexed terms are reciprocals of such numbers, so they decrease to \displaystyle \frac{1}{e} and stay in \displaystyle \left(\frac{1}{e}, 1\right). So every tail has supremum e and infimum \displaystyle \frac{1}{e}.
- For even n, \displaystyle s_n = \frac{1}{n} and \displaystyle t_n = -\frac{1}{n+1}; for odd n, \displaystyle s_n = \frac{1}{n+1} and \displaystyle t_n = -\frac{1}{n}, because the positive terms sit at even indices and decrease, and likewise the negative ones. Both tend to 0, so \limsup g_n = \liminf g_n = 0 and g_n \to 0 by Theorem 7(c).
- Odd-indexed terms are \displaystyle 2n + \frac{2}{n}, even-indexed terms \displaystyle \frac{2}{n}. The odd-indexed terms pass every bound, so \{f_n\} is unbounded and diverges. A convergent subsequence is bounded by some B, so it holds no odd index n with 2n > B; from some k on, \displaystyle f_{n_k} = \frac{2}{n_k} \le \frac{2}{k}, so its limit is 0, the limit of the even-indexed terms. Theorem 6 fails because \{f_n\} is not bounded.
- The peaks are every n \ge 3. The ratio \displaystyle \frac{a_{n+1}}{a_n} = \frac{(n + 1)^2}{2n^2} is less than 1 exactly when n \ge 3, so the terms \displaystyle \frac{1}{2}, 1, \frac{9}{8} rise and then decrease; n = 1, 2 are below a_3. The proof takes all the peaks: \displaystyle \frac{9}{8}, 1, \frac{25}{32}, \ldots, decreasing to 0 because the ratio is at most \displaystyle \frac{8}{9} for n \ge 3.
- Indices 2, 6, 7, 8, 9, \ldots, limit 1. The only peak is n = 1: b_1 = 2, while (n - 3)^2 < n^2 + 1 makes every later term less than 1. The derivative of \displaystyle \frac{(x - 3)^2}{x^2 + 1} is \displaystyle \frac{2(x - 3)(3x + 1)}{(x^2 + 1)^2} > 0 for x > 3, so the terms increase from n = 3 on, and b_6 > b_2. The proof starts at n_1 = 2 with \displaystyle b_2 = \frac{1}{5}; b_3 = 0, \displaystyle b_4 = \frac{1}{17} and \displaystyle b_5 = \frac{2}{13} are smaller and \displaystyle b_6 = \frac{9}{37} is larger, so n_2 = 6, and then each index is the next.
- The terms repeat with period 12. Left half first: [-1, 1], [-1, 0], \displaystyle \left[-1, -\frac{1}{2}\right], \displaystyle \left[-1, -\frac{3}{4}\right], \displaystyle \left[-1, -\frac{7}{8}\right], the last holding -1 but not \displaystyle -\frac{\sqrt{3}}{2} \approx -0.866; the limit is -1 = \liminf h_n. Right half first: [-1, 1], [0, 1], \displaystyle \left[\frac{1}{2}, 1\right], \displaystyle \left[\frac{3}{4}, 1\right], \displaystyle \left[\frac{7}{8}, 1\right]; the limit is 1 = \limsup h_n.
- r_q \approx 0.7321, 0.1962, 0.0526. Expanding gives (2p + 3q)^2 - 3(p + 2q)^2 = p^2 - 3q^2, so these pairs all have p^2 - 3q^2 = -2, and q \mapsto p + 2q increases q. Then \displaystyle q\sqrt{3} - p = \frac{2}{q\sqrt{3} + p} lies in (0, 1), so \displaystyle r_q = \frac{2}{q\sqrt{3} + p} < \frac{2}{q}, which tends to 0 along the flashes q = 1, 3, 11, 41, \ldots
- Proof. If a_n \to L, every subsequence converges to L by Theorem 2 and is a further subsequence of itself. If a_n \not\to L, the negation theorem gives \varepsilon > 0 and m_1 < m_2 < \cdots with \lvert a_{m_j} - L \rvert \ge \varepsilon; every further subsequence of \{a_{m_j}\} keeps this property, so none converges to L. \square
- Proof. Let \{a_n\} be nondecreasing with a_{n_k} \to L. For each n and every k \ge n, n_k \ge k \ge n, so a_{n_k} \ge a_n, and the order limit theorem gives a_n \le L. Given \varepsilon > 0, take K with a_{n_K} > L - \varepsilon; for n > n_K, L - \varepsilon < a_{n_K} \le a_n \le L. So a_n \to L; for nonincreasing sequences use \{-a_n\}. This uses only the ordered field rules and the order limit theorem, so it holds in any ordered field F. If every bounded sequence in F has a convergent subsequence, a nondecreasing sequence bounded above (and below by its first term) has one, so it converges, and Theorem 2 of The monotone convergence theorem gives the least-upper-bound property. \square
- Proof. Since s_n \to S, choose N with s_N < S + \varepsilon; for n \ge N, a_n \le s_n \le s_N < S + \varepsilon. The proof of Theorem 7(b) found \lvert a_k - S \rvert < \varepsilon, so a_k > S - \varepsilon, for infinitely many k. Conversely, let S' have both properties and \varepsilon > 0. The first gives s_n \le S' + \varepsilon for large n, so \limsup a_n \le S' + \varepsilon. The second gives, for each n, some k \ge n with a_k > S' - \varepsilon, so s_n > S' - \varepsilon and \limsup a_n \ge S' - \varepsilon. As \varepsilon is arbitrary, \limsup a_n = S'. \square
- Proof. For every k \ge n, a_k + b_k \le s_n(a) + s_n(b), where s_n(a) and s_n(b) are the tail suprema, so the tail supremum of \{a_n + b_n\} is at most s_n(a) + s_n(b). The sum law and the order limit theorem give the inequality of the limits. \square With a_n = (-1)^n and b_n = (-1)^{n+1}, the left side is 0 and the right side is 2.
- Proof. If x_n^2 + y_n^2 \le M^2 for every n, then \lvert x_n \rvert \le M and \lvert y_n \rvert \le M. Theorem 5 gives x_{n_k} \to x. The sequence \{y_{n_k}\} is bounded, so Theorem 5 gives y_{n_{k_j}} \to y. The indices n_{k_j} increase, so this is a subsequence of the points, and x_{n_{k_j}} \to x by Theorem 2. \square