Limits of sequences, epsilon and N
Calculus defined the limit of a sequence and then used the product, quotient and order laws on trust. What does it take to prove that a sequence does not converge, and why do products, quotients and inequalities survive the passage to the limit?
Where you are 5 in Sequences and series
Why this exists
Convergence of sequences defined convergence, the sequence form of the definition in The epsilon delta definition of a limit, and proved the sum law, the squeeze theorem and that a convergent sequence is bounded. For the product and quotient laws it pointed to the function case. Analysis cannot leave those pointers in place: every later theorem in this course is proved by passing to a limit, multiplying sequences, dividing them, comparing them, and each step needs a proved theorem behind it.
This lesson negates the definition, proves the product and quotient laws in full, proves that limits respect \le, and shows that only the tail of a sequence matters.
Definition
A sequence of real numbers is a function from the positive integers to \mathbb{R}. Its value at n is the term a_n, and the sequence is written \{a_n\} or \{a_n\}_{n=1}^{\infty}; its set of terms is \{a_n : n \ge 1\}. The sequence is bounded if there is B with \lvert a_n \rvert \le B for every n.
Convergence. The sequence \{a_n\} converges to the real number L, written \lim_{n \to \infty} a_n = L or a_n \to L, if for every \varepsilon > 0 there is a number N such that
A sequence that converges to some real number is convergent; otherwise it diverges.
The cutoff N is chosen after \varepsilon and may depend on it. It may be any real number; an integer above it exists by the The Archimedean property and density of the rationals, and raising N keeps the condition true, so a proof may always take N to be a positive integer.
Eventually, infinitely often. A statement about n holds eventually if there is N such that it holds for every n > N, and for infinitely many n if the set of n where it holds is infinite. So a_n \to L says: for every \varepsilon > 0, eventually \lvert a_n - L \rvert < \varepsilon.
Tails. For an integer K \ge 0, the K-tail of \{a_n\} is the sequence \{a_{n+K}\}_{n=1}^{\infty}: the terms a_{K+1}, a_{K+2}, \ldots, with the first K dropped.
Intuition
Draw the terms as dots (n, a_n) with a band of half-width \varepsilon around the line y = L. Convergence says every band, however thin, eventually holds every dot. Failure to converge to L is the opposite: one band that dots keep leaving, however far out one starts looking. A single dot outside proves nothing, since it may come before the cutoff; what defeats L is a supply of escaping dots that never runs out (Figure 1).
For a product, the error splits in two:
The first piece is a small number times b_n, which stays bounded because it converges; the second is a fixed number times a small one. A quotient needs more: \displaystyle \frac{1}{b_n} is huge where b_n is near 0, so the proof first shows that when B \ne 0 the terms eventually keep a safe distance from 0 (Figure 2).
Results
Theorem 1 (negation). The sequence \{a_n\} does not converge to L if and only if there is \varepsilon > 0 such that \lvert a_n - L \rvert \ge \varepsilon for infinitely many n. It diverges if and only if this holds for every real number L, with \varepsilon allowed to depend on L.
Theorem 2 (uniqueness). A sequence converges to at most one number.
Theorem 3 (tails). For every integer K \ge 0, \{a_n\} converges to L if and only if its K-tail converges to L. Consequently, if a_n = b_n for every n > M, the two sequences both converge, to the same limit, or both diverge.
Theorem 4 (product law). If a_n \to A and b_n \to B, then a_n b_n \to AB. In particular c\,a_n \to cA for every constant c, and with the sum law, a_n - b_n \to A - B.
Theorem 5 (away from zero). If b_n \to B and B \ne 0, there is N_0 with \displaystyle \lvert b_n \rvert > \frac{\lvert B \rvert}{2} for every n > N_0. In particular b_n \ne 0 eventually.
Theorem 6 (quotient law). If a_n \to A, b_n \to B and B \ne 0, then \displaystyle \frac{1}{b_n} \to \frac{1}{B} and \displaystyle \frac{a_n}{b_n} \to \frac{A}{B}. The quotients are taken from an index beyond which b_n \ne 0, which Theorem 5 supplies; by Theorem 3 the starting index does not affect the limit.
Theorem 7 (sign persistence). If a_n \to A and A > c, then a_n > c eventually. If A < c, then a_n < c eventually.
Theorem 8 (order limit theorem). If a_n \to A, b_n \to B and a_n \le b_n eventually, then A \le B. In particular, if \alpha \le a_n \le \beta eventually, then \alpha \le A \le \beta.
Why it is true
Theorem 1. Proof. Write the definition with quantifiers and negate it one symbol at a time, as in Negating a quantified statement. The statement “\forall \varepsilon > 0, \exists N, \forall n > N: \lvert a_n - L \rvert < \varepsilon” has negation
It remains to see that “for every N there is n > N with property P” says the same as “P holds for infinitely many n”. If the set of such n were finite, its largest element (or 0, if it is empty) would be an N with no such n beyond it. If the set is infinite, it has elements beyond every N, because only finitely many positive integers are at most N. Divergence is this statement for every L. \square
Theorem 2. Proof. Suppose a_n \to L and a_n \to L', and let \varepsilon > 0. There are N_1 and N_2 with \displaystyle \lvert a_n - L \rvert < \frac{\varepsilon}{2} for n > N_1 and \displaystyle \lvert a_n - L' \rvert < \frac{\varepsilon}{2} for n > N_2. For one integer n beyond both, the Triangle inequality gives \lvert L - L' \rvert \le \lvert L - a_n \rvert + \lvert a_n - L' \rvert < \varepsilon. A nonnegative number below every positive number is 0, so L = L'. \square
Theorem 3. Proof. Write t_n = a_{n+K}. Suppose a_n \to L, and let N serve \varepsilon for \{a_n\}. If n > N then n + K > N, so \lvert t_n - L \rvert < \varepsilon; the same N serves the tail. Conversely, suppose t_n \to L and N serves \varepsilon for \{t_n\}. If n > N + K then n - K > N, so \lvert a_n - L \rvert = \lvert t_{n-K} - L \rvert < \varepsilon; the cutoff N + K serves \{a_n\}. For the consequence, take an integer K \ge M: the two sequences have the same K-tail. \square
Theorem 4. Proof. A convergent sequence is bounded (Convergence of sequences), so there is M > 0 with \lvert b_n \rvert \le M for every n. Let \varepsilon > 0. Choose N_1 with \displaystyle \lvert a_n - A \rvert < \frac{\varepsilon}{2M} for n > N_1, and N_2 with \displaystyle \lvert b_n - B \rvert < \frac{\varepsilon}{2(\lvert A \rvert + 1)} for n > N_2; the +1 keeps the denominator positive when A = 0. For n > \max\{N_1, N_2\}, the splitting above and the triangle inequality give
For c\,a_n, take \{b_n\} to be the constant sequence c. For the difference, a_n - b_n = a_n + (-1)b_n. \square
Theorem 5. Proof. Use \displaystyle \varepsilon = \frac{\lvert B \rvert}{2} > 0 in the definition: there is N_0 with \displaystyle \lvert b_n - B \rvert < \frac{\lvert B \rvert}{2} for n > N_0. For such n, \displaystyle \lvert B \rvert \le \lvert B - b_n \rvert + \lvert b_n \rvert < \frac{\lvert B \rvert}{2} + \lvert b_n \rvert, so \displaystyle \lvert b_n \rvert > \frac{\lvert B \rvert}{2}. \square
Theorem 6. Proof. Let N_0 be as in Theorem 5, and let \varepsilon > 0. Choose N_1 with \displaystyle \lvert b_n - B \rvert < \frac{\varepsilon \lvert B \rvert^2}{2} for n > N_1. For n > \max\{N_0, N_1\} we have \displaystyle \frac{1}{\lvert b_n \rvert} < \frac{2}{\lvert B \rvert}, so
Then \displaystyle \frac{a_n}{b_n} = a_n \cdot \frac{1}{b_n} \to A \cdot \frac{1}{B} by Theorem 4. \square
Theorem 7. Proof. If A > c, use \varepsilon = A - c: eventually a_n > A - \varepsilon = c. If A < c, use \varepsilon = c - A: eventually a_n < A + \varepsilon = c. \square
Theorem 8. Proof. Suppose instead A > B, and let \displaystyle m = \frac{A + B}{2}, so B < m < A. By Theorem 7, eventually a_n > m and eventually b_n < m; also a_n \le b_n eventually. An n beyond all three cutoffs has b_n < m < a_n \le b_n, which is impossible. So A \le B. For the second statement, compare \{a_n\} with the constant sequences \alpha and \beta. \square
Worked examples
A cutoff for a difference of square roots
Prove from the definition that \displaystyle \sqrt{n + 1} - \sqrt{n} \to 0. Find the cutoff the proof gives for \varepsilon = 0.01, and the smallest integer cutoff.
SolutionNeither \displaystyle \sqrt{n + 1} nor \displaystyle \sqrt{n} converges, so no law applies. Multiply by the conjugate:
The inequality holds because \displaystyle \sqrt{n + 1} > \sqrt{n}. Let \varepsilon > 0 and \displaystyle N = \frac{1}{4\varepsilon^2}. If n > N then \displaystyle \sqrt{n} > \frac{1}{2\varepsilon}, so \displaystyle \lvert a_n - 0 \rvert < \frac{1}{2\sqrt{n}} < \varepsilon. That proves a_n \to 0.
For \varepsilon = 0.01 the proof gives N = 2500. The terms decrease, since \displaystyle \sqrt{n + 1} + \sqrt{n} increases with n, and a_n < 0.01 exactly when \displaystyle \sqrt{n + 1} + \sqrt{n} > 100. At n = 2500 the sum is \displaystyle \sqrt{2501} + 50 > 100; at n = 2499 it is \displaystyle 50 + \sqrt{2499} < 100. So the smallest integer cutoff is N = 2499; the definition asks only for some cutoff.
A sequence with two values
Prove that a_n = \cos \dfrac{2n\pi}{3} diverges.
SolutionThe terms are \displaystyle -\frac{1}{2}, -\frac{1}{2}, 1, repeating: a_n = 1 when 3 divides n, and \displaystyle a_n = -\frac{1}{2} otherwise. By Theorem 1 it is enough to give, for every real L, an \varepsilon > 0 with \lvert a_n - L \rvert \ge \varepsilon for infinitely many n. One \varepsilon serves every L. By the triangle inequality,
so one of the two distances is at least \displaystyle \frac{3}{4}. If \displaystyle \lvert 1 - L \rvert \ge \frac{3}{4}, every multiple of 3 is an n with \displaystyle \lvert a_n - L \rvert \ge \frac{3}{4}; otherwise every n not divisible by 3 is. Both sets are infinite, so \displaystyle \varepsilon = \frac{3}{4} defeats every L, and the sequence diverges. Figure 1 shows \displaystyle L = \frac{1}{2}.
No larger \varepsilon serves every L: at \displaystyle L = \frac{1}{4} every term is exactly \displaystyle \frac{3}{4} away, so any wider band around \displaystyle \frac{1}{4} holds them all.
Square roots pass to the limit
Prove: if a_n \ge 0 for every n and a_n \to A, then A \ge 0 and \displaystyle \sqrt{a_n} \to \sqrt{A}.
SolutionThe constant sequence 0 satisfies 0 \le a_n, so Theorem 8 gives A \ge 0; without this, \displaystyle \sqrt{A} might not exist. Let \varepsilon > 0.
Case A = 0. Choose N with a_n < \varepsilon^2 for n > N. Since the square root is increasing on [0, \infty), \displaystyle \sqrt{a_n} < \varepsilon.
Case A > 0. Multiply by the conjugate, using \displaystyle \sqrt{a_n} + \sqrt{A} \ge \sqrt{A} > 0:
Choose N with \displaystyle \lvert a_n - A \rvert < \varepsilon\sqrt{A} for n > N; then \displaystyle \left\lvert \sqrt{a_n} - \sqrt{A} \right\rvert < \varepsilon. \square
The n-th root of a constant
Prove that c^{1/n} \to 1 for every c > 0. For c = 5 and \varepsilon = 0.01, compare the cutoff from the proof with the smallest one.
SolutionCase c > 1. Then c^{1/n} > 1, for otherwise c = (c^{1/n})^n \le 1. Write c^{1/n} = 1 + d_n with d_n > 0. Every term in the binomial expansion of (1 + d_n)^n is nonnegative, so keeping the first two,
that is, \displaystyle 0 < d_n \le \frac{c - 1}{n}. The bound tends to 0, so the squeeze theorem gives d_n \to 0, and c^{1/n} = 1 + d_n \to 1.
Case c = 1. Every term is 1.
Case 0 < c < 1. Then \displaystyle \frac{1}{c} > 1, so \displaystyle \left(\frac{1}{c}\right)^{1/n} \to 1 by the first case, and \displaystyle c^{1/n} = 1 \big/ \left(\frac{1}{c}\right)^{1/n}. That limit is 1 \ne 0, so Theorem 6 gives c^{1/n} \to 1.
For c = 5 the bound \displaystyle \frac{4}{n} < 0.01 gives N = 400. Exactly, 5^{1/n} < 1.01 when \displaystyle n > \frac{\ln 5}{\ln 1.01} \approx 161.75, so the smallest integer cutoff is 161.
Leah’s beehive: averages of a convergent sequence
A sensor in Leah’s beehive records the temperature of the brood nest once a day, and her monitor shows the average of all readings so far. Malik models the daily readings as \displaystyle a_n = 95 + \frac{4}{n} degrees Fahrenheit, which tend to 95. Prove that the averages \displaystyle c_n = \frac{1}{n}(a_1 + a_2 + \cdots + a_n) tend to 95 too; in fact, for every sequence, a_n \to L implies c_n \to L. Compare a_{30} with c_{30}.
SolutionLet a_n \to L and \varepsilon > 0. Choose a positive integer K with \displaystyle \lvert a_k - L \rvert < \frac{\varepsilon}{2} for every k > K, and let S = \sum_{k=1}^{K} \lvert a_k - L \rvert, a fixed number once K is chosen. For n > K, split the sum at K:
Let \displaystyle N = \max\left\{K, \frac{2S}{\varepsilon}\right\}. For n > N, \displaystyle \frac{S}{n} < \frac{\varepsilon}{2}, so \lvert c_n - L \rvert < \varepsilon. \square
The first K terms are not dropped; their total pull S is divided by n, which outgrows it. For Malik’s model, \displaystyle c_n = 95 + \frac{4}{n}\left(1 + \frac{1}{2} + \cdots + \frac{1}{n}\right), and after 30 days a_{30} \approx 95.13 while c_{30} \approx 95.53: the average still remembers the warm early readings.
Boundaries
- The product and quotient laws need convergent factors. With \displaystyle a_n = \frac{1}{n} and b_n = n^2 the product n diverges; with b_n = n it is 1. A product can even converge when neither factor does: (-1)^n \cdot (-1)^n = 1.
- The quotient law needs B \ne 0. \displaystyle \frac{1/n}{1/n^2} = n diverges, while \displaystyle \frac{1/n}{1/n} = 1 converges.
- Strict inequalities are not preserved. \displaystyle \frac{1}{n} > 0 for every n, but the limit is 0. Theorem 8 concludes only A \le B, and it needs both sequences to converge.
- Finitely many terms, not infinitely many. The sequence equal to 1 at perfect squares and 0 elsewhere differs from the zero sequence infinitely often, and it diverges: every L is at least \displaystyle \frac{1}{2} from 0 or from 1, and each value occurs infinitely often.
Where this goes next
The monotone convergence theorem proves that a monotone bounded sequence converges, a fact equivalent to the least-upper-bound property. Subsequences and Bolzano-Weierstrass finds convergent subsequences of bounded sequences, and Cauchy sequences decides convergence without naming the limit. Limits of functions, epsilon and delta carries the pattern to functions, and Series convergence, proved applies these theorems to partial sums.
Exercises
Cutoffs from the definition
- Prove from the definition that \dfrac{4n - 1}{2n + 5} \to 2, and find the smallest integer N with \lvert a_n - 2 \rvert < 0.01 for every n > N.
- Show that \dfrac{n}{n^2 + 9} < \dfrac{1}{n}, and use it to prove \dfrac{n}{n^2 + 9} \to 0 with N = \dfrac{1}{\varepsilon}. For \varepsilon = 0.02, find the smallest integer cutoff.
Proving that a limit fails
- Let a_n = \dfrac{3n + 1}{n + 2}. Find the least n_0 such that \lvert a_n - 2.9 \rvert \ge 0.05 for every n \ge n_0, and conclude that a_n does not converge to 2.9.
- Let a_n = n for odd n and a_n = 0 for even n. Use Theorem 1 with \varepsilon = 1 to show that a_n does not converge to 0, then prove that \{a_n\} diverges.
- Write in symbols, and then in words, the negations of “\{a_n\} is bounded” and of “\{a_n\} converges”.
The laws in use
- Let \displaystyle a_n = 2 - \frac{1}{n} and \displaystyle b_n = 4 + \frac{3}{n}. Show that \lvert a_n - 2 \rvert < \delta and \lvert b_n - 4 \rvert < \delta give \lvert a_n b_n - 8 \rvert < \delta^2 + 6\delta. Take \displaystyle \delta = \frac{1}{1000} and state the cutoff this gives for \varepsilon = 0.01; then find the smallest integer cutoff.
- Let b_n = \dfrac{5n - 12}{n + 1}, so b_n \to 5. (a) Find the smallest integer N_0 with \displaystyle \lvert b_n \rvert > \frac{5}{2} for every n > N_0. (b) For \varepsilon = 0.001, find the cutoff N_1 that the proof of Theorem 6 asks of \{b_n\}. (c) Find the smallest integer cutoff for \displaystyle \left\lvert \frac{1}{b_n} - \frac{1}{5} \right\rvert < 0.001.
- The sequence a_n = \dfrac{n^2 - 4}{n^2 - 3n + 2} is defined for n \ge 3. Find its limit, naming the theorems used.
- Find \lim_{n \to \infty} \dfrac{7^{1/n} + 3}{2 - 5^{-1/n}}.
- Let a_n = 4 + (-1)^n. Show that \{a_n\} diverges, yet the averages \displaystyle c_n = \frac{1}{n}(a_1 + \cdots + a_n) converge, and find their limit.
- Let a_n = \dfrac{n^2 - 50n}{n^2 + 1}. Theorem 7 says \displaystyle a_n > \frac{1}{2} eventually. Find the smallest integer N with \displaystyle a_n > \frac{1}{2} for every n > N.
Limits and order
- Suppose a_n \to A and 2 \le a_n < 5 for every n > 10. Find every possible value of A, with a sequence for each.
Proofs
- Prove that if a_n \to 0 and \{b_n\} is bounded, then a_n b_n \to 0. Explain why Theorem 4 does not give this.
- Suppose one number N has the property that, for every \varepsilon > 0, \lvert a_n - L \rvert < \varepsilon for every n > N. Prove that a_n = L for every n > N.
- Prove that a_n \to L if and only if both a_{2n} \to L and a_{2n-1} \to L.
- Prove that if a_n \to A and b_n \to B, then \max\{a_n, b_n\} \to \max\{A, B\}.
- Let E \subseteq \mathbb{R} be nonempty and bounded above, with s = \sup E. Prove that some sequence of points of E converges to s, and that every convergent sequence of points of E has limit at most s.
Answers
- N = 547. \displaystyle a_n - 2 = \frac{4n - 1 - 4n - 10}{2n + 5} = -\frac{11}{2n + 5}, so \lvert a_n - 2 \rvert < \varepsilon exactly when \displaystyle 2n + 5 > \frac{11}{\varepsilon}; given \varepsilon > 0, \displaystyle N = \frac{1}{2}\left(\frac{11}{\varepsilon} - 5\right) works. For \varepsilon = 0.01 this is n > 547.5: n = 548 gives \displaystyle \frac{11}{1101} < 0.01, and n = 547 gives \displaystyle \frac{11}{1099} > 0.01.
- N = 49. Since n^2 + 9 > n^2, \displaystyle \frac{n}{n^2 + 9} < \frac{1}{n}, which is below \varepsilon when \displaystyle n > \frac{1}{\varepsilon}. For \varepsilon = 0.02 the proof gives N = 50. The terms decrease from n = 3 on (the numerator of a_{n+1} - a_n is 9 - n - n^2 < 0), \displaystyle a_{50} = \frac{50}{2509} < 0.02 and \displaystyle a_{49} = \frac{49}{2410} > 0.02.
- n_0 = 98. \displaystyle a_n = 3 - \frac{5}{n + 2}, so \displaystyle a_n - 2.9 = 0.1 - \frac{5}{n + 2}, which is at least 0.05 exactly when n + 2 \ge 100. At n = 97, \displaystyle 0 < 0.1 - \frac{5}{99} < 0.05. With \varepsilon = 0.05, infinitely many n have \lvert a_n - 2.9 \rvert \ge \varepsilon, so by Theorem 1, a_n \not\to 2.9.
- Every odd n has \lvert a_n - 0 \rvert = n \ge 1, and there are infinitely many odd n, so by Theorem 1, a_n \not\to 0. The sequence is not bounded, since for every B an odd n > B has a_n > B; a convergent sequence is bounded, so \{a_n\} diverges.
- Bounded is “\exists B, \forall n: \lvert a_n \rvert \le B”; its negation is “\forall B, \exists n: \lvert a_n \rvert > B”, for every B some term exceeds B in size. Converges is “\exists L, \forall \varepsilon > 0, \exists N, \forall n > N: \lvert a_n - L \rvert < \varepsilon”; its negation is “\forall L, \exists \varepsilon > 0, \forall N, \exists n > N: \lvert a_n - L \rvert \ge \varepsilon”, which is Theorem 1.
- The proof gives N = 3000; the smallest cutoff is N = 198. a_n b_n - 8 = (a_n - 2)b_n + 2(b_n - 4) and \lvert b_n \rvert < 4 + \delta, so \lvert a_n b_n - 8 \rvert < \delta(4 + \delta) + 2\delta = \delta^2 + 6\delta, which is 0.006001 < 0.01 at \displaystyle \delta = \frac{1}{1000}. The factors need \displaystyle \frac{1}{n} < \frac{1}{1000} and \displaystyle \frac{3}{n} < \frac{1}{1000}, so n > 3000. Directly, \displaystyle a_n b_n - 8 = \frac{2n - 3}{n^2}; for n \ge 2 this is below 0.01 exactly when n^2 - 200n + 300 > 0, that is \displaystyle n > 100 + \sqrt{9700} \approx 198.49, and at n = 1 its size is 1.
- (a) N_0 = 5. \displaystyle \lvert b_n - 5 \rvert = \frac{17}{n + 1}, so for n \ge 6, \displaystyle b_n \ge 5 - \frac{17}{7} = \frac{18}{7} > \frac{5}{2}, while \displaystyle b_5 = \frac{13}{6} < \frac{5}{2}. (b) N_1 = 1359: the proof needs \displaystyle \frac{17}{n + 1} < \frac{0.001 \cdot 5^2}{2} = \frac{1}{80}, that is n + 1 > 1360. (c) 682. For n \ge 3, b_n > 0 and \displaystyle \left\lvert \frac{1}{b_n} - \frac{1}{5} \right\rvert = \frac{17}{5(5n - 12)}, below 0.001 exactly when 5n - 12 > 3400, that is n > 682.4; at n = 1 and n = 2 the distance is far above 0.001.
- 1. For n \ge 3, \displaystyle \frac{(n - 2)(n + 2)}{(n - 1)(n - 2)} = \frac{n + 2}{n - 1} = \frac{1 + 2/n}{1 - 1/n}. The top tends to 1 by the sum law, the bottom to 1 \ne 0, so Theorem 6 gives 1. Starting at n = 3 is allowed by Theorem 3.
- 4. By Example 4, 7^{1/n} \to 1 and \displaystyle 5^{-1/n} = \left(\frac{1}{5}\right)^{1/n} \to 1. The top tends to 4 and the bottom to 1 \ne 0 (each bottom term exceeds 1, since 5^{-1/n} < 1), so Theorem 6 gives \displaystyle \frac{4}{1} = 4.
- Limit 4. The terms alternate 3, 5, 3, 5, \ldots, and every L is at least 1 from 3 or from 5, so \varepsilon = 1 and Theorem 1 show divergence. The sum of the first n terms is 4n - 1 for odd n and 4n for even n, so \displaystyle c_n = 4 - \frac{1}{n} or c_n = 4, and \displaystyle \lvert c_n - 4 \rvert \le \frac{1}{n}. The converse of Example 5 fails.
- N = 100. \displaystyle a_n > \frac{1}{2} exactly when 2n^2 - 100n > n^2 + 1, that is n^2 - 100n - 1 > 0, or \displaystyle n > 50 + \sqrt{2501} \approx 100.01. At n = 100, \displaystyle a_n = \frac{5000}{10001} < \frac{1}{2}.
- Every A with 2 \le A \le 5. Theorem 8 with constant sequences gives 2 \le A \le 5; the strict a_n < 5 yields only A \le 5. For 2 \le A < 5 the constant sequence A works, and for A = 5 the sequence \displaystyle 5 - \frac{1}{n}, which lies in [2, 5) for every n \ge 1.
- Let \lvert b_n \rvert \le M for every n, with M > 0 (raise M if needed). Given \varepsilon > 0, choose N with \displaystyle \lvert a_n \rvert < \frac{\varepsilon}{M} for n > N. Then \lvert a_n b_n - 0 \rvert \le M \lvert a_n \rvert < \varepsilon. Theorem 4 needs \{b_n\} to converge, and a bounded sequence need not, for example b_n = (-1)^n.
- Fix n > N. The number d = \lvert a_n - L \rvert is nonnegative and below every \varepsilon > 0. If d > 0, taking \varepsilon = d gives d < d, which is false. So d = 0 and a_n = L. This is why N comes after \varepsilon in the definition.
- Suppose a_n \to L, and let N serve \varepsilon. For every k > N, 2k > N and 2k - 1 \ge k > N, so \lvert a_{2k} - L \rvert < \varepsilon and \lvert a_{2k-1} - L \rvert < \varepsilon. Conversely, let N_1 serve \varepsilon for \{a_{2k}\} and N_2 for \{a_{2k-1}\}, and let M = \max\{N_1, N_2\}. If n > 2M, either n = 2k with k > M \ge N_1, or n = 2k - 1 with \displaystyle k > M + \frac{1}{2} > N_2; either way \lvert a_n - L \rvert < \varepsilon, so N = 2M serves \{a_n\}.
- For real x and y, \displaystyle \max\{x, y\} = \frac{1}{2}(x + y + \lvert x - y \rvert): if x \ge y the right side is \displaystyle \frac{2x}{2}, otherwise \displaystyle \frac{2y}{2}. By the difference law a_n - b_n \to A - B, and since \big\lvert \lvert u \rvert - \lvert v \rvert \big\rvert \le \lvert u - v \rvert, any cutoff for that limit also serves \lvert a_n - b_n \rvert \to \lvert A - B \rvert. The sum and constant-multiple laws give the limit \displaystyle \frac{1}{2}(A + B + \lvert A - B \rvert), which is \max\{A, B\}.
- For each n \ge 1, \displaystyle s - \frac{1}{n} < s is not an upper bound of E (the approximation property in Least upper bounds and completeness), so there is e_n \in E with \displaystyle s - \frac{1}{n} < e_n \le s. Both outer sequences tend to s, so the squeeze theorem gives e_n \to s. If x_n \in E for every n and x_n \to x, then x_n \le s for every n, since s is an upper bound, and Theorem 8 against the constant sequence s gives x \le s.