If φ(n) → a and ψ(n) → b then φ + ψ → a + b, φψ → ab and φ/ψ → a/b (when b ≠ 0); with these rules the limit of any rational function of n can be read off.
Proving every limit straight from the definition would be exhausting. §§63–68 proves, once and for all, that limits respect the arithmetic operations. After that, 2n2+4n+43n2−12n+5→23 needs no δ at all: it follows from 1/n→0 and a few rules. The proofs are also the first place where the definition is used as a tool rather than checked by hand, and the device they use, splitting the tolerance, appears in almost every later proof in analysis.
The idea
If ϕ(n) is nearly a and ψ(n) is nearly b, surely ϕ(n)+ψ(n) is nearly a+b. The only question is how nearly. The error of the sum is at most the sum of the errors:
∣(ϕ(n)+ψ(n))−(a+b)∣≤∣ϕ(n)−a∣+∣ψ(n)−b∣.
So if we want the sum within δ of a+b, it is enough to have each term within 21δ of its own limit. Each of those happens for large n, say from n1 and from n2 on, so both happen from n0=max(n1,n2) on.
ExploreSplit the tolerance
Try this.With “δ/2 each”, move δ and watch n1, n2 and n0=max(n1,n2). Then switch to “δ each” on the pair with errors of the same sign.
ϕ(n)=1+n1→1, ψ(n)=2+n1→2
φ(n) with band 1 ± 0.1; n₁ = 11
nφ
ψ(n) with band 2 ± 0.1; n₂ = 11
nψ
φ(n) + ψ(n) with band 3 ± 0.2; n₀ = max(n₁, n₂) = 11
nφ+ψ
n ≥ threshold, inside band
n ≥ threshold, outside
n₁, n₂, n₀
Tolerance for each term
Pair
Every term of the sum from n0=11 to n=4000 lies within δ of a+b=3. This is guaranteed: ∣ϕ+ψ−(a+b)∣≤∣ϕ−a∣+∣ψ−b∣<21δ+21δ.
What to notice
With half the tolerance each, the sum is always inside its band from n0 on: that is the triangle inequality at work. With the whole δ for each term, two errors of the same sign, each just under δ, add up to nearly 2δ, and the sum escapes. Halving is not a trick of style; it is what makes the argument correct.
Products need more care, because an error gets multiplied by the other factor. Writing ϕ(n)=a+ϕ1(n) and ψ(n)=b+ψ1(n), where ϕ1,ψ1→0 are the errors,
ϕ(n)ψ(n)−ab=aψ1(n)+bϕ1(n)+ϕ1(n)ψ1(n).
Each of the three terms on the right is small when the errors are small, so the product tends to ab. Quotients come from products once we know 1/ψ(n)→1/b, and the key point there is that ψ(n) stays away from 0 when b=0.
Precise statements
TheoremLimit of a sumHardy: Theorem I, §63
If ϕ(n)→a and ψ(n)→b, then ϕ(n)+ψ(n)→a+b.
TheoremLimit of a productHardy: Theorem II, §65
If ϕ(n)→a and ψ(n)→b, then ϕ(n)ψ(n)→ab. In particular kϕ(n)→ka for every constant k.
TheoremLimit of a reciprocalHardy: Theorem III, §66
If ϕ(n)→a and a=0, then 1/ϕ(n)→1/a.
Since ϕ(n)→a=0, ϕ(n) can be 0 for at most finitely many n (the proof shows ∣ϕ(n)∣>21∣a∣ for large n), and finitely many terms do not matter, so 1/ϕ(n) makes sense where it counts.
TheoremLimit of a quotientHardy: Theorem IV, §66
If ϕ(n)→a and ψ(n)→b with b=0, then ϕ(n)/ψ(n)→a/b.
TheoremRational functions of limitsHardy: Theorem V, §67
Let R=P/Q, where P and Q are polynomials in several variables. If ϕ(n)→a, ψ(n)→b, χ(n)→c,…, and Q(a,b,c,…)=0, then
R{ϕ(n),ψ(n),χ(n),…}→R(a,b,c,…).
TheoremLimit of a rational function of nHardy: §68
Let S(n)=b0nq+b1nq−1+⋯+bqa0np+a1np−1+⋯+ap with a0=0, b0=0. Then S(n)→0 if p<q; S(n)→a0/b0 if p=q; and if p>q, S(n)→+∞ or −∞ according as a0/b0 is positive or negative.
Proofs
ProofTheorem I: limit of a sum
Strategy
Given δ, ask each sequence to come within 21δ of its limit, take the later of the two thresholds, and add the two errors with the triangle inequality.
Let δ>0; get two thresholds
Since ϕ(n)→a, applying the definition with 21δ (also a positive number) gives n1 with ∣ϕ(n)−a∣<21δ for n≥n1. Similarly n2 with ∣ψ(n)−b∣<21δ for n≥n2.
Let n0=max(n1,n2)
For n≥n0 both inequalities hold.
Add the errors
For n≥n0, by the triangle inequality, ∣ϕ(n)+ψ(n)−(a+b)∣≤∣ϕ(n)−a∣+∣ψ(n)−b∣<21δ+21δ=δ.
∎
ProofTheorem II: limit of a product
Strategy
Split ϕψ−ab into three pieces using the errors ϕ1=ϕ−a, ψ1=ψ−b. Make both errors smaller than a number η chosen so that the three pieces together are below δ.
The identity
ϕψ−ab=(a+ϕ1)(b+ψ1)−ab=aψ1+bϕ1+ϕ1ψ1, so ∣ϕψ−ab∣≤∣a∣∣ψ1∣+∣b∣∣ϕ1∣+∣ϕ1∣∣ψ1∣.
Choose η=min{1,δ/(∣a∣+∣b∣+1)}
Then η>0 and η≤1. As in Theorem I there is an n0 such that ∣ϕ1(n)∣<η and ∣ψ1(n)∣<η for all n≥n0.
Estimate
For n≥n0: ∣ϕψ−ab∣<∣a∣η+∣b∣η+η⋅η≤(∣a∣+∣b∣+1)η≤δ, using η2≤η (as η≤1). The first inequality is strict because ∣ϕ1∣∣ψ1∣<η2, so this works even when a or b is 0.
The special case
A constant k is a sequence tending to k, so kϕ(n)→ka.
∎
Hardy’s own version takes ∣ϕ1∣<δ/(3∣b∣) and ∣ψ1∣<δ/(3∣a∣), assumes δ<3∣a∣∣b∣, and leaves the cases a=0 or b=0 to the reader; the choice of η above covers all cases at once.
ProofTheorem III: limit of a reciprocal
Strategy
ϕ1−a1=∣a∣∣ϕ∣∣ϕ−a∣. The top is small; the danger is a small denominator. So first keep ϕ(n) away from 0, then make the top small.
Keep ϕ(n) away from 0
Apply the definition with δ=21∣a∣>0: for n≥n1, ∣ϕ(n)−a∣<21∣a∣, so ∣ϕ(n)∣≥∣a∣−∣ϕ(n)−a∣>21∣a∣. In particular ϕ(n)=0.
Bound the error
For n≥n1: ϕ(n)1−a1=∣a∣∣ϕ(n)∣∣ϕ(n)−a∣<∣a∣22∣ϕ(n)−a∣.
Make it less than δ
Given δ>0, choose n2 with ∣ϕ(n)−a∣<21∣a∣2δ for n≥n2. For n≥max(n1,n2) the error is less than δ.
∎
Theorem IV is now immediate: ϕ/ψ=ϕ⋅(1/ψ)→a⋅(1/b) by Theorems III and II. Theorem V follows by building R in stages: each monomial Aϕpψq⋯ tends to Aapbq⋯ by Theorem II applied finitely many times, so P tends to P(a,b,…) by Theorem I applied finitely many times; the same holds for Q, and Theorem IV finishes the job because Q(a,b,…)=0.
The rational function of n in §68 is the most important application. Try it first.
ExploreOnly the top degrees matter
Try this.Set p=q=2 and look at the first 20 terms, then 600. Then try p=3,q=2 with a0 negative, and p=1,q=3.
S(n)=2n2+4n+4n2−12n+5
nS(n)
S(n)
(a₀/b₀) n^(p−q)
the limit
Since p=q: S(n)→a0/b0=0.5. The first terms can point the wrong way (the lower coefficients are chosen to mislead), but for large n the quotient behaves like the dashed curve (a0/b0)np−q. At n=60: S(n)=0.38756.
What to notice
The lower coefficients (such as the −12n on top) can make the early terms point the wrong way, but they lose in the end. Only p−q and the sign of a0/b0 decide the behaviour, and the terms hug the dashed curve (a0/b0)np−q.
Proof§68: limit of a rational function of n
Strategy
Divide the top by np and the bottom by nq. What remains is a rational function of 1/n, which tends to a0/b0 by Theorem V, multiplied by np−q, whose behaviour we know.
Rewrite
S(n)=np−qT(n), where T(n)=b0+b1/n+⋯+bq/nqa0+a1/n+⋯+ap/np.
T(n)→a0/b0
T(n)=R(1/n) for the rational function R(x)=(a0+a1x+⋯+apxp)/(b0+b1x+⋯+bqxq). Since 1/n→0 and the denominator at x=0 is b0=0, Theorem V gives T(n)→R(0)=a0/b0.
p≤q
If p<q, np−q=1/nq−p→0 (Examples XXIII, 1), so S(n)→0⋅a0/b0=0 by Theorem II. If p=q, np−q=1 and S(n)=T(n)→a0/b0.
p>q, a0/b0>0
For large n, T(n)>21a0/b0=c>0 (use the definition with δ=c). Then S(n)>cnp−q≥cn, which exceeds any Δ once n>Δ/c. So S(n)→+∞. If a0/b0<0, apply this to −S(n).
∎
When the limits are infinite or do not exist
Theorems I–IV assume that both limits exist. §64 works out what can be said about ϕ+ψ in the other cases, and the pattern is worth knowing. Some combinations are determined; others are indeterminate, meaning the behaviour of the sum depends on the particular sequences.
ϕ
ψ
ϕ+ψ
limit
→±∞, or oscillates (finitely or infinitely)
behaves like ψ
→+∞
→+∞, or oscillates finitely
→+∞
→+∞
→−∞
anything at all
→+∞
oscillates infinitely
→+∞ or oscillates infinitely
oscillates finitely
oscillates finitely
limit or oscillates finitely
oscillates finitely
oscillates infinitely
oscillates infinitely
oscillates infinitely
oscillates infinitely
anything at all
For example, with ϕ→+∞ and ψ→−∞: n+(−n)→0, n2−n→+∞, n−n2→−∞, (n+(−1)n)−n oscillates finitely, and (n2+(−1)nn)−n2 oscillates infinitely. Products have a similar table (§65), with “tends to 0” as a separate case, since 0× (anything bounded) →0.
Find the behaviour of (n+1n−1)2, (−1)n(n+1n−1)2, nn2+1, (−1)nnn2+1.
Solution
(n−1)/(n+1) has p=q=1 and a0/b0=1, so it tends to 1, and its square to 1 (Theorem II). With the factor (−1)n, the even terms tend to 1 and the odd ones to −1; the sequence is bounded by 1, so it oscillates finitely (as for (−1)n, no l is within 21 of both 1 and −1). (n2+1)/n has p=2>q=1 and a0/b0=1>0, so it tends to +∞. With (−1)n, the sizes tend to +∞ and the signs alternate: it oscillates infinitely.
ExerciseWhich tend to a limit?●●○Hardy: Examples XXVI, 2
For even n, cos221nπ=1 and sin221nπ=0; for odd n it is the other way round.
Solution
For even n the three terms are 1, 1/n and n/n=1. For odd n they are 1/n, 1/n2 and 1/n2. So the first oscillates finitely (even terms 1, odd terms tending to 0), the second tends to 0 (all terms are at most 1/n), and the third oscillates finitely (even terms 1, odd terms tending to 0). Only the second tends to a limit.
ExerciseShifting a rational function●●○Hardy: Examples XXVI, 3
With S(n) the general rational function of §68, show that S(n+1)/S(n)→1.
Hint
Write S(n)=np−qT(n) with T(n)→a0/b0=0.
Solution
S(n)S(n+1)=(nn+1)p−qT(n)T(n+1) for large n (when S(n)=0). Here (n+1)/n=1+1/n→1, so its (p−q)-th power tends to 1 (Theorems II and III, applied ∣p−q∣ times). Also T(n+1)→a0/b0 and T(n)→a0/b0=0, so their quotient tends to 1 by Theorem IV. By Theorem II the product tends to 1.
ExerciseInfinite plus bounded●○○Hardy: §64, 2
Prove that if ϕ(n)→+∞ and ψ(n) oscillates finitely, then ϕ(n)+ψ(n)→+∞.
Solution
Finite oscillation gives K with ∣ψ(n)∣<K, so ϕ(n)+ψ(n)>ϕ(n)−K. Given Δ, choose n0 with ϕ(n)>Δ+K for n≥n0 (Definition II with Δ+K). Then ϕ(n)+ψ(n)>Δ for n≥n0. The same proof works whenever ψ is bounded, in particular when ψ tends to a limit.
ExerciseInfinite times finitely oscillating●●○Hardy: §65 (i)
Let ϕ(n)=n. Find the behaviour of ϕ(n)ψ(n) for ψ(n)=2+(−1)n, −2−(−1)n and (−1)n, and explain why these are the only three possibilities when ϕ→+∞ and ψ oscillates finitely.
Solution
n(2+(−1)n)≥n→+∞; n(−2−(−1)n)≤−n→−∞; (−1)nn oscillates infinitely. In general ϕψ cannot tend to a limit or oscillate finitely: since ψ oscillates finitely it does not tend to 0, so there are ε>0 and infinitely many n with ∣ψ(n)∣≥ε (by the negation theorem of “For large values of n”); at those n, ∣ϕ(n)ψ(n)∣≥εϕ(n), which is unbounded. An unbounded sequence tends to ±∞ or oscillates infinitely.
Check yourself
Check yourselfChoose one
In the proof of Theorem I, why do we ask for ∣ϕ(n)−a∣<21δ rather than <δ?
Why
The triangle inequality bounds the error of the sum by the sum of the errors. With δ each we would only get an error below 2δ. Definition I may be applied with any positive number in place of δ, including 21δ; that freedom is what the proof uses.
Check yourselfTrue or false?
If ϕ(n)+ψ(n) tends to a limit, then ϕ(n) and ψ(n) both tend to limits.
Why
False. ϕ(n)=(−1)n and ψ(n)=−(−1)n both oscillate, but their sum is 0. Theorem I runs in one direction only.
Check yourselfEnter a number
What is lim2n2+4n+43n2−12n+5?
Why
Here p=q=2, so the limit is the ratio of the leading coefficients, 3/2. Dividing top and bottom by n2 gives (3−12/n+5/n2)/(2+4/n+4/n2)→3/2.
Check yourselfChoose one
ϕ(n)→+∞ and ψ(n)→−∞. What can you say about ϕ(n)+ψ(n)?
Why
∞−∞ is indeterminate: n−n→0, n2−n→+∞, n−n2→−∞, (n+(−1)n)−n oscillates finitely and (n2+(−1)nn)−n2 oscillates infinitely.
Hardy’s language
Summary
Summary
Limits respect +, −, × and ÷ (when the limit of the denominator is not 0), and therefore every rational function of convergent sequences.
The method: split the tolerance (21δ each), take the larger threshold (max(n1,n2)), and combine the errors with the triangle inequality.
For reciprocals, first keep the sequence away from 0: ∣ϕ(n)∣>21∣a∣ for large n.
A rational function of n behaves like (a0/b0)np−q: it tends to 0, to a0/b0, or to ±∞ as p<q, p=q or p>q.
When a limit is infinite or does not exist, some combinations are determined and others (∞−∞, two infinite oscillations) can do anything.