Chapter IV · Limits of functions of a positive integral variable

The definition of a limit

§§58–61 35 min

What it means, exactly, for ϕ(n)\phi(n) to tend to ll or to +∞+\infty, and the small print that the definition hides.

Why this matters

Everything in analysis (continuity, derivatives, integrals, infinite series) is built on one idea: a quantity getting as close as we like to a value. In §58 Hardy turns the vague phrase “ϕ(n)\phi(n) is nearly ll when nn is large” into a definition precise enough to prove things with. This lesson is about that definition: how to read it, how to use it, and how it can fail.

The idea

Take the sequence ϕ(n)=1/n\phi(n) = 1/n. Everyone agrees that it “tends to 00”, but what exactly is being claimed? Not that some term equals 00: none does. Not that the terms are small: 1/10001/1000 is small for one purpose and huge for another.

Hardy’s answer, prepared in §56, is a challenge. Someone names a tolerance δ\delta, say 0.010.01. We reply with a place n0n_0 in the sequence from which on every term is within δ\delta of 00: here n0=101n_0 = 101 will do, because 1/n<0.011/n < 0.01 as soon as n>100n > 100. The challenger names a smaller δ\delta; we find a later n0n_0. The claim “1/n1/n tends to 00” means that we can always answer, however small δ\delta is.

Explore Answer the challenge for 1/n1/n

Try this. Set δ\delta to about 0.10.1 and find the smallest n0n_0 for which no term from n0n_0 on is orange. Then make δ\delta ten times smaller and do it again. (Use the menu to show more terms when you need them.)

ϕ(n)=1n\phi(n) = \frac{1}{n} , candidate limit l=0l = 0
nφ(n)
  • band l ± δ
  • l
  • n₀ (drag)
  • n ≥ n₀, good
  • n ≥ n₀, bad
10 terms with n ≥ n₀ fall outside the band: n = 1, 2, 3, 4, 5, ….
What to notice

Each smaller δ\delta needs a larger n0n_0, roughly 1/δ1/\delta of them. But some n0n_0 always works: the green band never stays empty to the right. That is what “tends to 00” means.

Two features of the game matter, and both are easy to miss.

  • The order of the moves. First δ\delta is chosen, then n0n_0. Our n0n_0 may depend on δ\delta, which is why Hardy writes it as n0(δ)n_0(\delta). A single n0n_0 that worked for every δ\delta at once would be a much stronger (and usually false) claim.
  • “From n0n_0 on”, not “at n0n_0”. It is not enough that one term is inside the band. Every term with n≥n0n \ge n_0 must be.

Precise statements

Here is Hardy’s definition in modern dress. The notation ∣ϕ(n)−l∣|\phi(n) - l| is the distance between ϕ(n)\phi(n) and ll; Hardy calls it their difference “taken positively”.

Definition Limit of a sequence Hardy: Definition I, §58

The sequence ϕ(n)\phi(n) tends to the limit ll as n→∞n \to \infty if, for every positive number δ\delta, however small, there is a number n0(δ)n_0(\delta) such that

∣ϕ(n)−l∣<δfor all n≥n0(δ).|\phi(n) - l| < \delta \quad \text{for all } n \ge n_0(\delta).

We then write lim⁡n→∞ϕ(n)=l\lim_{n\to\infty} \phi(n) = l, or simply ϕ(n)→l\phi(n) \to l.

In §59 Hardy reads the definition as a picture, and the widget above draws exactly that picture. Plot the points (n,ϕ(n))(n, \phi(n)) and draw the lines y=l−δy = l - \delta and y=l+δy = l + \delta: this is the green band. The definition says that, however narrow the band, we can draw a vertical line x=n0x = n_0 (the purple line) such that every point on or to the right of it lies inside the band.

Sequences like n2n^2 do not approach any number, but they do something just as definite: they eventually exceed any bound. §60 gives this its own definition, with a large number Δ\Delta in place of the small δ\delta.

Definition Tending to +∞+\infty Hardy: Definition II, §60

The sequence ϕ(n)\phi(n) tends to +∞+\infty if, for every number Δ\Delta, however large, there is a number n0(Δ)n_0(\Delta) such that

ϕ(n)>Δfor all n≥n0(Δ).\phi(n) > \Delta \quad \text{for all } n \ge n_0(\Delta).

We write ϕ(n)→+∞\phi(n) \to +\infty (or ϕ(n)→∞\phi(n) \to \infty).

The symbol ∞\infty here is not a number. “ϕ(n)→+∞\phi(n) \to +\infty” is a single phrase, defined as a whole by the sentence above, just as “ϕ(n)→l\phi(n) \to l” is.

Our first result is the first of Hardy’s “points concerning the definitions” (§61): only the tail of a sequence matters.

Theorem Finitely many terms do not matter Hardy: §61 (1)

Let ψ(n)=ϕ(n)\psi(n) = \phi(n) for every nn except finitely many. If ϕ(n)→l\phi(n) \to l then ψ(n)→l\psi(n) \to l, and if ϕ(n)→+∞\phi(n) \to +\infty then ψ(n)→+∞\psi(n) \to +\infty.

The other points in §61 are warnings, and each one corrects a natural misreading.

  1. Infinitely many changes can matter. Change 1/n1/n to 11 whenever nn is a multiple of 100100. However far out we put n0n_0, there are later multiples of 100100, and for δ≤1\delta \le 1 they all lie outside the band. The new sequence has no limit.
  2. The test is for every n≥n0n \ge n_0. For that same sequence we can certainly find a single n0n_0 (say n0=101n_0 = 101) at which ∣ϕ(n0)∣<δ|\phi(n_0)| < \delta. That proves nothing; the multiples of 100100 beyond it still fail.
  3. Signs need the absolute value. For (−1)n/n(-1)^n/n the terms fall on both sides of the limit 00, so the condition must be ∣ϕ(n)∣<δ|\phi(n)| < \delta, not ϕ(n)<δ\phi(n) < \delta.
  4. A limit need not be a value, and a value need not be the limit. 1/n1/n never equals its limit 00. The constant sequence 0,0,0,…0, 0, 0, \dots equals its limit at every term. The sequence (sin⁡12nπ)/n(\sin \tfrac12 n\pi)/n equals its limit 00 at every even nn and at no odd nn.
  5. Large is not the same as tending to infinity. (−1)nn(-1)^n n is enormous in size for large nn, but it changes sign at every step, so it never stays above Δ=0\Delta = 0. A sequence that tends to +∞+\infty is eventually positive.
Explore Infinitely many exceptions

Try this. This is 1/n1/n with every hundredth term changed to 11. Find an n0n_0 that works for δ=0.5\delta = 0.5 with only the first 40 terms shown. Then show 250 terms.

ϕ(n)={1100∣n1/notherwise\phi(n) = \begin{cases}1 & 100\mid n\\ 1/n & \text{otherwise}\end{cases} , candidate limit l=0l = 0
nφ(n)
  • band l ± δ
  • l
  • n₀ (drag)
  • n ≥ n₀, good
  • n ≥ n₀, bad
2 terms with n ≥ n₀ fall outside the band: n = 1, 2.
What to notice

Every apparent success is undone by the next multiple of 100100. The terms in between do tend to 00, but the definition demands every term from n0n_0 on, so no n0n_0 works for any δ≤1\delta \le 1.

Proofs

A proof that ϕ(n)→l\phi(n) \to l is a strategy for the game: a rule that produces a winning n0n_0 from any δ\delta the challenger names. Here is the simplest example, which Hardy works out at the end of §58.

Proof 1/n→01/n \to 0
Strategy

Given an arbitrary δ>0\delta > 0, we must name an n0n_0 and show that 1/n<δ1/n < \delta for every n≥n0n \ge n_0. Since 1/n1/n gets smaller as nn grows, it is enough to make 1/n0<δ1/n_0 < \delta, that is, n0>1/δn_0 > 1/\delta.

Let δ>0\delta > 0 be given

The proof must work for every positive δ\delta, so we may not assume anything about it except δ>0\delta > 0.

Choose n0=[1/δ]+1n_0 = [1/\delta] + 1

Here [x][x] is the integral part of xx (the greatest integer not exceeding xx). Since [x]>x−1[x] > x - 1, we have n0>1/δn_0 > 1/\delta. This is the smallest choice that works, but any larger integer would do as well.

Check every n≥n0n \ge n_0

If n≥n0n \ge n_0 then n>1/δn > 1/\delta, and since both sides are positive, 1/n<δ1/n < \delta. Also 1/n>01/n > 0, so ∣1/n−0∣=1/n<δ|1/n - 0| = 1/n < \delta.

Conclude

Every δ>0\delta > 0 has a matching n0(δ)n_0(\delta), which is exactly what Definition I requires. Hence lim⁡1/n=0\lim 1/n = 0.

The rule n0(δ)=[1/δ]+1n_0(\delta) = [1/\delta] + 1 is a function of δ\delta. Drawing its graph shows Hardy’s remark that “the smaller δ\delta is, the larger, as a rule, will be n0(δ)n_0(\delta)”. It grows without bound as δ\delta shrinks, but it is finite for every δ>0\delta > 0.

Explore The winning rule n0(δ)n_0(\delta) for 1/n1/n

Try this. Drag δ\delta from 11 down towards 0.050.05 and watch the step function climb.

δn₀(δ)
  • smallest n₀ for 1/n
  • 1/δ
  • your δ
For δ=0.15\delta = 0.15 the smallest choice is n0=[1/δ]+1=7n_0 = [1/\delta] + 1 = 7. Halve δ and n₀ roughly doubles; but it is always finite.
What to notice

n0(δ)n_0(\delta) jumps at δ=1/2,1/3,1/4,…\delta = 1/2, 1/3, 1/4, \dots and stays close to the dashed curve 1/δ1/\delta. A proof only needs some valid n0n_0, not the smallest one, so a cruder rule such as n0=[2/δ]+1n_0 = [2/\delta] + 1 would prove the same thing.

The proof of our theorem uses the same strategy, with one new idea: push n0n_0 past every term that was changed.

Proof Finitely many terms do not matter
Strategy

Use the n0n_0 that works for ϕ\phi, but if necessary move it to the right of every altered term. Beyond that point, ψ\psi and ϕ\phi are the same sequence.

Name the altered terms

Only finitely many nn have ψ(n)≠ϕ(n)\psi(n) \ne \phi(n), so they have a largest member; call it NN. (If there are none, take N=0N = 0.) For every n>Nn > N we have ψ(n)=ϕ(n)\psi(n) = \phi(n).

Take any δ>0\delta > 0 and use the limit of ϕ\phi

Since ϕ(n)→l\phi(n) \to l, there is an n0n_0 with ∣ϕ(n)−l∣<δ|\phi(n) - l| < \delta for all n≥n0n \ge n_0.

Let n1=max⁡(n0,N+1)n_1 = \max(n_0, N + 1)

If n≥n1n \ge n_1 then n≥n0n \ge n_0, so ∣ϕ(n)−l∣<δ|\phi(n) - l| < \delta; and n>Nn > N, so ψ(n)=ϕ(n)\psi(n) = \phi(n). Together, ∣ψ(n)−l∣<δ|\psi(n) - l| < \delta for all n≥n1n \ge n_1.

The case +∞+\infty is the same

Replace “∣ϕ(n)−l∣<δ|\phi(n) - l| < \delta” by “ϕ(n)>Δ\phi(n) > \Delta” throughout. The argument does not change.

Notice why this argument cannot handle infinitely many changes: the step “they have a largest member” fails, and with it the choice of n1n_1.

Worked examples and exercises

Try each one before opening the hints. Hardy’s examples are statements to be proved, so “show” is always implied.

Exercise Tending to −∞-\infty ●○○ Hardy: §60

Hardy leaves the definition of “ϕ(n)→−∞\phi(n) \to -\infty” to the reader. Write it down, and use it to prove that −n2→−∞-n^2 \to -\infty.

Hint
Mirror Definition II: “however large Δ\Delta is, ϕ(n)\phi(n) is eventually below −Δ-\Delta.”
Hint
For −n2<−Δ-n^2 < -\Delta you need n2>Δn^2 > \Delta. When Δ>0\Delta > 0 that means n>Δn > \sqrt{\Delta}.

Solution

Definition. ϕ(n)→−∞\phi(n) \to -\infty if for every number Δ\Delta, however large, there is an n0(Δ)n_0(\Delta) such that ϕ(n)<−Δ\phi(n) < -\Delta for all n≥n0(Δ)n \ge n_0(\Delta). Equivalently, −ϕ(n)→+∞-\phi(n) \to +\infty.

Proof for −n2-n^2. Let Δ\Delta be given. If Δ≤0\Delta \le 0 then n0=1n_0 = 1 works, since −n2≤−1<0≤−Δ-n^2 \le -1 < 0 \le -\Delta. If Δ>0\Delta > 0, take n0=[Δ]+1n_0 = [\sqrt{\Delta}] + 1, so that n0>Δn_0 > \sqrt\Delta. For n≥n0n \ge n_0 we get n>Δn > \sqrt{\Delta}, so n2>Δn^2 > \Delta and −n2<−Δ-n^2 < -\Delta.

Exercise ●○○ Hardy: Examples XXIII, 1

Let kk be a rational number. Show that nk→+∞n^k \to +\infty if k>0k > 0, that nk→0n^k \to 0 if k<0k < 0, and that nk→1n^k \to 1 if k=0k = 0. For k=4k = 4, how large must nn be to make n4>10 000n^4 > 10\,000?

Hint
For k>0k > 0: nk>Δn^k > \Delta exactly when n>Δ1/kn > \Delta^{1/k} (with Δ>0\Delta > 0).
Hint
For k<0k < 0 write nk=1/n∣k∣n^k = 1/n^{|k|} and ask when this is less than δ\delta.

Solution

k>0k > 0. Let Δ>0\Delta > 0 (for Δ≤0\Delta \le 0 any n0n_0 works, since nk>0n^k > 0). Since x↦xkx \mapsto x^{k} is increasing on the positive numbers, nk>Δn^k > \Delta if and only if n>Δ1/kn > \Delta^{1/k}. So n0=[Δ1/k]+1n_0 = [\Delta^{1/k}] + 1 works, and nk→+∞n^k \to +\infty.

k<0k < 0. Let δ>0\delta > 0. Now 0<nk=1/n∣k∣0 < n^k = 1/n^{|k|}, and 1/n∣k∣<δ1/n^{|k|} < \delta if and only if n∣k∣>1/δn^{|k|} > 1/\delta, that is, n>δ−1/∣k∣n > \delta^{-1/|k|}. So n0=[δ−1/∣k∣]+1n_0 = [\delta^{-1/|k|}] + 1 works, and nk→0n^k \to 0.

k=0k = 0. Then nk=1n^k = 1 for every nn, a constant sequence, and n0=1n_0 = 1 works for every δ\delta.

k=4k = 4. n4>104n^4 > 10^4 exactly when n>10n > 10, so n≥11n \ge 11 is needed. In the same way n4>108n^4 > 10^8 exactly when n≥101n \ge 101.

Exercise ●○○ Hardy: Examples XXIII, 5

Let ϕ(n)=1 000 000/n\phi(n) = 1\,000\,000/n and ψ(n)=n/1 000 000\psi(n) = n/1\,000\,000. Show that ϕ(n)→0\phi(n) \to 0 and ψ(n)→+∞\psi(n) \to +\infty, even though ϕ(n)>ψ(n)\phi(n) > \psi(n) for every n<1 000 000n < 1\,000\,000.

Hint
Copy the proof for 1/n1/n, with 106/δ10^6/\delta in place of 1/δ1/\delta.

Solution

Given δ>0\delta > 0, take n0=[106/δ]+1n_0 = [10^6/\delta] + 1. For n≥n0n \ge n_0 we have n>106/δn > 10^6/\delta, so 0<106/n<δ0 < 10^6/n < \delta. Hence ϕ(n)→0\phi(n) \to 0.

Given Δ>0\Delta > 0, take n0=[106Δ]+1n_0 = [10^6\Delta] + 1. For n≥n0n \ge n_0 we have n>106Δn > 10^6\Delta, so n/106>Δn/10^6 > \Delta. Hence ψ(n)→+∞\psi(n) \to +\infty.

Finally, ϕ(n)>ψ(n)\phi(n) > \psi(n) means 1012>n210^{12} > n^2, that is, n<106n < 10^6. The first million terms say nothing about the limit. That is the theorem on altering finitely many terms seen from another side: only the tail counts.

Exercise ●●○ Hardy: Examples XXIII, 6

Decide how each of these sequences behaves as n→∞n \to \infty, and prove it:

1n−(−1)n,n−(−1)n,n{1−(−1)n}.\frac{1}{n - (-1)^n}, \qquad n - (-1)^n, \qquad n\{1 - (-1)^n\}.
Hint
In the first two, n−(−1)nn - (-1)^n is either n−1n - 1 or n+1n + 1. So it is at least n−1n - 1.
Hint
Write out the first few terms of the third: 2,0,6,0,10,0,…2, 0, 6, 0, 10, 0, \dots. Which definition does it break, and how?
Solution

First. For n≥2n \ge 2 we have n−(−1)n≥n−1≥1n - (-1)^n \ge n - 1 \ge 1, so 0<ϕ(n)≤1/(n−1)0 < \phi(n) \le 1/(n-1). Given δ>0\delta > 0, take n0=[1/δ]+2n_0 = [1/\delta] + 2. Then for n≥n0n \ge n_0 we have n−1>1/δn - 1 > 1/\delta, so ϕ(n)≤1/(n−1)<δ\phi(n) \le 1/(n-1) < \delta. Hence ϕ(n)→0\phi(n) \to 0.

Second. n−(−1)n≥n−1n - (-1)^n \ge n - 1. Given Δ\Delta, take n0=max⁡(1,[Δ]+2)n_0 = \max(1, [\Delta] + 2). For n≥n0n \ge n_0, n−1>Δn - 1 > \Delta, so the term exceeds Δ\Delta. Hence it tends to +∞+\infty.

Third. The terms are 00 for even nn and 2n2n for odd nn. They do not tend to +∞+\infty: with Δ=1\Delta = 1, every even n≥n0n \ge n_0 gives a term 0<10 < 1, whatever n0n_0 is. They do not tend to a limit ll either: with δ=1\delta = 1, the odd terms 2n2n exceed l+1l + 1 as soon as n>(l+1)/2n > (l+1)/2, so infinitely many terms lie outside the band, whatever n0n_0 is. (In the language of the next lesson, the sequence oscillates infinitely.)

Exercise ●○○ Hardy: Examples XXIII, 7

Let θ\theta be any real number. Show that (sin⁡nθπ)/n→0(\sin n\theta\pi)/n \to 0.

Hint
You know nothing useful about sin⁡nθπ\sin n\theta\pi except that it lies between −1-1 and 11.

Solution

Since ∣sin⁡nθπ∣≤1|\sin n\theta\pi| \le 1, we have ∣ϕ(n)∣≤1/n|\phi(n)| \le 1/n for every nn. Given δ>0\delta > 0, take n0=[1/δ]+1n_0 = [1/\delta] + 1. For n≥n0n \ge n_0, ∣ϕ(n)−0∣≤1/n<δ|\phi(n) - 0| \le 1/n < \delta. The same n0n_0 works for every θ\theta: bounding the hard part by something simple is the main technique in this chapter.

Exercise When one n0n_0 serves every δ\delta ●●○ Hardy: §58

Show that a single NN works for every δ>0\delta > 0 (that is, ∣ϕ(n)−l∣<δ|\phi(n) - l| < \delta for all n≥Nn \ge N and all δ>0\delta > 0) if and only if ϕ(n)=l\phi(n) = l for every n≥Nn \ge N.

Hint
One direction is immediate. For the other, fix n≥Nn \ge N: what do you know about the number ∣ϕ(n)−l∣|\phi(n) - l|?

Solution

If ϕ(n)=l\phi(n) = l for all n≥Nn \ge N, then ∣ϕ(n)−l∣=0<δ|\phi(n) - l| = 0 < \delta for every such nn and every δ>0\delta > 0.

Conversely, fix n≥Nn \ge N and let d=∣ϕ(n)−l∣≥0d = |\phi(n) - l| \ge 0. By hypothesis d<δd < \delta for every δ>0\delta > 0. If d>0d > 0, then taking δ=d\delta = d gives d<dd < d, which is impossible. So d=0d = 0, that is, ϕ(n)=l\phi(n) = l. Thus the only sequences for which n0n_0 need not depend on δ\delta are those that are eventually constant. For all other convergent sequences, n0n_0 really is a function of δ\delta.

Check yourself

Check yourself Choose one

Which statement says that ϕ(n)→l\phi(n) \to l?

Why

The order matters: δ\delta comes first and n0n_0 may depend on it. The first option fixes one n0n_0 for all δ\delta, which forces the sequence to be eventually constant (last exercise). The third only asks for one term near ll; the sequence 1/n1/n with 11 at multiples of 100100 satisfies it for l=1l = 1 without tending to 11. The fourth says every term equals ll.

Check yourself Enter a number

For ϕ(n)=1/n\phi(n) = 1/n, l=0l = 0 and δ=0.004\delta = 0.004, what is the smallest n0n_0 that works?

Why

1/n<0.0041/n < 0.004 means n>250n > 250. The term 1/250=0.0041/250 = 0.004 is not less than 0.0040.004, so the smallest n0n_0 is 251=[1/δ]+1251 = [1/\delta] + 1.

Check yourself True or false?

If ϕ(n)→l\phi(n) \to l, then ϕ(n)≠l\phi(n) \neq l for every nn.

Why

False. A limit may be a value, even infinitely often: 1/n1/n with every hundredth term replaced by 00 still tends to 00, and a constant sequence equals its limit everywhere. What is true is that the limit need not be a value at all, as 1/n1/n shows.

Check yourself True or false?

(−1)n n→+∞(-1)^n\, n \to +\infty, because its terms become arbitrarily large.

Why

False. Take Δ=0\Delta = 0. Every odd nn gives a negative term, so no n0n_0 makes all later terms exceed 00. The sizes ∣ϕ(n)∣=n|\phi(n)| = n do tend to +∞+\infty, but the sequence itself does neither that nor anything else definite: it oscillates.

Check yourself Choose one

Let ϕ(n)=1\phi(n) = 1 when nn is a multiple of 100100 and ϕ(n)=1/n\phi(n) = 1/n otherwise. Which is correct?

Why

Whatever n0n_0 is chosen, there is a multiple of 100100 beyond it, where ∣ϕ(n)−0∣=1>12|\phi(n) - 0| = 1 > \tfrac12. The second option is Hardy’s point (3): finding one good n0n_0-th term proves nothing. The last option fails for the same reason with the roles swapped: the terms 1/n1/n stay far from 11.

Hardy’s language

A limit is not a value of the function: it is something quite distinct from these values, though it is defined by its relations to them and may possibly be equal to some of them.

— G. H. Hardy, §61

Summary

Summary
  • ϕ(n)→l\phi(n) \to l means: for every δ>0\delta > 0 there is an n0(δ)n_0(\delta) with ∣ϕ(n)−l∣<δ|\phi(n) - l| < \delta for all n≥n0(δ)n \ge n_0(\delta).
  • Picture: however narrow the band l±δl \pm \delta, all points from some vertical line onwards lie inside it.
  • ϕ(n)→+∞\phi(n) \to +\infty means: for every Δ\Delta there is an n0(Δ)n_0(\Delta) with ϕ(n)>Δ\phi(n) > \Delta for all n≥n0(Δ)n \ge n_0(\Delta). The symbol ∞\infty is only part of a phrase.
  • To prove a limit, give a rule for n0n_0 in terms of δ\delta and check it works; it need not be the smallest.
  • Changing finitely many terms never matters; changing infinitely many can. A limit need not be a value, and may be one.