“As x approaches a, ϕ(x) approaches l” is the sentence on which continuity, derivatives and most of calculus rest. In §§93–95 Hardy makes it precise. He reaches it from the last lesson by a change of variable, ends with the two-quantifier definition that every analysis course now uses, and adds the one-sided limits ϕ(a−0) and ϕ(a+0), which we need to describe jumps.
The idea
Hardy’s route is a substitution. If ϕ(x)→l as x→∞, put y=1/x. Large positive x means small positive y, so the statement becomes: by taking y positive and small enough, we can make ψ(y)=ϕ(1/y) as close to l as we please. That is a statement about y approaching 0 from the right, and once it is written out we can forget x and use it for any function ψ. Shifting by a (put x=a+y) then gives limits at any point a.
The game has the same shape as before, but the second move is different. The opponent names a tolerance ε>0, a horizontal band l±ε. We answer with a distance δ>0, a vertical strip around a. We win if every x in the strip other than a itself gives a point of the graph inside the band.
ExploreFind δ for x2 near 1
Try this.Set ε=0.5 and drag the purple handle (or use the δ slider) until the verdict is green. Halve ε and repeat a few times. Compare each working δ with ε/3.
f(x)=x2, claim limx→1f(x)=1 at a=1
xy
y = f(x)
band L ± ε
strip of width δ (drag)
escapes the band
This δ fails. For example x=1.5 gives ∣f(x)−1∣=1.25≥ε. Try a smaller δ — or ask whether any δ can work.
What to notice
Each smaller ε needs a smaller δ, and δ=ε/3 always works once ε≤3. That is no accident: ∣x2−1∣=∣x−1∣∣x+1∣, and near x=1 the factor ∣x+1∣ stays below 3, so a strip of half-width ε/3 cannot climb out of a band of half-width ε. The proof below turns this into an argument.
Two details of the game are essential.
x=a is excluded. The strip is 0<∣x−a∣<δ. The limit describes the values neara, never the value ata, which may be different or may not exist at all. The next lesson is about exactly this.
δ may depend on ε (and on a), just as n0 depended on δ in Chapter IV.
Precise statements
From here on we write the tolerance as ε and the closeness of x to a as δ, which is the modern habit. Hardy uses the same two letters the other way round; see “Hardy’s language” below.
DefinitionLimit as x→aHardy: §94
ϕ(x)tends to the limit l as x tends to a if for every ε>0 there is a δ>0 such that
∣ϕ(x)−l∣<εwhenever 0<∣x−a∣<δ.
We write limx→aϕ(x)=l, or ϕ(x)→l as x→a. The case a=0 is Hardy’s definition in §93.
DefinitionOne-sided limitsHardy: §§93–94
ϕ(x)→las x→a from the right (Hardy: “by values greater than a”, written x→a+0) if for every ε>0 there is a δ>0 with ∣ϕ(x)−l∣<ε whenever a<x<a+δ. The limit from the left uses a−δ<x<a. The two one-sided limits, when they exist, are written
ϕ(a+0)=x→a+0limϕ(x),ϕ(a−0)=x→a−0limϕ(x).
Tending to ±∞ is defined by the same recipe with a large Δ: ϕ(x)→∞ as x→a if for every Δ there is a δ>0 with ϕ(x)>Δ whenever 0<∣x−a∣<δ; and similarly on one side. If none of these happens, ϕ(x)oscillates as x→a (or as x→a+0), finitely or infinitely according as it is or is not bounded near a.
TheoremTwo-sided means both one-sided, and equalHardy: §94
ϕ(x)→l as x→a if and only if ϕ(a−0) and ϕ(a+0) both exist and both equal l.
Hardy closes the section with monotone functions, which can never oscillate near a point.
TheoremMonotone functions have one-sided limitsHardy: §95
If ϕ(x) increases steadily in an interval (a−h,a+h), then ϕ(a−0) and ϕ(a+0) exist, and
ϕ(a−0)≤ϕ(a)≤ϕ(a+0).
The same holds, with the inequalities reversed, for a steadily decreasing function.
A jump of a monotone function is therefore always a clean gap between ϕ(a−0) and ϕ(a+0), with ϕ(a) somewhere in the gap.
ExploreOne side at a time
Try this.Here is [x] near a=1. Make ε small (say 0.2). With the claimed limit L=1, can any δ win? Move L to 0 and try again. Then look for an L that works.
f(x)=[x], claim limx→1f(x)=1 at a=1
xy
y = f(x)
band L ± ε
strip of width δ (drag)
escapes the band
This δ fails. For example x=0.5 gives ∣f(x)−1∣=1≥ε. Try a smaller δ — or ask whether any δ can work.
What to notice
With L=1 every strip fails on its left half, where [x]=0; with L=0 it fails on the right half, where [x]=1. No L survives a small ε, so limx→1[x] does not exist. Yet each half on its own is perfectly behaved: ϕ(1−0)=0 and ϕ(1+0)=1, and [x], being steadily increasing, obeys ϕ(1−0)≤ϕ(1)=1≤ϕ(1+0).
Proofs
Here is a model ε–δ proof. It is a rule that produces a winning δ from any ε, like the rule n0=[1/δ]+1 in Chapter IV.
Proofx2→1 as x→1
Strategy
Factor the distance: ∣x2−1∣=∣x−1∣∣x+1∣. The first factor is what we control. The second is not small, but it is bounded near 1; bound it first, then choose δ.
Let ε>0 be given; first insist on δ≤1
If ∣x−1∣<1 then 0<x<2, so ∣x+1∣<3.
Choose δ=min(1,ε/3)
Then for 0<∣x−1∣<δ we have both ∣x+1∣<3 and ∣x−1∣<ε/3.
Check
∣x2−1∣=∣x−1∣∣x+1∣<(ε/3)⋅3=ε. Every ε has its δ, so x2→1.
∎
ProofTwo-sided means both one-sided, and equal
Strategy
The punctured strip 0<∣x−a∣<δ is the union of a left strip and a right strip. A δ that works for the whole strip works for each half; conversely, take the smaller of the two one-sided δ’s.
If ϕ(x)→l, both one-sided limits are l
Given ε, the δ for the two-sided limit gives ∣ϕ(x)−l∣<ε on a<x<a+δ and on a−δ<x<a, which is what the one-sided definitions ask.
Conversely, suppose ϕ(a−0)=ϕ(a+0)=l
Given ε, there are δ1>0 with ∣ϕ(x)−l∣<ε on a−δ1<x<a, and δ2>0 with the same on a<x<a+δ2.
Take δ=min(δ1,δ2)
If 0<∣x−a∣<δ, then x lies in one of the two half-strips, so ∣ϕ(x)−l∣<ε.
∎
ProofMonotone functions have one-sided limits
Strategy
This is the monotone-limit theorem of the last lesson in a new setting. For x approaching a from the left, the values ϕ(x) climb and never pass ϕ(a), so they should tend to their upper bound. Hardy gets this from §92 by putting y=1/(a−x); here is the direct argument.
Let l be the upper bound of ϕ(x) for a−h<x<a
These values are bounded above by ϕ(a), since ϕ increases steadily. So the upper bound l exists and l≤ϕ(a).
Given ε>0, find x1<a with ϕ(x1)>l−ε
l−ε is less than the upper bound, so some value exceeds it.
Take δ=a−x1
If x1<x<a then l−ε<ϕ(x1)≤ϕ(x)≤l, so ∣ϕ(x)−l∣<ε. Hence ϕ(a−0)=l≤ϕ(a).
The right side is symmetric
For a<x<a+h, the values are bounded below by ϕ(a); their lower bound is ϕ(a+0), by the same argument, and it is at least ϕ(a).
∎
Worked examples and exercises
Exercise●○○Hardy: Examples XXXV, 2–3
Let m be an integer. Show that xm→0 as x→0 when m>0. When m<0, show that xm→+∞ as x→+0, and that as x→−0 it tends to −∞ or +∞ according as m is odd or even.
Hint
For m>0: if ∣x∣<1 then ∣xm∣≤∣x∣.
Hint
For m=−k<0: xm=1/xk, and 1/xk>Δ when 0<x<Δ−1/k.
Solution
m>0. Given ε>0, take δ=min(1,ε). If 0<∣x∣<δ then ∣x∣<1, so ∣xm∣=∣x∣m≤∣x∣<ε.
m=−k, k>0, from the right. Given Δ>0, take δ=Δ−1/k. If 0<x<δ then xk<δk=1/Δ, so xm=1/xk>Δ. (For Δ≤0 any δ works.) So xm→+∞ as x→+0.
From the left. Put x=−y with y>0: xm=(−1)k/yk=(−1)mym. As x→−0, y→+0 and ym→+∞; so xm→+∞ if m is even and −∞ if m is odd. For odd negative m, then, xm tends to +∞ on one side and −∞ on the other, so it tends to neither as x→0.
Exercise●○○Hardy: Examples XXXVI, 22
Let m be an integer. Show that as x→m+0, [x]→m and x−[x]→0, while as x→m−0, [x]→m−1 and x−[x]→1.
Hint
Which integer is [x] when m<x<m+1? When m−1<x<m?
Solution
For m<x<m+1 we have [x]=m exactly, so with δ=1 every such x gives ∣[x]−m∣=0<ε; and x−[x]=x−m, which is less than ε once x<m+ε, so δ=min(1,ε) works. For m−1<x<m we have [x]=m−1 exactly, and x−[x]=x−m+1, which is within ε of 1 once x>m−ε. So [m−0]=m−1=m=[m+0]: the one-sided limits differ, and limx→m[x] does not exist.
Exercise●●○Hardy: Examples XXXVI, 6
Let a,b>0. Show that as x→+0
ax[xb]→ab,xb[ax]→0.
How do the two functions behave as x→−0?
Hint
Use t−1<[t]≤t for every real t.
Hint
For the second function, what is [x/a] when 0<x<a? When −a<x<0?
Solution
First function, x>0. From b/x−1<[b/x]≤b/x, multiplying by x/a>0,
ab−ax<ax[xb]≤ab.
So the distance from b/a is less than x/a, which is less than ε when 0<x<aε. Limit b/a.
First function, x<0. Now x/a<0 and the inequalities reverse: b/a≤(x/a)[b/x]<b/a−x/a=b/a+∣x∣/a. Again the limit is b/a.
Second function. For 0<x<a, [x/a]=0, so the function is identically 0 there and tends to 0. For −a<x<0, [x/a]=−1, so the function equals −b/x=b/∣x∣, which tends to +∞ as x→−0. So the second function has different behaviour on the two sides and no limit as x→0.
ExerciseA limit, if it exists, is unique●●○Hardy: §94
Show that ϕ(x) cannot tend to two different limits l and l′ as x→a. (Assume, as always, that ϕ is defined at points arbitrarily close to a on at least one side.)
Hint
Take ε=21∣l−l′∣ and use both definitions at one point x.
Solution
Suppose l=l′ and put ε=21∣l−l′∣>0. There are δ1 and δ2 with ∣ϕ(x)−l∣<ε for 0<∣x−a∣<δ1 and ∣ϕ(x)−l′∣<ε for 0<∣x−a∣<δ2. Pick any x with 0<∣x−a∣<min(δ1,δ2) at which ϕ is defined. Then ∣l−l′∣≤∣l−ϕ(x)∣+∣ϕ(x)−l′∣<2ε=∣l−l′∣, a contradiction. (This is why the hypothesis matters: if there were no such x, every l would satisfy the definition vacuously.)
Check yourself
ExploreIs there any limit at all?
Try this.This is sin(1/x) near 0. The opponent names three tolerances. Move L anywhere you like and try to survive the rounds.
f(x)=sinx1, claim limx→0f(x)=0 at a=0Round 1 of 3
Opponent: “I name ε=0.5. Find δ so that ∣f(x)−L∣<ε whenever 0<∣x−a∣<δ.”
xy
y = f(x)
band L ± ε
strip of width δ (drag)
escapes the band
What to notice
Every strip around 0, however thin, contains whole waves of the graph, so it contains points where sin(1/x)=1 and points where it is −1. No band narrower than 1 can hold both. The function oscillates finitely as x→0, from either side.
Check yourselfChoose one
Let ψ(x)=0 for x=0 and ψ(0)=1. What is limx→0ψ(x)?
Why
The definition only looks at 0<∣x∣<δ, and there ψ(x)=0 exactly. The value ψ(0)=1 is invisible to the limit. Saying the limit fails confuses “has a limit” with “is continuous”, which is the subject of a later lesson.
Check yourselfTrue or false?
If ϕ(a+0) exists, then limx→aϕ(x) exists.
Why
False. Both one-sided limits must exist and be equal. [x] at a=1 has ϕ(1+0)=1 but ϕ(1−0)=0.
Check yourselfChoose one
Which statement says that ϕ(x)→l as x→a?
Why
The tolerance comes first and the strip answers it. The first option fixes one strip for all tolerances, which forces ϕ(x)=l on that whole punctured strip. The third only says ϕ is bounded near a (sin(1/x) satisfies it with l=0). The fourth asks for a single good point.
Check yourselfEnter a number
For ϕ(x)=3x+2 at a=1 (limit 5) and ε=0.03, what is the largest δ that works?
Why
∣ϕ(x)−5∣=3∣x−1∣, which is less than 0.03 exactly when ∣x−1∣<0.01. So δ=ε/3=0.01; any smaller δ also works, and any larger one lets in points such as x=1+0.0101, where the distance is 0.0303.
Hardy’s language
It will of course be understood that ϕ(a − 0) has no meaning other than that of a conventional abbreviation for the limit on the left-hand side.
— G. H. Hardy, §95
Summary
Summary
limx→aϕ(x)=l: for every ε>0 there is δ>0 with ∣ϕ(x)−l∣<ε whenever 0<∣x−a∣<δ. The point x=a is excluded.
One-sided limits use only a<x<a+δ or a−δ<x<a; the two-sided limit exists exactly when ϕ(a−0) and ϕ(a+0) exist and agree.
To prove a limit, bound the awkward factor near a first, then choose δ in terms of ε (e.g. δ=min(1,ε/3) for x2 at 1).
A steadily increasing function has both one-sided limits at every point, and ϕ(a−0)≤ϕ(a)≤ϕ(a+0).
Hardy’s letters are swapped relative to modern usage: his δ is the tolerance, his ϵ the width of the strip.