Unit 1, lesson 5

The Formal Definition of a Limit

The epsilon-delta definition of a limit, why the order of its quantifiers matters, how to choose delta in a proof, and how to show that a number is not the limit.

  • Theory part 9
  • Practice part 10
  • Formal Definitions

Theory, part 9

Practice, part 10

“As close as we like” and “close enough” are fine for building intuition, but they cannot settle a disagreement. Think of a machinist who is told a part must be within 0.010.01 millimeters of spec. The machine has to be set precisely enough to hit that tolerance, and a tighter tolerance needs a tighter setting. The epsilon-delta definition turns this kind of promise into exact mathematics.

Key terms

Epsilon-delta definition

Definition Let ff be defined on an open interval containing aa, except possibly at aa, and let LL be a real number. Then lim⁡x→af(x)=L\lim_{x \to a} f(x) = L means that for every ε>0\varepsilon > 0 there exists δ>0\delta > 0 such that for all xx,

0<∣x−a∣<δ  ⟹  ∣f(x)−L∣<ε.0 < \lvert x - a \rvert < \delta \implies \lvert f(x) - L \rvert < \varepsilon.

In plain words The number ε\varepsilon is the output tolerance someone demands, and δ\delta is an input tolerance you supply. The limit is LL when you can always meet the demand, whatever ε\varepsilon is.

Example To prove lim⁡x→3(2x−1)=5\lim_{x \to 3} (2x - 1) = 5, let ε>0\varepsilon > 0 and choose δ=ε/2\delta = \varepsilon / 2. If 0<∣x−3∣<δ0 < \lvert x - 3 \rvert < \delta, then

∣(2x−1)−5∣=2∣x−3∣<2δ=ε.\lvert (2x - 1) - 5 \rvert = 2 \lvert x - 3 \rvert < 2\delta = \varepsilon.

Every epsilon before there exists delta

Definition The statement has the form “for every ε>0\varepsilon > 0 there exists δ>0\delta > 0”. The order matters. The number δ\delta is chosen after ε\varepsilon is given, so δ\delta may depend on ε\varepsilon and on aa. If the order were swapped to “there exists δ>0\delta > 0 such that for every ε>0\varepsilon > 0”, the condition would force ∣f(x)−L∣<ε\lvert f(x) - L \rvert < \varepsilon for all ε>0\varepsilon > 0, that is f(x)=Lf(x) = L for every xx in a punctured neighborhood of aa. That is a much stronger and different statement.

In plain words First the challenge, then the answer. Each tolerance gets its own δ\delta, and smaller tolerances usually need smaller δ\delta.

Example For lim⁡x→3(2x−1)=5\lim_{x \to 3} (2x - 1) = 5 the rule δ=ε/2\delta = \varepsilon / 2 gives

ε\varepsilon110.10.10.010.01
δ\delta0.50.50.050.050.0050.005

No single δ\delta works for every ε\varepsilon here, which is fine, because the definition never asks for one.

The condition on x

Definition The hypothesis 0<∣x−a∣<δ0 < \lvert x - a \rvert < \delta has two parts. The right inequality ∣x−a∣<δ\lvert x - a \rvert < \delta keeps xx within δ\delta of aa. The left inequality 0<∣x−a∣0 < \lvert x - a \rvert says x≠ax \neq a. So the condition describes exactly the punctured neighborhood (a−δ,a)∪(a,a+δ)(a - \delta, a) \cup (a, a + \delta), and the value f(a)f(a) is never tested.

In plain words Only inputs near aa count, and aa itself never counts.

Example Let g(x)=x+1g(x) = x + 1 for x≠1x \neq 1 and g(1)=5g(1) = 5. We claim lim⁡x→1g(x)=2\lim_{x \to 1} g(x) = 2. Given ε>0\varepsilon > 0, take δ=ε\delta = \varepsilon. If 0<∣x−1∣<δ0 < \lvert x - 1 \rvert < \delta, then x≠1x \neq 1, so

∣g(x)−2∣=∣(x+1)−2∣=∣x−1∣<ε.\lvert g(x) - 2 \rvert = \lvert (x + 1) - 2 \rvert = \lvert x - 1 \rvert < \varepsilon.

At x=1x = 1 itself ∣g(1)−2∣=3\lvert g(1) - 2 \rvert = 3, but that input is excluded, so it does no harm.

All nearby domain points

Definition In the definition, the implication

0<∣x−a∣<δ  ⟹  ∣f(x)−L∣<ε0 < \lvert x - a \rvert < \delta \implies \lvert f(x) - L \rvert < \varepsilon

must hold for every xx in the domain of ff that satisfies 0<∣x−a∣<δ0 < \lvert x - a \rvert < \delta, not merely for some of them. Under the standard convention of this course, every point of the punctured neighborhood is such an xx once δ\delta is small, because ff is defined on an open interval around aa except possibly at aa.

In plain words It is not enough to find a few inputs near aa that land close to LL. Every input of the function in the punctured neighborhood has to land close to LL.

Example Take h(x)=sin⁡(π/x)h(x) = \sin (\pi/x) near 00. For every positive whole number nn the input x=1/nx = 1/n gives h(x)=sin⁡(nπ)=0h(x) = \sin(n\pi) = 0, so some inputs in every punctured neighborhood of 00 land exactly on 00. But the inputs x=2/(4n+1)x = 2/(4n + 1) give

h(24n+1)=sin⁡(2nπ+π2)=1,h\left(\frac{2}{4n + 1}\right) = \sin\left(2n\pi + \frac{\pi}{2}\right) = 1,

and they also come arbitrarily close to 00. With ε=1/2\varepsilon = 1/2 no δ\delta works for L=0L = 0, so 00 is not the limit.

Accumulation point

Definition A number aa is an accumulation point of a set DD if every interval (a−δ,a+δ)(a - \delta, a + \delta) with δ>0\delta > 0 contains a point of DD other than aa. The point aa itself may or may not belong to DD. Some advanced texts extend the definition of lim⁡x→af(x)=L\lim_{x \to a} f(x) = L to a function with any domain DD by testing only inputs x∈Dx \in D, and then they take limits only at accumulation points of DD. Under the convention of this course this condition holds automatically, since DD contains a whole punctured neighborhood of aa.

In plain words The function must have inputs arbitrarily close to aa. Otherwise there is nothing to approach aa through, and every number LL would pass the test for no reason, which would break uniqueness.

Example The number 00 is an accumulation point of (0,1](0, 1], even though 00 is not in the set, and also of the set {1,12,13,… }\{1, \frac{1}{2}, \frac{1}{3}, \dots\}. The number 55 is not an accumulation point of [0,1]∪{5}[0, 1] \cup \{5\}, since the interval (4,6)(4, 6) contains no point of the set other than 55.

Choosing delta

Definition A proof of lim⁡x→af(x)=L\lim_{x \to a} f(x) = L needs a rule that turns each ε>0\varepsilon > 0 into some δ>0\delta > 0 satisfying the definition. A standard method is to factor ∣f(x)−L∣\lvert f(x) - L \rvert as ∣x−a∣\lvert x - a \rvert times another factor, bound that factor by first requiring δ≤1\delta \le 1 (or some other fixed number), and then take δ\delta to be the minimum of the two requirements. Any smaller positive δ\delta also works.

In plain words Do the scratch work backwards from ∣f(x)−L∣<ε\lvert f(x) - L \rvert < \varepsilon, then write the proof forwards. When a factor depends on xx, trap xx in a small interval first.

Example To prove lim⁡x→3x2=9\lim_{x \to 3} x^2 = 9, start from the factorization

∣x2−9∣=∣x−3∣ ∣x+3∣.\lvert x^2 - 9 \rvert = \lvert x - 3 \rvert \, \lvert x + 3 \rvert.

If ∣x−3∣<1\lvert x - 3 \rvert < 1, then 2<x<42 < x < 4 and ∣x+3∣<7\lvert x + 3 \rvert < 7. So take

δ=min⁡(1,ε7).\delta = \min\left(1, \frac{\varepsilon}{7}\right).

If 0<∣x−3∣<δ0 < \lvert x - 3 \rvert < \delta, then both requirements hold, so

∣x2−9∣=∣x−3∣ ∣x+3∣<ε7⋅7=ε.\lvert x^2 - 9 \rvert = \lvert x - 3 \rvert \, \lvert x + 3 \rvert < \frac{\varepsilon}{7} \cdot 7 = \varepsilon.

Negating the definition

Definition The statement lim⁡x→af(x)=L\lim_{x \to a} f(x) = L is false exactly when there exists ε>0\varepsilon > 0 such that for every δ>0\delta > 0 there is an xx with

0<∣x−a∣<δand∣f(x)−L∣≥ε.0 < \lvert x - a \rvert < \delta \quad \text{and} \quad \lvert f(x) - L \rvert \ge \varepsilon.

To negate, each “for every” becomes “there exists”, each “there exists” becomes “for every”, and the implication “if 0<∣x−a∣<δ0 < \lvert x - a \rvert < \delta then ∣f(x)−L∣<ε\lvert f(x) - L \rvert < \varepsilon” becomes ”0<∣x−a∣<δ0 < \lvert x - a \rvert < \delta and ∣f(x)−L∣≥ε\lvert f(x) - L \rvert \ge \varepsilon”.

In plain words To show LL is not the limit, find one tolerance that cannot be met. However small δ\delta is, some input inside it lands at least ε\varepsilon away from LL.

Example Let H(x)=0H(x) = 0 for x<0x < 0 and H(x)=1H(x) = 1 for x≥0x \ge 0. To show that 11 is not the limit at 00, take ε=1/2\varepsilon = 1/2. For any δ>0\delta > 0 the input x=−δ/2x = -\delta/2 satisfies 0<∣x∣<δ0 < \lvert x \rvert < \delta and ∣H(x)−1∣=1≥1/2\lvert H(x) - 1 \rvert = 1 \ge 1/2. In fact no LL works. Take ε=1/2\varepsilon = 1/2 again. If some δ\delta worked, the inputs ±δ/2\pm \delta/2 would give the outputs 00 and 11, both within 1/21/2 of LL, and the triangle inequality would give

1=∣1−0∣≤∣1−L∣+∣L−0∣<12+12=1,1 = \lvert 1 - 0 \rvert \le \lvert 1 - L \rvert + \lvert L - 0 \rvert < \frac{1}{2} + \frac{1}{2} = 1,

which is impossible.

Check yourself

  1. For lim⁡x→2(5x+3)=13\lim_{x \to 2} (5x + 3) = 13 and ε=0.01\varepsilon = 0.01, find a δ\delta that works.
Show the answer

Since ∣(5x+3)−13∣=5∣x−2∣\lvert (5x + 3) - 13 \rvert = 5 \lvert x - 2 \rvert, any positive δ≤0.002\delta \le 0.002 works, for example δ=0.002\delta = 0.002.

  1. Prove that lim⁡x→1(4−3x)=1\lim_{x \to 1} (4 - 3x) = 1.
Show the answer

Let ε>0\varepsilon > 0 and take δ=ε/3\delta = \varepsilon / 3. If 0<∣x−1∣<δ0 < \lvert x - 1 \rvert < \delta, then

∣(4−3x)−1∣=3∣x−1∣<3δ=ε.\lvert (4 - 3x) - 1 \rvert = 3 \lvert x - 1 \rvert < 3\delta = \varepsilon.
  1. For lim⁡x→3x2=9\lim_{x \to 3} x^2 = 9 and ε=1\varepsilon = 1, does δ=0.2\delta = 0.2 work?
Show the answer

No. The input x=3.19x = 3.19 satisfies ∣x−3∣<0.2\lvert x - 3 \rvert < 0.2, but x2=10.1761x^2 = 10.1761 is more than 11 away from 99. The rule from the lesson gives δ=min⁡(1,1/7)=1/7\delta = \min(1, 1/7) = 1/7, which does work.

How this lesson connects

It also uses quantifiers and inequalities.