The epsilon-delta definition of a limit and the challenge-and-response way of reading it, proving limits for linear and quadratic functions with the backwards scratch work and forwards proof, the epsilon-delta forms of one-sided and infinite limits, and using the definition to prove a limit law.
Subject: Calculus I · 67 slides · symbolic lesson
Open the interactive version of this deck
Title
Calculus I · Section 2.5
Turning getting close into something you can prove
Objectives
Every limit so far has been found by looking, by tables, or by the limit laws. This lesson says exactly what the word limit means, so that a limit can be proved and not just believed.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — learning objectives 2.5.1 to 2.5.4
So far you have found limits by drawing graphs, by building tables, and by applying the limit laws. All of that rested on an informal sentence about getting close, and none of it was ever proved. This section replaces that sentence with a definition precise enough to prove things with.
The definition is famous for being hard on first contact, so the lesson takes it slowly. First you learn to read it, as a game between someone setting a tolerance and you answering with a window. Then you prove limits of lines, where the answer is a single division, and of curves, where one extra idea is needed.
Finally you use the definition to do what the informal version never could: prove a limit law, prove that a limit fails to exist, and state exactly what one-sided and infinite limits mean.
Warm-up
Discussion prompt
Section 2.2 said the limit of f at a is L if f(x) gets arbitrarily close to L as x gets close to a. Name one phrase in that sentence that could not be checked by a calculation, and say what number is missing from it.
Write your answer before you reveal. Most people circle the word close, and that is exactly right. Close to what degree? Within a tenth, a millionth? The informal sentence never says.
There are really two missing numbers. One measures how close the output is to L, and the other measures how close the input is to a. The precise definition names them epsilon and delta, and its entire content is how those two numbers are related.
Keep this in mind as you read the definition on the coming slides: every symbol in it is there to fill one of the gaps you just identified.
Section
Part 1
Intuition
Is 3.01 close to 3? For a carpenter, yes. For a chip designer, no. Closeness has no meaning until someone names a tolerance.
\[ |f(x) - 3| < 0.1 \qquad |f(x) - 3| < 0.001 \qquad |f(x) - 3| < 10^{-9} \]
A limit must survive every tolerance anyone could name, not one chosen in advance. That is the whole idea of the definition: whatever tolerance is demanded, inputs close enough to a deliver it.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — introduction
The point of this slide is that closeness is relative to a purpose. A value of 3.01 is excellent for a bookshelf and useless for a microprocessor. So a statement like f of x is close to 3 means nothing until someone says how close.
A limit is supposed to be a fact about the function itself, not about anybody's purpose. The way to make it independent of purpose is to make it work for every purpose at once: whatever tolerance anyone demands, you can meet it by taking inputs close enough to a.
That idea, every tolerance can be met, is the definition in plain words. The rest of the lesson turns it into symbols and then into proofs.
Concept
Figure (svg): A number line from minus 1 to 7 with points a at 2 and b at 5 marked, and a double-headed arrow between them labelled with the distance 3.
\[ \text{distance from } a \text{ to } b = |a - b| = |b - a| \]
Two statements carry the whole definition, and both are distances:
\[ |f(x) - L| < \varepsilon \quad\text{the output is within } \varepsilon \text{ of } L \]
\[ 0 < |x - a| < \delta \quad\text{the input is within } \delta \text{ of } a, \text{ but not } a \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — Quantifying Closeness
Everything in the precise definition is a distance, and on the number line the distance between two numbers is the absolute value of their difference. The picture shows 2 and 5: the gap is 3 whichever one you subtract from the other.
Read the two key inequalities as distance statements. The absolute value of f of x minus L less than epsilon says the output is less than epsilon away from L. The absolute value of x minus a less than delta says the input is less than delta away from a.
The extra zero on the left of the input statement is new. It says the distance is positive, which means x is not a itself. You will see why that matters on the next two slides.
Concept
\[ |f(x) - L| < \varepsilon \iff -\varepsilon < f(x) - L < \varepsilon \]
\[ \iff L - \varepsilon < f(x) < L + \varepsilon \]
Figure (svg): The rising curve y equals f of x crosses a horizontal band between the heights L minus epsilon and L plus epsilon; the part of the curve inside the band is highlighted.
Undo the absolute value, then add L to all three parts. The output sits in a band of half-height epsilon centred on L.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — the first equivalence
An absolute value less than epsilon means the quantity inside is between minus epsilon and epsilon. Add L to all three parts and you get f of x trapped between L minus epsilon and L plus epsilon.
Geometrically that is a horizontal band of total height two epsilon centred on the line at height L. The figure highlights in green the part of the graph inside the band. A point of the graph satisfies the output inequality exactly when it is in that band.
When you picture the definition, picture this band first. The challenger chooses how thick it is, and it can be made as thin as they like.
Concept
\[ 0 < |x - a| < \delta \iff a - \delta < x < a + \delta \text{ and } x \ne a \]
Figure (svg): A number line with an open interval from a minus delta to a plus delta shaded; the endpoints are open circles and the centre a is also an open circle, because x may not equal a.
The right-hand inequality makes a window of half-width delta around a. The left-hand one, zero less than the distance, removes a itself: a limit never asks what happens at a, only near it.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — the second equivalence
The input condition packs two facts into one line. The right-hand inequality says x is within delta of a, so x lies in an open interval centred on a. The left-hand inequality, zero less than the distance, says x is not equal to a.
The figure shows this as a window with open ends and an open circle at the centre. The hole matters because a limit describes what happens near a, not at a. The function might be undefined at a, or defined with some unrelated value, and neither should affect the limit.
This is the same principle you met in Section 2.2, when you learned to ignore the value at the point. The definition builds it in with one small zero.
Concept
limit (precise definition) — Let f be defined for all x other than a on an open interval containing a, and let L be a real number. The limit of f at a is L if, for every positive epsilon, there is a positive delta such that whenever x is within delta of a and not equal to a, f(x) is within epsilon of L.
\[ \lim_{x\to a} f(x) = L \]
\[ \forall \varepsilon > 0 \;\exists \delta > 0: \quad 0 < |x - a| < \delta \;\Longrightarrow\; |f(x) - L| < \varepsilon \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — Definition
Here it is in full. Read it slowly, phrase by phrase: for every positive epsilon, there exists a positive delta, such that if x is within delta of a and not equal to a, then f of x is within epsilon of L.
The symbolic line underneath says the same thing in shorthand. The upside-down A means for every, and the backwards E means there exists. These two symbols are the most important part of the definition, because their order is what makes it work.
Do not worry if it does not feel natural yet. The next few slides read it three different ways, as annotations, as a game, and as a picture, and by the end of this part it should feel like a precise version of something you already believed.
Notation
Annotate
On: \( \forall \varepsilon > 0 \;\exists \delta > 0: \; 0 < |x - a| < \delta \Rightarrow |f(x) - L| < \varepsilon \)
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — Table 2.9
Step through the annotations in order. The first phrase is a universal quantifier: the claim is about every positive epsilon, so in a proof you cannot pick a convenient epsilon. You start by saying let epsilon be positive, and treat it as an unknown positive number.
The second phrase is an existential quantifier. It is your job to produce a delta, and the only thing you may build it from is epsilon. The book calls this a scavenger hunt, and the scratch work you will learn soon is how you hunt.
The last two annotations are the conditional. The window condition is the hypothesis and the band condition is the conclusion, and the conclusion must hold for every x in the window, not just a convenient one.
Intuition
| move | who makes it | in symbols |
|---|---|---|
| 1. name a tolerance | the challenger, trying to break the claim | ε > 0, as small as they like |
| 2. reply with a window | you, knowing ε | δ > 0, a formula in ε |
| 3. test any input | the challenger again | any x with 0 < |x − a| < δ |
| 4. referee | the inequality | is |f(x) − L| < ε ? |
The limit is L when you can win every round. Winning one round, or a hundred, proves nothing; a proof is a rule that wins them all at once.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 169 — Table 2.9, translation column
Many people find the definition clicks when it is read as a game. A sceptic who doubts that the limit is L names a tolerance. You reply with a window. The sceptic then picks any input in your window and checks whether the output is inside the tolerance.
You win the round if every input in your window passes. The limit is L exactly when you can win every round, however small the sceptic makes the tolerance. That is why a table of values, which plays a few rounds, can suggest a limit but never prove it.
Notice the order in the table. You choose delta after seeing epsilon, but before seeing x. That order is the reason delta may depend on epsilon and may never depend on x, a rule that will matter a great deal in Part 3.
Picture it
Figure (svg): Three panels of the same falling curve with limit L at a. The horizontal band around L narrows from epsilon 0.8 to 0.4 to 0.2, and the yellow window around a narrows with it, from about 0.68 to 0.37 to 0.19.
Same curve, three challenges. Each panel shows the band the challenger names and the window that answers it. The window is set by whichever side of the curve leaves the band first.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 170 — Figure 2.39
This is the book's Figure 2.39, drawn from a real falling curve. In each panel the orange lines are the band the challenger named and the yellow window on the axis is your answer. As epsilon drops from 0.8 to 0.4 to 0.2, delta drops from about 0.68 to 0.37 to 0.19.
Look at how the window is set. On each side of a, the curve leaves the band at some point, and those two exit points are generally different distances from a, because the curve bends. The window must fit inside both, so delta is the smaller distance. You will compute exactly this in Example 2.41.
The limit exists because this process never fails. No matter how thin the band, some window answers it.
Ranking
Put in order
Put the moves of the epsilon-delta game in the order the definition makes them.
Why: Epsilon comes first and is out of your control; delta is chosen knowing epsilon; x is chosen after delta, so delta cannot depend on x; the output inequality is what gets proved.
Drag the moves into order before checking. The order is not a matter of style; it is the logical structure of the definition.
Epsilon is named first and you have no control over it. Delta comes second, chosen knowing epsilon. Then x is chosen, anywhere in the window. Finally the output inequality is checked. Because x arrives after delta, a formula for delta that contains x is meaningless: at the moment delta is chosen, there is no x yet.
Prediction
\[ \lim_{x\to 2}(3x - 1) = 5 \quad \text{is proved with } \delta = \frac{\varepsilon}{3} \]
Predict first
Would the choice δ = ε/10 also give a valid proof?
Correct: Yes: every x in the smaller window is also in the larger one
Why: If every x within ε/3 of 2 lands in the band, then in particular every x within ε/10 of 2 does, since those x are a subset. Any positive delta smaller than a working one also works. The definition asks for SOME delta, never the best one.
Commit to an answer first. This is a question students often get wrong, because they think of delta as the answer rather than an answer.
If every input within epsilon over 3 of 2 produces an output in the band, then certainly every input within epsilon over 10 does, since those inputs are among the first set. Shrinking a working window can never let a bad input in.
This fact is used constantly. It is why a proof may take a minimum of two numbers, and why being generous with a smaller delta than necessary is never an error. The definition asks whether some delta exists, not what the largest one is.
Section
Part 2
Concept
To FIND delta, start from what you want and solve for the input distance. To PROVE the limit, run the chain the other way, from the window to the band.
\[ \text{scratch: } |f(x) - L| < \varepsilon \;\leadsto\; \cdots \;\leadsto\; |x - a| < \text{something in } \varepsilon \]
\[ \text{proof: } 0 < |x - a| < \delta \;\Longrightarrow\; \cdots \;\Longrightarrow\; |f(x) - L| < \varepsilon \]
The scratch work is how you discover delta. It is not part of the proof, and a proof that only shows the scratch work has proved the wrong direction.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 170-171 — Example 2.39 and its Analysis
Every epsilon-delta proof is really two pieces of work. The first is private: you start from the inequality you want, the output within epsilon of L, and manipulate it until the input distance stands alone. Whatever it is less than suggests delta.
The second is the proof, and it runs in the opposite direction. You assume the input is in the window and derive that the output is in the band. That is the direction the definition's if-then points.
Keep the two separate on paper. A common way to lose all the marks on a proof is to hand in only the scratch work, which shows that the band implies the window: the reverse of what was asked. The trap later in this part shows the wrong version side by side with the right one.
Worked example
Find a delta that will prove the limit below. This is the scratch work.
\[ \lim_{x\to 1}(2x + 1) = 3 \]
Write down the target
Why: The output must land within epsilon of 3.
\[ |(2x + 1) - 3| < \varepsilon \]
Simplify inside the bars
Why: The constants cancel to leave 2x minus 2.
\[ |2x - 2| < \varepsilon \]
Factor out the input distance
Why: The 2 comes out of the absolute value as its own absolute value.
\[ |2|\,|x - 1| < \varepsilon \]
Divide by 2
Why: Now the input distance stands alone.
\[ |x - 1| < \frac{\varepsilon}{2} \]
Read off delta
Why: The window that guarantees the target has half-width epsilon over 2.
\[ \delta = \frac{\varepsilon}{2} \]
Check with one challenge
Why: Take epsilon 0.1, so delta is 0.05, and an input inside the window, 1.04: the output misses 3 by 0.08, inside the tolerance.
\[ \varepsilon = 0.1,\; x = 1.04: \quad |2(1.04) + 1 - 3| = 0.08 < 0.1 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 170 — Example 2.39
This is the scratch work. Start from the goal, the output within epsilon of 3, and simplify. The constants cancel, the 2 factors out, and dividing by 2 leaves the input distance on its own, less than epsilon over 2.
Reading that off, a window of half-width epsilon over 2 around 1 should do the job. For a line, the scratch work is always this short: the output error is a constant times the input distance, so you divide epsilon by that constant.
The check plays one round of the game with real numbers. Epsilon is a tenth, delta is five hundredths, and the input 1.04 is inside the window. Its output misses 3 by 0.08, which is inside the tolerance. One round proves nothing on its own, but it catches arithmetic slips before you write the proof.
Picture it
Figure (svg): The line y equals 2x plus 1 crossing a band from 3 minus epsilon to 3 plus epsilon; vertical lines drop from the crossings to x equals 1 minus epsilon over 2 and 1 plus epsilon over 2, and that window is marked on the x-axis.
\[ 2x + 1 = 3 \pm \varepsilon \iff x = 1 \pm \frac{\varepsilon}{2}, \qquad \delta = \min\left\{\tfrac{\varepsilon}{2}, \tfrac{\varepsilon}{2}\right\} = \tfrac{\varepsilon}{2} \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 171 — Figure 2.40
This is the geometric way to find the same delta. Draw the band from 3 minus epsilon to 3 plus epsilon, and find where the line crosses each edge. Solving 2x plus 1 equal to each edge gives 1 minus epsilon over 2 and 1 plus epsilon over 2.
The window runs between those two crossings, and it is symmetric, so delta is the minimum of two equal numbers, epsilon over 2. A line is the only kind of graph where the two gaps are always equal, because its slope never changes.
Hold on to the minimum in the formula. Here it looks pointless, but for a curve the two gaps differ, and the minimum is what keeps the window inside the band on both sides.
Worked example
Now the proof itself, in the order the definition demands.
Let epsilon be arbitrary
Why: Signals that nothing is assumed about its size.
\[ \text{Let } \varepsilon > 0. \]
Choose delta
Why: The value found in the scratch work.
\[ \text{Choose } \delta = \frac{\varepsilon}{2}. \]
Assume x is in the window
Why: This is the if-part of the implication.
\[ \text{Assume } 0 < |x - 1| < \delta. \]
Rewrite the output error
Why: Simplify and factor, exactly as before but now as equalities.
\[ |(2x + 1) - 3| = |2x - 2| = 2|x - 1| \]
Use the assumption
Why: Here the window enters: the input distance is below delta.
\[ 2|x - 1| < 2\delta \]
Use the choice of delta
Why: Two times epsilon over two.
\[ 2\delta = 2 \cdot \frac{\varepsilon}{2} = \varepsilon \]
Check the chain reads start to finish
Why: Each link is an equality or uses the assumption, so the conclusion holds for every x in the window and every positive epsilon.
\[ |(2x + 1) - 3| < \varepsilon \;\therefore\; \lim_{x\to 1}(2x + 1) = 3 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 171-172 — Example 2.39, final version
Now the proof itself, which is what you would actually hand in. It opens with let epsilon be positive, announces the choice of delta, and assumes x is in the window. Everything after that is a chain from the output error down to epsilon.
Look at where each assumption is used. The equalities just rewrite the output error. The one strict inequality is where the window assumption enters, replacing the input distance by delta. The final equality is where the choice of delta enters.
The check step reads the whole chain from start to finish. Every link is either an equality or an honest use of the assumption, so the output error is below epsilon for every x in the window, and epsilon was arbitrary. That is the definition, satisfied.
Concept
\[ \text{Let } \varepsilon > 0 \;\to\; \text{Choose } \delta \;\to\; \text{Assume } 0 < |x-a| < \delta \;\to\; |f(x) - L| < \varepsilon \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 172 — Problem-Solving Strategy
The book summarises Example 2.39 as a Problem-Solving Strategy, and it is worth memorising as a template. Four of its five lines are the same in every proof; only the choice of delta and the middle chain change.
Line one fixes an arbitrary tolerance. Line two announces delta, found beforehand in scratch work. Line three assumes the window. Line four is the real work, a chain ending in epsilon that uses the assumption somewhere. Line five states what was proved.
Using the template has a practical benefit: it forces you to write the proof in the forward direction, so you cannot accidentally prove the converse.
Worked example
Fill in the template for this limit.
\[ \lim_{x\to -1}(4x + 1) = -3 \]
Fill in the input distance
Why: The distance from x to minus 1 is x plus 1.
\[ |x - (-1)| = |x + 1| \]
Rewrite the output error
Why: Subtracting minus 3 adds 3.
\[ |(4x + 1) - (-3)| = |4x + 4| \]
Factor
Why: Out comes the input distance.
\[ |4x + 4| = 4|x + 1| \]
Use the window
Why: The input distance is less than delta.
\[ 4|x + 1| < 4\delta \]
Choose delta to finish the chain
Why: Four times delta should equal epsilon.
\[ \delta = \frac{\varepsilon}{4} \;\Longrightarrow\; 4\delta = \varepsilon \]
Check with one challenge
Why: Epsilon 0.2 gives delta 0.05; the input minus 0.96 is inside the window and misses minus 3 by 0.16.
\[ x = -0.96: \quad |4(-0.96) + 1 + 3| = 0.16 < 0.2 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 172-173 — Example 2.40
The only new feature here is that a is negative. The input distance from minus 1 is the absolute value of x minus minus 1, which is the absolute value of x plus 1. Getting that sign right is most of the difficulty.
The output error simplifies the same way: subtracting minus 3 adds 3, leaving 4x plus 4, which factors as 4 times x plus 1. The slope is 4, so delta is epsilon over 4, exactly as the rule for lines predicts.
The check uses epsilon 0.2 and delta 0.05. The input minus 0.96 is inside the window, and its output, minus 2.84, is 0.16 from minus 3. Inside the tolerance, as promised.
Fill the middle
\[ \lim_{x\to 2}(3x - 2) = 4 \]
Fill in the blanks
|(3x - 2) - 4| = |3x - 6| = 3\,|x - 2| < 3\,\delta = \varepsilon \quad\text\varepsilon/3\quad \delta = ___
Why: Three x minus 6 factors as 3 times x minus 2, so the output error is three times the input distance. With the input distance below delta the error is below three delta, and three delta equals epsilon when delta is epsilon over 3.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 173 — Checkpoint 2.27
Fill in the three blanks before checking. This is the book's checkpoint, with the template already written out, so all that is left is the chain.
Subtracting 4 from 3x minus 2 gives 3x minus 6, which factors as 3 times x minus 2. So the output error is three times the input distance, which is less than three delta. To make three delta equal epsilon, choose delta to be epsilon over 3.
If you got the slope right, you got the whole problem right. For any line the multiplier in the chain is the absolute value of the slope, and delta is epsilon divided by it.
Concept
\[ f(x) = mx + b, \quad m \ne 0, \quad L = ma + b \]
\[ |f(x) - L| = |mx - ma| = |m|\,|x - a| \]
\[ |m|\,|x - a| < |m|\,\delta = \varepsilon \iff \delta = \frac{\varepsilon}{|m|} \]
Figure (svg): Two panels with the same band of half-width one half around height 2. Left: a line of slope one half, whose window around x equals 2 has half-width 1. Right: a line of slope 3, whose window has half-width one sixth.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 170-173 — Examples 2.39 and 2.40 generalised
Here the three examples so far are done once for every line at the same time. The limit of mx plus b at a is ma plus b, and subtracting it leaves m times x minus a. The output error is the absolute value of m times the input distance.
So the output error is below epsilon whenever the input distance is below epsilon over the absolute value of m. The figure shows why the slope belongs in the denominator. With the same band in both panels, the shallow line stays inside it over a wide window, while the steep line shoots out almost at once.
The slope must not be zero here. A horizontal line never leaves the band at all, so any delta works, and the formula would divide by zero.
Estimation
Figure (svg): The line y equals 3x minus 7 through the point (3, 2), with a horizontal band from 0.5 to 3.5 and the matching window from 2.5 to 3.5 marked on the x-axis.
\[ \lim_{x\to 3} f(x) = 2, \qquad \varepsilon = 1.5 \text{ and } \varepsilon = 3 \]
Predict first
From the graph, which pair of deltas works for ε = 1.5 and ε = 3?
Correct: δ = 0.5 and δ = 1
Why: The line passes through (3, 2) and (4, 5), so its slope is 3 and f(x) = 3x − 7. Delta is epsilon over the slope: 1.5 over 3 is 0.5, and 3 over 3 is 1. Copying epsilon as delta forgets that the line is steep.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 180 — Exercises 184 and 185
Pick your answer before revealing. The graph is the one printed with Exercises 184 and 185, and reading it carefully is the whole exercise.
Find two grid points on the line. It passes through 3 comma 2, the limit point, and through 4 comma 5, so it rises 3 for each unit across. Its slope is 3, and the rule you just derived gives delta as epsilon over 3: a half for epsilon 1.5, and 1 for epsilon 3.
The figure shows the first case: the band of half-height 1.5 around 2 and the window from 2.5 to 3.5. The most tempting wrong answer copies epsilon as delta, which would be right only for a line of slope 1.
Trap
A write-up that starts from the goal:
\[ |(2x + 1) - 3| < \varepsilon \]
\[ \Rightarrow 2|x - 1| < \varepsilon \]
\[ \Rightarrow |x - 1| < \tfrac{\varepsilon}{2} = \delta \]
Wrong as a proof. It assumes the conclusion.
That chain shows the band forces the window: the converse. The definition needs the window to force the band. Start from the assumption and end at the target.
\[ |x - 1| < \delta \Rightarrow \]
\[ |(2x + 1) - 3| < 2\delta = \varepsilon \]
The left column looks like a proof and is written in good faith, but read the arrows. It starts from the output inequality and ends at the input inequality. That shows that if the output is in the band then the input is in the window, which is the converse of the definition.
For a line the converse happens to be true as well, which is why this error survives so long. For other functions it is not, and a proof in the wrong direction proves nothing at all. The fix is simple: use the scratch work to find delta, then write a fresh chain that starts from the assumption about x.
Section
Part 3
Concept
\[ \lim_{x\to 2} x^2 = 4 \qquad |x^2 - 4| = |x - 2|\,|x + 2| \]
The input distance factors out as before, but it drags along a second factor that changes with x. For a line that factor was the constant slope.
\[ |x - 2|\,|x + 2| < \varepsilon \iff |x - 2| < \frac{\varepsilon}{|x + 2|} \]
The right-hand side is not allowed as delta: it contains x, and x is chosen after delta. The whole art of nonlinear proofs is getting rid of that loose factor.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 173-174 — Examples 2.41 and 2.42
Try the line method on x squared at 2. The output error factors as the input distance times the absolute value of x plus 2. For a line, the second factor was a fixed number, the slope. Here it changes with x.
Dividing through gives the input distance less than epsilon over the absolute value of x plus 2, and it is tempting to call that delta. It is not allowed. Delta must be chosen before x is known, so it cannot contain x.
The book offers two ways around this. The geometric way solves directly for where the curve meets the band. The algebraic way puts a ceiling on the loose factor. Both are coming next.
Worked example
Find the window for the limit of x squared at 2 by solving for where the curve meets the band.
\[ \lim_{x\to 2} x^2 = 4, \quad \text{band } 4 - \varepsilon < x^2 < 4 + \varepsilon \]
Solve at the lower edge
Why: Near 2 the input is positive, so take the positive root. This needs epsilon at most 4.
\[ x^2 = 4 - \varepsilon \iff x = \sqrt{4 - \varepsilon} \]
Solve at the upper edge
Why: Same idea on the other side.
\[ x^2 = 4 + \varepsilon \iff x = \sqrt{4 + \varepsilon} \]
Measure both gaps from 2
Why: Left gap and right gap.
\[ 2 - \sqrt{4 - \varepsilon} \qquad \sqrt{4 + \varepsilon} - 2 \]
Take the smaller gap
Why: A symmetric window must fit on both sides.
\[ \delta = \min\left\{2 - \sqrt{4 - \varepsilon},\; \sqrt{4 + \varepsilon} - 2\right\} \]
Figure (svg): The parabola y equals x squared near x equals 2, with a band from 3 to 5. The band edges meet the curve at root 3 and root 5; the left gap from root 3 to 2 is about 0.268 and the right gap from 2 to root 5 is about 0.236.
Check at epsilon equal to 1
Why: The gaps are 0.268 and 0.236, so delta is 0.236; an input just inside the window, 2.23, squares to 4.973, inside the band below 5.
\[ \delta = \min\{0.268,\, 0.236\} = 0.236, \quad 2.23^2 = 4.9729 < 5 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 173 — Example 2.41 and Figure 2.41
This is the geometric route. Instead of manipulating inequalities, solve for the two inputs where the parabola meets the edges of the band: the square root of 4 minus epsilon on the left, and the square root of 4 plus epsilon on the right.
Measure each from 2. Because the parabola gets steeper as x grows, it leaves the band sooner on the right, so the right gap is smaller. The window has to be symmetric, so delta is the smaller of the two gaps. That is where the minimum in the formula comes from.
The figure puts real numbers on it with epsilon equal to 1: the left gap is about 0.268 and the right gap about 0.236. Using the left gap would let inputs just below 2.268 through, and their squares exceed 5. The check confirms that 2.23, inside the correct window, squares to 4.973, safely in the band.
Concept
Example 2.41 assumed epsilon is at most 4, because the square root of 4 minus epsilon needs it. Why is that fair?
\[ \delta \text{ works for } \varepsilon = 4 \;\Longrightarrow\; \text{the same } \delta \text{ works for every } \varepsilon > 4 \]
A wider band is easier to hit: any window that lands every output within 4 of L lands it within 10 of L too. So the hard cases are the small epsilons, and you may ignore large ones.
The reverse is never allowed. Assuming epsilon is at least some positive number throws away exactly the small tolerances that make the limit meaningful.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 173 — Example 2.41, step 2
Example 2.41 quietly assumed epsilon is at most 4, because otherwise the square root of 4 minus epsilon is not a real number. This slide explains why that is legitimate.
A large epsilon is an easy challenge. If a window keeps every output within 4 of the limit, it certainly keeps every output within 10. So if you can answer every epsilon up to 4, you can answer the bigger ones with the same window you used for 4. The interesting cases are the small ones, and you have not lost any of them.
The opposite move is never allowed. Assuming epsilon is at least a tenth would skip exactly the tiny tolerances the limit is about. Cap epsilon from above if it helps; never bound it from below.
Worked example
With delta chosen, show the window lands in the band.
Assume the window and drop the hole
Why: The proof never needs x different from 2.
\[ |x - 2| < \delta \iff -\delta < x - 2 < \delta \]
Replace delta by the gap on each side
Why: Delta is no bigger than either gap.
\[ -(2 - \sqrt{4 - \varepsilon}) < x - 2 < \sqrt{4 + \varepsilon} - 2 \]
Add 2 to all three parts
Why: The twos cancel on the outside.
\[ \sqrt{4 - \varepsilon} < x < \sqrt{4 + \varepsilon} \]
Square all three parts
Why: Allowed because every part is positive.
\[ 4 - \varepsilon < x^2 < 4 + \varepsilon \]
Subtract 4 from all three parts
Why: This is the band written with a bar.
\[ -\varepsilon < x^2 - 4 < \varepsilon \iff |x^2 - 4| < \varepsilon \]
Check the step that needed care
Why: Squaring preserves the order only for positive numbers; with epsilon at most 4, the left end is at least 0, and the endpoints square back exactly to the band's edges.
\[ \left(\sqrt{4 - \varepsilon}\right)^2 = 4 - \varepsilon, \quad \left(\sqrt{4 + \varepsilon}\right)^2 = 4 + \varepsilon \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 173-174 — Example 2.41, steps 4 and 5
Now the proof in the forward direction. Start from the window and drop the condition that x is not 2; the proof does not need it. Unwrap the absolute value into minus delta less than x minus 2 less than delta.
The clever step is the second one. Delta is no bigger than the left gap, so minus delta is no smaller than minus the left gap; delta is no bigger than the right gap. Replacing delta on each side by its own gap widens the interval. Adding 2 puts x between the two square roots.
Squaring all three parts is safe only because all three are positive, and that is what the check step verifies. The endpoints square back exactly to the band's edges, and subtracting 4 gives the output within epsilon of 4.
Worked example
\[ \lim_{x\to 9} \sqrt{x} = 3, \quad \text{band } 3 - \varepsilon < \sqrt{x} < 3 + \varepsilon, \quad 0 < \varepsilon \le 3 \]
Square the band's edges
Why: The root is increasing, so squaring the positive edges gives the matching inputs.
\[ (3 - \varepsilon)^2 < x < (3 + \varepsilon)^2 \]
Measure the left gap
Why: Expand and subtract from 9.
\[ 9 - (3 - \varepsilon)^2 = 6\varepsilon - \varepsilon^2 \]
Measure the right gap
Why: Expand and subtract 9.
\[ (3 + \varepsilon)^2 - 9 = 6\varepsilon + \varepsilon^2 \]
Take the smaller
Why: Subtracting epsilon squared makes the left one smaller.
\[ \delta = \min\{9 - (3 - \varepsilon)^2,\; (3 + \varepsilon)^2 - 9\} = 6\varepsilon - \varepsilon^2 \]
Figure (svg): The curve y equals root x from 0 to 17 with a band from 2.2 to 3.8 around height 3. The band edges meet the curve at x equals 4.84 and 14.44; the gaps to 9 are 4.16 on the left and 5.44 on the right.
Check at epsilon equal to 1
Why: Delta is 5, so the window runs from 4 to 14; the roots of 4 and 14 are 2 and 3.742, both within 1 of 3.
\[ \delta = 5: \quad \sqrt{4} = 2, \quad \sqrt{14} \approx 3.742, \quad \text{both in } (2, 4) \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 174 — Checkpoint 2.28
This is the same geometric method on a different curve. Square the band's edges to get the inputs where the root curve enters and leaves the band, then measure each from 9.
The left gap is 6 epsilon minus epsilon squared and the right gap is 6 epsilon plus epsilon squared, so the left one is smaller. This is the reverse of the parabola: the root curve flattens as x grows, so it leaves the band sooner on the left. The figure shows this with epsilon 0.8, where the gaps are 4.16 and 5.44.
Epsilon is capped at 3 so that 3 minus epsilon stays positive. The check takes epsilon 1, where delta is 5 and the window runs from 4 to 14. The square roots of the endpoints, 2 and about 3.742, both lie within 1 of 3.
Concept
\[ \lim_{x\to -1}(x^2 - 2x + 3) = 6 \qquad |x^2 - 2x - 3| = |x + 1|\,|x - 3| \]
Decide in advance that delta will be at most 1. Then x stays between minus 2 and 0, and on that stretch the loose factor has a ceiling.
\[ |x + 1| < 1 \Rightarrow -2 < x < 0 \Rightarrow -5 < x - 3 < -3 \Rightarrow |x - 3| < 5 \]
Figure (svg): The graph of y equals the absolute value of x minus 3, for x from minus 3 to 1.5. On the shaded window from minus 2 to 0 the graph stays between the dashed lines at heights 3 and 5.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 174-175 — Example 2.42
Here is the second way around the loose factor. Before choosing delta, decide that it will be no bigger than 1. Then x is within 1 of minus 1, so x lies between minus 2 and 0.
On that short stretch, the loose factor, the absolute value of x minus 3, cannot be large. Subtracting 3 puts x minus 3 between minus 5 and minus 3, so its absolute value is less than 5. The figure shows the graph of the loose factor staying below the line at height 5 over the whole shaded window.
That is the move: the loose factor is replaced by a fixed number, 5, and the problem becomes a line problem with slope 5.
Worked example
\[ \lim_{x\to -1}(x^2 - 2x + 3) = 6 \]
Let epsilon be arbitrary and choose delta
Why: Two demands: at most 1 to fence the loose factor, at most epsilon over 5 to finish.
\[ \text{Let } \varepsilon > 0. \quad \text{Choose } \delta = \min\left\{1, \tfrac{\varepsilon}{5}\right\}. \]
Assume the window
Why: Delta is below both numbers, so both bounds hold.
\[ 0 < |x + 1| < \delta \Rightarrow |x + 1| < 1 \text{ and } |x + 1| < \tfrac{\varepsilon}{5} \]
Bound the loose factor
Why: From the first bound, as on the previous slide.
\[ -1 < x + 1 < 1 \Rightarrow -5 < x - 3 < -1 \Rightarrow |x - 3| < 5 \]
Factor the output error
Why: Six subtracted, then factored.
\[ |(x^2 - 2x + 3) - 6| = |x + 1|\,|x - 3| \]
Use both bounds
Why: Epsilon over five times five.
\[ |x + 1|\,|x - 3| < \frac{\varepsilon}{5} \cdot 5 = \varepsilon \]
Check with epsilon equal to 1
Why: Delta is 0.2. The window's extremes, minus 1.19 and minus 0.81, give errors 0.796 and 0.724, both under 1.
\[ x = -1.19: 0.7961 < 1, \qquad x = -0.81: 0.7239 < 1 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 174-175 — Example 2.42
The choice of delta makes two promises at once. It is at most 1, which fences x in and gives the ceiling 5 on the loose factor. It is at most epsilon over 5, which finishes the chain. Taking the minimum is how you make both promises true.
After that, the proof is one line. The output error factors into the input distance times the loose factor, the first is less than epsilon over 5, the second is less than 5, and the product is less than epsilon.
The check uses epsilon 1, so delta is 0.2, and tries inputs near both ends of the window. Both errors come in under 1, with room to spare, which is typical: the algebraic delta is usually smaller than the best possible, and that is perfectly fine.
Notation
Annotate
On: \( \delta = \min\left\{1,\; \frac{\varepsilon}{5}\right\} \)
Work through the four annotations. The minimum is the device that lets one number satisfy two conditions. Whichever of 1 and epsilon over 5 is smaller, delta is at most both.
The 1 is arbitrary. The book says so directly: any positive number would work, and a different choice gives a different ceiling on the loose factor. The slider a few slides on lets you see the effect of changing it.
The 5 is not arbitrary; it is earned from the 1. And epsilon over 5 is the line rule in disguise: once the loose factor is fenced below 5, it acts like a slope of 5.
Step zero
\[ \lim_{x\to 1} x^2 = 1, \qquad |x^2 - 1| = |x - 1|\,|x + 1| \]
Discussion prompt
Checkpoint 2.29 will choose a preliminary bound of 1 on delta. Before any proof: if the input distance from 1 is less than 1, what ceiling does that put on the loose factor, and so what should delta be?
Write your answer before revealing. The aim of this slide is to make the preliminary bound a habit you apply before writing anything else.
The loose factor is the absolute value of x plus 1. If the input distance from 1 is less than 1, x is between 0 and 2, so x plus 1 is between 1 and 3, and its absolute value is below 3. The ceiling is 3.
With the ceiling in hand, delta is the minimum of 1 and epsilon over 3, and the proof on the next slide writes itself.
Worked example
\[ \lim_{x\to 1} x^2 = 1 \]
Let epsilon be arbitrary and choose delta
Why: The ceiling 3 comes from the previous slide.
\[ \text{Let } \varepsilon > 0. \quad \text{Choose } \delta = \min\left\{1, \tfrac{\varepsilon}{3}\right\}. \]
Assume the window
Why: Then both bounds hold.
\[ 0 < |x - 1| < \delta \]
Bound the loose factor
Why: Add 2 to each part of minus 1 less than x minus 1 less than 1.
\[ -1 < x - 1 < 1 \Rightarrow 1 < x + 1 < 3 \Rightarrow |x + 1| < 3 \]
Factor and finish
Why: Epsilon over three times three.
\[ |x^2 - 1| = |x - 1|\,|x + 1| < \frac{\varepsilon}{3} \cdot 3 = \varepsilon \]
Check with epsilon equal to one half
Why: Delta is one sixth, about 0.167; the input 1.16 is in the window and its square misses 1 by 0.3456.
\[ \delta = \tfrac16, \quad 1.16^2 - 1 = 0.3456 < 0.5 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 175 — Checkpoint 2.29
This is the template with the fence from the previous slide. Let epsilon be positive, choose delta as the minimum of 1 and epsilon over 3, and assume the window.
The first bound gives x plus 1 between 1 and 3, so its absolute value is below 3. The second bound makes the input distance below epsilon over 3. The output error is their product, below epsilon.
The check uses epsilon one half. Then delta is one sixth, the input 1.16 is inside the window, and its square misses 1 by 0.3456, well within one half.
Worked example
\[ \lim_{x\to 2}(x^2 + 2x) = 8 \]
Factor the output error
Why: Two is a root of x squared plus 2x minus 8.
\[ |x^2 + 2x - 8| = |x - 2|\,|x + 4| \]
Fence the loose factor
Why: With the input distance below 1, x is between 1 and 3.
\[ |x - 2| < 1 \Rightarrow 5 < x + 4 < 7 \Rightarrow |x + 4| < 7 \]
Choose delta
Why: At most 1 for the fence, at most epsilon over 7 for the finish.
\[ \delta = \min\left\{1, \tfrac{\varepsilon}{7}\right\} \]
Run the chain
Why: Under the assumption zero less than the input distance less than delta.
\[ |x^2 + 2x - 8| = |x - 2|\,|x + 4| < \frac{\varepsilon}{7}\cdot 7 = \varepsilon \]
Check with epsilon equal to 1
Why: Delta is one seventh, about 0.143. At the inputs 2.14 and 1.86 the errors are 0.860 and 0.820, both under 1.
\[ x = 2.14: 0.8596 < 1, \qquad x = 1.86: 0.8204 < 1 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 180 — Exercise 192
This exercise from the section's set follows the same pattern, so try it yourself before reading the steps. Subtract 8 and factor: since 2 is a root of x squared plus 2x minus 8, the factor x minus 2 comes out, leaving x plus 4 as the loose factor.
Fence it in with a preliminary bound of 1. Then x is between 1 and 3, x plus 4 is between 5 and 7, and the ceiling is 7. Delta is the minimum of 1 and epsilon over 7.
Notice that the ceiling is the value of the loose factor at the far edge of the fence, not at the point 2. Using 6, the value at 2, would be a mistake, because just to the right of 2 the factor is bigger than 6. The check at both edges of the window confirms the errors stay below 1.
Trap
Scratch work for Example 2.42, stopped too soon:
\[ |x + 1|\,|x - 3| < \varepsilon \]
\[ \delta = \frac{\varepsilon}{|x - 3|} \]
Wrong. Delta is chosen before x.
In the game, x is picked after delta is announced, so delta may use epsilon and nothing else. Replace the loose factor by a number, its ceiling on a preliminary window.
\[ |x - 3| < 5 \Rightarrow \]
\[ \delta = \min\left\{1, \tfrac{\varepsilon}{5}\right\} \]
This is the most common error in nonlinear proofs. The scratch work reaches the input distance less than epsilon over the loose factor, and it is tempting to stop there and call that delta.
Go back to the game. Delta is announced before the challenger picks x, so a formula with x in it cannot be announced; it is not a number yet. Worse, as x varies, the supposed window would change size, which is not what a window is.
The fix is always the same: replace the loose factor by a number that is bigger than it everywhere on a preliminary window, and build that preliminary window into delta with a minimum.
Tweak it
Parameter explorer
Example 2.42 used a preliminary bound of 1. With a bound k instead, the loose factor is below k + 4, so δ = min(k, ε/(k + 4)). The curve shows δ against ε. Slide k: what changes for small ε, and what for large ε?
\[ \delta(\varepsilon) = \min\left\{{k},\; \frac{\varepsilon}{{k} + 4}\right\} \]
Start at k equal to 1, which is Example 2.42. The graph shows delta as a function of epsilon: a straight rising line for small epsilon, then a flat cap at height 1 once epsilon passes 5.
Now slide k. Making k larger raises the cap but makes the rising part slightly less steep, because the ceiling on the loose factor, k plus 4, grows. Making k smaller lowers the cap and steepens the line a little.
For small epsilon, which is the part that matters, the choice of k hardly changes anything. That is why the book calls the choice of 1 arbitrary. Every positive k gives a correct proof, and 1 is simply easy to work with.
Error analysis
Annotate
On: \( \text{Choose } \delta = \frac{\varepsilon}{5}. \;\text{Then } |x+1|\,|x-3| < \frac{\varepsilon}{5}\cdot 5 = \varepsilon \)
Read the argument and try to spot the flaw before revealing the notes. The chain looks exactly like Example 2.42, except that delta is just epsilon over 5.
The flaw is that the bound 5 on the loose factor was never earned. It is only true on the fenced window, and without the minimum nothing keeps x on that window. For a small epsilon you might not notice, but the definition must work for every epsilon.
The annotations give a concrete failure. With epsilon 100, delta is 20, and the input 15 is in the window. There the loose factor is 12, not less than 5, and the output misses 6 by 192. A proof that fails for one epsilon is not a proof.
Sorting
\[ \lim_{x\to 2} x^2 = 4, \quad \varepsilon = 1, \quad \text{largest valid } \delta = \sqrt5 - 2 \approx 0.236 \]
Sort into buckets
For each δ, does every x with 0 < |x − 2| < δ satisfy |x² − 4| < 1?
Sort each window before checking. You know from Example 2.41 that the widest valid window for epsilon 1 has half-width about 0.236, the right-hand gap.
Every delta at or below 0.236 works, including the algebraic choice 0.2 from the minimum of 1 and epsilon over 5, which is more cautious than it needs to be. Every delta above 0.236 fails, because it lets in inputs just to the right of the square root of 5, whose squares exceed 5.
The item worth dwelling on is 0.268. It is a genuine gap from the picture, but it is the left one, the larger one. Choosing it would satisfy the left side and fail on the right, which is exactly why the minimum is needed.
Section
Part 4
Concept
triangle inequality — For any real numbers a and b, the absolute value of the sum is at most the sum of the absolute values.
\[ |a + b| \le |a| + |b| \]
Figure (svg): Two number lines. Top: an arrow a of length 3 followed by an arrow b of length 2 in the same direction, ending at 5, so the distance travelled equals the distance from 0. Bottom: an arrow of length 3 followed by one of length 2 pointing back, ending at 1, so the net distance is less than the path length.
It is the tool for splitting one error into two: if two quantities are each close to their targets, their sum is close to the sum of the targets.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 175 — Definition, the triangle inequality
The triangle inequality says the absolute value of a sum is at most the sum of the absolute values. The figure shows it as walking along the number line: walk a, then walk b. The distance from where you started can never exceed the total length you walked.
If both steps go the same way, as in the top line, the two are equal. If the second step goes back, as in the bottom line, you end up closer to the start than the total distance walked. Nothing can make the net distance bigger than the path.
In limit proofs the triangle inequality is how you split one error into two smaller errors you already control. That is exactly what the next proof does.
Worked example
\[ \lim_{x\to a} f(x) = L, \;\lim_{x\to a} g(x) = M \;\Longrightarrow\; \lim_{x\to a}\big(f(x) + g(x)\big) = L + M \]
Let epsilon be arbitrary
Why: The target is the sum within epsilon of L plus M.
\[ \text{Let } \varepsilon > 0. \]
Spend half the budget on f
Why: The limit of f is L, so a window exists for the tolerance epsilon over 2.
\[ 0 < |x - a| < \delta_1 \Rightarrow |f(x) - L| < \tfrac{\varepsilon}{2} \]
Spend the other half on g
Why: Likewise for g with its own window.
\[ 0 < |x - a| < \delta_2 \Rightarrow |g(x) - M| < \tfrac{\varepsilon}{2} \]
Choose the smaller window
Why: Inside it, both promises hold.
\[ \delta = \min\{\delta_1, \delta_2\} \]
Regroup the error
Why: Pair each function with its own limit.
\[ |(f + g) - (L + M)| = |(f - L) + (g - M)| \]
Apply the triangle inequality
Why: Split the error in two.
\[ \le |f - L| + |g - M| < \tfrac{\varepsilon}{2} + \tfrac{\varepsilon}{2} = \varepsilon \]
Check on an example
Why: Take f(x) = x and g(x) = 2x at 1, epsilon 0.3: delta one is 0.15, delta two is 0.075, so delta is 0.075; at 1.07 the sum misses 3 by 0.21.
\[ x = 1.07: \quad |1.07 + 2.14 - 3| = 0.21 < 0.3 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 175 — Proof of the sum law
This is the payoff of the precise definition: a limit law, proved rather than assumed. You are given that f tends to L and g tends to M, and you must show that f plus g tends to L plus M.
The key idea is budgeting. You need the total error below epsilon, so ask f for an error below half of epsilon and g for an error below the other half. You are allowed to, because the hypothesis about f holds for every tolerance, including epsilon over 2. Each gives a window, and the smaller window satisfies both.
Then regroup the error so each function sits next to its own limit, and the triangle inequality splits it into two pieces, each less than half of epsilon. The check runs the proof with real functions: with f equal to x and g equal to 2x at 1, delta comes out as 0.075, and the input 1.07 misses by 0.21, inside the tolerance of 0.3.
Trap
A sum-law attempt that asks each function for the full tolerance:
\[ |f - L| < \varepsilon, \quad |g - M| < \varepsilon \]
\[ \Rightarrow |(f+g) - (L+M)| < 2\varepsilon \]
Not what was asked: the target was epsilon.
The definition fixes the tolerance first. Since the limits of f and g hold for every tolerance, ask each for half of it, then add.
\[ \tfrac{\varepsilon}{2} + \tfrac{\varepsilon}{2} = \varepsilon \]
The left column asks each function for the full tolerance epsilon and ends with an error below two epsilon. Many students then say that is small enough. It is not what the definition asked, and on a proof it is marked wrong.
The repair costs nothing, because the hypotheses hold for every tolerance. Decide the target first and share it out: half to f, half to g. With three functions you would give each a third. Planning the budget before you start is the habit that makes these proofs come out exactly to epsilon.
Concept
Negate the definition one piece at a time: every becomes some, some becomes every, and the conclusion fails.
| definition | its opposite |
|---|---|
| for every ε > 0, | there exists ε > 0 so that |
| there exists δ > 0 so that | for every δ > 0, |
| if 0 < |x − a| < δ, then |f(x) − L| < ε | there is an x with 0 < |x − a| < δ and |f(x) − L| ≥ ε |
\[ \exists \varepsilon > 0 \;\forall \delta > 0 \;\exists x: \; 0 < |x - a| < \delta, \; |f(x) - L| \ge \varepsilon \]
The limit does not exist when this holds for every candidate L.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 176 — Table 2.10
To prove a limit fails, you need the precise opposite of the definition. The table negates it line by line. For every epsilon becomes there is some epsilon. There exists a delta becomes for every delta. And the conclusion becomes there is an input in the window whose output misses L by at least epsilon.
Read the negation as the game with the roles reversed. Now you are the challenger. You pick one tolerance, and you must show that every window the other side offers contains a bad input.
And because a limit that does not exist has no value at all, you must defeat every candidate L, not just one.
Worked example
\[ \lim_{x\to 0}\frac{|x|}{x} \text{ does not exist} \]
Take any candidate and one tolerance
Why: The negation lets you choose epsilon; one half will do for every L.
\[ L \in \mathbb{R}, \quad \varepsilon = \tfrac12 \]
Let delta be arbitrary, and split on the sign of L
Why: Either L is at least 0 or it is negative.
\[ \delta > 0; \quad L \ge 0 \text{ or } L < 0 \]
If L is at least 0, use a point on the left branch
Why: Half of delta to the left of 0 is inside the window.
\[ x = -\tfrac{\delta}{2}: \quad |x| = \tfrac{\delta}{2} < \delta, \quad \tfrac{|x|}{x} = -1 \]
Measure the miss
Why: Minus 1 is at least 1 away from any nonnegative L.
\[ |-1 - L| = 1 + L \ge 1 > \tfrac12 \]
If L is negative, use the right branch
Why: Half of delta to the right.
\[ x = \tfrac{\delta}{2}: \quad \tfrac{|x|}{x} = 1, \quad |1 - L| = 1 + |L| > \tfrac12 \]
Figure (svg): The graph of the absolute value of x over x: the line y equals minus 1 for negative x and y equals 1 for positive x, with open circles at x equals 0. A band from 0.2 to 1.2 around a candidate L of 0.7 contains the right branch but not the left.
Check that nothing was left out
Why: Every L falls in one case, every delta was arbitrary, and each time an x in the window missed by more than one half: exactly the negated definition.
\[ \forall L \;\exists \varepsilon = \tfrac12 \;\forall \delta \;\exists x: \; |f(x) - L| \ge \tfrac12 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 176-177 — Example 2.43
The function is minus 1 to the left of 0 and 1 to the right. Its two branches are 2 apart, so a tolerance of one half should be impossible to meet with any L. The proof makes that precise.
Take any candidate L, fix epsilon at one half, and let delta be any window. If L is 0 or more, the left branch at minus 1 is at least 1 away from L, and the input minus delta over 2 is in the window, so it is a bad input. If L is negative, the right branch at 1 is more than 1 away, and delta over 2 is the bad input.
The figure shows one candidate, L equal to 0.7. Its band holds the right branch but not the left, and the point half a window to the left of 0 falls outside. The check confirms every case is covered, which is what the negated definition requires.
Counterexample
\[ \exists \delta > 0 \;\forall \varepsilon > 0: \quad 0 < |x - a| < \delta \Rightarrow |f(x) - L| < \varepsilon \]
Discussion prompt
This reads almost like the definition, with the two quantifiers swapped. Find a limit it gets wrong: a function whose limit at 1 is plainly 3 but which fails this statement. Then say which functions DO satisfy it.
Write your answer before revealing. This slide shows why the order for every epsilon, there exists delta cannot be changed.
With the quantifiers swapped, one window must work for every tolerance at the same time. Then every output in that window is within every positive distance of L. The only number within every positive distance of L is L itself, so the function must equal L throughout the window.
That kills almost every limit you know, including the line 2x plus 1 at 1, which obviously tends to 3. The swapped statement describes functions that are constant near a, which is a completely different idea. In the real definition, delta is allowed to shrink as epsilon shrinks, and that freedom is everything.
Section
Part 5
Concept
Only the input condition changes. From the right, x must lie just above a; from the left, just below.
\[ \lim_{x\to a^{+}} f(x) = L: \quad 0 < x - a < \delta \Rightarrow |f(x) - L| < \varepsilon \]
\[ \lim_{x\to a^{-}} f(x) = L: \quad -\delta < x - a < 0 \Rightarrow |f(x) - L| < \varepsilon \]
The absolute value is gone from the input, because the sign of x minus a is now known. Each still starts: for every positive epsilon there exists a positive delta.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 177 — Definition, limits from the right and left
One-sided limits need only a small change. From the right, the input must be above a and within delta of it, so x minus a is between 0 and delta. From the left, x minus a is between minus delta and 0.
The output condition does not change at all, and neither do the quantifiers. The absolute value disappears from the input condition simply because you already know which side of a the input is on.
These definitions are what make limits at the edge of a domain meaningful. The square root of x minus 4 is only defined from 4 onward, so its limit at 4 can only be taken from the right, as the next example does.
Worked example
\[ \lim_{x\to 4^{+}} \sqrt{x - 4} = 0 \]
Scratch: solve the target for the input
Why: Square both sides; everything is nonnegative.
\[ \sqrt{x - 4} < \varepsilon \iff 0 < x - 4 < \varepsilon^2 \]
Let epsilon be arbitrary and choose delta
Why: The scratch work names it.
\[ \text{Let } \varepsilon > 0. \quad \text{Choose } \delta = \varepsilon^2. \]
Assume the right-hand window
Why: Only inputs above 4.
\[ 0 < x - 4 < \delta = \varepsilon^2 \]
Take square roots
Why: The root is increasing, so it keeps the order.
\[ 0 < \sqrt{x - 4} < \varepsilon \]
Write it as the definition wants
Why: The limit is 0.
\[ \left|\sqrt{x - 4} - 0\right| < \varepsilon \]
Figure (svg): The curve y equals the square root of x minus 4, starting at x equals 4. A band from minus 1.2 to 1.2 is drawn; the curve leaves the band at x equals 5.44, so the right-hand window runs from 4 to 5.44, of width epsilon squared equals 1.44.
Check with epsilon equal to 0.1
Why: Delta is 0.01; the input 4.009 is inside and its output 0.0949 is under 0.1.
\[ \sqrt{4.009 - 4} = \sqrt{0.009} \approx 0.0949 < 0.1 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 177-178 — Example 2.44 and Figure 2.42
The scratch work is quick: the root of x minus 4 is below epsilon exactly when x minus 4 is below epsilon squared. So delta should be epsilon squared, the book's clear choice.
The proof runs forwards from the right-hand window. Taking square roots keeps the order because the square root is an increasing function, and the output within epsilon of 0 is what was wanted.
The figure shows the geometric version with epsilon 1.2: the curve climbs out of the band where the root equals 1.2, at x equal to 5.44, a distance 1.44 from 4. Notice that delta is much smaller than epsilon when epsilon is small. With epsilon a tenth, delta is a hundredth, and the check confirms that an input inside that tiny window works.
Worked example
\[ \lim_{x\to 1^{-}} \sqrt{1 - x} = 0 \]
Write the target
Why: The limit is 0.
\[ \sqrt{1 - x} < \varepsilon \]
Square
Why: Both sides are nonnegative.
\[ 0 < 1 - x < \varepsilon^2 \]
Rewrite as a left-hand window
Why: Multiply through by minus 1, which reverses the inequalities.
\[ -\varepsilon^2 < x - 1 < 0 \]
Read off delta
Why: This is the left-hand window with half-width epsilon squared.
\[ \delta = \varepsilon^2 \]
Check with epsilon equal to 0.2
Why: Delta is 0.04; the input 0.97 is inside, and the root of 0.03 is 0.173, under 0.2.
\[ \sqrt{1 - 0.97} = \sqrt{0.03} \approx 0.1732 < 0.2 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 178 — Checkpoint 2.30
This is the mirror image of Example 2.44. The root of 1 minus x is only defined for x at most 1, so the limit is taken from the left.
Squaring the target gives 1 minus x between 0 and epsilon squared. Multiplying by minus 1 reverses the inequalities and gives x minus 1 between minus epsilon squared and 0, which is exactly the left-hand window. So delta is epsilon squared again.
The check uses epsilon 0.2, so delta is 0.04. The input 0.97 is 0.03 to the left of 1, inside the window, and the square root of 0.03 is about 0.173, inside the tolerance.
Picture it
Figure (svg): Two panels of y equals one over x squared. Left: a horizontal line at height M equals 4, crossed at x equals plus and minus one half. Right: the line at M equals 9, crossed at plus and minus one third. The window between the crossings is marked on each x-axis.
For an infinite limit there is no L to be close to. The challenger names a height M instead, as large as they like, and you must find a window in which the graph is above it.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 178 — Figure 2.43
When a function grows without bound near a, there is no L to be close to, so the band makes no sense. The definition replaces it with a bar: the challenger names a height M, and you must find a window in which the whole graph lies above that height.
The two panels are the book's Figure 2.43 drawn for one over x squared. With the bar at 4 the window has half-width one half; with the bar at 9 it shrinks to one third. The higher the bar, the narrower the window, just as a thinner band needed a narrower window before.
Saying the limit is infinity is therefore a precise statement, not a way of saying the limit fails. It is a limit that does not exist as a number but exists in this stronger, specific sense.
Notation
Annotate
On: \( \lim_{x\to a} f(x) = +\infty: \; \forall M > 0 \;\exists \delta > 0: \; 0 < |x - a| < \delta \Rightarrow f(x) > M \)
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 178 — Definition, infinite limits
Go through the annotations. The shape is the same as before: for every challenge there exists a window. Only the challenge and the conclusion change. The challenge M is a large positive number, and the conclusion is that f of x is bigger than M.
There is no absolute value in the conclusion, because the graph must be above the bar, not near it. For a negative infinite limit, the conclusion flips to f of x less than minus M, a floor as deep as the challenger likes.
Once you see the pattern, you can write down the definition for any variation, which is what Exercises 176 to 179 and the table a few slides on ask you to do.
Worked example
\[ \lim_{x\to 0} \frac{1}{x^2} = +\infty \]
Scratch: solve the target for the input
Why: Both sides positive, so take reciprocals and flip.
\[ \frac{1}{x^2} > M \iff x^2 < \frac1M \]
Take square roots
Why: The root of x squared is the absolute value of x.
\[ |x| < \frac{1}{\sqrt{M}} \]
Let M be arbitrary and choose delta
Why: The window from the scratch work.
\[ \text{Let } M > 0. \quad \text{Choose } \delta = \frac{1}{\sqrt M}. \]
Assume the window and run forwards
Why: Square, then take reciprocals, which reverses the order.
\[ 0 < |x| < \frac{1}{\sqrt M} \Rightarrow 0 < x^2 < \frac1M \Rightarrow \frac{1}{x^2} > M \]
Check with M equal to ten thousand
Why: Delta is 0.01; the input 0.009 is inside and its output is about 12346, over the bar.
\[ \frac{1}{0.009^2} \approx 12{,}346 > 10{,}000 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 180 — Exercise 196
The scratch work solves the target inequality for x. Both sides are positive, so taking reciprocals flips the inequality: x squared is below one over M. Taking square roots gives the absolute value of x below one over the square root of M.
The proof runs that chain forwards from the window. Squaring gives x squared below one over M, and taking reciprocals, which flips the inequality again, gives one over x squared above M.
The check takes a bar of ten thousand. Delta is a hundredth, the input 0.009 is inside the window, and its output is about 12346, above the bar. Notice that the window shrinks only like one over a square root as the bar rises.
Comparison
Comparison matrix
| statement | the challenge | input window | what must follow |
|---|---|---|---|
| lim f(x) = L as x → a | ε > 0 | 0 < |x − a| < δ | |f(x) − L| < ε |
| lim f(x) = L as x → a⁺ | ε > 0 | 0 < x − a < δ | |f(x) − L| < ε |
| lim f(x) = L as x → a⁻ | ε > 0 | −δ < x − a < 0 | |f(x) − L| < ε |
| lim f(x) = +∞ as x → a | M > 0 | 0 < |x − a| < δ | f(x) > M |
| lim f(x) = −∞ as x → a | M > 0 | 0 < |x − a| < δ | f(x) < −M |
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, pp. 177-178 — the definitions; Exercises 176-179
Fill in the blanks before checking. Every row has the same skeleton, for every challenge there exists delta such that the window implies the conclusion, so you only need to know which piece changes.
For one-sided limits the input window changes and the rest stays. For infinite limits the challenge becomes M and the conclusion becomes a comparison with M or minus M. Nothing changes in two places at once.
Seeing all five definitions side by side is the best way to remember them, and it is also exactly the skill Exercises 176 to 179 test, where the letters are changed to make sure you understand the structure rather than the symbols.
Matching
Match the pairs
Why: One-sided limits drop the absolute value and keep one side of a; the infinite limit uses M in place of ε; the two-sided linear limit divides ε by the slope 4. Each window is the delta you found on an earlier slide.
Match each limit to the window you found for it earlier in the lesson. The shape of the window tells you the kind of limit before you look at any numbers.
Windows with no absolute value, lying on one side of a, belong to one-sided limits: epsilon squared to the right of 4, and epsilon squared to the left of 1. A window whose size is written with M belongs to the infinite limit. And the line with slope 4 divides epsilon by 4.
Estimation
\[ \lim_{x\to 2} -\frac{1}{(x - 2)^2} = -\infty, \qquad M = 100 \]
Predict first
Which window guarantees f(x) < −100?
Correct: δ = 0.1
Why: −1/(x − 2)² < −M is the same as (x − 2)² < 1/M, so |x − 2| < 1/√M = 1/10. For example x = 2.09 gives −1/0.0081 ≈ −123.5, below −100. The choice 0.01 also works (smaller is always fine), but 0.1 is the one the scratch work produces.
\[ -\frac{1}{(x-2)^2} < -M \iff |x - 2| < \frac{1}{\sqrt M} = 0.1 \]
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 180 — Exercise 198
Pick an answer first. This is a negative infinite limit, so the conclusion is that f of x is below minus M.
Multiplying minus one over x minus 2 squared less than minus M by minus 1 reverses it: one over x minus 2 squared is greater than M. From there it is exactly Exercise 196, shifted to 2: the input distance must be below one over the square root of M, which for M equal to 100 is a tenth.
As always, a smaller window also works, so a hundredth is not wrong, just unnecessarily cautious. The answers 1 and 10 fail badly: at the input 3, the function is only minus 1.
Real world
Figure (svg): The area function A equals s squared for side lengths from 11.4 to 12.6 centimetres. A band of areas from 136 to 152 around 144 meets the curve at side lengths 11.662 and 12.329; the gaps to 12 are 0.338 and 0.329.
Discussion prompt
A machine cuts a square of Aerogel with target area 144 square centimetres, and the area may be off by at most 8. How accurately must each side be cut? Name the f, a, L, epsilon and delta.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 180 — Exercise 199
Try the question before revealing. The area of a square of side s is s squared, and the target is 144 square centimetres, so the target side is 12. The tolerance of 8 on the area is epsilon.
Solving for the sides whose squares lie between 136 and 152 gives a side between about 11.662 and 12.329 centimetres. The gaps to 12 are 0.338 and 0.329, and as with any curve, delta is the smaller one: cut each side to within about a third of a centimetre.
This is the definition doing real work. The customer specifies a tolerance on the output, the area. The engineer needs a tolerance on the input, the side. The epsilon-delta definition is precisely the promise that such an input tolerance exists for every output tolerance.
Section
Part 6
Pattern
Figure (svg): A flow diagram of an epsilon-delta proof: scratch work first, starting from the target inequality and solving for the distance from x to a; if a loose factor remains, bound it with a preliminary delta of 1 and take a minimum; then the write-up: let epsilon be positive, choose delta, assume the window, chain to the target.
This is the procedure for every proof in this lesson. The diagram shows its two passes. On the left, scratch work that you do not hand in, starting from the target and ending with a delta. On the right, the proof itself, running forwards through the template.
The one decision point is the loose factor. If the output error is a constant times the input distance, divide epsilon by the constant. If a factor still depends on x, fence it in with a preliminary bound of 1, find its ceiling, and take a minimum.
One-sided and infinite limits use the same procedure with the window or the conclusion changed, exactly as their definitions say. If you remember nothing else, remember the order: let epsilon, choose delta, assume the window, reach epsilon.
Check
Check your understanding
Which statement is exactly what the limit of f at a equal to L means?
Answer: B
Why: Epsilon is arbitrary and comes first; delta is found in response; the window implies the band. Every other option changes the order of the quantifiers or the direction of the implication.
Read all four carefully; they differ by the order of two phrases or the direction of one arrow. The correct one names epsilon first, finds delta second, and has the window implying the band.
Each wrong option is a real mistake worth recognising. Swapping the quantifiers demands a constant function. Letting epsilon be chosen after delta makes the statement almost empty. And reversing the implication is the scratch-work direction, not the definition.
Check
Check your understanding
To prove that the limit of 10x + 3 as x approaches 1 is 13, which δ works for every ε > 0?
Answer: A
Why: The output error is |10x + 3 − 13| = 10|x − 1|, which is below 10δ, and 10δ equals ε when δ = ε/10. The slope divides epsilon.
Subtract the limit and factor: 10x plus 3 minus 13 is 10 times x minus 1. The slope is 10, so delta is epsilon over 10.
The option with 13 in it is worth thinking about. The limit value disappears the moment you subtract it; only the slope survives into the output error. For every line, delta is epsilon over the absolute value of the slope, and nothing else about the line matters.
Check
Check your understanding
For the limit of x² as x approaches 3, equal to 9, with a preliminary bound δ ≤ 1, which choice of δ is correct?
Answer: B
Why: |x² − 9| = |x − 3||x + 3|. With |x − 3| < 1, x lies between 2 and 4, so |x + 3| < 7. Then |x² − 9| < 7δ ≤ ε when δ = min(1, ε/7).
Factor the output error into the input distance times the absolute value of x plus 3. With the preliminary bound of 1, x is between 2 and 4, so x plus 3 is between 5 and 7. The ceiling is 7, and delta is the minimum of 1 and epsilon over 7.
The tempting wrong answers are all versions of errors from this lesson: using the value of the loose factor at the point, which is 6; putting x into delta; and computing the fence on the wrong interval, which gives 5.
Explain it to yourself
Discussion prompt
In a quadratic proof, delta is the minimum of 1 and epsilon over K. Explain in two or three sentences what each of the two numbers is doing, and what goes wrong if you drop the 1.
Write your explanation before revealing. If you can explain the minimum clearly, you understand nonlinear epsilon-delta proofs.
The two numbers do two different jobs. The 1 is about the loose factor: it keeps x close enough to a that the factor has a known ceiling. The epsilon over K is about the input distance: it makes that distance small enough that K times it is below epsilon. The minimum makes both true at once.
Without the 1, the ceiling on the loose factor is an assumption, not a fact, and for large epsilon it is false. The find-the-error slide showed a concrete input where the argument collapses.
Exit ticket
\[ \lim_{x\to -1} \frac{3}{(x + 1)^2} = +\infty \]
Discussion prompt
Exercise 197. Find delta in terms of M, write the proof in the template, and check it with M = 300.
OpenStax Calculus Volume 1, §2.5 The Precise Definition of a Limit §2.5, p. 180 — Exercise 197
This combines the scratch work, the template, and an infinite limit in one problem. Solve the target inequality for the input distance, write the four lines of the proof, and check with a bar of 300.
Three over x plus 1 squared is above M exactly when x plus 1 squared is below 3 over M, so delta is the square root of 3 over M. In the proof, the assumption makes x plus 1 squared below 3 over M, so three over it is above M.
With M equal to 300, delta is a tenth; the input minus 0.91 is inside the window and gives about 370, over the bar. If your delta matched, you have the whole lesson.
Recap
| statement | precise meaning |
|---|---|
| lim f(x) = L as x → a | ∀ε > 0 ∃δ > 0: 0 < |x − a| < δ ⇒ |f(x) − L| < ε |
| from the right / left | input window 0 < x − a < δ, or −δ < x − a < 0 |
| lim f(x) = ±∞ | ∀M > 0 ∃δ > 0: 0 < |x − a| < δ ⇒ f(x) > M (or < −M) |
| limit does not exist | for every L: ∃ε > 0 ∀δ > 0 ∃x in the window with |f(x) − L| ≥ ε |
\[ \text{line: } \delta = \frac{\varepsilon}{|m|} \qquad \text{quadratic: } \delta = \min\left\{1, \frac{\varepsilon}{K}\right\} \]
Next, Section 3.1 uses a limit to define the derivative, and the laws you can now prove are what make derivatives computable.
Stewart, Calculus: Early Transcendentals 8e, §2.4 The Precise Definition of a Limit §2.4, pp. 104-113 — the same material in Stewart
The table collects the four precise statements from the lesson. Each one has the same skeleton, a challenge that can be made as demanding as anyone likes, and a window that answers it. Only the input window and the conclusion vary.
For proofs, remember the two formulas at the bottom. A line needs delta equal to epsilon over the slope. A curve needs a preliminary bound, a ceiling K on the loose factor, and the minimum of 1 and epsilon over K.
In the next chapter the derivative is defined as a limit. Every derivative rule you will learn rests on limit laws like the sum law you proved here, so this definition is the foundation under all of differential calculus.
Want this taught 1-on-1? Alexander tutors Calculus I — $55/session, free consultation.