Visual walkthrough — Uniform continuity — difference from pointwise

---

Step 1 — The picture that starts everything: two dots and a box

WHAT. Draw the graph of some function $f$. Pick two points on the horizontal axis and call them $x$ and $y$. "$f(x)$" just means "the height of the curve above the point $x$". So above $x$ and $y$ sit two heights, $f(x)$ and $f(y)$.

Two numbers now matter:

• the horizontal gap $|x-y|$ — how far apart the dots are sideways. The two vertical bars $|\cdot |$ mean "distance", i.e. "always a positive amount, ignore direction".
• the vertical gap $|f(x)-f(y)|$ — how far apart the curve's heights are up–down.

WHY. Continuity is a promise about a trade: "keep the sideways gap small and I promise the up–down gap stays small." To reason about it we must first name these two gaps and see them as two sides of a box. That box is the whole story.

PICTURE. The red box below has width $|x-y|$ and height $|f(x)-f(y)|$. Continuity controls the height by controlling the width.

---

Step 2 — Two dials: $\epsilon$ (the goal) and $\delta$ (the effort)

WHAT. Two Greek letters enter. They are just names for two positive numbers:

• $\epsilon$ (epsilon) — the height tolerance you are challenged to meet: "make the output gap smaller than $\epsilon$". It is handed to you by an opponent; you don't get to choose it.
• $\delta$ (delta) — the width you are allowed to use: "as long as the input gap is below $\delta$". You get to choose this.

The whole game: someone names $\epsilon$; you must answer with a $\delta$ so that $\underset{\text{dots within width }\delta }{\underbrace{|x-y|<\delta }}⟹\underset{\text{heights within tolerance }\epsilon }{\underbrace{|f(x)-f(y)|<\epsilon }}.$ Read the arrow $⟹$ as "guarantees". The symbol $<$ is "strictly less than".

WHY these two dials and not one? Because continuity is a cause-and-effect promise. One dial ($\epsilon$) sets the effect you want; the other ($\delta$) is the cause you're allowed to control. You cannot state the promise with a single number — you need the target and the effort.

PICTURE. Think of an $\epsilon$-tall horizontal band around the height $f(x)$, and a $\delta$-wide vertical band around $x$. The rule says: any dot inside the $\delta$-wide band must land inside the $\epsilon$-tall band.

---

Step 3 — The ONE quantifier swap (drawn, not just stated)

WHAT. Now we ask the decisive question: when you choose $\delta$, are you allowed to look at where the point is? To answer precisely we write the full rule for both flavours, keeping every piece — the tolerance $\forall \epsilon$, the width $\exists \delta$, the point sweep, and the closeness implication:

$$\text{Pointwise: }\forall x_0\forall \epsilon >0\exists \delta >0\forall y:|x_0-y|<\delta ⟹|f(x_0)-f(y)|<\epsilon$$ $$\text{Uniform: }\forall \epsilon >0\exists \delta >0\forall x\forall y:|x-y|<\delta ⟹|f(x)-f(y)|<\epsilon$$

Reading each piece: $\forall$ = "for every", $\exists$ = "there is", $\epsilon$ = the height tolerance (Step 2), $\delta$ = the width you supply, and the arrow $⟹$ carries the same closeness promise in both — "if inputs are within $\delta$ then outputs are within $\epsilon$." Only the ordering of the boxed pieces differs:

• Pointwise: you first stand at a fixed point $x_0$, then (given $\epsilon$) choose $\delta$. Because $\exists \delta$ sits after $x_0$, your width may be re-chosen at every point — write it $\delta (\epsilon ,x_0)$, "$\delta$ allowed to depend on both $\epsilon$ and the location $x_0$".
• Uniform: $\exists \delta$ sits before the points $x,y$ are revealed, so one width must serve all pairs at once — write it $\delta (\epsilon )$: depends on the tolerance only, never on location.

WHY this matters visually. On a curve whose steepness changes, "one $\delta$ for the steep part" and "one $\delta$ for the flat part" are different sizes. Pointwise lets you switch $\delta$ region by region; uniform forbids switching. That single freedom-vs-restriction is what we now watch break.

PICTURE. Left panel: a flat region needs a wide $\delta$; a steep region needs a narrow $\delta$. Pointwise picks each locally (both are fine). Uniform must pick ONE red $\delta$ — and the narrowest one wins.

---

Step 4 — Why steepness controls the height gap: the slope factor

WHAT. Take the concrete curve $f(x)=x^2$. Compute the output gap and factor it:

$$|f(x)-f(y)|=|x^2-y^2|=\underset{\text{amplifier}}{\underbrace{|x+y|}}\cdot \underset{\text{input gap}}{\underbrace{|x-y|}}.$$

Term by term:

• $|x-y|$ is the input gap you control (Step 1).
• $|x+y|$ is an amplifier: it multiplies the input gap to produce the output gap. Where $|x+y|$ is large, a tiny input gap gets blown up into a large output gap.

WHY factor? Because $a^2-b^2=(a+b)(a-b)$ is the one algebra move that separates "the gap I control" from "the multiplier I don't". Without factoring, $|x^2-y^2|$ is opaque; after factoring, the danger — the growing $|x+y|$ — is staring at us.

PICTURE. The output gap (height of the red box) is the input gap stretched by the local steepness $|x+y|$. Slide the pair rightward and the same-width box grows taller and taller.

---

Step 5 — The break: no single $\delta$ survives on all of $\mathbb{R}$

WHAT. To force $|x^2-y^2|<\epsilon$ we would need

$$|x-y|<\frac{\epsilon }{|x+y|}.$$

As $|x+y|\to \infty$ (i.e. as we walk to the right forever), the allowed width $\frac{\epsilon }{|x+y|}\to 0$. A single fixed $\delta >0$ cannot stay below a target that is heading to $0$. So uniform continuity fails on $\mathbb{R}$.

Let's make the failure explicit — this is the negation of uniform continuity in action:

WHY this construction? We engineered two points close in input but far in output by parking them where the amplifier is large. That is precisely what "not uniformly continuous" means: for some $\epsilon$, no $\delta$ works.

PICTURE. The narrow red input gap sits far right where the parabola is steep; its box is taller than the $\epsilon =1$ line no matter how narrow we make it.

---

Step 6 — The rescue: cap the amplifier by staying on $[0,5]$

WHAT. Restrict $x^2$ to the closed interval $[0,5]$. Now both $x,y\in [0,5]$, so the amplifier is bounded:

$$|x+y|\le 5+5=10.$$

Therefore

$$|x^2-y^2|=|x+y||x-y|\le 10|x-y|.$$

Choose $\delta =\frac{\epsilon }{10}$. Then $|x-y|<\delta$ gives $|x^2-y^2|\le 10\cdot \frac{\epsilon }{10}=\epsilon$. Crucially $\delta$ contains no $x$ — one width serves the entire interval. That is uniformity.

WHY it works now. On $[0,5]$ the curve cannot get infinitely steep; the amplifier tops out at $10$. Capping the amplifier caps how much a fixed width can be stretched — so a fixed width is enough. This is the Heine–Cantor Theorem in miniature: on a compact (closed and bounded) interval, continuity is automatically upgraded to uniform continuity. The engine is the Lebesgue number lemma — cover by local $\delta$-balls, extract a finite subcover, take the minimum $\delta$; a minimum of finitely many positive numbers is still positive. On unbounded $\mathbb{R}$ that "minimum" becomes an infimum of infinitely many shrinking numbers, which can be $0$ — and Step 5 is exactly that collapse.

PICTURE. Chop $[0,5]$ into local balls each with its own $\delta$; finitely many cover the interval; the smallest red $\delta$ works everywhere.

---

Step 7 — The degenerate warning: bounded but NOT closed is not enough

WHAT. One might think "$[0,5]$ worked because it was small/bounded — so any bounded interval works." It does not. Consider $f(x)=\frac{1}{x}$ on the open interval $(0,1)$. It is bounded as a set, and $f$ is continuous there, yet near $0$ the curve shoots to infinity. Let us find the amplifier exactly, the same way we factored $x^2$. Put the two fractions over a common denominator:

$$\left|\frac{1}{x}-\frac{1}{y}\right|=\left|\frac{y-x}{xy}\right|=\underset{\text{amplifier}}{\underbrace{\frac{1}{xy}}}\cdot \underset{\text{input gap}}{\underbrace{|x-y|}}.$$

Term by term: $|x-y|$ is again the input gap you control; the amplifier is now $\frac{1}{xy}$, and as $x,y\to 0^{+}$ this blows up ($xy\to 0$ so $1/(xy)\to \infty$).

To force the output gap below $\epsilon$ we would need $|x-y|<\epsilon xy$, and the right-hand side collapses to $0$ near the origin. Concretely, fix $\epsilon =1$ and any candidate $\delta >0$; pick

$$x=\delta ,y=\frac{\delta }{2}(\text{both near }0\text{ for small }\delta ),|x-y|=\frac{\delta }{2}<\delta ,$$

yet

$$\left|\frac{1}{x}-\frac{1}{y}\right|=\left|\frac{1}{\delta }-\frac{2}{\delta }\right|=\frac{1}{\delta },$$

which exceeds $\epsilon =1$ as soon as $\delta <1$. Two points within the width, output gap huge — exactly the Step 5 mechanism, now driven by $1/(xy)\to \infty$ instead of $|x+y|\to \infty$.

WHY it fails. $(0,1)$ is bounded but not closed: the troublesome endpoint $0$ is missing, so nothing stops the amplifier $1/(xy)$ from becoming arbitrarily large as we approach it. Compactness needs closed and bounded (via the Heine–Cantor Theorem); drop "closed" and the rescue of Step 6 collapses.

PICTURE. Two dots near $0$ with a tiny sideways gap; their heights on the $1/x$ spike are miles apart — the box is arbitrarily tall for arbitrarily thin width.

---

Step 8 — The other degenerate case: uniform WITHOUT bounded slope ($\sqrt{x}$)

WHAT. Symmetric warning in the other direction: "uniform continuity requires bounded slope." Not so. Take $f(x)=\sqrt{x}$ on $[0,\infty )$. Its slope is $f^{\prime }(x)=\frac{1}{2\sqrt{x}}\to \infty$ as $x\to 0^{+}$ — unbounded — so it is not Lipschitz and the MVT shortcut doesn't apply. Yet it is uniformly continuous, and here is the clean derivation of the bound that saves it.

Assume $x\ge y\ge 0$ (relabel if needed). Multiply the output gap by the conjugate $\frac{\sqrt{x}+\sqrt{y}}{\sqrt{x}+\sqrt{y}}$ — the standard trick that clears square roots:

$$\sqrt{x}-\sqrt{y}=\frac{(\sqrt{x}-\sqrt{y})(\sqrt{x}+\sqrt{y})}{\sqrt{x}+\sqrt{y}}=\frac{x-y}{\sqrt{x}+\sqrt{y}}.$$

Now compare with $\sqrt{x-y}$. Since $x\ge y\ge 0$ we have $\sqrt{x}\ge \sqrt{x-y}$ (bigger inside, bigger root) and $\sqrt{y}\ge 0$, so the denominator satisfies $\sqrt{x}+\sqrt{y}\ge \sqrt{x-y}$. A bigger denominator makes the fraction smaller:

$$\sqrt{x}-\sqrt{y}=\frac{x-y}{\sqrt{x}+\sqrt{y}}\le \frac{x-y}{\sqrt{x-y}}=\sqrt{x-y}.$$

In symmetric form: $|\sqrt{x}-\sqrt{y}|\le \sqrt{|x-y|}$. The output gap is bounded by the square root of the input gap — gentler than linear, yet still shrinking to $0$ as the input gap shrinks, uniformly in location. Choose $\delta ={\epsilon }^2$:

$$|x-y|<{\epsilon }^2⟹|\sqrt{x}-\sqrt{y}|\le \sqrt{|x-y|}<\sqrt{{\epsilon }^2}=\epsilon .$$

Again $\delta ={\epsilon }^2$ has no $x$ in it — uniform.

WHY show this. To fence off the whole space of possibilities: Lipschitz (bounded slope) is sufficient but not necessary. A curve can be infinitely steep at a point and still be uniformly continuous, provided the steepness is "integrable" enough that the square-root bound holds.

PICTURE. Near $0$ the $\sqrt{x}$ curve is vertical, yet the output gap is capped by $\sqrt{\text{input gap}}$ — the box height never outruns the tolerance.

---

The one-picture summary

Everything above compresses into a single comparison: is the amplifier bounded on the whole domain?

• Bounded → one $\delta$ works → uniform (right box of $x^2$ on $[0,5]$; also $\sqrt{x}$ via the ​ bound).
• Unbounded → $\delta$ must shrink to $0$ → not uniform ($x^2$ on $\mathbb{R}$; $1/x$ near $0$).

Recall Feynman retelling — the whole walkthrough in plain words

Draw a curve. Put two dots on it. A box appears between them: its width is how far apart they are sideways, its height is how far apart the curve's values are up–down. Continuity is a promise: keep the box narrow and I'll keep it short. Someone shouts a height limit $\epsilon$; you answer with a width $\delta$ that keeps every box under that height. The only question that separates the two flavours is: do you get to choose a different width in each region (pointwise), or must you pick one width for the entire page (uniform)? For $x^2$, the local steepness is $|x+y|$ — it stretches your box taller the farther right you go. On the whole line it grows without limit, so any single width eventually produces a box that's too tall: continuous but not uniformly so. Chop the line down to $[0,5]$ and the steepness can't exceed $10$; now one width ($\delta =\epsilon /10$) tames every box — that's the compactness rescue (Heine–Cantor). But watch the traps: $(0,1)$ is bounded yet not closed, and $1/x$ still explodes near the missing edge $0$ (its amplifier $1/(xy)\to \infty$), so "small interval" isn't enough; and $\sqrt{x}$ is infinitely steep at $0$ yet stays uniform because its box height only grows like the square root of the width. One idea rules them all: uniform continuity ⇔ the amplifier is bounded across the whole domain.

---

Connections

• Continuity (pointwise) — the weaker promise (Step 3) this walkthrough strengthens.
• Heine–Cantor Theorem — the compactness rescue of Step 6.
• Compactness / Lebesgue number lemma — the finite-subcover + minimum-$\delta$ engine.
• Lipschitz continuity — bounded amplifier as a sufficient test (Steps 6, 8).
• Mean Value Theorem — turns bounded slope into a Lipschitz bound.
• Cauchy sequences — uniform continuity preserves Cauchyness; pointwise need not.

Concept Map