Measure theory — Lebesgue measure (intro): Visual walkthrough

Step 1 — The question: what is the "length" of a set of points?

WHY this is hard. For an interval $[a, b]$ the answer is obvious: the size is $b - a$ , the distance from the left end $a$ to the right end $b$ . But for a fog of scattered points there are no "ends" to subtract. We need a machine that works on any $A$ and still gives $b - a$ on intervals.

PICTURE. The green blob is our mystery set $A$ . Below it sits the one case we already trust: the interval $[a, b]$ with its honest length $b - a$ .

Figure — Measure theory — Lebesgue measure (intro)

Step 2 — The one honest move: cover the set with intervals

WHY covers. Intervals are the only things whose length we know ( $ℓ = b - a$ ). So the trick is: approximate the unknown blob using things we DO know. If the intervals blanket $A$ completely, then the total length of the intervals is surely at least as big as any honest "length of $A$ " could be. We never undershoot with a cover — that is the key one-directional guarantee.

PICTURE. Chalk-blue intervals $I_{1}, I_{2}, I_{3}, \dots$ laid over the green blob. Each point of $A$ falls under some interval. Notice the intervals can overlap and can stick out past $A$ — both are allowed.

The cost of a cover is the total chalk we spent: $cost = \sum_{n = 1}^{\infty} ℓ (I_{n}) = \sum_{n = 1}^{\infty} (b_{n} - a_{n}) .$

Step 3 — Squeeze: take the tightest cover (the infimum)

WHY infimum and not minimum? There may be no single "cheapest" cover — you can often make the cost a hair smaller by shrinking intervals a tiny bit more, forever. So there is no minimum. Instead we take the infimum ( $in f$ ): the greatest number that no cover can go below — the "floor" that costs approach but might never touch. That floor is the honest size of $A$ .

PICTURE. Three covers of the same blob: fat/expensive (pink), medium (yellow), tight/cheap (blue). An arrow shows the costs marching downward toward a floor line — that floor is $μ^{*} (A)$ .

Step 4 — First sanity check: a single point costs nothing

WHY it must be zero. A point has no width. We can trap it in a tiny interval of any width we like and then shrink that width toward $0$ .

PICTURE. The point $p$ (a chalk dot) inside a shrinking interval $(p - ε, p + ε)$ ; the width $2 ε$ collapses onto the dot.

${p} \subseteq (p - ε, p + ε), ℓ = (p + ε) - (p - ε) = 2 ε .$

$ε$ ("epsilon") is any tiny positive number we choose.
The single interval already covers $p$ , so this is a legal cover of cost $2 ε$ .

Since $μ^{*} ({p}) \leq 2 ε$ for every $ε > 0$ , and $μ^{*} \geq 0$ always, the floor is forced to be $μ^{*} ({p}) = 0$ . ✔

Step 5 — The $ε / 2^{n}$ magic: even $Q$ costs nothing

WHY it works. $Q$ is countable, so we can line the fractions up: $q_{1}, q_{2}, q_{3}, \dots$ . Now give each fraction its own tiny interval, but make them shrink geometrically — the $n$ -th one gets width $ε / 2^{n}$ . A shrinking geometric series has a finite total, and we can make that total as small as we like.

PICTURE. The fractions as dots along the line; interval $n$ has width $ε / 2^{n}$ so the widths halve down the row: $ε /2, ε /4, ε /8, \dots$ . A stacked bar shows the widths adding up to just $ε$ .

$Q \subseteq ⋃_{n = 1}^{\infty} (q_{n} - \frac{ε}{2 ^{n + 1}}, q_{n} + \frac{ε}{2 ^{n + 1}}), width of interval n = \frac{ε}{2 ^{n}} .$

$total cost = \sum_{n = 1}^{\infty} \frac{ε}{2 ^{n}} = ε = 1 (\frac{1}{2} + \frac{1}{4} + \frac{1}{8} + \dots) = ε .$

Each term $ε / 2^{n}$ is the width covering the single fraction $q_{n}$ .
The geometric sum $\frac{1}{2} + \frac{1}{4} + \frac{1}{8} + \dots$ equals exactly $1$ — this is why the whole cover costs a mere $ε$ .

Since $μ^{*} (Q) \leq ε$ for every $ε > 0$ , we get $μ^{*} (Q) = 0$ . See Cantor set for an uncountable set that is also measure zero — showing "big count" and "big measure" are unrelated.

Step 6 — Why $μ^{*}$ isn't good enough yet: the leak

WHY this is fatal for our wishlist. Our wishlist demanded exact additivity: disjoint pieces' sizes must sum exactly to the whole. With only " $\leq$ ", $μ^{*}$ fails wishlist item 3. The culprit sets are pathological monsters like the Vitali set (built with the Axiom of Choice).

PICTURE. A set $A$ chopped by a wall into $A \cap E$ (blue) and $A \cap E^{c}$ (pink). For a nice $E$ the two chalk bars stack up exactly to the $A$ -bar. For a bad $E$ the pink+blue bars overshoot the $A$ -bar — that gap is the leak.

For any $E$ , sub-additivity gives the free direction $μ^{*} (A) \leq μ^{*} (A \cap E) + μ^{*} (A \cap E^{c}) .$

$A \cap E$ = the part of $A$ inside $E$ ; $A \cap E^{c}$ = the part outside ( $E^{c}$ is the complement of $E$ ).
The " $\leq$ " always holds because covers of the two halves glue into a cover of $A$ .

The missing " $\geq$ " is exactly what separates good sets from leaky ones.

Step 7 — Carathéodory's filter: keep only the clean-splitting sets

WHY this exact test. We saw " $\leq$ " is automatic. So demanding equality is really demanding the extra " $\geq$ " — i.e. $E$ never causes a leak against any possible probe $A$ . Sets passing this test can be added up exactly, which is precisely wishlist item 3 restored.

PICTURE. The universe of all sets as a big board; inside it, a fenced region $M$ = the measurable sets (intervals, Borel sets, and more) glowing; the Vitali monster sits outside the fence.

The rewards, from the parent note: the measurable sets $M$ form a Sigma-algebra (closed under complement and countable unions), they include every interval and hence every Borel set, and on $M$ our size machine finally becomes countably additive: $μ (⨆_{n} E_{n}) = \sum_{n} μ (E_{n}) (disjoint E_{n}) .$

This is the payoff that powers the Lebesgue integral and rescues integrals that Riemann integration cannot do.

The one-picture summary

The full pipeline on one board: blob → cover it → squeeze the cost to a floor ( $μ^{*}$ ) → filter out leaky sets (Carathéodory) → get exact additivity ( $μ$ ).

Recall Feynman retelling — say it in plain words

I want to know how much line a blob $A$ takes up. I only trust the length of an interval, which is right end minus left end. So I blanket the blob with a list of tiny intervals — a cover — and add their lengths. That total is never too small, only possibly too big, so I keep shrinking the cover and take the very bottom of all those totals: the infimum. That floor is the outer measure $μ^{*}$ , and it works for every set. I check it on a point (cover with width $2 ε$ , shrink to $0$ → size $0$ ) and on all the fractions (give fraction $n$ a width $ε / 2^{n}$ ; the halving widths add to just $ε$ → size $0$ , even though fractions are everywhere). But $μ^{*}$ has a flaw: splitting a monster set into two disjoint parts can make the parts' sizes overshoot the whole — a leak. So I keep only sets $E$ that split every probe set exactly (Carathéodory). On those good sets, sizes of disjoint pieces add up perfectly. That perfect adding is the whole point — it is what makes the Lebesgue measure and the Lebesgue integral work.

Recall Quick self-test

Why do we cover with intervals rather than measure the blob directly? ::: Intervals are the only things whose length ( $b - a$ ) we already trust; covers over-estimate, never under-estimate.
Why infimum and not minimum? ::: You can usually shrink a cover slightly forever, so no single cheapest cover exists; the infimum is the floor those costs approach.
Why is $μ^{*} (Q) = 0$ despite density? ::: Countability lets us give fraction $n$ a width $ε / 2^{n}$ ; the geometric widths total only $ε \to 0$ . Density $\neq =$ measure.
What does Carathéodory's test add on top of $μ^{*}$ ? ::: The " $\geq$ " direction — no measurement leak — turning sub-additivity into exact countable additivity on $M$ .