Here is a puzzle that sounds like it should be trivial. Take the set of all numbers strictly between 0 and 1. Which of them is the biggest? Try 0.9 — but 0.99 is bigger. Try 0.999 — beaten again. Whatever element you nominate, the number halfway between it and 1 is still in the set and still larger. The set has no biggest member. And yet it obviously has a perfect ceiling: the number 1 sits exactly on top of it, so close that the set touches its hair, even though 1 itself stays outside.

That ceiling deserves a name and a definition, because "the best possible bound" turns out to be one of the most important numbers in analysis. It is the least upper bound, or supremum — and the promise that it always exists is precisely what separates \mathbb{R} from \mathbb{Q}. We saw that the rationals have gaps: a bounded set of rationals can crowd up against a boundary that simply isn't there in \mathbb{Q}. The completeness axiom is the precise promise that \mathbb{R} never does this — every crowd has a ceiling, and the ceiling is a real number.

First, the vocabulary. Fix a nonempty set S \subseteq \mathbb{R}. A number u is an upper bound of S if it is at least as large as every element:

A set may have many upper bounds — if u works, so does anything bigger, so the upper bounds form an infinite ray stretching right. The interesting question is where that ray starts. The supremum \sup S is the smallest of all upper bounds — the least upper bound, the left edge of the ray, the ceiling lowered until it just touches. The mirror notion, the greatest lower bound, is the infimum \inf S: the largest of all lower bounds, the floor raised until it just grazes the set from below.

The two-part test for a supremum

Saying "s is the least upper bound" packs two separate demands: s must bound the set, and nothing smaller may. Unfolding them gives the working characterisation every proof in analysis uses. We claim: s = \sup S exactly when both

Condition (ii) is the useful, quantified way of saying "nothing smaller works": step even a hair's breadth \varepsilon below s, and some element of S pokes above you — so you are no longer an upper bound. Let's see why the two conditions are exactly equivalent to being the supremum.

Step 1 — why (i): s must be an upper bound. The supremum is by definition an upper bound (the least one), so it certainly bounds S from above. That is exactly (i).

Step 2 — why (ii): nothing smaller can bound S. Suppose, for some \varepsilon > 0, there were no element of S exceeding s - \varepsilon. Then every x \in S satisfies x \le s - \varepsilon, making s - \varepsilon an upper bound smaller than s. That contradicts s being the least upper bound. So such an \varepsilon cannot exist: for every \varepsilon > 0 some x \in S pokes above s - \varepsilon. That is (ii).

Step 3 — the converse. Conversely, if s satisfies (i) and (ii), then (i) makes it an upper bound, and (ii) says no number below it is an upper bound (anything s - \varepsilon < s is beaten by some element of S). So s is the least upper bound — s = \sup S. The two conditions together are equivalent to being the supremum, and from now on "check (i) and (ii)" is how we prove a number is a sup.

Three worked examples

Example 1: \sup\,(0, 1) = 1 — and 1 is not in the set

Back to the opening puzzle: S = (0, 1) = \{x : 0 < x < 1\}. We run the two-part test on the candidate s = 1.

Check (i): every x \in S satisfies x < 1 \le 1, so 1 is an upper bound. Immediate.

Check (ii): given \varepsilon > 0, we must find an element of (0,1) exceeding 1 - \varepsilon. If \varepsilon \ge 1, any element (say \tfrac12) works, since 1 - \varepsilon \le 0. If \varepsilon < 1, take the midpoint x = 1 - \tfrac{\varepsilon}{2}: it lies in (0,1), and x > 1 - \varepsilon. Done.

So \sup\,(0,1) = 1. But — and this is the crucial observation — 1 \notin (0,1). The set has a supremum and no maximum. A maximum is an element of the set that beats all the others; it exists precisely when \sup S \in S, and then \max S = \sup S. The supremum is the more robust notion: closed interval [0,1] and open interval (0,1) have the same supremum, but only the first has a maximum.

Example 2: \sup\{1 - \tfrac{1}{n}\} = 1 — the ε-argument in action

Let S = \{\, 1 - \tfrac1n : n = 1, 2, 3, \dots \,\} = \{0,\ \tfrac12,\ \tfrac23,\ \tfrac34,\ \dots\}. The terms climb toward 1 without ever reaching it. We verify \sup S = 1 against the test.

Check (i): for every n \ge 1 we have \tfrac1n > 0, so 1 - \tfrac1n < 1. Thus 1 is an upper bound.

Check (ii): given any \varepsilon > 0, we must find n with 1 - \tfrac1n > 1 - \varepsilon, i.e. \tfrac1n < \varepsilon. By the Archimedean property (proved below — from completeness itself!) there is an integer n > \tfrac1\varepsilon, and for that n we get \tfrac1n < \varepsilon. So some element of S lands within \varepsilon of 1.

Both halves hold, so \sup S = 1. Note again 1 \notin S: a supremum need not be attained. (Here \inf S = 0, and that one is attained, at n = 1 — so S has a minimum but no maximum.)

Example 3: \sup\{q \in \mathbb{Q} : q^2 < 2\} = \sqrt{2} — the poster child of completeness

Now the example the whole subject is built around. Let S = \{q \in \mathbb{Q} : q^2 < 2\} — all rationals whose square stays below 2. This set is nonempty (1 \in S) and bounded above (by 2, say, since q > 2 forces q^2 > 4). Working inside \mathbb{R}, its supremum is \sqrt{2}: every element satisfies q < \sqrt2 (that's (i)), and because the rationals are dense, every window (\sqrt2 - \varepsilon,\ \sqrt2) contains a rational, whose square is then below 2 (that's (ii)).

Here is the punchline. S is a set of rationals, described entirely in rational terms — yet its supremum, \sqrt2, is irrational. Inside \mathbb{Q} this set has upper bounds galore but no least one: for any rational upper bound you offer, a slightly smaller rational still bounds S. The candidates for "least upper bound" spiral down toward a point that \mathbb{Q} simply doesn't contain. That failure is exactly the "gap" in the rationals — and the completeness axiom is the declaration that \mathbb{R} has plugged every such gap.

The axiom — and what it buys you

Everything above defined the supremum and computed a few. The deep statement is that in \mathbb{R} you never have to wonder whether one exists:

Note what the axiom does not say. It does not tell you what \sup S equals, and it does not put it inside S. It is a pure existence statement — and that is precisely its power. Analysis is full of numbers you can describe but not compute ("the largest value f attains", "the first time the curve crosses zero"). Completeness is the licence to say let s be that number and get on with the proof.

The axiom at work: monotone sequences converge

Here is a first taste of the licence in use. Let (a_n) be an increasing sequence that is bounded above — say a_1 \le a_2 \le a_3 \le \cdots \le M. Intuitively it must "pile up" against something. Completeness makes the something a number.

Sketch. The set of terms \{a_n : n \ge 1\} is nonempty and bounded above, so by the axiom it has a supremum s = \sup_n a_n. Given \varepsilon > 0, condition (ii) hands us some term a_N > s - \varepsilon. But the sequence is increasing, so every later term satisfies s - \varepsilon < a_N \le a_n \le s for all n \ge N. That is exactly the definition of a_n \to s: the sequence converges, and it converges to the supremum of its terms. One application of the axiom, one use of the two-part test — and out falls the monotone convergence theorem, the workhorse behind infinite series, decimal expansions, and Newton's method alike.

See the least upper bound

Below, the dotted points form a set S living on the number line, crowding rightward toward 4. Every point from \sup S rightward is an upper bound (the solid bar); the supremum is the left edge of that bar — the smallest number still bounding S, the place where the ray of upper bounds begins.

Now drag the \varepsilon slider and watch the window (\sup S - \varepsilon,\ \sup S]. However thin you make it, it always catches part of S — the set refuses to leave a gap below its supremum. That is condition (ii) made visible: step any distance \varepsilon below the sup and you land strictly inside the crowd, no longer above it. Compare with a point like 5: it bounds S too, but the window (5-\varepsilon, 5] would be empty for small \varepsilon — an upper bound with daylight beneath it is not the least one.