The Epsilon Characterization of Suprema

The epsilon characterization of the supremum is not a museum piece; it is the load-bearing technique behind much of analysis and its applications.

• Limits and convergence proofs. Whenever a proof shows $\lim_{n\to\infty} a_n = L$ by producing, for each $\varepsilon>0$, an index $N$ with $|a_n - L|<\varepsilon$ for $n\ge N$, the underlying logical shape is identical to the epsilon criterion here; monotone bounded sequences converge to their supremum precisely by this argument.
• Riemann and Lebesgue integration. The Riemann integral is defined via suprema of lower sums and infima of upper sums; proving two such quantities are equal, or computing them, is done by producing partitions whose sums fall within $\varepsilon$ of the claimed value — a direct instance of the epsilon criterion.
• Optimization and variational problems. In showing a claimed optimal value is truly the extremum (e.g., in convex optimization, calculus of variations, or numerical analysis), one certifies optimality by exhibiting feasible points or sequences approaching the bound arbitrarily closely.
• Metric space and functional analysis. The notion of the distance from a point to a set, $d(x,S) = \inf_{s\in S} d(x,s)$, and best-approximation problems in normed spaces, are analyzed via the epsilon criterion to show a candidate distance is genuinely minimal.
• Computer science and numerical methods. Algorithms that compute approximate suprema/infima (e.g., in linear programming relaxations, machine learning loss-bound arguments, or root-finding via bisection) implicitly rely on the same "arbitrarily close" certification to argue correctness and convergence rates.
• Probability and statistics. Essential suprema of random variables and tight concentration bounds are established by producing events of small but positive probability arbitrarily close to the claimed bound, mirroring this exact structure.

Ask the tutor: Can you show me why the Archimedean Property itself needs this exact epsilon argument to prove?

Picture the real line with the set $S$ scattered along it, bounded above, and a candidate ceiling $M$ sitting above all of it. The condition "$M$ is an upper bound" says nothing can poke out above $M$. The subtlety is how tight a ceiling $M$ is. If you could slide the ceiling down a little — say to $M - \delta$ for some fixed $\delta > 0$ — and still have every point of $S$ underneath, then $M$ was not the least upper bound; $M - \delta$ (or something between) works too. The epsilon criterion is precisely the denial of this possibility, quantified over all conceivable gap sizes at once. It says: for every proposed gap $\varepsilon>0$, however small, the set $S$ has a point living inside the strip $(M-\varepsilon, M]$. If even one $\varepsilon_0>0$ failed — meaning no point of $S$ exceeds $M - \varepsilon_0$ — then $M-\varepsilon_0$ itself would be a smaller upper bound, contradicting leastness. So the criterion is not an approximation of leastness; it is leastness, restated so you can act on it. The picture to hold in mind is a sequence of shrinking windows just below $M$, each of which must be hit by $S$.

Example 1 — an infinite, non-symmetric set defined by a rational function. Let

We claim $\sup S = 3$. Step 1: $3$ is an upper bound. For $n \in \mathbb{N}$,

which is always true. Since $n+2>0$, the equivalence (multiplying both sides by $n+2$) is valid without flipping the inequality. So every element of $S$ is $\le 3$; $3$ is an upper bound. Step 2: the epsilon condition. Let $\varepsilon > 0$ be given. We must find $n \in \mathbb{N}$ with

Compute the gap directly:

So the condition $\frac{3n-1}{n+2} > 3-\varepsilon$ is equivalent to $\frac{7}{n+2} < \varepsilon$, i.e. $n+2 > \frac{7}{\varepsilon}$, i.e.

By the Archimedean Property of $\mathbb{R}$ (which itself is proved using this same epsilon-of-supremum machinery, and which we may cite here as a standard consequence available once suprema are understood), there exists a natural number $n$ exceeding any given real number, so in particular exceeding $\frac{7}{\varepsilon}-2$. Choose such an $n$. Then $\frac{3n-1}{n+2} > 3-\varepsilon$, as required. Since $\varepsilon>0$ was arbitrary, the epsilon condition holds, and by the theorem, $\sup S = 3$. Note that $3 \notin S$ (no finite $n$ makes the equality $\frac{7}{n+2}=0$ hold), so this genuinely uses the epsilon criterion rather than simply exhibiting a maximum element.

Example 2 — a set defined by a bound on a two-variable condition (bounded region). Let

We claim $\sup S = \sqrt5$. Step 1: $\sqrt5$ is an upper bound. If $x \in S$ then $x>0$ and $x^2 < 5 = (\sqrt5)^2$. Since both $x$ and $\sqrt5$ are positive and the squaring function is strictly increasing on $[0,\infty)$ (if $0 \le a < b$ then $a^2 < b^2$, and conversely $a^2<b^2$ with $a,b\ge0$ forces $a<b$ — this is a routine consequence of factoring $b^2-a^2=(b-a)(b+a)$ with $b+a>0$ when not both zero), $x^2 < (\sqrt5)^2$ with $x,\sqrt5 > 0$ gives $x < \sqrt5$. So $\sqrt5$ is an upper bound for $S$. Step 2: the epsilon condition. Let $\varepsilon>0$. We may assume $\varepsilon < \sqrt5$ (if the condition holds for small $\varepsilon$ it automatically holds for larger $\varepsilon$, since the required point only needs to clear a lower bar $\sqrt5-\varepsilon$; so WLOG shrink $\varepsilon$). We must produce a rational $x>0$ with $x^2<5$ and $x > \sqrt5 - \varepsilon$. Consider $y := \sqrt5 - \tfrac{\varepsilon}{2}$, a real number with $0 < y < \sqrt5$. By the Density of the Rationals in the Reals (again, a theorem whose own proof uses this same epsilon technique, cited here only as a downstream consequence, not circularly assumed for this theorem), there is a rational number $x$ with

Then $x > 0$ (since $y>0$ forces $x>y>0$), and $x < \sqrt5$ gives $x^2 < 5$ by the same monotonicity of squaring on positives used above. So $x \in S$. Moreover $x > y = \sqrt5 - \varepsilon/2 > \sqrt5 - \varepsilon$. This $x$ satisfies the epsilon condition for this $\varepsilon$. Since $\varepsilon>0$ was arbitrary, the epsilon criterion holds, and by the theorem $\sup S = \sqrt5$. This example is important precisely because $\sqrt5 \notin S$ — the supremum is irrational while $S$ consists only of rationals, so there is no "obvious largest element" to point to; the epsilon criterion is the only available route.