The Epsilon-N Definition of a Limit

The $\varepsilon$–$N$ definition is not a historical curiosity; it is the operational engine behind essentially all quantitative claims in mathematics and its applications. In numerical analysis, convergence proofs for iterative algorithms (Newton's method, gradient descent, fixed-point iteration) are $\varepsilon$–$N$ arguments that translate directly into error bounds and stopping criteria for real software. In probability and statistics, the Law of Large Numbers and consistency proofs for estimators are convergence-in-probability statements built on the same tail-control logic, just replacing $|a_n - L|$ with a probability of deviation. In cryptography and complexity theory, asymptotic statements ("negligible function," "polynomial-time convergence of an algorithm's success probability") are $\varepsilon$–$N$-style bounds on sequences indexed by security parameter. In physics and engineering, convergence of Fourier series, perturbative expansions, and numerical simulations to a true solution is justified by exactly this formalism, ensuring that truncating an infinite process after finitely many steps still gives answers within any desired tolerance. Within pure mathematics, this definition is the seed from which grow Uniqueness of Limits and Boundedness, the Algebraic Limit Theorems, the definition of continuity via sequences, and ultimately the entire $\varepsilon$–$\delta$ machinery of Limits of Functions. Understanding it cold — being able to produce an $N$ from an $\varepsilon$ on demand — is the single clearest signal that a student has crossed from computational calculus into rigorous analysis.

Ask the tutor: Can you show me an $\varepsilon$–$N$ proof for a sequence involving a square root, like $a_n = \sqrt{n+1} - \sqrt{n} \to 0$?

Think of $\varepsilon$ as a tolerance you are handed by an adversary who wants to make your life difficult. They pick $\varepsilon$ as small as they like — say $\varepsilon = 0.001$, or $\varepsilon = 10^{-100}$. Your job is to respond with a threshold index $N$, and you win the round if every term $a_n$ with $n \geq N$ lands within distance $\varepsilon$ of $L$. Convergence means you can always win, no matter how small an $\varepsilon$ they throw at you. Crucially, $N$ is allowed to depend on $\varepsilon$: smaller tolerances typically force you to wait longer (larger $N$) before the sequence settles down. The definition says nothing about how fast $N$ must grow, or about how the early terms $a_1, \dots, a_{N-1}$ behave — they can do anything, even leave the $\varepsilon$-band and come back. Only the tail of the sequence, from $N$ onward, is constrained. Geometrically, fix a horizontal band of height $2\varepsilon$ centered at $L$ on the vertical axis. Convergence to $L$ means that for every such band, however thin, the graph of $(n, a_n)$ eventually enters the band and never leaves it again.

The following theorem records the basic logical structure of the definition — that it is equivalent to a statement purely about the tail of the sequence — and is stated here explicitly because every subsequent proof in analysis relies on manipulating this tail freely.

This theorem is what licenses the informal paraphrase "converges to $L$" as "eventually all terms are as close to $L$ as we like, with only finitely many exceptions for each tolerance" — and it shows this paraphrase is not an approximation of the definition but logically equivalent to it.

Example 1 (a linear sequence with a shift). Prove directly from the $\varepsilon$–$N$ definition that

Scratch work. We need, given $\varepsilon > 0$, to find $N$ so that $n \geq N$ forces $|a_n - 3| < \varepsilon$. Compute:

So $|a_n - 3| = \dfrac{1}{n+2}$, which is positive for all $n \in \mathbb{N}$. We want $\dfrac{1}{n+2} < \varepsilon$, i.e. $n+2 > \dfrac1\varepsilon$, i.e. $n > \dfrac1\varepsilon - 2$. So any integer $N$ with $N > \dfrac1\varepsilon - 2$ will work; by the Archimedean property such $N$ exists. Formal proof. Let $\varepsilon > 0$ be given. By the Archimedean Property of $\mathbb{R}$, there exists $N \in \mathbb{N}$ with $N > \dfrac{1}{\varepsilon} - 2$. Suppose $n \geq N$. Then

and since $n+2 > 0$ and $\varepsilon > 0$, this inequality may be inverted to give

Hence

Since $\varepsilon>0$ was arbitrary and $N$ was produced depending only on $\varepsilon$, we conclude $a_n \to 3$. $\blacksquare$ Note the asymmetry deliberately built in here: the sequence is not centered at a nice number like $0$, and the algebra $a_n - 3 = \tfrac{1}{n+2}$ (not $\tfrac1n$) forces genuine care in solving the inequality for $n$, rather than allowing a lucky shortcut.

Example 2 (a sequence with alternating sign, testing absolute values carefully). Prove directly that

Scratch work. We must bound $|b_n - 0| = |b_n|$. By the triangle inequality,

This is messier than a single power of $n$, so we simplify further using $n+4 \le n+4n = 5n$ for all $n\ge 1$ (since $4 \le 4n$ when $n \ge 1$). Hence for $n \geq 1$,

So it suffices to force $\dfrac5n < \varepsilon$, i.e. $n > \dfrac{5}{\varepsilon}$. Formal proof. Let $\varepsilon > 0$. By the Archimedean Property, choose $N \in \mathbb{N}$ with $N > \dfrac{5}{\varepsilon}$. Let $n \geq N$ (so in particular $n \geq 1$). By the triangle inequality,

Since $n \geq 1$ we have $4 \leq 4n$, so $n + 4 \leq n + 4n = 5n$, giving

Since $n \geq N > \dfrac{5}{\varepsilon} > 0$, we have $n > \dfrac5\varepsilon$, and inverting (both sides positive) gives $\dfrac{5}{n} < \varepsilon$. Chaining these inequalities,

As $\varepsilon > 0$ was arbitrary, $b_n \to 0$. $\blacksquare$ This example illustrates the standard technique of over-estimating the error term with a cruder but easier-to-invert bound (here $\tfrac{n+4}{n^2} \le \tfrac{5}{n}$), rather than solving the exact inequality $\tfrac{n+4}{n^2}<\varepsilon$, which would require the quadratic formula. This trade-off — a looser bound purchased for algebraic simplicity — is used throughout analysis.