This page is where analysis begins. In calculus you computed limits; from now on you will prove them. That changes everything. Phrases like "gets closer and closer" or "infinitely close" were good enough to launch the idea of a limit, but they are not statements you can verify or refute — they are vibes. The ε–N definition below is the moment vagueness becomes a provable claim: a sentence built only from inequalities and quantifiers, so that "the sequence converges" is either demonstrably true or demonstrably false, with nothing infinitely small anywhere in sight. Everything you will meet later — continuity, derivatives, integrals, sums of series — is this one definition wearing different costumes. Learn to read it, wield it and negate it here, and the rest of analysis opens up. (For a gentler first meeting with the same idea, see the series-track page; this is the rigorous version.)

A sequence (a_n) = a_1, a_2, a_3, \dots is an infinite list of reals — formally, a function \mathbb{N} \to \mathbb{R}. We say (a_n) converges to L, written a_n \to L, when the terms get — and stay — arbitrarily close to L. Formally:

Read it as a game against an adversary: they name a tolerance \varepsilon (how close is "close enough"); you must produce a cutoff N such that every term from a_N onward sits within \varepsilon of L. If you can always win, no matter how tiny their \varepsilon, then a_n \to L. A sequence is bounded if all its terms fit in some band: |a_n| \le M for a single M and all n.

Dissecting the sentence: the order of quantifiers is the content

The definition is three quantifiers deep, and the meaning lives entirely in their order. Read it left to right as moves in the game:

• \forall\, \varepsilon > 0 — the adversary moves first. They may pick 0.1, or 10^{-12}, or something crueller. You must handle every choice.
• \exists\, N — you move second, and because you move second, your N is allowed to depend on \varepsilon. Smaller tolerance, bigger cutoff — that trade-off is the entire substance of a convergence proof. Think of a winning strategy as a function \varepsilon \mapsto N(\varepsilon).
• \forall\, n \ge N — beyond the cutoff there are no exceptions. Not "most terms", not "infinitely many terms": all of them, forever after, trapped in the band (L - \varepsilon,\ L + \varepsilon).

Notice what the definition does not care about: the first N - 1 terms. They may be enormous, negative, or ridiculous — convergence is a property of the tail. And notice what it quietly replaced: no term ever needs to equal L, and nothing is "infinitely close" to anything. Every clause is a finite, checkable inequality.

Worked proof 1: \tfrac1n \to 0, with all the bookkeeping

Every ε–N proof has two phases. Scratch work (private): start from the target inequality and solve for n. Write-up (public): present the moves in the order the quantifiers demand — ε first, then your N, then the verification.

Scratch work. We want |\tfrac1n - 0| = \tfrac1n < \varepsilon. Rearranging: n > \tfrac1\varepsilon. So any cutoff past 1/\varepsilon will do. But N must be a natural number, and 1/\varepsilon almost never is — hence the ceiling function: N = \lceil 1/\varepsilon \rceil + 1 is a natural number strictly bigger than 1/\varepsilon. (That such an integer exists at all is the Archimedean property of \mathbb{R} — even this humble step leans on the structure of the reals.)

Write-up. Let \varepsilon > 0 be given. Set

Then for every n \ge N,

Try it with numbers: if the adversary plays \varepsilon = 0.01, your recipe answers N = 101, and indeed \tfrac1{101}, \tfrac1{102}, \dots all lie below 0.01. Play \varepsilon = 10^{-6} and it answers N = 10^6 + 1. One formula, every tolerance beaten.

Worked proof 2: \tfrac{2n+1}{n+3} \to 2, and the art of the lazy bound

Scratch work. First simplify the distance to the limit — this algebraic step is where most of the real work happens:

We could solve \tfrac{5}{n+3} < \varepsilon exactly, but here is the professional habit: bound it by something simpler. Since n + 3 > n, we have \tfrac{5}{n+3} < \tfrac{5}{n}, and \tfrac5n < \varepsilon as soon as n > 5/\varepsilon. The definition asks for some N that works — never the smallest one — so a generous, easy bound is just as valid as a sharp, painful one.

Write-up. Let \varepsilon > 0. Set N = \lceil 5/\varepsilon \rceil + 1, so N > 5/\varepsilon. Then for all n \ge N,

The template generalises: compute or bound |a_n - L| by a simple decreasing expression in n, then choose N to crush that expression below \varepsilon. That two-step rhythm carries you through most of a first analysis course.

Worked proof 3: limits are unique — the ε/2 trick

We have been saying "the limit", which silently assumes a sequence cannot converge to two different values. That needs proof — and the proof introduces the single most-reused argument pattern in analysis: split your tolerance in half and spend the halves separately.

Claim. If a_n \to L and a_n \to M, then L = M.

Proof. Let \varepsilon > 0 be arbitrary. Since a_n \to L, there is an N_1 with |a_n - L| < \varepsilon/2 for all n \ge N_1. Since a_n \to M, there is an N_2 with |a_n - M| < \varepsilon/2 for all n \ge N_2. Pick any single n \ge \max\{N_1, N_2\} — one term inside both tails — and route from L to M through it with the triangle inequality:

So |L - M| < \varepsilon for every positive \varepsilon. The only non-negative real smaller than every positive number is 0, hence |L - M| = 0 and L = M. \blacksquare

Two things to savour. First, the halves: we wanted \varepsilon at the end, so we ordered \varepsilon/2 from each hypothesis — the definition lets us demand any tolerance, so we demand exactly what the triangle inequality will spend. Second, the closing move "less than every \varepsilon > 0, hence zero" is itself a classic. (Many books run the same proof as a contradiction: assume L \ne M, play \varepsilon = |L - M|/2, and watch the two ε-bands — now disjoint — both claim the entire tail. Same engine, different bodywork.)

Saying "no": negating the definition properly

To prove a sequence diverges you must know precisely what failure looks like — and the negation of a quantified sentence is mechanical: push the \neg through, flipping each quantifier (\forall \leftrightarrow \exists) and negating the inner inequality at the end. "a_n \not\to L" becomes:

In game terms the roles reverse: now you commit to one fixed escape distance \varepsilon up front, and then defeat every proposed cutoff N by exhibiting a later term that still strays. Note what it is not: not "all terms are far from L", and not "some term is far from L". It is: infinitely often, some term escapes by a fixed amount. ("Diverges" outright means this holds for every candidate L.)

Worked example. a_n = (-1)^n converges to no L whatsoever. Fix any L and play \varepsilon = 1. The points +1 and -1 are distance 2 apart, so by the triangle inequality L cannot be within 1 of both: |1 - L| + |{-1} - L| \ge 2 forces |1 - L| \ge 1 or |{-1} - L| \ge 1. Now given any N, both an even and an odd index lie beyond it, so some n \ge N has |a_n - L| \ge 1. Every cutoff is defeated; the sequence diverges. \blacksquare

Convergent \Rightarrow bounded

Our first theorem links the two notions: if a sequence converges, it cannot run off to infinity. The proof is a clean ε–N argument and a model of the genre — a tail controlled by the limit, a head controlled by finiteness.

Step 1 — pin the tail with one specific \varepsilon. Convergence holds for every \varepsilon, so we are free to pick a convenient one. Take \varepsilon = 1. Then there is an N with

Step 2 — bound the tail. For those n \ge N, the triangle inequality turns closeness-to-L into a bound on |a_n| itself:

So every term from a_N onward is bounded by 1 + |L|. The infinite tail is under control.

Step 3 — handle the finite head. Only finitely many terms remain: a_1, a_2, \dots, a_{N-1}. A finite list of numbers always has a largest absolute value — call it

Step 4 — combine head and tail. Take the larger of the two bounds:

Every term is now covered — the head by K, the tail by 1 + |L| — so |a_n| \le M for all n. The sequence is bounded. \blacksquare

Notice the strategy echoing the "Watch out!" box: the tail is where convergence lives, and the head is finite, so it can always be mopped up with a \max. The converse fails: a_n = (-1)^n is bounded (|a_n| = 1) but, as we proved above, converges to nothing. So bounded does not force convergence — but it does force the next, deeper, fact.

Bolzano–Weierstrass

A subsequence picks out infinitely many terms in order: a_{n_1}, a_{n_2}, \dots with n_1 < n_2 < \cdots. The Bolzano–Weierstrass theorem says a bounded sequence, even one that never converges, must contain a convergent subsequence. Here is the bisection sketch.

Step 1 — trap the sequence in an interval. Boundedness puts every term in a closed interval I_0 = [-M, M].

Step 2 — bisect, keeping the busy half. Cut I_0 in half. At least one half contains infinitely many terms of the sequence (two halves cannot both hold only finitely many, or the whole would be finite). Call that half I_1; its length is M.

Step 3 — repeat forever. Bisect I_1, again keep a half with infinitely many terms, getting I_2 of length M/2; and so on. This yields nested closed intervals

Step 4 — completeness delivers the point. By the completeness of \mathbb{R} (nested closed intervals whose lengths shrink to 0 meet in exactly one point), there is a unique L \in \bigcap_k I_k.

Step 5 — extract the subsequence. From each I_k pick a term a_{n_k} with n_1 < n_2 < \cdots (possible since each I_k holds infinitely many). Since a_{n_k} and L both lie in I_k, whose length \to 0, we get |a_{n_k} - L| \to 0. So a_{n_k} \to L: a convergent subsequence. \blacksquare

A bounded sequence and a convergent subsequence

The sequence below is our star witness a_n = (-1)^n\big(1 + \tfrac1n\big): it oscillates and never converges as a whole (run the negation argument with \varepsilon = 1 to check), yet it is bounded by 2 — so Bolzano–Weierstrass promises a convergent subsequence hiding inside it. Toggle the highlight to pick out one such subsequence — the even-indexed terms a_2, a_4, a_6, \dots drifting down to +1 — from the full bounded sequence. Can you write the ε–N proof that this subsequence really does converge to 1? (Its distance from 1 is exactly \tfrac1n, so proof 1 above hands you the recipe.)

See it explained