Here is an awkward feature of the definition of convergence: to verify a_n \to L you must already know L — every estimate is a comparison against it. But in analysis the limit is very often the thing you are trying to construct. You sum a series, iterate Newton's method, take successive decimal approximations — and you want to certify "this settles down" before you have any name for where it settles.

The Cauchy criterion is exactly that certificate. It never mentions a limit. Instead it asks that the terms huddle together among themselves, ever more tightly: a sequence (a_n) is Cauchy when

Read it quantifier by quantifier, because one detail carries all the weight. For every tolerance \varepsilon there is a cutoff N beyond which — and here is the detail — two indices are quantified, not one. The condition |a_m - a_n| < \varepsilon must hold for every pair m, n \ge N: the millionth term against the billionth, the billionth against the trillionth, any tail term against any other. The whole tail collapses into a single window of width \varepsilon.

Set the two definitions side by side and the shift in viewpoint is plain:

Convergence is an external condition — it measures the sequence against a landmark outside itself. Cauchy is an internal one — the sequence is judged only by its own mutual distances. The headline result of this page is that in \mathbb{R} the two are exactly equivalent — and that this equivalence is not a triviality but a restatement of what makes \mathbb{R} special.

Worked example: a_n = \tfrac1n is Cauchy, straight from the definition

Before any theorems, drive the definition once by hand. Claim: a_n = \tfrac1n is Cauchy — and we will prove it without ever mentioning the limit 0.

Scratch work. Take any m, n \ge N. Both \tfrac1m and \tfrac1n lie in the interval \left(0, \tfrac1N\right], and two numbers in an interval of length \tfrac1N can differ by at most \tfrac1N:

So it suffices to force \tfrac1N < \varepsilon, i.e. N > \tfrac1\varepsilon — and the Archimedean property hands us such an N.

The proof, cleanly. Let \varepsilon > 0. Choose N \in \mathbb{N} with N > \tfrac1\varepsilon. Then for all m, n \ge N,

so (\tfrac1n) is Cauchy. \blacksquare Notice the shape of the argument: bound |a_m - a_n| by something that depends only on N, then make that something small. That template — "trap the whole tail in a shrinking box" — is how nearly every direct Cauchy proof runs, and at no point did the number 0 appear.

Convergent \Rightarrow Cauchy

One direction is easy and holds everywhere. If the terms approach a single limit, they must approach each other: two people walking to the same lamppost end up standing next to one another. The engine is the triangle inequality, run with an \tfrac\varepsilon2 + \tfrac\varepsilon2 split.

Step 1 — convergence, but aimed at \tfrac\varepsilon2. Suppose a_n \to L. Convergence holds for every tolerance, so apply it to the half-tolerance \tfrac\varepsilon2: there is an N with

Step 2 — compare two tail terms via the limit. Take any m, n \ge N. Insert and remove L — a zero — then apply the triangle inequality:

Step 3 — add the two halves. Both m and n exceed N, so each term on the right is below \tfrac\varepsilon2:

That is the Cauchy condition. So every convergent sequence is Cauchy — and notice we never used any special feature of \mathbb{R}; this half is true in \mathbb{Q}, in any metric space, anywhere distances make sense. \blacksquare

Cauchy \Rightarrow convergent (in \mathbb{R})

The reverse is the deep direction, and it is where completeness enters. It is false in \mathbb{Q} (see the hole-detector vignette below) and true in \mathbb{R}. The argument runs in three moves.

Step 1 — a Cauchy sequence is bounded. Apply the Cauchy condition with \varepsilon = 1: past some N, all terms sit within 1 of a_N, hence inside the bounded window (a_N - 1,\ a_N + 1); the finitely many earlier terms are bounded too. So the whole sequence is bounded — the same head/tail argument as for convergent sequences.

Step 2 — extract a convergent subsequence. Being bounded, by Bolzano–Weierstrass the sequence has a subsequence a_{n_k} \to L for some L \in \mathbb{R}. This is precisely the step completeness powers — L exists because \mathbb{R} has no gaps.

Step 3 — promote the subsequence to the whole sequence. A Cauchy sequence with a convergent subsequence converges (to the same limit). Given \varepsilon > 0, choose N so that m, n \ge N \Rightarrow |a_m - a_n| < \tfrac\varepsilon2, and pick a subsequence index n_k \ge N with |a_{n_k} - L| < \tfrac\varepsilon2. Then for all n \ge N,

So a_n \to L. The bunching of the tail, anchored to the subsequence's limit, drags the entire sequence to L. \blacksquare

The theorem's third bullet quietly redefines the star of the previous page. "Every bounded set has a supremum", "every bounded monotone sequence converges", "every Cauchy sequence converges" — over an Archimedean ordered field these are three faces of one property, completeness. The Cauchy formulation is the most portable of the three: it never mentions order, so it survives the journey to \mathbb{C}, to \mathbb{R}^n, to spaces of functions — which is why it is the version analysts carry everywhere.

Watch the tail bunch within \varepsilon

Below is the Cauchy sequence a_n = 2 + \tfrac{\sin n}{n} (it converges to 2, but pretend you don't know that). Drag \varepsilon to set a window (L - \varepsilon,\ L + \varepsilon); the tail eventually enters — and never leaves — that window. Past that cutoff N, every pair of terms differs by less than 2\varepsilon: the Cauchy condition, on screen.

Two things to notice as you play. First, shrink \varepsilon and the cutoff slides right — a tighter huddle takes longer to form — but a cutoff always exists: that is the \forall\varepsilon\,\exists N of the definition. Second, the early terms wander well outside the band and the sequence is Cauchy anyway: like convergence, the Cauchy property is a statement about tails. Changing the first million terms of a sequence changes nothing.