Cauchy Sequences and Completeness

Suppose a sequence of rational numbers gets tighter together, term after term. Yet you cannot name the limit. It happens to be irrational. Does the sequence still "converge"? This question exposes a subtlety hiding inside the very idea of a limit. The definition of convergence requires knowing the limit in advance. But often we only know that terms are clustering. Cauchy sequences solve this problem. They define a notion of "clustering" that refers only to the terms themselves. It never refers to a limit. The remarkable fact — and the centerpiece of this article — is this. In the real numbers, this internal clustering condition is equivalent to convergence. This equivalence is not free. It is precisely where the completeness of $\mathbb{R}$ enters analysis. This is what distinguishes $\mathbb{R}$ from $\mathbb{Q}$ . A sequence of rationals can be Cauchy without converging to a rational number. But every Cauchy sequence of reals converges to a real number. This single theorem — the Cauchy Criterion — lets us prove a sequence converges. We never need to exhibit its limit. The idea reaches far beyond a technical curiosity. It underlies the construction of $\mathbb{R}$ itself from $\mathbb{Q}$ (as equivalence classes of Cauchy sequences). It is the standard tool for proving convergence of series and iterative algorithms. It also generalizes to the notion of a complete metric space. This concept is indispensable in functional analysis, differential equations, and numerical methods. This article gives the precise definition of a Cauchy sequence. It proves the Cauchy Criterion for $\mathbb{R}$ in full rigor. This uses the Bolzano-Weierstrass Theorem. It also shows, with an explicit counterexample, why completeness of the ambient space cannot be dropped. It previews how this machinery feeds directly into the Cauchy Criterion for Series. It also connects to the theory of Limit Superior and Limit Inferior.

Key takeaways

A sequence is Cauchy if its terms eventually become arbitrarily close to each other, not just to some fixed limit.
Every convergent sequence (in any metric space) is automatically Cauchy — this direction is easy and needs no completeness.
In $\mathbb{R}$ , the converse also holds: every Cauchy sequence converges. This is the Cauchy Criterion, and it fails in $\mathbb{Q}$ .
The proof of the converse combines boundedness of Cauchy sequences with Bolzano-Weierstrass Theorem to extract a convergent subsequence, then shows the whole sequence must converge to the same limit.
"Cauchy" is a completeness-free, self-contained test for convergence — useful exactly when you cannot guess the limit.
The definition generalizes verbatim to any metric space, where completeness becomes the defining property of spaces (like $\mathbb{R}^n$ , or Banach spaces) where the equivalence still holds.
Key idea: Convergence and the Cauchy property coincide in $\mathbb{R}$ , and this equivalence is what "completeness" means.

Where this lives today

The Cauchy criterion is the working definition of convergence whenever an explicit limit is unavailable. In numerical analysis, iterative root-finding and fixed-point methods (Newton's method, as in Example 1) are proven to converge by showing the iterates form a Cauchy sequence, which is often far easier than guessing the exact limit. In the theory of series, the Cauchy Criterion for partial sums underlies tests like the Cauchy Criterion for Series and the ratio/root tests. In functional analysis, Banach spaces are defined exactly as normed vector spaces that are complete with respect to their norm — Cauchy sequences of functions converge to functions in the same space — which is the bedrock of the existence theory for differential equations (Picard iteration) and of Fourier analysis. In probability theory, "Cauchy in probability" or "Cauchy almost surely" versions of this concept underlie almost-sure convergence theorems for martingales and random series. In computer science and cryptography, Cauchy-sequence-style arguments justify convergence of iterative algorithms (gradient descent, fixed-point iteration in cryptographic protocols) without needing closed-form limits, and $p$ -adic number systems in cryptography are literally built, like $\mathbb{R}$ itself, as completions of $\mathbb{Q}$ with respect to a different (non-Archimedean) metric — the same completion machinery pioneered by the real-number construction from Cauchy sequences.

Ask the tutor: Can you show me how the Cauchy Criterion is used to prove the Ratio Test for series convergence?

(a_n)_{n=1}^\infty

Intuition

Ordinary convergence $a_n \to L$ says: pick any tolerance $\varepsilon$ , and eventually every term is within $\varepsilon$ of the target $L$ . This definition needs $L$ as an input. The Cauchy condition instead compares terms to each other: eventually, any two terms (however far apart in index) are within $\varepsilon$ of each other. It's a statement purely about the internal geometry of the sequence, with no reference to an external point. Picture the sequence as a cloud of points on the number line, indexed by time. Convergence says the cloud collapses onto a single point $L$ . The Cauchy condition says the cloud's diameter shrinks to zero as time goes on — the points bunch together, but a priori we don't know they're bunching around any particular real number. The theorem below says: on the real line, a shrinking-diameter cloud of points must be collapsing onto some point, and that point is a real number. This fails on $\mathbb{Q}$ because the "hole" the cloud collapses toward might not be a rational number — the cloud shrinks in diameter, but there's no rational point at its center.

Figure

Two sequences: one collapses onto a real limit (top), one shrinks in diameter but the "hole" it targets is irrational (bottom), illustrating why completeness of the ambient space matters

The main results

(a_n)

\varepsilon = 1

(a_n)

\varepsilon > 0

\mathbb{R}

\Rightarrow

(a_n)

(a_n)

Worked examples

Example 1 (proving convergence without knowing the limit). Define $(a_n)$ recursively by

a_1 = 1, \qquad a_{n+1} = \frac{1}{2}\left(a_n + \frac{3}{a_n}\right) \quad (n \ge 1).

This is Newton's method approximating $\sqrt{3}$ . We show $(a_n)$ is Cauchy directly from the recursion, without assuming we know $\sqrt 3$ is the limit. First, $a_1 = 1 > 0$ , and if $a_n > 0$ then $a_{n+1} = \frac12(a_n + 3/a_n) > 0$ ; by induction $a_n > 0$ for all $n$ . Next, for $n \ge 1$ ,

a_{n+1}^2 - 3 = \frac{1}{4}\left(a_n + \frac{3}{a_n}\right)^2 - 3 = \frac{1}{4a_n^2}\left(a_n^2-3\right)^2 \ge 0,

so $a_{n+1}^2 \ge 3$ for every $n \ge 1$ , i.e. $a_n \ge \sqrt3$ for all $n \ge 2$ . Now for $n \ge 2$ ,

a_{n+1} - a_n = \frac{1}{2}\left(\frac{3}{a_n} - a_n\right) = \frac{3 - a_n^2}{2a_n} \le 0,

using $a_n^2 \ge 3$ . So $(a_n)_{n\ge 2}$ is monotone decreasing and bounded below by $\sqrt3 > 0$ ; in particular it's bounded. We can go further and show it is Cauchy using a quantitative estimate. From $a_{n+1} - \sqrt3 = \frac{(a_n-\sqrt3)^2}{2a_n}$ (a direct algebraic identity: expand $\frac{1}{2}(a_n+3/a_n) - \sqrt3 = \frac{a_n^2 - 2\sqrt3 a_n+3}{2a_n} = \frac{(a_n-\sqrt3)^2}{2a_n}$ ), and since $a_n \ge \sqrt3$ for $n\ge2$ ,

0 \le a_{n+1} - \sqrt3 \le \frac{(a_n - \sqrt3)^2}{2\sqrt3}.

With $a_2 = \frac12(1+3) = 2$ , so $a_2 - \sqrt3 \le 2 - 1 = 1$ (since $\sqrt3 > 1$ ), a quick induction shows $a_n - \sqrt 3 \to 0$ doubly-exponentially fast, so certainly $a_n - a_m \to 0$ as $m,n\to\infty$ : the sequence is Cauchy. By the Cauchy Criterion, $(a_n)$ converges to some real $L$ ; passing to the limit in the recursion $L = \frac12(L + 3/L)$ gives $L^2 = 3$ , and since $a_n \ge \sqrt3>0$ , $L = \sqrt3$ . The point of this example: we established convergence (via the Cauchy property, monotonicity and boundedness together) before pinning down that the limit is $\sqrt3$ — exactly the situation where the Cauchy criterion is essential, since $\sqrt3$ was not given to us in advance as a number to test against.

Example 2 (a divergence proof via the contrapositive). Let $a_n = \sum_{k=1}^n \frac{1}{k} = 1 + \frac12 + \cdots + \frac1n$ (the harmonic partial sums). We show $(a_n)$ is not Cauchy, hence diverges, using the corollary above. Take $\varepsilon_0 = \tfrac12$ . Given any $N$ , set $m = N$ and $n = 2N$ . Then

a_n - a_m = \sum_{k=N+1}^{2N} \frac{1}{k}.

This sum has exactly $N$ terms, each term is $\ge \frac{1}{2N}$ (since $k \le 2N$ ), so

a_{2N} - a_N \ge N \cdot \frac{1}{2N} = \frac12 = \varepsilon_0.

Thus for every $N$ we exhibited $m=N, n=2N \ge N$ with $|a_n - a_m| \ge \varepsilon_0 = \tfrac12$ . By the contrapositive corollary, $(a_n)$ is not Cauchy, and therefore (by the Cauchy Criterion) $(a_n)$ diverges. This reproves the classical fact that the harmonic series diverges, and directly foreshadows the Cauchy Criterion for Series, which restates exactly this test in terms of partial sums.

\Rightarrow

Scope, edge cases & common traps

Situation	What holds	What can fail
Sequence in $\mathbb{R}$	Cauchy $\iff$ convergent	—
Sequence in $\mathbb{Q}$	Convergent $\Rightarrow$ Cauchy always	Cauchy $\not\Rightarrow$ convergent (in $\mathbb{Q}$ )
Sequence in a general metric space	Convergent $\Rightarrow$ Cauchy always	Cauchy $\Rightarrow$ convergent only if space is complete
Checking finitely many $m,n$	Irrelevant	Definition needs all $m,n \ge N$ simultaneously, not just consecutive terms
$\|a_{n+1}-a_n\|\to 0$	Necessary for Cauchy	Not sufficient — does not imply Cauchy (see below)

Common trapthe classic error: mistaking "successive terms get close" for the Cauchy condition

A very common mistake is to check only that consecutive terms get close, i.e. $|a_{n+1} - a_n| \to 0$ , and conclude the sequence is Cauchy. This is false in general. Take $a_n = \sum_{k=1}^n \frac1k$ , the harmonic partial sums from Example 2. Here $|a_{n+1} - a_n| = \frac{1}{n+1} \to 0$ , yet we proved $(a_n)$ is not Cauchy and diverges. The Cauchy definition quantifies over all pairs $m, n \ge N$ — not just adjacent indices — and this quantifier order (∀ε ∃N ∀m,n≥N) cannot be weakened to comparing only neighbors. Always check the full definition, or use a rigorous sufficient condition (e.g., that the tail sums are bounded by a genuinely summable/convergent quantity), never the consecutive-gap shortcut alone.

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered

Augustin-Louis Cauchy

1789–1857 · Early 19th-century France, in the aftermath of the French Revolution and during the rise of rigorous analysis at the École Polytechnique

Cauchy was one of the principal architects of modern rigorous calculus, formalizing limits, continuity, and convergence with the $\varepsilon$ -style arguments used throughout analysis today. His 1821 Cours d'Analyse introduced what we now call the Cauchy criterion for convergence of sequences and series, replacing informal 18th-century reasoning about infinitesimals with precise inequalities. He also made foundational contributions to complex analysis (Cauchy's integral theorem), differential equations, and elasticity theory, publishing an extraordinary volume of work — over 800 papers — across nearly every branch of mathematics and mathematical physics of his time.

History

- 1789 — Cauchy born in Paris, shortly after the storming of the Bastille.

- 1821 — Publishes Cours d'Analyse, stating the Cauchy convergence criterion for sequences and series in rigorous form.

- 1829 — Develops further foundations of complex analysis and real analysis while teaching at the École Polytechnique.

- 1872 — Eduard Heine and Georg Cantor, building on Cauchy's criterion, use equivalence classes of Cauchy sequences of rationals to construct the real numbers rigorously.

- 1920s — The abstract notion of a complete metric space, modeled directly on the Cauchy criterion, becomes central to the emerging field of functional analysis (Banach, Fréchet).

What exactly is the difference between "convergent" and "Cauchy"?

Convergent means there is a specific real number $L$ such that terms get arbitrarily close to $L$ ; the definition explicitly references $L$ . Cauchy means terms get arbitrarily close to each other, with no reference to any particular target value. In $\mathbb{R}$ , the Cauchy Criterion proves these are logically equivalent notions, so in practice "Cauchy" is often used precisely when you want to establish convergence without first knowing (or being able to compute) what the limit is. In general metric spaces they are not equivalent: convergent always implies Cauchy, but Cauchy implies convergent only if the space is complete.

Why does the proof need Bolzano–Weierstrass rather than something simpler?

The Cauchy condition alone only tells you the terms cluster together; it says nothing directly about where they cluster. Bolzano–Weierstrass guarantees that any bounded sequence in $\mathbb{R}$ has a convergent subsequence, which supplies a candidate limit $L$ . The rest of the proof (Step 3) is the genuinely Cauchy-specific argument: it uses the Cauchy property to show that because all terms eventually bunch within $\varepsilon/2$ of each other, and infinitely many of them (the subsequence) bunch within $\varepsilon/2$ of $L$ , the entire tail must be within $\varepsilon$ of $L$ . Both ingredients are necessary; neither alone suffices.

Is every bounded sequence Cauchy?

No. Boundedness is necessary for a Cauchy sequence (we proved this), but it is far from sufficient. The sequence $a_n = (-1)^n$ is bounded (between $-1$ and $1$ ) but not Cauchy: taking $\varepsilon_0 = 1$ , for any $N$ we can pick consecutive $m=N, n=N+1$ with $|a_n - a_m| = 2 \ge 1$ . This sequence oscillates forever and never settles down, so it is bounded but neither Cauchy nor convergent — it has two different subsequential limits, $1$ and $-1$ , which by the theorem's Step 3 argument (uniqueness of the limit target) already rules out convergence, hence rules out being Cauchy.

Does the Cauchy Criterion give a way to compute the limit?

No — and that is precisely its value. The Cauchy Criterion is an existence theorem: it certifies that a real number $L$ exists such that $a_n \to L$ , without constructing or naming $L$ . This is exactly what makes it useful for cases like Example 1, where establishing convergence first (via Cauchy-ness, or via monotone + bounded arguments) is a separate step from later identifying the limit (e.g., by passing to the limit inside a recursive formula, or via other structural facts).

How does this relate to the construction of the real numbers themselves?

One standard construction of $\mathbb{R}$ (due to Cantor, Heine, and Méray) builds real numbers directly as Cauchy sequences of rationals, where two Cauchy sequences $(q_n)$ and $(r_n)$ are declared equivalent if $q_n - r_n \to 0$ . Each equivalence class is defined to be a "real number." Under this construction, the fact that $\mathbb{R}$ (built this way) is complete becomes a theorem to prove, and the Cauchy Criterion in this article can be viewed as the statement, once $\mathbb{R}$ is constructed via the least-upper-bound property instead, that both constructions agree.

Does the Cauchy criterion generalize to sequences of functions or vectors?

Yes. In $\mathbb{R}^n$ (with the Euclidean metric) the identical statement holds: Cauchy sequences of vectors converge, because $\mathbb{R}^n$ is complete (this follows by applying the one-dimensional Cauchy Criterion coordinatewise). In spaces of functions, "Cauchy in the sup-norm" sequences of continuous functions converge uniformly to a continuous function precisely because the space of bounded continuous functions with the sup norm is a complete metric space (a Banach space). This generalization is the entire basis for existence proofs via Picard iteration and for foundational results in Fourier series and PDE theory.

What is the relationship to the Cauchy Criterion for Series?

A series $\sum a_k$ converges exactly when its sequence of partial sums $s_n = \sum_{k=1}^n a_k$ converges, so applying the Cauchy Criterion to $(s_n)$ directly yields the Cauchy Criterion for Series: $\sum a_k$ converges iff for every $\varepsilon>0$ there is $N$ such that for all $n>m\ge N$ , $|\sum_{k=m+1}^n a_k| < \varepsilon$ . Example 2's divergence proof for the harmonic series is a direct instance of this specialization.