Consider the sequence of decimal approximations to $\sqrt{2}$: \begin{align*} 1, \quad 1.4, \quad 1.41, \quad 1.414, \quad 1.4142, \quad 1.41421, \quad \ldots \end{align*} Each term is rational, and the terms are getting closer to each other: the distance $|a_n - a_m|$ between any two late terms is less than $10^{-\min(n,m)}$. In $\mathbb{R}$, this sequence converges to $\sqrt{2}$. But in $\mathbb{Q}$, the sequence has no limit — the number $\sqrt{2}$ does not exist in $\mathbb{Q}$, so the sequence approaches a "gap" in the rational numbers. The terms huddle together ever more tightly, yet there is no rational number for them to converge to. This phenomenon — a sequence whose terms get arbitrarily close to each other, yet which fails to converge — reveals that convergence depends not only on the sequence itself but on the *completeness* of the ambient space. A Cauchy sequence is one whose terms eventually cluster together; a *complete* space is one where this clustering always produces a limit. The distinction between Cauchy and convergent, invisible in $\mathbb{R}$ (where the two notions coincide), is the central structural divide of analysis. It separates $\mathbb{R}$ from $\mathbb{Q}$, [Banach spaces](/page/Banach%20Space) from [normed spaces](/page/Normed%20Vector%20Space), and spaces where limits can be taken freely from spaces where they cannot. ## Definition The standard definition of convergence $a_n \to a$ requires knowing the [limit](/page/Limit) $a$ in advance: for every $\varepsilon > 0$, all sufficiently late terms satisfy $|a_n - a| < \varepsilon$. But in many situations — summing an infinite [series](/page/Series), solving a differential equation by iteration, constructing $\mathbb{R}$ from $\mathbb{Q}$ — the limit is precisely what we are trying to *find*. We need a criterion for convergence that refers only to the terms of the [sequence](/page/Sequence) themselves, with no mention of the limit. Cauchy's insight was to replace "close to the limit $a$" with "close to each other": a sequence should converge if, eventually, *any two* terms are within $\varepsilon$ of each other. [definition:Cauchy Sequence] Let $(X, d)$ be a metric space. A sequence $(a_n)_{n=1}^\infty$ in $X$ is a **Cauchy sequence** if for every $\varepsilon > 0$ there exists $N \in \mathbb{N}$ such that \begin{align*} d(a_n, a_m) < \varepsilon \quad \text{for all } n, m \geq N. \end{align*} [/definition] The quantifier structure is crucial: we demand that *all* pairs $n, m \geq N$ satisfy the bound, not just consecutive terms. The sequence $a_n = \sum_{k=1}^n 1/k$ (partial sums of the harmonic series) has $|a_{n+1} - a_n| = 1/(n+1) \to 0$, so consecutive terms get close, yet the sequence diverges to $+\infty$. The Cauchy condition catches this: $|a_{2n} - a_n| = \sum_{k=n+1}^{2n} 1/k \geq n \cdot 1/(2n) = 1/2$, so the sequence is not Cauchy. ## Relationship to Convergence Every convergent sequence is Cauchy — this is immediate from the triangle inequality. If $a_n \to a$, then for any $\varepsilon > 0$, choose $N$ with $d(a_n, a) < \varepsilon/2$ for $n \geq N$; then $d(a_n, a_m) \leq d(a_n, a) + d(a, a_m) < \varepsilon$ for $n, m \geq N$. The fundamental question is the *converse*: is every Cauchy sequence convergent? This depends entirely on the space. [quotetheorem:172] This is the **completeness of $\mathbb{R}$** — arguably the most important single fact in real analysis. It says that the Cauchy condition, which can be checked without knowing the limit, is *equivalent* to convergence in $\mathbb{R}$. The proof has three steps: (i) every Cauchy sequence is bounded (take $\varepsilon = 1$ to bound all but finitely many terms); (ii) by the [Bolzano-Weierstrass Theorem](/theorems/628), the bounded sequence has a convergent subsequence $a_{n_k} \to a$; (iii) the Cauchy condition forces the *full* sequence to converge to the same limit $a$, because $|a_n - a| \leq |a_n - a_{n_k}| + |a_{n_k} - a|$, and both terms can be made small for large $n$ and $k$. ### Failure in $\mathbb{Q}$: Why Completeness Matters The same theorem fails catastrophically in $\mathbb{Q}$. Define the sequence \begin{align*} a_1 = 1, \qquad a_{n+1} = \frac{1}{2}\left(a_n + \frac{2}{a_n}\right). \end{align*} This is Newton's method applied to $x^2 - 2 = 0$. One can verify that $|a_{n+1} - \sqrt{2}| \leq |a_n - \sqrt{2}|^2 / (2a_n)$, so the sequence converges quadratically to $\sqrt{2}$ in $\mathbb{R}$. Since it converges in $\mathbb{R}$, it is Cauchy in $\mathbb{R}$, hence also Cauchy in $\mathbb{Q}$ (the Cauchy condition refers only to distances between terms, and all terms are rational). But the limit $\sqrt{2} \notin \mathbb{Q}$ — the Cauchy sequence has no limit in $\mathbb{Q}$. This is not an isolated pathology. The rational numbers are riddled with such gaps: for every irrational number $\alpha$, one can construct a Cauchy sequence of rationals converging to $\alpha$ (e.g., the decimal truncations). The completeness of $\mathbb{R}$ is the assertion that *all* these gaps have been filled. In fact, $\mathbb{R}$ can be *constructed* from $\mathbb{Q}$ by the completion process — adjoining limits of Cauchy sequences — making the Cauchy sequence the foundational tool for building the real numbers. ## Completeness The [General Principle of Convergence](/theorems/172) is not a property of $\mathbb{R}$ alone — it is a property of the *metric*. Some [metric spaces](/page/Metric%20Space) have it, others do not. This motivates the following definition. [definition:Complete Metric Space] A metric space $(X, d)$ is **complete** if every Cauchy sequence in $X$ converges to a limit in $X$. [/definition] Completeness says that the space has "no gaps" — every sequence that *should* converge (in the sense that its terms cluster together) *does* converge to a point already in the space. [example:Complete And Incomplete Spaces] **Complete:** - $\mathbb{R}$ and $\mathbb{R}^n$ with the Euclidean metric (the [General Principle of Convergence](/theorems/172) and its multi-dimensional extension). - $C([a, b])$ with the supremum metric $d_\infty(f, g) = \|f - g\|_\infty$ — the uniform limit of [continuous](/page/Continuity) [functions](/page/Function) is continuous. This makes $C([a, b])$ a [Banach space](/page/Banach%20Space). - $\ell^p$ for $1 \leq p \leq \infty$ — the space of $p$-summable sequences with the $\ell^p$ norm. - Any closed subset of a complete metric space (by the [Completeness and Closedness theorem](/theorems/287): closed subsets of complete spaces are complete). **Incomplete:** - $\mathbb{Q}$ with the usual metric — the $\sqrt{2}$ example above. - $(0, 1)$ with the usual metric — the sequence $a_n = 1/n$ is Cauchy but converges to $0 \notin (0, 1)$. - $C([0, 1])$ with the $L^1$ metric $d_1(f, g) = \int_0^1 |f - g|$ — a sequence of continuous functions can converge in $L^1$ to a discontinuous function (e.g., piecewise linear approximations to the step function). The completion of $(C([0, 1]), d_1)$ is the Lebesgue space $L^1([0, 1])$. - The space of polynomials with the supremum norm on $[0, 1]$ — the [Weierstrass approximation theorem](/theorems/480) says polynomials are dense in $C([0, 1])$, so their completion under $\|\cdot\|_\infty$ is $C([0, 1])$ itself. [/example] ### Completeness and Closedness There is a tight relationship between completeness and closedness: in a complete metric space, a subset is complete (in the inherited metric) if and only if it is closed. [quotetheorem:287] The forward direction (complete implies closed) holds in any metric space: if $x_n \in N$ and $x_n \to x$ in $X$, then $(x_n)$ is Cauchy in $N$; completeness of $N$ gives a limit $y \in N$; [uniqueness of limits](/theorems/625) forces $x = y \in N$. The reverse (closed in complete implies complete) uses the fact that a Cauchy sequence in $N$ is also Cauchy in $X$; completeness of $X$ gives a limit $x \in X$; closedness of $N$ forces $x \in N$. This theorem is used constantly: to show that a subset of a Banach space is itself a Banach space, it suffices to show it is closed. This is typically much easier than verifying completeness directly. For instance, $C([a, b])$ is a closed subset of the complete space $(\ell^\infty([a, b]), \|\cdot\|_\infty)$ — closed because uniform limits of continuous functions are [continuous](/page/Uniform%20Convergence) — and this immediately gives completeness. ## The Completion Every metric space, no matter how many gaps it has, can be "filled in" to produce a complete space. This is the **completion** construction, which adjoins limits of all Cauchy sequences. [definition:Completion Of A Metric Space] Let $(X, d)$ be a metric space. A **completion** of $X$ is a pair $(\hat{X}, \iota)$ where $\hat{X}$ is a complete metric space and $\iota: X \to \hat{X}$ is an isometric embedding (i.e., $d_{\hat{X}}(\iota(x), \iota(y)) = d(x, y)$) such that $\iota(X)$ is dense in $\hat{X}$. [/definition] The completion is unique up to isometry: if $(\hat{X}_1, \iota_1)$ and $(\hat{X}_2, \iota_2)$ are both completions, there is a unique isometry $\phi: \hat{X}_1 \to \hat{X}_2$ with $\phi \circ \iota_1 = \iota_2$. **Construction.** The completion is built from equivalence classes of Cauchy sequences. Define an [equivalence relation](/page/Equivalence%20Relation) on the [set](/page/Set) of all Cauchy sequences in $X$: $(a_n) \sim (b_n)$ if $d(a_n, b_n) \to 0$. Set $\hat{X} = \{\text{Cauchy sequences in } X\} / {\sim}$, with metric $d_{\hat{X}}([(a_n)], [(b_n)]) = \lim_{n \to \infty} d(a_n, b_n)$ (the limit exists because $d(a_n, b_n)$ is a Cauchy sequence of real numbers, by the triangle inequality, and $\mathbb{R}$ is complete). The embedding $\iota$ sends $x$ to the class of the constant sequence $(x, x, x, \ldots)$. This construction is how $\mathbb{R}$ is built from $\mathbb{Q}$: a real number *is* an equivalence class of Cauchy sequences of rationals. The algebraic operations on $\mathbb{R}$ (addition, multiplication, ordering) are inherited termwise from $\mathbb{Q}$ and shown to be well-defined on equivalence classes. The completeness of $\mathbb{R}$ is then a *theorem* of the construction, not an axiom. [example:Completions In Practice] The completion of $(\mathbb{Q}, |\cdot|)$ is $\mathbb{R}$. The completion of $(C([0,1]), \|\cdot\|_{L^p})$ for $1 \leq p < \infty$ is the [Lebesgue space](/page/Lebesgue%20Integral) $L^p([0,1])$. The completion of the space of step functions under the $L^1$ norm is again $L^1$ — this is one route to constructing the Lebesgue integral. The completion of the polynomials under $\|\cdot\|_\infty$ on $[0, 1]$ is $C([0, 1])$ (by Weierstrass approximation). In each case, the completion process reveals what "objects" the original space is "missing": $\mathbb{Q}$ is missing the irrationals; $C([0,1])$ under the $L^1$ norm is missing the discontinuous [integrable](/page/Integral) functions; the polynomials under $\|\cdot\|_\infty$ are missing the non-polynomial continuous functions. [/example] ## Cauchy Sequences and Series An infinite series $\sum_{j=1}^\infty a_j$ converges if and only if its partial sums $S_N = \sum_{j=1}^N a_j$ form a convergent sequence. The Cauchy criterion for series — the analogue of the General Principle of Convergence — avoids computing the sum explicitly. The sequence of partial sums $(S_N)$ is Cauchy if and only if for every $\varepsilon > 0$, there exists $N$ such that $|S_m - S_n| = |\sum_{j=n+1}^m a_j| < \varepsilon$ for all $m > n \geq N$. This is the condition that the "tails" of the series become uniformly small. Setting $m = n + 1$ gives the *necessary* condition $|a_{n+1}| < \varepsilon$ for large $n$ — which is the familiar fact that $a_n \to 0$ is necessary for convergence. But the converse fails: the harmonic series $\sum 1/n$ has $a_n \to 0$ yet diverges, because the tails $\sum_{j=n+1}^{2n} 1/j \geq 1/2$ do not tend to zero. Absolute convergence ($\sum |a_j| < \infty$) implies convergence in any complete space, via the Cauchy criterion: $|\sum_{j=n+1}^m a_j| \leq \sum_{j=n+1}^m |a_j|$, and the right side tends to zero because the partial sums of $\sum |a_j|$ are Cauchy. In incomplete spaces, even absolute convergence can fail to produce a limit — this is another way to see the incompleteness of $\mathbb{Q}$ (the series $\sum 1/n! = e$ converges absolutely in $\mathbb{R}$ but has no limit in $\mathbb{Q}$). In fact, the converse characterises completeness: a [normed vector space](/page/Normed%20Vector%20Space) is a [Banach space](/page/Banach%20Space) if and only if every absolutely convergent series converges. This gives a practical test for completeness that avoids working with arbitrary Cauchy sequences. ## Cauchy Sequences of Functions When the elements of a sequence are *functions*, the Cauchy condition takes different forms depending on the metric. **Uniform Cauchy sequences.** A sequence of functions $f_n: E \to \mathbb{R}$ is **uniformly Cauchy** if for every $\varepsilon > 0$, there exists $N$ such that $|f_n(x) - f_m(x)| < \varepsilon$ for all $x \in E$ and all $n, m \geq N$. This is the Cauchy condition in the supremum metric. Since $\mathbb{R}$ is complete, a uniformly Cauchy sequence converges pointwise to a limit function $f$; the uniformity ensures that the convergence is also uniform: $\|f_n - f\|_\infty \to 0$. Moreover, if each $f_n$ is continuous, the limit $f$ is continuous — this is the [uniform limit theorem](/theorems/258), and it is the reason that $C([a, b])$ is complete under $\|\cdot\|_\infty$. **Pointwise Cauchy sequences.** A weaker condition: for each fixed $x$, the numerical sequence $f_n(x)$ is Cauchy. By completeness of $\mathbb{R}$, $f_n(x)$ converges for each $x$, defining a pointwise limit $f(x) = \lim f_n(x)$. But this limit need not be continuous, even if every $f_n$ is: the sequence $f_n(x) = x^n$ on $[0, 1]$ converges pointwise to the discontinuous function $f(x) = 0$ for $x \in [0, 1)$ and $f(1) = 1$. **$L^p$ Cauchy sequences.** In the $L^p$ norm, $\|f_n - f_m\|_{L^p} \to 0$ means the functions agree "on average" but may differ on sets of small measure. The Riesz-Fischer theorem states that $L^p$ is complete: every $L^p$-Cauchy sequence converges in $L^p$ to some $f \in L^p$. The proof extracts a subsequence converging pointwise a.e. (using the summability of $\|f_{n_{k+1}} - f_{n_k}\|_{L^p}$) and verifies $L^p$ convergence to the pointwise limit. The distinction between these modes of convergence — uniform, pointwise, $L^p$ — is one of the central themes of analysis. The Cauchy condition provides a uniform framework: in each case, "Cauchy" means "the terms eventually cluster together in the relevant metric," and completeness determines whether a limit exists. ## Applications of Completeness ### The Banach Fixed-Point Theorem The most powerful application of completeness is the [Banach Fixed-Point Theorem](/theorems/270): every contraction on a complete metric space has a unique fixed point. The proof constructs the fixed point as the limit of the Cauchy sequence $x, T(x), T^2(x), \ldots$, which is Cauchy because $d(T^n(x), T^{n+1}(x)) \leq \lambda^n d(x, T(x))$ with $\lambda < 1$. Without completeness, the sequence might converge to a point outside the space, and the fixed point would not exist. This theorem is the engine behind the [Picard-Lindelöf theorem](/theorems/69) (existence and uniqueness for ODEs), the [inverse function theorem](/page/Inverse%20Function%20Theorem) (in Banach space form), and many iterative numerical methods. In each case, the argument reduces to: (i) define a map on a complete metric space, (ii) verify the contraction condition, (iii) invoke Banach to get a unique fixed point. Step (iii) requires completeness — and the examples where it fails (contractions on incomplete spaces with no fixed point, such as $T(x) = x/2$ on $(0, 1)$) show that completeness is not a luxury. ### The [Baire Category Theorem](/theorems/630) Complete metric spaces satisfy the Baire Category Theorem: the intersection of countably many dense [open sets](/page/Open%20Set) is dense. This has striking consequences: in a complete metric space, a "generic" element (in the sense of Baire category) often has properties that seem pathological. For instance, the Baire Category Theorem applied to $C([0, 1])$ with the supremum norm shows that the set of nowhere-[differentiable](/page/Derivative) continuous functions is a dense $G_\delta$ — "most" continuous functions are nowhere differentiable. Without completeness, the Baire Category Theorem fails, and such density results have no guarantee. ### Construction of the Real Numbers The Cauchy sequence construction of $\mathbb{R}$ from $\mathbb{Q}$ is one of the two standard foundations for real analysis (the other being Dedekind cuts). Both constructions produce the same complete ordered field, but the Cauchy sequence approach generalises: the completion of any metric space uses the same idea (equivalence classes of Cauchy sequences), while Dedekind cuts rely on the ordering of $\mathbb{Q}$ and do not extend to general metric spaces.