Many constructions in analysis proceed by approximation: we build a sequence of objects that get progressively closer to a desired target and then argue that the sequence converges to the target. The [Cauchy criterion](/page/Cauchy%20Sequence) gives a purely intrinsic way to detect that a sequence "wants to converge" — the terms eventually cluster together — without referring to any candidate limit. But an intrinsic criterion is only useful if the ambient space guarantees that the limit actually exists within it. When it does not, approximation arguments collapse. [example: A Cauchy Sequence With No Rational Limit] Consider the [metric space](/page/Metric%20Space) $(\mathbb{Q}, d)$ where $d(x, y) = |x - y|$ is the standard absolute-value metric. Define the [sequence](/page/Sequence) $(x_n)_{n=1}^\infty$ in $\mathbb{Q}$ by the following recursion: \begin{align*} x_1 = 1, \qquad x_{n+1} = \frac{1}{2}\left(x_n + \frac{2}{x_n}\right). \end{align*} This is the Babylonian method (Newton's method applied to $f(t) = t^2 - 2$). The first several terms are: \begin{align*} x_1 = 1, \quad x_2 = \frac{3}{2}, \quad x_3 = \frac{17}{12}, \quad x_4 = \frac{577}{408}. \end{align*} Each $x_n$ is rational (the recursion involves only arithmetic operations on rationals). To see how fast the iterates approach $\sqrt{2}$, expand $x_{n+1}^2 - 2$ directly from the recursion $x_{n+1} = \frac{1}{2}(x_n + 2/x_n)$: \begin{align*} x_{n+1}^2 - 2 = \frac{1}{4}\left(x_n + \frac{2}{x_n}\right)^2 - 2 = \frac{1}{4}\left(x_n^2 + 4 + \frac{4}{x_n^2}\right) - 2 = \frac{x_n^4 - 4x_n^2 + 4}{4x_n^2} = \frac{(x_n^2 - 2)^2}{4x_n^2}. \end{align*} Since $x_n \ge 1$ for all $n \ge 1$, this gives $|x_{n+1}^2 - 2| \le \frac{1}{4}|x_n^2 - 2|^2$. The error contracts quadratically, so $(x_n)$ is a Cauchy sequence in $\mathbb{Q}$. Yet $(x_n)$ does not converge in $\mathbb{Q}$. If $x_n \to L$ with $L \in \mathbb{Q}$, then passing to the limit in the recursion yields $L = \frac{1}{2}(L + 2/L)$, hence $L^2 = 2$. But there is no rational number whose square is $2$. The sequence is "trying" to converge to $\sqrt{2}$, which does not exist in $\mathbb{Q}$. [/example] The example above is not pathological — it is the generic situation in $\mathbb{Q}$. The rationals are riddled with gaps: every irrational number is the limit of a Cauchy sequence of rationals that fails to converge within $\mathbb{Q}$. The real numbers $\mathbb{R}$ were constructed precisely to fill these gaps, and the key property that distinguishes $\mathbb{R}$ from $\mathbb{Q}$ is **completeness**. Completeness is the bridge between the intrinsic criterion (being Cauchy) and the extrinsic conclusion (convergence to a limit in the space). It is the hypothesis that makes iterative constructions, density arguments, and approximation procedures work throughout analysis, from the existence of solutions to differential equations to the foundations of functional analysis. ## Definition The central idea is that a [metric space](/page/Metric%20Space) is complete when the Cauchy criterion is not only necessary for [convergence](/page/Convergence%20(Real%20Sequences)) (which it always is) but also sufficient. [definition: Complete Metric Space] Let $(X, d)$ be a metric space. We say that $(X, d)$ is **complete** if every [Cauchy sequence](/page/Cauchy%20Sequence) in $X$ converges to a limit in $X$. That is, for every sequence $(x_n)_{n=1}^\infty$ in $X$ satisfying \begin{align*} \forall \varepsilon > 0, \; \exists N \in \mathbb{N} \; \text{such that } m, n \ge N \implies d(x_n, x_m) < \varepsilon, \end{align*} there exists $x \in X$ with $d(x_n, x) \to 0$ as $n \to \infty$. [/definition] The definition is purely metric — it depends on the distance function $d$, not on any algebraic or topological structure beyond what $d$ induces. A metric space that is not complete is called **incomplete**. [remark: Completeness and the Metric] Different metrics on the same underlying set can produce different completeness outcomes. For instance, the open interval $(0, 1)$ is incomplete under the standard metric $d(x,y) = |x-y|$, because the Cauchy sequence $x_n = 1/n$ has no limit in $(0,1)$. However, the same set $(0,1)$ can be equipped with a different metric — for example, one transported from $\mathbb{R}$ via the homeomorphism $t \mapsto \tan(\pi(t - 1/2))$ — under which it becomes complete. Completeness is a property of the pair $(X, d)$, not of $X$ alone. [/remark] ### Immediate Examples The most fundamental examples of complete and incomplete spaces orient the theory. The real line $(\mathbb{R}, |\cdot|)$ is complete: this is the content of the Cauchy completeness axiom, which is equivalent to the least upper bound property. Every construction of $\mathbb{R}$ from $\mathbb{Q}$ — whether by Dedekind cuts or Cauchy completion — is designed to achieve exactly this property. Euclidean space $(\mathbb{R}^n, d)$ with $d(x, y) = |x - y| = \left(\sum_{i=1}^n (x_i - y_i)^2\right)^{1/2}$ is complete: a sequence $(x_k)$ in $\mathbb{R}^n$ is Cauchy if and only if each component sequence $(x_{k,i})$ is Cauchy in $\mathbb{R}$, and convergence in $\mathbb{R}^n$ is equivalent to componentwise convergence. The rational numbers $(\mathbb{Q}, |\cdot|)$ are not complete, as the opening example demonstrates. Nor is the open interval $((0,1), |\cdot|)$: the sequence $x_n = 1/n$ is Cauchy but converges to $0 \notin (0,1)$. A more subtle class of examples arises in function spaces, which we treat in a dedicated section below. ## Completeness and Closed Subsets A recurring question in analysis is: when does a subset of a complete metric space inherit completeness? A bounded subset, or even a dense subset, need not be complete — the rationals are dense in $\mathbb{R}$ but not complete. The correct criterion turns out to be closedness: a subset of a complete space is complete if and only if it is closed. This connection between an analytic property (completeness) and a topological property (closedness) is one of the most useful structural results in the theory. [quotetheorem:287] The two directions have different flavors. For part (1), the reasoning is: if $(y_n)$ is a sequence in $Y$ converging to some $x \in X$, then $(y_n)$ is Cauchy in $Y$, so by completeness of $Y$, the limit lies in $Y$; hence $Y$ contains all its limit points and is closed. For part (2), a Cauchy sequence in $Y$ is also Cauchy in $X$, so it converges to some $x \in X$ by completeness of $X$; since $Y$ is closed and the sequence lies in $Y$, the limit $x$ must belong to $Y$. The interplay between the two parts creates a powerful identification: **in a complete metric space, the complete subsets are exactly the closed subsets.** This is used constantly in functional analysis. For instance, to prove that a subset of a [Banach space](/page/Banach%20Space) is itself a Banach space (under the induced norm), it suffices to verify closedness. [example: Closed vs. Non-Closed Subsets of $\mathbb{R}$] The closed interval $[0, 1]$ is a closed subset of the complete metric space $\mathbb{R}$. By the theorem, $[0, 1]$ is complete under the standard metric. The open interval $(0, 1)$ is not closed in $\mathbb{R}$ — the sequence $x_n = 1/n$ converges to $0 \notin (0, 1)$. Accordingly, $(0, 1)$ is not complete: the same sequence $x_n = 1/n$ is Cauchy in $(0, 1)$ but has no limit there. The set $\mathbb{Q} \cap [0, 1]$ is not closed in $\mathbb{R}$ (e.g., $\sqrt{2}/2 \in [0,1] \setminus \mathbb{Q}$ is a limit point of rationals). Correspondingly, $\mathbb{Q} \cap [0,1]$ is not complete. The set $\{0\} \cup \{1/n : n \in \mathbb{N}\}$ is closed in $\mathbb{R}$ (the only limit point is $0$, which belongs to the set). Hence it is complete. [/example] The completeness-closedness characterisation also provides a convenient method for proving completeness of new spaces: embed them as closed subsets of spaces already known to be complete. ## The Completion of a Metric Space The failure of $\mathbb{Q}$ to be complete is resolved by constructing $\mathbb{R}$ — a larger, complete space in which $\mathbb{Q}$ embeds densely. A natural question is whether this procedure generalises: given an arbitrary incomplete metric space, can we always enlarge it to a complete one in a canonical way, without introducing unnecessary extra points? The answer is yes, and the construction mirrors the passage from $\mathbb{Q}$ to $\mathbb{R}$. The idea is to identify the "missing" limits of Cauchy sequences and formally adjoin them to the space. [definition: Completion of a Metric Space] Let $(X, d)$ be a metric space. A **completion** of $(X, d)$ is a pair $(\hat{X}, \hat{d})$ together with an isometric embedding $\iota: X \to \hat{X}$ (meaning $\hat{d}(\iota(x), \iota(y)) = d(x, y)$ for all $x, y \in X$) such that: 1. $(\hat{X}, \hat{d})$ is a complete metric space. 2. The image $\iota(X)$ is dense in $\hat{X}$: for every $\hat{x} \in \hat{X}$ and every $\varepsilon > 0$, there exists $x \in X$ with $\hat{d}(\iota(x), \hat{x}) < \varepsilon$. [/definition] The construction proceeds by taking the set of all Cauchy sequences in $X$, declaring two Cauchy sequences $(x_n)$ and $(y_n)$ equivalent when $d(x_n, y_n) \to 0$, and defining the metric on equivalence classes by $\hat{d}([(x_n)], [(y_n)]) = \lim_{n \to \infty} d(x_n, y_n)$. The original space $X$ embeds via the map $x \mapsto [(x, x, x, \ldots)]$ — the equivalence class of the constant sequence. [quotetheorem:1002] The uniqueness statement is a rigidity result: there is essentially one way to fill in the gaps of a metric space. This justifies speaking of "the" completion of $(X, d)$. [explanation: Why Completions Matter] The completion construction is not merely a curiosity about $\mathbb{Q}$ and $\mathbb{R}$. It underlies several major constructions in analysis: - The passage from $(\mathbb{Q}, |\cdot|)$ to $(\mathbb{R}, |\cdot|)$ is the prototypical example. - The [Lebesgue spaces](/page/Lebesgue%20Integral) $L^p$ can be viewed as completions of the space of continuous functions (or step functions) under the $L^p$ norm. - The Sobolev spaces $W^{k,p}(U)$ arise as completions of smooth functions under the [Sobolev norm](/page/Sobolev%20Space). - The completion of a [normed vector space](/page/Normed%20Vector%20Space) is a [Banach space](/page/Banach%20Space); the completion of an inner product space is a [Hilbert space](/page/Hilbert%20Space). In each case, the underlying pattern is the same: begin with a space of "nice" objects (rationals, smooth functions, simple functions), equip it with a norm that captures the relevant notion of approximation, and pass to the completion to obtain a space in which limits always exist. The resulting space inherits the algebraic structure (addition, scalar multiplication) from the original, because these operations are uniformly continuous and therefore extend uniquely to the completion. [/explanation] ## Completeness in Function Spaces The most important applications of completeness arise not in finite-dimensional spaces like $\mathbb{R}^n$ (where completeness is essentially immediate) but in infinite-dimensional function spaces. Whether a given function space is complete depends sensitively on the choice of metric or norm, and several of the most important theorems in analysis establish completeness for specific function spaces. ### The Supremum Metric and Bounded Functions The simplest function space arises from bounded functions equipped with the uniform metric. The completeness of this space reflects the fact that uniform limits of bounded functions are bounded. [quotetheorem:288] This theorem is powerful because $S$ carries no structure at all — it is simply a set. The completeness of the target space $(Y, e)$ is the essential ingredient: a Cauchy sequence $(f_n)$ in the uniform metric converges pointwise (by completeness of $Y$) and the uniform Cauchy condition forces the convergence to be uniform, hence the limit is bounded. When $S$ is a [compact](/page/Compact%20Space) topological space and $Y = \mathbb{R}$ (or $\mathbb{C}$), write $\ell_\infty(S, \mathbb{R})$ for the space of all bounded functions $f: S \to \mathbb{R}$ equipped with the supremum metric $\mathcal{D}$. The subspace $C(S) \subset \ell_\infty(S, \mathbb{R})$ of continuous functions is closed under uniform limits. By the completeness-closedness characterisation, $C(S)$ equipped with the supremum norm $\|f\|_\infty = \sup_{s \in S} |f(s)|$ is therefore a complete metric space — in fact, a Banach space. This is the standard function space in which the Arzelà-Ascoli theorem, the Stone-Weierstrass theorem, and the theory of ODEs operate. [example: Incompleteness Under the $L^1$ Metric] Consider the space $C([0, 1])$ of continuous functions $f: [0, 1] \to \mathbb{R}$, but now equipped with the $L^1$ metric: \begin{align*} d_1(f, g) = \int_0^1 |f(x) - g(x)| \, dx. \end{align*} This metric space is NOT complete. Define the sequence $(f_n)_{n=1}^\infty$ of continuous functions by: \begin{align*} f_n(x) = \begin{cases} 0 & \text{if } 0 \le x \le \frac{1}{2}, \\ n\left(x - \frac{1}{2}\right) & \text{if } \frac{1}{2} < x \le \frac{1}{2} + \frac{1}{n}, \\ 1 & \text{if } \frac{1}{2} + \frac{1}{n} < x \le 1. \end{cases} \end{align*} Each $f_n$ is continuous and transitions linearly from $0$ to $1$ in an interval of width $1/n$ around $x = 1/2$. The sequence is Cauchy in $d_1$: for $m > n$, \begin{align*} d_1(f_n, f_m) = \int_0^1 |f_n(x) - f_m(x)| \, dx \le \frac{1}{n}, \end{align*} since $f_n$ and $f_m$ agree outside the interval $[1/2, 1/2 + 1/n]$, which has length $1/n$, and both functions take values in $[0, 1]$. If $(f_n)$ converged in $d_1$ to some continuous function $f$, then $f$ would need to satisfy $f(x) = 0$ for $x < 1/2$ and $f(x) = 1$ for $x > 1/2$ (since $f_n$ converges pointwise to the step function $\mathbb{1}_{(1/2, 1]}$ at every point $x \neq 1/2$). But no continuous function on $[0,1]$ can jump from $0$ to $1$ at $x = 1/2$. The "limit" of this Cauchy sequence is a discontinuous function that does not belong to $C([0, 1])$. The completion of $(C([0, 1]), d_1)$ is the Lebesgue space $L^1([0, 1])$, which includes such discontinuous functions (defined up to sets of measure zero). [/example] ### $L^p$ Spaces The most important function spaces in analysis — the Lebesgue spaces $L^p$ — are complete. This fact, sometimes called the Riesz-Fischer theorem, is fundamental to the entire theory of integration, functional analysis, and partial differential equations. [quotetheorem:892] The completeness of $L^p$ spaces is what makes them the correct setting for integration theory and PDEs. In a [Sobolev space](/page/Sobolev%20Space) $W^{k,p}(U)$, completeness follows from the completeness of $L^p$: the Sobolev norm controls each weak derivative in $L^p$ separately, and a Cauchy sequence in $W^{k,p}$ produces Cauchy sequences of each derivative in $L^p$, which converge by the completeness of $L^p$. A key point in the proof for $1 \le p < \infty$ is that one does not simply argue pointwise. Instead, one extracts a subsequence $(f_{n_k})$ that converges rapidly enough (say $\|f_{n_{k+1}} - f_{n_k}\|_p < 2^{-k}$), constructs the series $g = |f_{n_1}| + \sum_{k=1}^\infty |f_{n_{k+1}} - f_{n_k}|$, uses the monotone convergence theorem to show $g \in L^p$, and then concludes that the telescoping series converges absolutely a.e. and in $L^p$. The full sequence converges because it is Cauchy and has a convergent subsequence. ## The Baire Category Theorem Completeness has consequences that go far beyond convergence of individual sequences. One of the deepest is the **Baire Category Theorem**, which constrains how a complete metric space can be decomposed into "small" pieces. The question it addresses is: can a complete metric space be written as a countable union of "negligible" sets? Here "negligible" means **nowhere dense** — a set whose closure has empty interior, so that no open ball is contained in the closure. Examples of nowhere dense sets include finite subsets of $\mathbb{R}$, the Cantor set, and the graph of a continuous function in the plane. [quotetheorem:630] The two forms are logically equivalent (take complements), but they serve different purposes. The dense $G_\delta$ form is used to prove existence results: if each $G_n$ represents the set of functions satisfying some "good" approximation property, then the intersection — the set of functions satisfying all such properties simultaneously — is still dense, hence nonempty. The category form is used to prove impossibility results: a complete space cannot be "exhausted" by countably many small pieces. The hypothesis of completeness is essential. The rational numbers $\mathbb{Q}$ can be written as $\mathbb{Q} = \bigcup_{q \in \mathbb{Q}} \{q\}$, a countable union of singletons, each of which is nowhere dense (since every open interval in $\mathbb{Q}$ contains points other than $q$). This is precisely the kind of decomposition that the Baire Category Theorem forbids for complete spaces. [example: Continuous Nowhere-Differentiable Functions] One of the most striking applications of the Baire Category Theorem shows that "most" continuous functions are nowhere differentiable. Consider the Banach space $C([0,1])$ with the supremum norm. For each $n \in \mathbb{N}$, define the set: \begin{align*} A_n = \left\{ f \in C([0,1]) : \exists\, x \in [0,1] \text{ such that } |f(x+h) - f(x)| \le n|h| \text{ for all small } h \right\}. \end{align*} The set $A_n$ consists of functions that have a "bounded difference quotient" (hence are at least Lipschitz-like at some point). Each $A_n$ is closed: if $f_k \to f$ uniformly with each $f_k \in A_n$, the witnessing points $x_k$ have a convergent subsequence $x_{k_j} \to x$ (by compactness of $[0,1]$), and the bound $|f(x+h) - f(x)| \le n|h|$ passes to the uniform limit, so $f \in A_n$. Each $A_n$ has empty interior: given any $f \in A_n$ and $\varepsilon > 0$, a sufficiently rapidly oscillating piecewise-linear perturbation of $f$ with amplitude less than $\varepsilon$ produces a function in $B(f, \varepsilon)$ whose difference quotients exceed $n$ at every point, placing it outside $A_n$. Hence each $A_n$ is nowhere dense. The set of functions differentiable at some point is contained in $\bigcup_{n=1}^\infty A_n$, which by the Baire Category Theorem is a **meagre** (first category) set. Its complement — the set of continuous functions that are differentiable at NO point — is a dense $G_\delta$ set in $C([0,1])$. In particular, continuous nowhere-differentiable functions are not rare curiosities; they are the generic case. Smooth functions, though dense, form a meagre set. [/example] The Baire Category Theorem is the foundation for three cornerstones of functional analysis: the [Uniform Boundedness Principle](/page/Linear%20Operators%20on%20Banach%20Spaces) (Banach-Steinhaus), the Open Mapping Theorem, and the Closed Graph Theorem. Each of these results relies on completeness to prevent a Banach space from being decomposed into countably many "thin" pieces. ## Fixed Points and the Contraction Mapping Principle Completeness is the engine that drives iterative methods for solving equations. The paradigm is simple: define a map $T$ on a metric space such that the solution to the equation $x = T(x)$ is a fixed point of $T$, construct the sequence $x_0, T(x_0), T^2(x_0), \ldots$ of iterates, show it is Cauchy, and invoke completeness to conclude that it converges to a fixed point. Without completeness, the iteration might converge to a "ghost" — a point that does not exist in the space. [quotetheorem:270] The hypothesis of completeness is used at a single but critical juncture: the sequence of iterates $(x_n)$ is shown to be Cauchy (because $d(x_{n+1}, x_n) \le \lambda^n d(x_1, x_0)$, and the geometric series $\sum \lambda^n$ converges), and completeness converts this into convergence to a limit $\bar{x} \in X$. The contraction condition then shows that $T(\bar{x}) = \bar{x}$ (by continuity of $T$) and that the fixed point is unique. Without completeness, the theorem fails. Consider the contraction $T: (0, 1) \to (0, 1)$ defined by $T(x) = x/2$. This is a contraction with $\lambda = 1/2$, but the unique fixed point of $T$ on $\mathbb{R}$ is $\bar{x} = 0$, which does not belong to $(0, 1)$. The iterates $x_n = x_0 / 2^n$ converge to $0$ in $\mathbb{R}$, but since $(0,1)$ is incomplete, the Contraction Mapping Theorem does not apply there and indeed the conclusion fails. The contraction condition $\lambda < 1$ is also essential: the map $T: [1, \infty) \to [1, \infty)$ defined by $T(x) = x + 1/x$ satisfies $|T(x) - T(y)| \le |x - y|$ for $x, y \ge 1$ (it is non-expansive), but it has no fixed point, since $T(x) > x$ for all $x \ge 1$. [example: Picard-Lindelöf via the Contraction Mapping Theorem] The Contraction Mapping Theorem provides the standard proof of existence and uniqueness for ordinary differential equations. Consider the initial value problem: \begin{align*} y'(t) = f(t, y(t)), \qquad y(t_0) = y_0, \end{align*} where $f: [t_0 - a, t_0 + a] \times \overline{B}(y_0, b) \to \mathbb{R}^n$ is continuous and Lipschitz in the second variable with constant $L$: $|f(t, u) - f(t, v)| \le L|u - v|$. The integral reformulation is $y(t) = y_0 + \int_{t_0}^t f(s, y(s)) \, ds$, so a solution is a fixed point of the operator: \begin{align*} (Ty)(t) = y_0 + \int_{t_0}^t f(s, y(s)) \, ds. \end{align*} Set $\delta = \min(a, b/M, 1/(2L))$ where $M = \sup |f|$, and let $X = \{y \in C([t_0 - \delta, t_0 + \delta]; \mathbb{R}^n) : \|y - y_0\|_\infty \le b\}$ with the supremum metric. The space $X$ is a closed subset of the complete space $(C([t_0-\delta, t_0+\delta]; \mathbb{R}^n), \|\cdot\|_\infty)$, hence complete. For $y, z \in X$: \begin{align*} |(Ty)(t) - (Tz)(t)| &\le \int_{t_0}^t |f(s, y(s)) - f(s, z(s))| \, ds \le L|t - t_0| \cdot \|y - z\|_\infty \le L\delta \, \|y - z\|_\infty. \end{align*} Since $L\delta \le 1/2 < 1$, the operator $T$ is a contraction on the complete metric space $X$, so it has a unique fixed point — the unique local solution to the ODE. [/example] For a deeper treatment of the contraction mapping principle and its extensions (including the Banach fixed point theorem with error estimates), see the [Contraction Mapping Principle](/page/Contraction%20Mapping%20Principle) page. ## Compactness and Completeness The relationship between [compactness](/page/Compact%20Space) and completeness is a source of frequent confusion. Both properties concern the behavior of sequences, both guarantee "limits exist," and in $\mathbb{R}^n$ they are closely related (the Heine-Borel theorem characterises compact sets as those that are closed and bounded). But in general metric spaces, the two properties are logically independent in one direction and related in the other. **Compactness implies completeness, but not conversely.** A compact metric space is always complete: if $(x_n)$ is a Cauchy sequence, then compactness (sequential compactness in metric spaces) provides a convergent subsequence $x_{n_k} \to x$, and the Cauchy condition forces the entire sequence to converge to $x$. Completeness, however, does not imply compactness. The real line $\mathbb{R}$ is complete but not compact (the sequence $x_n = n$ has no convergent subsequence). Infinite-dimensional Banach spaces are complete but their closed unit balls are never compact (by Riesz's lemma). The precise characterisation of compactness in metric spaces involves a strengthening of completeness: [quotetheorem:316] Completeness ensures that Cauchy sequences converge; total boundedness ensures that every sequence has a Cauchy subsequence (by a diagonal argument). Together, they yield sequential compactness, which coincides with compactness for metric spaces. Completeness alone is not enough because it does not prevent sequences from "escaping" (like $x_n = n$ in $\mathbb{R}$); total boundedness alone is not enough because it does not prevent limits from being "missing" (like in $\mathbb{Q} \cap [0,1]$, which is totally bounded but not complete). [example: Total Boundedness Without Completeness] The metric space $(\mathbb{Q} \cap [0,1], |\cdot|)$ is totally bounded: for any $\varepsilon > 0$, finitely many rational points $0, \varepsilon, 2\varepsilon, \ldots$ cover $[0,1]$, hence also $\mathbb{Q} \cap [0,1]$, with balls of radius $\varepsilon$. But it is not compact: the sequence of rational approximations to $\sqrt{2}/2$ defined by $x_n = \lfloor n\sqrt{2}/2 \rfloor / n$ is contained in $\mathbb{Q} \cap [0,1]$ and is Cauchy, but its limit $\sqrt{2}/2$ is irrational and does not belong to the space. [/example] ### The Cantor Intersection Property Another consequence of completeness that closely parallels compactness is the Cantor intersection property for decreasing sequences of closed sets. In a compact space, any decreasing sequence of nonempty closed sets has nonempty intersection. In a complete metric space, the same conclusion holds provided the diameters of the sets shrink to zero. [quotetheorem:624] The diameter condition is essential and has no analogue in the compact setting. In $\mathbb{R}$, the decreasing sequence $F_n = [n, \infty)$ consists of nonempty closed sets in a complete space, but $\bigcap_{n=1}^\infty F_n = \varnothing$ because $\operatorname{diam}(F_n) = \infty$ for all $n$. Similarly, $F_n = [0, 1 + 1/n]$ has $\bigcap F_n = [0, 1]$, which is nonempty but not a single point, because the diameters converge to $1$, not $0$. ## Completeness Is Not Topological A subtlety that distinguishes completeness from most properties encountered in point-set topology is that completeness depends on the metric, not just on the topology it induces. [quotetheorem:293] This means that asking "is $X$ complete?" is not well-defined for a topological space $X$ — it depends on which metric generates the topology. A topological space is called **completely metrisable** if it admits at least one complete metric compatible with its topology. The open interval $(0, 1)$ is completely metrisable (because it is homeomorphic to $\mathbb{R}$, which carries a complete metric), even though its standard metric is not complete. The rationals $\mathbb{Q}$ are NOT completely metrisable — this is a consequence of the Baire Category Theorem: $\mathbb{Q}$ is a countable union of singletons (each nowhere dense), so any metric on $\mathbb{Q}$ compatible with the usual topology must fail to be complete. [explanation: When Does Completeness Transfer?] While homeomorphisms do not preserve completeness, **uniform isomorphisms** (bijections where both the map and its inverse are uniformly continuous) do preserve completeness. If $f: (X, d_X) \to (Y, d_Y)$ is a uniform homeomorphism and $(X, d_X)$ is complete, then $(Y, d_Y)$ is complete. This is because uniform continuity guarantees that Cauchy sequences map to Cauchy sequences, and bijectivity ensures that limits are preserved. In particular, Lipschitz equivalence (a stronger condition than uniform equivalence) preserves completeness. The metrics $d_1(x,y) = |x - y|$ and $d_2(x, y) = \min(|x - y|, 1)$ on $\mathbb{R}$ are uniformly equivalent (both generate the same uniform structure), so $(\mathbb{R}, d_2)$ is complete even though $d_2$ is a bounded metric. Isometries, being the strongest notion of equivalence, preserve completeness as a special case. This is why the isometric embedding in the definition of completion is the right notion for constructing complete extensions. [/explanation] ## Standard Arguments in Complete Metric Spaces Completeness appears as a hypothesis in a wide range of arguments across analysis and topology. The following are the standard patterns in which completeness is invoked. ### The Cauchy-and-Subsequence Argument The most elementary pattern: to show that a sequence $(x_n)$ converges in a complete metric space, it suffices to show (a) the sequence is Cauchy, or (b) it has a Cauchy subsequence (since a Cauchy sequence with a convergent subsequence must itself converge). Variant (b) is particularly useful when combined with compactness of a related space. A typical instance: suppose $T: X \to X$ is a map on a complete metric space and we want to show that the iterates $T^n(x_0)$ converge. If we can prove $\sum_{n=0}^\infty d(T^{n+1}(x_0), T^n(x_0)) < \infty$ (an "absolute convergence" condition), then the sequence $(T^n(x_0))$ is Cauchy because for $m > n$: \begin{align*} d(T^m(x_0), T^n(x_0)) \le \sum_{k=n}^{m-1} d(T^{k+1}(x_0), T^k(x_0)), \end{align*} which is the tail of a convergent series, hence tends to zero. ### The Completeness-Closedness Method To show that a subset $Y$ of a known complete metric space $(X, d)$ is complete, show it is closed. This reduces a metric-analytic problem (completeness) to a topological one (closedness), which is often easier. Common strategies for proving closedness: - **Sequential closedness:** Take a sequence $(y_n)$ in $Y$ converging to $x \in X$ and show $x \in Y$. In a metric space, sequential closedness is equivalent to closedness. - **Continuous preimage:** Write $Y = f^{-1}(C)$ where $f: X \to Z$ is continuous and $C$ is closed in $Z$. The preimage of a closed set under a continuous map is closed. [example: Completeness of the Space of Lipschitz Functions] Let $(X, d_X)$ and $(Y, d_Y)$ be metric spaces with $Y$ complete, and let $x_0 \in X$ be a basepoint. Define the space of Lipschitz functions that vanish at $x_0$: \begin{align*} \operatorname{Lip}_0(X, Y) = \{ f: X \to Y \mid f(x_0) = y_0, \; \sup_{x \neq y} \frac{d_Y(f(x), f(y))}{d_X(x, y)} < \infty \}. \end{align*} This space embeds into the complete space $(\ell_\infty(X, Y), \mathcal{D})$ of bounded functions (the Lipschitz condition and the basepoint normalization imply boundedness). We show $\operatorname{Lip}_0(X, Y)$ is closed. Suppose $(f_n)$ is a sequence in $\operatorname{Lip}_0(X, Y)$ with $\mathcal{D}(f_n, f) \to 0$ for some bounded $f: X \to Y$. Each $f_n$ has Lipschitz constant at most $L_n$. Fix $x, y \in X$ with $x \neq y$. Then: \begin{align*} d_Y(f(x), f(y)) &\le d_Y(f(x), f_n(x)) + d_Y(f_n(x), f_n(y)) + d_Y(f_n(y), f(y)) \\ &\le 2\mathcal{D}(f_n, f) + L_n \, d_X(x, y). \end{align*} If the Lipschitz constants $L_n$ are uniformly bounded (say $L_n \le L$ for all $n$), then taking $n \to \infty$: \begin{align*} d_Y(f(x), f(y)) \le L \, d_X(x, y), \end{align*} so $f$ is Lipschitz with constant at most $L$. Also, $f(x_0) = \lim f_n(x_0) = y_0$. Hence $f \in \operatorname{Lip}_0(X, Y)$ and the space is closed, therefore complete. [/example] ### The Approximation-and-Completeness Strategy This is the pattern underlying the construction of solutions in PDE theory, the calculus of variations, and many existence proofs. The structure is: 1. **Construct approximate solutions.** Build a sequence $(u_n)$ of "approximate" objects (e.g., solutions to regularised problems, Galerkin approximations, or minimizing sequences). 2. **Derive uniform estimates.** Show that $(u_n)$ is bounded in a complete function space $X$ (e.g., a Sobolev space or $L^p$ space). 3. **Extract a convergent subsequence.** Use compactness (Rellich-Kondrachov, Banach-Alaoglu, Arzelà-Ascoli) or direct Cauchy estimates to extract a subsequence converging in some topology. 4. **Identify the limit.** Show that the limit solves the original problem by passing to the limit in the approximate equations. Completeness enters in step 3: the function space must be complete for the extracted subsequence to have a limit within the space. If the space were incomplete, the limit might fail to exist or might not belong to the function space, and the entire argument would break down. ### The Banach Fixed Point Strategy As illustrated in the Picard-Lindelöf example, this strategy converts an equation $x = T(x)$ into a fixed point problem on a complete metric space. The steps are: 1. **Choose the complete metric space.** Identify a closed, bounded, or otherwise constrained subset $X$ of a Banach or function space. 2. **Define the operator.** Reformulate the equation as $x = T(x)$ for some map $T: X \to X$. 3. **Verify the contraction property.** Show $d(T(x), T(y)) \le \lambda \, d(x, y)$ with $\lambda < 1$. This often requires choosing a sufficiently small time interval, domain, or parameter range. 4. **Apply the Contraction Mapping Theorem.** Conclude existence, uniqueness, and an explicit convergence rate. This strategy applies not only to ODEs but also to nonlinear integral equations, implicit function theorems, and perturbation problems in PDEs. ## References - Rudin, W., *Principles of Mathematical Analysis*, 3rd ed. (1976). - Munkres, J., *Topology*, 2nd ed. (2000). - Sutherland, W. A., *Introduction to Metric and Topological Spaces*, 2nd ed. (2009). - Kreyszig, E., *Introductory Functional Analysis with Applications* (1978). - Evans, L. C., *Partial Differential Equations*, 2nd ed. (2010).