Distances give a way to talk about nearness, but ordinary continuity only controls nearness after a base point has been chosen. For a function $f: X \to Y$ between metric spaces, continuity at $x_0$ says that points sufficiently close to $x_0$ are sent close to $f(x_0)$. The allowed radius may depend on $x_0$. Uniform continuity asks whether a single input scale works everywhere at once. The difference matters whenever a space has many regions with different behaviour. A function may behave gently near each individual point while still becoming too steep globally for one scale to control it. Uniform continuity is the form of continuity that survives approximation arguments, completion, interchange of limits, and numerical error estimates, because it gives a global rule: if two inputs are close enough, then their outputs are close, independently of where the inputs lie. [example: Squaring on an Unbounded Interval] Let $f:(0,\infty)\to\mathbb R$ be given by $f(x)=x^2$, with Euclidean metrics. First, $f$ is continuous at each fixed $a\in(0,\infty)$: if $|x-a|<\delta$ and $\delta\le 1$, then $x<a+1$, so \begin{align*} |f(x)-f(a)|=|x^2-a^2|=|x-a|\,|x+a|<\delta(2a+1). \end{align*} Thus choosing $\delta=\min\{1,\varepsilon/(2a+1)\}$ gives $|f(x)-f(a)|<\varepsilon$. We now show that no single input tolerance works everywhere. Take $\varepsilon_0=1$. Given any $\delta>0$, choose $n\in\mathbb N$ with $1/n<\delta$, and set $x_n=n$ and $y_n=n+1/n$. Then $x_n,y_n\in(0,\infty)$ and \begin{align*} |x_n-y_n|=\left|n-\left(n+\frac{1}{n}\right)\right|=\frac{1}{n}<\delta. \end{align*} For the outputs, \begin{align*} |f(y_n)-f(x_n)|=\left|\left(n+\frac{1}{n}\right)^2-n^2\right|=\left|n^2+2+\frac{1}{n^2}-n^2\right|=2+\frac{1}{n^2}>1. \end{align*} So for the fixed output tolerance $1$, every proposed input tolerance $\delta$ admits two points less than $\delta$ apart whose squared values are more than $1$ apart. Hence $f$ is not uniformly continuous on $(0,\infty)$. [/example] This example shows the core failure that uniform continuity is designed to prevent. Ordinary continuity lets the permissible input radius shrink as the base point moves. On an unbounded domain, that shrinking may have no positive lower bound. Uniform continuity rules out this drifting local scale. ## Definition The parent idea, [uniform continuity](/page/Uniform%20Continuity), is the demand that the same input tolerance works throughout the whole domain. In the metric-space setting, this becomes an epsilon-delta condition in which the metrics on the domain and codomain explicitly measure the two kinds of error. [definition: Uniform Continuity Between Metric Spaces] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces, and let $f:X\to Y$ be a function. The function $f$ is uniformly continuous from $(X,d_X)$ to $(Y,d_Y)$ if for every $\varepsilon > 0$ there exists $\delta > 0$ such that for all $x,y \in X$, \begin{align*} d_X(x,y) < \delta \implies d_Y(f(x),f(y)) < \varepsilon. \end{align*} [/definition] The order of the quantifiers is the whole point. The number $\delta$ is allowed to depend on $\varepsilon$, on the function, and on the two metrics, but not on $x$ or $y$. This single sentence is often the first place where global geometry enters analysis. Sometimes the same underlying set carries more than one metric. Uniform continuity is then not only a property of a formula for $f$, but a property of the chosen metrics. Changing either metric can change the answer. [example: Same Formula, Different Metrics] Let $X=(0,1)$ and let $f:X\to\mathbb R$ be given by $f(x)=1/x$. First equip $X$ with the [Euclidean metric](/page/Euclidean%20Metric). We show that $f$ is not uniformly continuous by fixing the output tolerance $\varepsilon_0=1/2$. For each $n\ge 2$, set $x_n=1/n$ and $y_n=1/(n+1)$. Then $x_n,y_n\in(0,1)$, and \begin{align*} |x_n-y_n|=\left|\frac{1}{n}-\frac{1}{n+1}\right|=\left|\frac{n+1-n}{n(n+1)}\right|=\frac{1}{n(n+1)}. \end{align*} Since $n(n+1)\to\infty$, we have $|x_n-y_n|\to0$. For the output values, \begin{align*} |f(x_n)-f(y_n)|=\left|\frac{1}{1/n}-\frac{1}{1/(n+1)}\right|=|n-(n+1)|=|-1|=1. \end{align*} Thus arbitrarily close Euclidean input pairs can have output distance $1$, so no single Euclidean input tolerance can force all output distances below $1/2$. Now equip the same set $X=(0,1)$ with the metric $d_X(x,y)=|1/x-1/y|$. For all $x,y\in X$, \begin{align*} |f(x)-f(y)|=\left|\frac{1}{x}-\frac{1}{y}\right|=d_X(x,y). \end{align*} Given $\varepsilon>0$, choose $\delta=\varepsilon$. If $d_X(x,y)<\delta$, then \begin{align*} |f(x)-f(y)|=d_X(x,y)<\delta=\varepsilon. \end{align*} So the same formula $f(x)=1/x$ is uniformly continuous for this metric. The example shows that uniform continuity is a property of the function together with the chosen domain and codomain metrics, not of the formula alone. [/example] This dependence on the metrics is not a nuisance; it is the feature that makes the definition portable. The same definition works for Euclidean domains, spaces of functions, discrete spaces, quotient-like constructions, and abstract completions. ## Local Continuity and Global Control The first structural question is how uniform continuity relates to the more familiar pointwise notion of continuity. Uniform continuity should imply ordinary continuity, since a global tolerance can be used at any chosen base point. The converse fails, and the failure is precisely the lack of a uniform lower bound on the local tolerances. ### Pointwise Continuity Before comparing the two notions, we isolate the pointwise version in metric language. This gives a common template for the rest of the page: the same metrics appear, but the quantifier over the base point is placed outside the choice of $\delta$. [definition: Continuity Between Metric Spaces] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces, let $x_0 \in X$, and let $f:X\to Y$ be a function. The function $f$ is continuous at $x_0$ if for every $\varepsilon>0$ there exists $\delta>0$ such that for all $x \in X$, \begin{align*} d_X(x,x_0)<\delta \implies d_Y(f(x),f(x_0))<\varepsilon. \end{align*} The function $f$ is continuous on $X$ if it is continuous at every $x_0 \in X$. [/definition] This definition allows the acceptable input radius to vary with the chosen point. The next question is what happens when that freedom is removed. If the stronger global radius exists for every output tolerance, then each individual point can use the same radius, so the pointwise continuity condition should follow. [quotetheorem:8495] This theorem is often used silently, but its value is conceptual: uniform continuity is stronger than continuity because it controls pairs of nearby points anywhere in the domain, not just points near one fixed centre. ### Sequential Failure Tests To see what the converse would need, it helps to express failure of uniform continuity using sequences. This turns a quantifier failure into a concrete diagnostic: find pairs of points whose distance tends to zero while their images do not approach each other. [quotetheorem:9181] The theorem explains why the opening examples were effective. They did not need to inspect every possible $\delta$ directly; they produced a sequence of input pairs defeating all possible choices at once. [example: The Reciprocal Near a Missing Endpoint] Let $f:(0,1)\to \mathbb R$ be given by $f(x)=1/x$, with Euclidean metrics. We show that input pairs can become arbitrarily close near $0$ while their images stay a fixed positive distance apart. For each $n\in\mathbb N$, set \begin{align*} x_n=\frac{1}{n} \end{align*} and \begin{align*} y_n=\frac{1}{n+1}. \end{align*} For $n\ge 1$, both $x_n$ and $y_n$ lie in $(0,1]$, and for $n\ge 2$ they lie in $(0,1)$. Their input distance is \begin{align*} |x_n-y_n|=\left|\frac{1}{n}-\frac{1}{n+1}\right|. \end{align*} Putting the fractions over the common denominator $n(n+1)$ gives \begin{align*} \left|\frac{1}{n}-\frac{1}{n+1}\right|=\left|\frac{n+1}{n(n+1)}-\frac{n}{n(n+1)}\right|. \end{align*} Therefore \begin{align*} |x_n-y_n|=\left|\frac{1}{n(n+1)}\right|=\frac{1}{n(n+1)}. \end{align*} Since $n(n+1)\to\infty$, we have $|x_n-y_n|\to 0$. For the output distance, \begin{align*} |f(x_n)-f(y_n)|=\left|\frac{1}{1/n}-\frac{1}{1/(n+1)}\right|. \end{align*} Using $\frac{1}{1/n}=n$ and $\frac{1}{1/(n+1)}=n+1$, this becomes \begin{align*} |f(x_n)-f(y_n)|=|n-(n+1)|. \end{align*} Hence \begin{align*} |f(x_n)-f(y_n)|=|-1|=1. \end{align*} Thus the input distances tend to $0$, but the output distances remain equal to $1$. The obstruction is not unboundedness of the interval, since $(0,1)$ is bounded; it is that the domain omits the endpoint where the reciprocal function becomes unbounded. [/example] This example separates two common intuitions. Bounded domain alone does not guarantee uniform continuity; compactness is the condition that prevents both escape to infinity and escape toward a missing boundary point. ## Moduli and Quantitative Continuity Uniform continuity says that a suitable $\delta$ exists for each $\varepsilon$, but applications often need to know how the output error depends on the input error. A modulus packages the dependence into a function. This makes uniform continuity quantitative and lets us compare stronger regularity conditions such as Lipschitz and Holder continuity. ### Moduli of Continuity When estimates are passed through an argument, repeatedly unpacking the epsilon-delta definition is inefficient. A single function measuring the worst allowed output error at each input scale gives a reusable object. This is the role of a modulus. [definition: Modulus of Continuity] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces, and let $f:X\to Y$ be a function. A modulus of continuity for $f$ is a function $\omega:[0,\infty)\to[0,\infty]$ such that $\lim_{t\to 0^+}\omega(t)=0$, and for all $x,y\in X$, \begin{align*} d_Y(f(x),f(y)) \le \omega(d_X(x,y)). \end{align*} [/definition] A modulus is a global error bound. The definition suggests that a modulus should be exactly the quantitative form of uniform continuity, but this needs a theorem because the original definition only asserts the existence of a tolerance for each target error. The next result confirms that no information is lost by replacing tolerances with a modulus. [quotetheorem:9182] This characterization is a bridge between qualitative topology and estimates. Many important classes of maps are introduced by specifying a particularly simple modulus, so the next step is to examine the linear case. ### Lipschitz and Holder Control A linear modulus is the most stable form of control: output distances are bounded by a constant multiple of input distances. This is stronger than uniform continuity and behaves well under composition, limits, and extension. [definition: Lipschitz Map Between Metric Spaces] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces. A function $f:X\to Y$ is Lipschitz if there exists $L\ge 0$ such that for all $x,y\in X$, \begin{align*} d_Y(f(x),f(y)) \le L d_X(x,y). \end{align*} [/definition] The constant $L$ is a global slope bound. To test uniform continuity, one must choose a single input tolerance that works everywhere in the domain. A Lipschitz estimate solves exactly that problem: making $d_X(x,y)$ smaller than $\varepsilon/L$ forces the output distance below $\varepsilon$, with the case $L=0$ handled separately. Thus a Lipschitz bound is a convenient certificate of uniform continuity. The same issue can be asked for a whole family of functions. A family $\mathcal F$ of maps from $X$ to $Y$ is uniformly equicontinuous if for every $\varepsilon>0$ there is one $\delta>0$ such that $d_X(x,y)<\delta$ implies $d_Y(f(x),f(y))<\varepsilon$ for every $f\in\mathcal F$ and all $x,y\in X$. Thus a common Lipschitz bound is not just a certificate for each individual map; it gives one shared input tolerance that works across the entire family. The useful criterion is therefore to replace many separate continuity estimates by one numerical slope bound. The question is whether a single Lipschitz constant for the family is enough to produce the single $\delta$ required in the definition of uniform equicontinuity. The next formal result records exactly this passage from a common slope estimate to a uniform choice of input scale. [quotetheorem:1097] The implication is not reversible. Some uniformly continuous functions have slopes that become unbounded while their values still change slowly enough at small scales. [example: Square Root Is Uniformly Continuous but Not Lipschitz] Let $f:[0,1]\to\mathbb R$ be given by $f(x)=\sqrt{x}$, with Euclidean metrics. We first prove the estimate \begin{align*} |\sqrt{x}-\sqrt{y}| \le |x-y|^{1/2} \end{align*} for all $x,y\in[0,1]$. If $x\ge y$, then $y^2\le xy$, so $y\le\sqrt{xy}$ because both sides are nonnegative. Hence \begin{align*} |\sqrt{x}-\sqrt{y}|^2=(\sqrt{x}-\sqrt{y})^2=x-2\sqrt{xy}+y\le x-2y+y=x-y=|x-y|. \end{align*} If $y\ge x$, the same argument with $x$ and $y$ exchanged gives \begin{align*} |\sqrt{x}-\sqrt{y}|^2\le y-x=|x-y|. \end{align*} Taking square roots gives $|\sqrt{x}-\sqrt{y}|\le |x-y|^{1/2}$. Now let $\varepsilon>0$ and choose $\delta=\varepsilon^2$. If $|x-y|<\delta$, then the estimate above gives \begin{align*} |f(x)-f(y)|=|\sqrt{x}-\sqrt{y}|\le |x-y|^{1/2}<\delta^{1/2}=\varepsilon. \end{align*} Thus $f$ is uniformly continuous on $[0,1]$, and the same estimate is recorded quantitatively by the modulus $\omega(t)=t^{1/2}$. The function is not Lipschitz on $[0,1]$. Suppose, for contradiction, that there were a constant $L\ge0$ such that \begin{align*} |\sqrt{x}-\sqrt{y}|\le L|x-y| \end{align*} for all $x,y\in[0,1]$. Taking $x=1/n$ and $y=0$ gives \begin{align*} \frac{|f(1/n)-f(0)|}{|1/n-0|}=\frac{|\sqrt{1/n}-0|}{1/n}=\frac{1/\sqrt{n}}{1/n}=\sqrt{n}. \end{align*} The Lipschitz inequality would force $\sqrt{n}\le L$ for every $n\in\mathbb N$, but choosing $n>L^2$ gives $\sqrt{n}>L$. Therefore no Lipschitz constant exists. [/example] The square-root example motivates a family of estimates intermediate between arbitrary uniform continuity and Lipschitz continuity. These estimates appear throughout analysis, especially in compactness and regularity theory. [definition: Holder Continuous Map Between Metric Spaces] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces. A function $f:X\to Y$ is Holder continuous with exponent $\alpha\in(0,1]$ if there exists $C\ge0$ such that for all $x,y\in X$, \begin{align*} d_Y(f(x),f(y)) \le C d_X(x,y)^\alpha. \end{align*} [/definition] Holder continuity with $\alpha=1$ is Lipschitz continuity. When $0<\alpha<1$, the estimate can tolerate unbounded local slopes, but it still gives a uniform way to turn a desired output error into an input tolerance. The key point is that the bound $C d_X(x,y)^\alpha$ is itself a modulus of continuity: choosing $d_X(x,y)$ small enough compared with $(\varepsilon/C)^{1/\alpha}$ forces the output distance below $\varepsilon$. [quotetheorem:8226] Quantitative continuity is useful because estimates can be chained. If one process introduces a controlled error and another process transforms that error with its own modulus, the combined process should still have global small-scale control. This motivates the closure result for composition. [quotetheorem:9183] This closure property is one reason uniform continuity is the correct notion for metric categories where maps are meant to preserve small-scale structure globally. ## Compactness and the Heine-Cantor Principle The central positive theorem about uniform continuity says that compactness turns local continuity into global continuity. The reason is that compactness prevents the local radii from degenerating indefinitely: finitely many local controls suffice to cover the whole space. ### Compact Domains A space can fail to support uniform estimates because points escape to infinity or because they approach a missing boundary point. Compactness rules out both types of escape in metric analysis. The formal definition is stated in terms of finite subcovers. [definition: Compact Metric Space] A [metric space](/page/Metric%20Space) $(X,d_X)$ is compact if every [open cover](/page/Open%20Cover) of $X$ has a [finite subcover](/page/Finite%20Subcover). [/definition] The natural question is whether compactness is strong enough to upgrade pointwise continuity into uniform continuity for every target metric space. Each point supplies a local radius, and compactness is exactly the hypothesis that can reduce infinitely many local pieces to finitely many. The [Heine-Cantor theorem](/theorems/280) is the resulting global control principle. [quotetheorem:954] The theorem does not say that compactness is necessary for a particular function to be uniformly continuous. Constant maps, Lipschitz maps, and many special functions can be uniformly continuous on noncompact domains. It says instead that compactness guarantees uniform continuity for every [continuous function](/page/Continuous%20Function) out of the space. [example: Sine on the Real Line] Let $f:\mathbb R\to\mathbb R$ be given by $f(x)=\sin x$, with Euclidean metrics. The domain is not compact: the open intervals $(-n,n)$ for $n\in\mathbb N$ cover $\mathbb R$, but any finite subcollection is contained in $(-N,N)$ for some largest $N$, and therefore does not contain $N+1$. For $x,y\in\mathbb R$, the trigonometric identity for a difference of sines gives \begin{align*} |\sin x-\sin y|=\left|2\cos\left(\frac{x+y}{2}\right)\sin\left(\frac{x-y}{2}\right)\right|. \end{align*} Using $|\cos t|\le 1$ and $|\sin t|\le |t|$ for real $t$, this yields \begin{align*} |\sin x-\sin y|=2\left|\cos\left(\frac{x+y}{2}\right)\right|\left|\sin\left(\frac{x-y}{2}\right)\right|\le 2\cdot 1\cdot \left|\frac{x-y}{2}\right|=|x-y|. \end{align*} Given $\varepsilon>0$, choose $\delta=\varepsilon$. If $|x-y|<\delta$, then \begin{align*} |f(x)-f(y)|=|\sin x-\sin y|\le |x-y|<\delta=\varepsilon. \end{align*} Thus $\sin x$ is uniformly continuous on the noncompact space $\mathbb R$; noncompactness permits failure of uniform continuity, but does not force it. [/example] Noncompactness allows failure; it does not force failure. A more refined statement asks which metric spaces have the property that every continuous real-valued function on them is uniformly continuous. That direction leads toward uniform spaces and special classes of metric spaces, but the compact case is the main theorem used in elementary analysis. ### Missing Boundary Points The [compactness theorem](/theorems/2748) is most memorable when contrasted with bounded noncompact domains. A bounded interval with an endpoint removed still permits sequences to approach a place where the function has no controlled limiting behaviour. [example: A Continuous Function on a Bounded Noncompact Domain] Let $X=(0,1)$ with the Euclidean metric and let $f:X\to\mathbb R$ be given by $f(x)=\sin(1/x)$. The function is continuous on $X$, because $x\mapsto 1/x$ is continuous at every point of $(0,1)$ and $t\mapsto \sin t$ is continuous on $\mathbb R$, so their composition is continuous. We show that this continuous function is not uniformly continuous. Fix $\varepsilon_0=1$. For each $n\in\mathbb N$, set \begin{align*} x_n=\frac{1}{2\pi n+\pi/2} \end{align*} and \begin{align*} y_n=\frac{1}{2\pi n+3\pi/2}. \end{align*} Both denominators are larger than $1$, so $x_n,y_n\in(0,1)$. Since $2\pi n+\pi/2<2\pi n+3\pi/2$, we have $x_n>y_n$, and therefore \begin{align*} |x_n-y_n|=\frac{1}{2\pi n+\pi/2}-\frac{1}{2\pi n+3\pi/2}. \end{align*} Putting the two fractions over a common denominator gives \begin{align*} |x_n-y_n|=\frac{(2\pi n+3\pi/2)-(2\pi n+\pi/2)}{(2\pi n+\pi/2)(2\pi n+3\pi/2)}. \end{align*} The numerator is $\pi$, so \begin{align*} |x_n-y_n|=\frac{\pi}{(2\pi n+\pi/2)(2\pi n+3\pi/2)}. \end{align*} Since $2\pi n+\pi/2>2\pi n$ and $2\pi n+3\pi/2>2\pi n$, we get \begin{align*} 0<|x_n-y_n|<\frac{\pi}{(2\pi n)(2\pi n)}=\frac{1}{4\pi n^2}. \end{align*} Thus, given any $\delta>0$, choosing $n$ so large that $1/(4\pi n^2)<\delta$ gives $|x_n-y_n|<\delta$. For the output values, \begin{align*} f(x_n)=\sin(2\pi n+\pi/2)=\sin(\pi/2)=1 \end{align*} by $2\pi$-periodicity of sine, while \begin{align*} f(y_n)=\sin(2\pi n+3\pi/2)=\sin(3\pi/2)=-1. \end{align*} Hence \begin{align*} |f(x_n)-f(y_n)|=|1-(-1)|=2>1=\varepsilon_0. \end{align*} So every proposed input tolerance $\delta$ admits two points of $(0,1)$ less than $\delta$ apart whose images are more than $1$ apart. Therefore $f$ is not uniformly continuous on the bounded noncompact domain $(0,1)$. [/example] The oscillation accumulates at a point not present in the domain. Uniform continuity would forbid this: close input pairs near the missing endpoint would have to produce close output pairs. ## Cauchy Sequences and Completions Uniform continuity is also the right condition for preserving Cauchy behaviour. Continuity preserves limits that already exist in the domain. Uniform continuity preserves the internal promise that a sequence is trying to converge, even before the limit has been supplied. ### Cauchy Behaviour To state this, we first recall the sequential object that belongs purely to the metric structure. A [Cauchy sequence](/page/Cauchy%20Sequence) is a sequence whose terms eventually become mutually close, rather than close to a named limit. [definition: Cauchy Sequence in a Metric Space] Let $(X,d_X)$ be a metric space. A sequence $(x_n)_{n=1}^{\infty}$ in $X$ is a Cauchy sequence if for every $\varepsilon>0$ there exists $N\in\mathbb N$ such that for all $m,n\ge N$, \begin{align*} d_X(x_m,x_n)<\varepsilon. \end{align*} [/definition] This definition is internal to $X$: it does not require knowing whether the [limit point](/page/Limit%20Point) is in $X$. A Cauchy sequence carries small-scale information about all sufficiently late pairs of terms, so a uniformly continuous function should transfer that information to the target space. [quotetheorem:1052] This theorem distinguishes uniform continuity from ordinary continuity. A continuous function can map a Cauchy sequence to a sequence that is not Cauchy if the sequence approaches a missing point where the function has no controlled limiting behaviour. [example: A Continuous Map That Does Not Preserve Cauchy Sequences] Let $f:(0,1)\to\mathbb R$ be given by $f(x)=1/x$. First note that $f$ is continuous at every $a\in(0,1)$. If $|x-a|<\delta$ and $\delta\le a/2$, then $x>a-a/2=a/2$, so \begin{align*} |f(x)-f(a)|=\left|\frac{1}{x}-\frac{1}{a}\right|=\left|\frac{a-x}{ax}\right|=\frac{|x-a|}{a x}<\frac{2|x-a|}{a^2}. \end{align*} Given $\varepsilon>0$, choosing $\delta=\min\{a/2,\varepsilon a^2/2\}$ gives $|f(x)-f(a)|<\varepsilon$. Now consider the sequence $x_n=1/n$ for $n\ge2$, which lies in $(0,1)$. If $m,n\ge N$, then $1/m\le 1/N$ and $1/n\le 1/N$, so \begin{align*} |x_m-x_n|=\left|\frac{1}{m}-\frac{1}{n}\right|\le \frac{1}{m}+\frac{1}{n}\le \frac{2}{N}. \end{align*} Thus, for any $\varepsilon>0$, choosing $N>2/\varepsilon$ gives $|x_m-x_n|<\varepsilon$ whenever $m,n\ge N$, so $(x_n)$ is Cauchy in $(0,1)$. Its image sequence is \begin{align*} f(x_n)=f\left(\frac{1}{n}\right)=\frac{1}{1/n}=n. \end{align*} This sequence is not Cauchy in $\mathbb R$: taking $\varepsilon_0=1$, for every $N$ the indices $m=N+2$ and $n=N$ satisfy $m,n\ge N$ and \begin{align*} |f(x_m)-f(x_n)|=|m-n|=|(N+2)-N|=2>1. \end{align*} So a continuous function can send a Cauchy sequence in its domain to a non-Cauchy sequence in the target; ordinary continuity alone does not preserve Cauchy behaviour. [/example] The preservation of Cauchy sequences is the mechanism behind extension to completions. If a uniformly continuous function is defined on a dense subspace, then points in the completion can be approached by Cauchy sequences from the original space, and the images form Cauchy sequences in the target. ### Dense Subspaces and Extension Extension from a smaller space to a larger one requires a way to approximate every new point by old points. Density supplies that approximation scheme, and uniform continuity makes the value assigned to a limiting point independent of the approximating sequence. [definition: Dense Subset of a Metric Space] Let $(X,d_X)$ be a metric space. A subset $A\subset X$ is dense in $X$ if for every $x\in X$ and every $r>0$, there exists $a\in A$ such that \begin{align*} d_X(x,a)<r. \end{align*} [/definition] Density says that $A$ can approximate every point of $X$ at every scale. The next problem is whether a function defined only on $A$ assigns compatible values to different approximating sequences. Uniform continuity gives exactly the compatibility needed for a well-defined extension into a complete target, which is the content of the [extension theorem](/theorems/59). [quotetheorem:964] The important point is that uniform continuity supplies a definition of $F(x)$ that does not depend on which sequence in $A$ approaches $x$. The same global epsilon-delta control that made the values well-defined also passes to the extension. [example: Extending a Function from $\mathbb Q$ to $\mathbb R$] Let $A=\mathbb Q\cap[-1,1]$, and define $f:A\to\mathbb R$ by $f(q)=q^2$, with the Euclidean subspace metric on $A$. For $q,r\in A$, we have $|q|\le 1$ and $|r|\le 1$, so by the triangle inequality, \begin{align*} |q+r|\le |q|+|r|\le 1+1=2. \end{align*} Also, \begin{align*} |f(q)-f(r)|=|q^2-r^2|=|(q-r)(q+r)|=|q-r|\,|q+r|. \end{align*} Using the bound on $|q+r|$, this gives \begin{align*} |f(q)-f(r)|\le 2|q-r|. \end{align*} Thus, given $\varepsilon>0$, choosing $\delta=\varepsilon/2$ ensures that whenever $q,r\in A$ and $|q-r|<\delta$, \begin{align*} |f(q)-f(r)|\le 2|q-r|<2\delta=\varepsilon. \end{align*} So $f$ is uniformly continuous on $A$. The set $A=\mathbb Q\cap[-1,1]$ is dense in $[-1,1]$, since every point of $[-1,1]$ can be approximated arbitrarily closely by rational numbers still lying in $[-1,1]$. By *Uniform Extension from a [Dense Subset](/page/Dense%20Subset)*, $f$ has a unique uniformly continuous extension $F:[-1,1]\to\mathbb R$. To identify it, let $x\in[-1,1]$ and choose a sequence $(q_n)$ in $A$ with $q_n\to x$. Then \begin{align*} |q_n^2-x^2|=|q_n-x|\,|q_n+x|. \end{align*} Since $q_n,x\in[-1,1]$, we have $|q_n+x|\le |q_n|+|x|\le 2$, so \begin{align*} |q_n^2-x^2|\le 2|q_n-x|. \end{align*} Because $q_n\to x$, the right-hand side tends to $0$, hence $q_n^2\to x^2$. Therefore the extension is \begin{align*} F(x)=x^2 \end{align*} for every $x\in[-1,1]$. This example shows how a uniformly controlled formula on a dense rational subspace determines its real-valued extension without any extra choices. [/example] This is the analytic reason dense subspaces are powerful. Once a construction is uniformly controlled on a dense set, it can often be completed without making arbitrary choices. ## Uniform Equivalence and Metric Structure Uniform continuity depends on metrics, so it is natural to ask when two metric spaces have the same uniform structure. Homeomorphisms preserve open sets and ordinary continuity. Uniform homeomorphisms preserve global small-scale control in both directions. ### Uniform Homeomorphisms A bijection may be continuous with continuous inverse while still distorting small distances too violently near infinity or near missing boundary points. Ordinary homeomorphism therefore does not guarantee that estimates transported through the map remain uniform. The stronger [equivalence relation](/page/Equivalence%20Relation) requires uniform control in both directions, so neither space loses its global small-scale structure when used as coordinates for the other. [definition: Uniform Homeomorphism] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces. A function $f:X\to Y$ is a uniform homeomorphism if $f$ is bijective, $f$ is uniformly continuous, and $f^{-1}:Y\to X$ is uniformly continuous. [/definition] Uniform homeomorphism means that each space can use the other as a coordinate system without losing uniform estimates. It is stricter than topological equivalence. [example: A Homeomorphism That Is Not Uniform in Both Directions] Let $f:(0,\infty)\to(0,\infty)$ be given by $f(x)=x^2$, with Euclidean metrics. The map is bijective: if $x>0$, then $\sqrt{x^2}=x$, and if $y>0$, then $f(\sqrt y)=(\sqrt y)^2=y$. Its inverse is therefore $f^{-1}(y)=\sqrt y$. The map $f$ is continuous at each $a>0$. If $|x-a|<\delta$ and $\delta\le 1$, then $x<a+1$, so $x+a<2a+1$. Hence \begin{align*} |f(x)-f(a)|=|x^2-a^2|=|(x-a)(x+a)|=|x-a|\,|x+a|<\delta(2a+1). \end{align*} Given $\varepsilon>0$, choosing $\delta=\min\{1,\varepsilon/(2a+1)\}$ gives $|f(x)-f(a)|<\varepsilon$. The inverse is uniformly continuous. For $u,v>0$, assume first that $u\ge v$. Since $v^2\le uv$ and both $v$ and $\sqrt{uv}$ are nonnegative, we have $v\le\sqrt{uv}$. Therefore \begin{align*} |\sqrt u-\sqrt v|^2=(\sqrt u-\sqrt v)^2=u-2\sqrt{uv}+v\le u-2v+v=u-v=|u-v|. \end{align*} If $v\ge u$, the same argument with $u$ and $v$ exchanged gives $|\sqrt u-\sqrt v|^2\le |u-v|$. Taking square roots yields \begin{align*} |\sqrt u-\sqrt v|\le |u-v|^{1/2}. \end{align*} Given $\varepsilon>0$, choose $\delta=\varepsilon^2$. If $|u-v|<\delta$, then \begin{align*} |f^{-1}(u)-f^{-1}(v)|=|\sqrt u-\sqrt v|\le |u-v|^{1/2}<\delta^{1/2}=\varepsilon. \end{align*} Thus $f^{-1}$ is uniformly continuous. However, $f$ is not uniformly continuous. Fix $\varepsilon_0=1$. Given any $\delta>0$, choose $n\in\mathbb N$ with $1/n<\delta$, and set $x_n=n$ and $y_n=n+1/n$. Then \begin{align*} |x_n-y_n|=\left|n-\left(n+\frac{1}{n}\right)\right|=\left|-\frac{1}{n}\right|=\frac{1}{n}<\delta. \end{align*} But the output distance is \begin{align*} |f(y_n)-f(x_n)|=\left|\left(n+\frac{1}{n}\right)^2-n^2\right|=\left|n^2+2+\frac{1}{n^2}-n^2\right|=2+\frac{1}{n^2}>1. \end{align*} So arbitrarily close input pairs can have squared values more than $1$ apart. Although $f$ is a homeomorphism and $f^{-1}$ is uniformly continuous, $f$ itself is not uniformly continuous; hence this homeomorphism is not a uniform homeomorphism. [/example] This example shows why uniform structure is finer than topology. The intervals $(0,\infty)$ and $(0,\infty)$ are topologically identical under $x\mapsto x^2$, but the map changes large-scale metric error too aggressively in one direction. ### Completeness as a Uniform Invariant Uniform homeomorphisms preserve the concepts built from Cauchy sequences. Completeness is not purely topological, because homeomorphic spaces can differ in whether Cauchy sequences converge. Since uniform homeomorphisms preserve Cauchy sequences in both directions, completeness becomes an invariant at the uniform level. [quotetheorem:9184] The theorem gives a useful test for proposed uniform equivalences. If one space is complete and the other is not, no uniform homeomorphism can exist between them, even when they are homeomorphic. [example: The Open Interval and the Real Line] Define $g:(-1,1)\to\mathbb R$ by $g(x)=\frac{x}{1-|x|}$. The map $g$ is a homeomorphism with inverse \begin{align*} g^{-1}(y)=\frac{y}{1+|y|}. \end{align*} Indeed, if $x\in(-1,1)$ and $y=g(x)$, then $|y|=|x|/(1-|x|)$, so \begin{align*} \frac{y}{1+|y|}=\frac{x/(1-|x|)}{1+|x|/(1-|x|)}=x. \end{align*} Conversely, for $y\in\mathbb R$ we have $|y|/(1+|y|)<1$, and \begin{align*} g\left(\frac{y}{1+|y|}\right)=\frac{y/(1+|y|)}{1-|y|/(1+|y|)}=y. \end{align*} The function $g$ is continuous on $(-1,1)$ because the denominator is positive there, and $g^{-1}$ is continuous by the same kind of algebraic formula with a positive denominator. We now show that $g$ is not uniformly continuous. For each $n\in\mathbb N$, set \begin{align*} x_n=\frac{n}{n+1} \end{align*} and \begin{align*} y_n=\frac{n+1}{n+2}. \end{align*} Then $x_n,y_n\in(-1,1)$, and \begin{align*} g(x_n)=n,\qquad g(y_n)=n+1, \end{align*} so $|g(y_n)-g(x_n)|=1$. Their input distance is \begin{align*} |y_n-x_n|=\frac{n+1}{n+2}-\frac{n}{n+1}=\frac{1}{(n+1)(n+2)}. \end{align*} Thus $|y_n-x_n|\to0$ while $|g(y_n)-g(x_n)|=1$ for every $n$. Taking $\varepsilon_0=1/2$, no single input tolerance can force all output distances below $\varepsilon_0$, so $g$ is not uniformly continuous. Hence this homeomorphism from $(-1,1)$ to $\mathbb R$ is not a uniform homeomorphism. A continuous change of coordinates can stretch the missing endpoints to infinity topologically, but it cannot remove the small-scale metric distortion near those endpoints. [/example] Uniform equivalence is therefore the right level of sameness for questions involving Cauchy sequences, completions, quantitative approximation, and global error bounds. ## Beyond and Connected Topics Uniform [continuity between metric spaces](/page/Continuity%20Between%20Metric%20Spaces) sits between elementary continuity and the broader theory of uniform spaces. It is the metric version of a structure that remembers which pairs of points are uniformly close, rather than only which sets are open. In later topology, uniform spaces generalize this idea beyond metrics while retaining the language needed for Cauchy filters and completions. The compactness theorem connects this page directly to [Cambridge IB Analysis and Topology](/page/Cambridge%20IB%20Analysis%20and%20Topology), where compact metric spaces, [sequential compactness](/page/Sequential%20Compactness), and continuity theorems form a central toolkit. Heine-Cantor is one of the main bridges between pointwise epsilon-delta reasoning and global compactness arguments, and it explains why compact domains behave differently from bounded but nonclosed intervals. The examples involving intervals, reciprocal functions, and oscillation build on the epsilon-delta foundations of [Cambridge IA Analysis Notes](/page/Cambridge%20IA%20Analysis%20Notes). At that level, the key skill is learning to read the quantifier order: continuity chooses a radius after fixing a point, while uniform continuity chooses one radius for all points. This distinction is also the starting point for understanding why Cauchy sequences and completions require uniform rather than merely topological control. In analysis of functions, uniform continuity becomes part of the compactness language behind equicontinuity and Arzela-Ascoli type theorems. This direction belongs naturally with [Cambridge II Analysis of Functions](/page/Cambridge%20II%20Analysis%20of%20Functions), where families of functions are controlled by shared moduli rather than by pointwise estimates alone. There is also a complex-analytic version of the same theme. Holomorphic functions are continuous, and on compact subsets they are uniformly continuous. In [Cambridge IB Complex Analysis](/page/Cambridge%20IB%20Complex%20Analysis), this compact-subset viewpoint is constantly used when estimates are made on contours, closed discs, and compact subsets of domains. ## References Androma, [Cambridge IB Analysis and Topology](/page/Cambridge%20IB%20Analysis%20and%20Topology). Androma, [Cambridge IA Analysis Notes](/page/Cambridge%20IA%20Analysis%20Notes). Androma, [Cambridge IB Complex Analysis](/page/Cambridge%20IB%20Complex%20Analysis). Androma, [Cambridge II Analysis of Functions](/page/Cambridge%20II%20Analysis%20of%20Functions). Walter Rudin, *Principles of Mathematical Analysis* (1976). James R. Munkres, *Topology* (2000). Stephen Abbott, *Understanding Analysis* (2015).