A sequence of functions can fail to converge for two different reasons. It may fail at a particular input, or it may converge at every input while the error remains large somewhere else. The uniform metric is designed to measure the second phenomenon: it asks for the largest possible output error over the whole domain, rather than checking each input in isolation. The construction is deliberately insensitive to algebra. The domain $S$ is only a set, while the codomain $(Y,e)$ is a [metric space](/page/Metric%20Space). This makes the uniform metric available before vector spaces, integrals, differentiability, or topology on the domain have entered the discussion. It is the metric structure on $Y$ that supplies the notion of error. The price for measuring the largest error is that the largest error must be finite. This is why bounded functions, or more precisely functions whose images are bounded in $Y$, are the natural arena. Without boundedness the same formula may take the value $\infty$, which is not allowed for an ordinary metric. [example: Escaping Error on an Unbounded Domain] Let $S=\mathbb R$, let $Y=\mathbb R$ with $e(a,b)=|a-b|$, and define $f,g:\mathbb R\to\mathbb R$ by $f(x)=0$ and $g(x)=x$. For each fixed $x\in\mathbb R$, the pointwise error is the finite real number \begin{align*} e(f(x),g(x))=|f(x)-g(x)|=|0-x|=|-x|=|x|. \end{align*} The collection of all pointwise errors is unbounded above. Indeed, if $M\ge 0$, choose $x=M+1$. Then $x\in\mathbb R$ and \begin{align*} |f(x)-g(x)|=|0-(M+1)|=M+1>M. \end{align*} Thus no finite real number bounds the set $\{|f(x)-g(x)|:x\in\mathbb R\}$ from above, so \begin{align*} \sup_{x\in\mathbb R}|f(x)-g(x)|=\infty. \end{align*} The largest-error formula therefore does not produce an ordinary finite distance for these two functions; restricting to bounded functions is what prevents this kind of error from escaping to infinity across the domain. [/example] The example shows why the word "uniform" is not cosmetic. A pointwise comparison can inspect every input separately and still miss the fact that the errors escape to infinity across the domain. The uniform metric turns that global failure into a single numerical distance. ## Bounded Function Spaces Before defining the distance between two functions, we need to say which functions are allowed. In a general metric codomain there is no preferred zero vector, so boundedness cannot mean $|f(s)| \le M$. It must mean that the image of the function has finite diameter. [definition: Bounded Function into a Metric Space] Let $S$ be a set and let $(Y,e)$ be a metric space. If $S=\varnothing$, the unique function $f:\varnothing\to Y$ is bounded by convention. If $S\neq\varnothing$, a function $f: S \to Y$ is bounded if there exists $y_0 \in Y$ and $M \ge 0$ such that \begin{align*} e(f(s),y_0) \le M \end{align*} for every $s \in S$. [/definition] In a metric space, the choice of base point is not part of the data. If the image of $f$ lies in a ball around one point of $Y$, then it lies in a possibly larger ball around any other fixed point of $Y$. For real-valued functions this recovers the usual condition that $|f(s)|$ is bounded above. We need a named ambient collection of exactly these functions, because the uniform distance will be finite only on that collection. [definition: Space of Bounded Functions] Let $S$ be a set and let $(Y,e)$ be a metric space. The space of bounded functions from $S$ to $Y$ is \begin{align*} B(S,Y) = \{f:S \to Y : f \text{ is bounded}\}. \end{align*} [/definition] The notation $B(S,Y)$ now gives a precise domain on which the largest-error formula can be finite. With that support in place, the page's central object can be defined directly. ## Definition Now the distance between two bounded functions is obtained by looking at every input at once. The formula records the worst output error, using the metric $e$ in the codomain. [definition: Uniform Metric on Bounded Functions] Let $S$ be a set and let $(Y,e)$ be a metric space. For $f,g \in B(S,Y)$, the uniform metric, also called the sup metric, is the function $D: B(S,Y) \times B(S,Y) \to [0,\infty)$ defined by \begin{align*} D(f,g)=\sup_{s \in S} e(f(s),g(s)). \end{align*} If $S=\varnothing$, the supremum over the empty index set is defined to be $0$. [/definition] The supremum is finite because the two images are bounded subsets of $Y$. If $f(S)$ lies in a ball of radius $M_f$ around $y_f$ and $g(S)$ lies in a ball of radius $M_g$ around $y_g$, then the triangle inequality gives \begin{align*} e(f(s),g(s)) \le M_f + e(y_f,y_g) + M_g \end{align*} for every $s \in S$. A supremum is used rather than a maximum because the worst error might be approached without being attained. This distinction matters as soon as the domain is not finite or compact. [remark: Supremum Need Not Be Maximum] The notation $\sup$ is essential. If $S=(0,1)$, $Y=\mathbb R$, $f(x)=x$, and $g(x)=0$, then \begin{align*} D(f,g)=\sup_{x \in (0,1)} |x| = 1, \end{align*} but there is no $x \in (0,1)$ at which $|f(x)-g(x)|=1$. [/remark] ## Metric Structure ### Completeness of the Uniform Metric The uniform distance turns the set of bounded functions into a metric space; once that structure is in place, the next natural question is whether this metric space has the same completeness behavior as the target space. We write $\ell_\infty(S,Y)$ for this same bounded function space $B(S,Y)$ when we want to emphasize the uniform, or sup-distance, metric on it. The issue is that a [Cauchy sequence](/page/Cauchy%20Sequence) of functions first produces only pointwise candidate values in $Y$. The theorem below records that completeness of $Y$ is exactly the condition that lets those pointwise limits assemble into another bounded function and that the convergence is still uniform. [quotetheorem:288] This result is stronger than the metric axioms alone. It says that a uniformly Cauchy sequence of bounded functions has a pointwise candidate limit and that the uniform control is strong enough to keep the limit bounded and force convergence back in the same space. Thus no new limiting objects are created outside $B(S,Y)$, provided the target metric space $Y$ itself has no missing limits. The empty domain produces a harmless convention issue: the supremum of an empty subset of $[0,\infty)$ is often taken to be $0$ in this setting. Most mathematical uses either assume $S$ is nonempty or adopt that convention explicitly. Finite domains strip the construction down to its simplest form. When there are only finitely many inputs, the supremum is a maximum and the distance is controlled by the largest coordinate-wise error. [example: Finite Domain] Let $S=\{1,2,3\}$ and let $Y=\mathbb R$ with the usual metric $e(a,b)=|a-b|$. For the functions $f,g:S\to\mathbb R$ defined by $f(1)=2$, $f(2)=-1$, $f(3)=4$, $g(1)=1$, $g(2)=5$, and $g(3)=2$, we compute the uniform distance by evaluating the pointwise errors at all three points of $S$. At $1$, \begin{align*} e(f(1),g(1))=|f(1)-g(1)|=|2-1|=|1|=1. \end{align*} At $2$, \begin{align*} e(f(2),g(2))=|f(2)-g(2)|=|-1-5|=|-6|=6. \end{align*} At $3$, \begin{align*} e(f(3),g(3))=|f(3)-g(3)|=|4-2|=|2|=2. \end{align*} Since the domain has exactly the three points $1,2,3$, the supremum over $S$ is the maximum of these three displayed values: \begin{align*} D(f,g)=\sup_{s\in S}e(f(s),g(s))=\max\{1,6,2\}=6. \end{align*} The uniform distance is therefore controlled by the second point of the domain, where the two function values differ by $6$. [/example] ### Normed Codomains When $Y$ is a [normed vector space](/page/Normed%20Vector%20Space), the metric can be expressed through a norm of the difference $f-g$. This is useful because many analytic arguments estimate the size of a function rather than the distance between two separately named functions. [definition: Sup Norm] Let $S$ be a set and let $(V,\|\cdot\|_V)$ be a normed [vector space](/page/Vector%20Space). The sup norm is the function $\|\cdot\|_\infty:B(S,V)\to[0,\infty)$ defined by \begin{align*} \|u\|_\infty = \sup_{s \in S} \|u(s)\|_V. \end{align*} If $S=\varnothing$, the supremum over the empty index set is defined to be $0$. [/definition] With this notation, the uniform metric associated to the norm metric on $V$ is \begin{align*} D(f,g)=\|f-g\|_\infty. \end{align*} This identity is often the bridge from metric-space language to functional analysis. The formula for $\|u\|_\infty$ is useful only if it behaves like a genuine norm rather than just a convenient size function. The possible obstruction is the triangle inequality: taking a supremum after adding functions must still be controlled by the separate suprema. Once that is checked, linear estimates, contraction arguments, and Banach-space methods can use bounded functions in the same way they use vectors. [quotetheorem:8711] The norm formulation adds linear structure to the same largest-error idea: the size of a bounded function is its largest output size, and the distance between two functions is the size of their difference. This matters because it lets one use the algebra of functions, not just the metric between them. The theorem is useful precisely when estimates are first obtained pointwise in $V$ and then need to be promoted to estimates for whole functions. If $\|u(s)\|_V\le M$ for every $s$, then $\|u\|_\infty\le M$; if $\|u(s)+v(s)\|_V\le\|u(s)\|_V+\|v(s)\|_V$ pointwise, then the same inequality survives after taking the supremum over $S$. The boundedness hypothesis is the limitation: without it, the displayed formula may be infinite, so the space would no longer be an ordinary normed vector space. This is why later approximation, perturbation, and fixed-point arguments are naturally stated inside $B(S,V)$ rather than on all functions $S\to V$: the sup norm turns uniform error bounds into ordinary norm estimates. ## Uniform Convergence ### Metric Convergence of Functions The uniform metric turns a qualitative convergence condition into ordinary metric convergence. This matters because sequences of functions are often studied by estimating a single error term that must work for all inputs. [definition: Uniform Convergence] Let $S$ be a set, let $(Y,e)$ be a metric space, and let $f_k:S\to Y$ be a sequence of functions. The sequence $(f_k)_{k=1}^\infty$ converges uniformly to $f:S\to Y$ if for every $\varepsilon>0$ there exists $N\in\mathbb N$ such that for all $k\ge N$ and all $s\in S$, \begin{align*} e(f_k(s),f(s))<\varepsilon. \end{align*} [/definition] Uniform convergence asks for one index $N$ that works across the entire domain. Pointwise convergence allows $N$ to depend on $s$, which is often too weak to preserve continuity, boundedness, or interchange of limiting operations. The previous definition was phrased without mentioning $D$, so it is not yet clear that the metric topology on $B(S,Y)$ has captured the same notion. The issue is whether the single quantified error bound in uniform convergence is exactly the same data as the statement that the supremum distance from $f_k$ to $f$ tends to zero. This equivalence is the bridge needed before using metric-space language for sequences of functions. Without it, a statement such as $D(f_k,f)\to0$ would be only a convenient shorthand; with it, convergence in the uniform metric is exactly uniform convergence as defined by the quantifiers above. The next point is therefore a compatibility test between two languages: the quantified language of uniform convergence and the metric language of convergence in $B(S,Y)$. The theorem records that these languages impose the same error control, so later metric-space arguments can be read directly as statements about uniform convergence. [quotetheorem:8712] The theorem is the main reason for introducing $D$. It lets the tools of metric spaces, such as Cauchy sequences, open balls, closures, and completeness, apply directly to uniform convergence of functions. Pointwise convergence can look convincing while missing a persistent moving error. The standard powers on an interval show the failure with a computation that the uniform metric detects immediately. [example: Pointwise but Not Uniform] Let $S=(0,1)$, let $Y=\mathbb R$, and define $f_k:(0,1)\to\mathbb R$ by $f_k(x)=x^k$. We first compute the pointwise limit. Fix $x\in(0,1)$ and write $x=1/(1+t)$ with $t=1/x-1>0$. For every $k\in\mathbb N$, Bernoulli's inequality gives $(1+t)^k\ge 1+kt$, so \begin{align*} 0\le x^k=\frac{1}{(1+t)^k}\le \frac{1}{1+kt}. \end{align*} Since $1/(1+kt)\to0$, the squeeze property gives $x^k\to0$. Thus $f_k$ converges pointwise to the zero function. The uniform distance from $f_k$ to $0$ is not small. For every $x\in(0,1)$, \begin{align*} |f_k(x)-0|=|x^k|=x^k<1. \end{align*} Hence $\sup_{x\in(0,1)} x^k\le 1$. Conversely, if $c<1$, choose $x\in(0,1)$ with $\max\{0,c\}^{1/k}<x<1$. Then \begin{align*} x^k>\max\{0,c\}\ge c. \end{align*} So every number below $1$ fails to be an upper bound for the set $\{x^k:x\in(0,1)\}$, while $1$ is an upper bound. Therefore \begin{align*} D(f_k,0)=\sup_{x\in(0,1)}|f_k(x)-0|=\sup_{x\in(0,1)}x^k=1. \end{align*} Since this distance equals $1$ for every $k$, it does not tend to $0$; the convergence is pointwise but not uniform. [/example] The example explains why the uniform metric is stronger than pointwise convergence. The error near $x=1$ does not disappear uniformly, even though each fixed point eventually sees the error go to zero. ### Uniform Neighbourhoods Metric convergence is controlled by metric balls, so it is useful to spell out what balls in the uniform metric mean. They are not local neighbourhoods around points of the domain; they are tubes around an entire function. [definition: Uniform Ball Around a Function] Let $S$ be a set, let $(Y,e)$ be a metric space, let $f\in B(S,Y)$, and let $r>0$. The open uniform ball of radius $r$ around $f$ is \begin{align*} B_D(f,r)=\{g\in B(S,Y):D(f,g)<r\}. \end{align*} [/definition] This description is especially useful for estimates: proving $g\in B_D(f,r)$ means producing a single bound $e(f(s),g(s))<r$ valid for every $s\in S$. ## Completeness and Closed Function Classes ### Cauchy Sequences of Functions The uniform metric is not only a way to define convergence. It also transfers completeness from the codomain to the function space. This is the mechanism behind many existence arguments: construct a Cauchy sequence of approximate functions, then take a uniform limit. [definition: Uniformly Cauchy Sequence] Let $S$ be a set, let $(Y,e)$ be a metric space, and let $(f_k)_{k=1}^\infty$ be a sequence in $B(S,Y)$. The sequence is uniformly Cauchy if for every $\varepsilon>0$ there exists $N\in\mathbb N$ such that for all $j,k\ge N$, \begin{align*} \sup_{s\in S} e(f_j(s),f_k(s))<\varepsilon. \end{align*} [/definition] This condition says that the functions eventually become close to each other with one uniform error tolerance across $S$. If $Y$ is complete, each pointwise sequence $f_k(s)$ has a limit in $Y$, and the uniform Cauchy condition prevents those pointwise limits from forming an unbounded function. A Cauchy condition is useful only if it produces a limit inside the space under discussion. Here there are two possible failures: the pointwise limits might not exist in $Y$, or they might exist but fail to assemble into a bounded function that is reached uniformly. Completeness of the codomain removes the first obstruction, and the uniform Cauchy bound is what controls the second. [quotetheorem:288] This theorem is a prototype for many completion arguments in analysis. A complete codomain allows pointwise limits to exist; the uniform metric upgrades those pointwise limits into convergence inside the same function space. The case $S=\mathbb N$ turns bounded functions into bounded sequences. This example connects the uniform metric with the classical sequence space $\ell^\infty$. [example: Bounded Real Sequences] Take $S=\mathbb N$ and $Y=\mathbb R$ with the usual metric. A function $a:\mathbb N\to\mathbb R$ is the same thing as a real sequence $(a_n)_{n\in\mathbb N}$, and bounded functions are exactly bounded real sequences, so $B(\mathbb N,\mathbb R)=\ell^\infty$. For $a,b\in \ell^\infty$, the uniform metric is \begin{align*} D(a,b)=\sup_{n\in\mathbb N}|a_n-b_n|. \end{align*} Define $a^{(k)}\in\ell^\infty$ by $a^{(k)}_n=1/n$ for $n\le k$ and $a^{(k)}_n=0$ for $n>k$, and define $a\in\ell^\infty$ by $a_n=1/n$. Both are bounded because $0\le a^{(k)}_n\le 1$ and $0<a_n\le 1$ for every $n\in\mathbb N$. For $n\le k$, \begin{align*} |a^{(k)}_n-a_n|=\left|\frac{1}{n}-\frac{1}{n}\right|=0. \end{align*} For $n>k$, \begin{align*} |a^{(k)}_n-a_n|=\left|0-\frac{1}{n}\right|=\frac{1}{n}. \end{align*} Hence the set of all pointwise errors is \begin{align*} \{|a^{(k)}_n-a_n|:n\in\mathbb N\}=\{0\}\cup\left\{\frac{1}{n}:n>k\right\}. \end{align*} If $n>k$, then $n\ge k+1$, so \begin{align*} \frac{1}{n}\le \frac{1}{k+1}. \end{align*} The value $1/(k+1)$ is attained at $n=k+1$, so it is the supremum of the displayed set. Therefore \begin{align*} D(a^{(k)},a)=\sup_{n\in\mathbb N}|a^{(k)}_n-a_n|=\frac{1}{k+1}. \end{align*} Since $1/(k+1)\to0$, the sequence of sequences $a^{(k)}$ converges uniformly to $a$ in $\ell^\infty$. The sup metric is measuring these sequences as functions on $\mathbb N$, so convergence means that the largest coordinate error tends to zero. [/example] ### Continuous Bounded Functions Completeness becomes more useful when combined with closed subspaces. In the real-valued setting, the natural candidate is the subspace of bounded functions that also respect the topology of the domain. The key question is whether the uniform limit process used in the bounded-function space can leave this continuous subspace. [definition: Bounded Continuous Functions] Let $(S,d)$ be a metric space. The space of bounded continuous real-valued functions on $S$ is \begin{align*} C_b(S,\mathbb R)=\{f:S\to \mathbb R : f \text{ is continuous and bounded}\}. \end{align*} [/definition] Boundedness alone is already preserved by uniform limits in the bounded-function metric. To see that $C_b(S,\mathbb R)$ is stable under the same limit operation, we also need continuity to survive uniform convergence. This is exactly the role of the real-valued [uniform limit theorem](/theorems/258) on a metric domain. [quotetheorem:258] Thus, inside the [complete metric space](/page/Complete%20Metric%20Space) of bounded real-valued functions on $S$, the subset $C_b(S,\mathbb R)$ is closed under uniform limits. The theorem fails for pointwise convergence, and this failure is one of the reasons the uniform metric is central in analysis. [example: Pointwise Limit Can Lose Continuity] Let $S=[0,1]$, let $Y=\mathbb R$, and define $f_k:[0,1]\to\mathbb R$ by $f_k(x)=x^k$. Each $f_k$ is continuous because it is the restriction to $[0,1]$ of a polynomial function. We compute the pointwise limit. If $x=0$, then $f_k(0)=0^k=0$ for every $k$. If $0<x<1$, write $x=1/(1+t)$ with $t=1/x-1>0$. Bernoulli's inequality gives $(1+t)^k\ge 1+kt$, so \begin{align*} 0\le x^k=\frac{1}{(1+t)^k}\le \frac{1}{1+kt}. \end{align*} Since $1/(1+kt)\to0$, we get $x^k\to0$. At $x=1$, \begin{align*} f_k(1)=1^k=1 \end{align*} for every $k$. Thus the pointwise limit is the function $f:[0,1]\to\mathbb R$ given by $f(x)=0$ for $0\le x<1$ and $f(1)=1$. This limit is not continuous at $1$. Indeed, take $\varepsilon=1/2$. For every $\delta>0$, set $x=1-\min\{\delta/2,1/2\}$. Then $0\le x<1$, $|x-1|=\min\{\delta/2,1/2\}<\delta$, and \begin{align*} |f(x)-f(1)|=|0-1|=1>\frac12. \end{align*} The uniform metric detects the same failure. For $0\le x<1$, \begin{align*} |f_k(x)-f(x)|=|x^k-0|=x^k\le 1. \end{align*} At $x=1$, \begin{align*} |f_k(1)-f(1)|=|1-1|=0. \end{align*} Hence $D(f_k,f)\le1$. Conversely, if $0<\varepsilon<1$, choose $x=(1-\varepsilon)^{1/k}$. Then $0<x<1$, so $f(x)=0$, and \begin{align*} |f_k(x)-f(x)|=\left|\left((1-\varepsilon)^{1/k}\right)^k-0\right|=1-\varepsilon. \end{align*} Therefore the supremum of the errors is at least $1-\varepsilon$ for every $0<\varepsilon<1$, and it is at most $1$, so \begin{align*} D(f_k,f)=\sup_{x\in[0,1]}|f_k(x)-f(x)|=1. \end{align*} Thus $D(f_k,f)$ equals $1$ for every $k\in\mathbb N$, so the convergence is pointwise but not uniform, and the pointwise limit can lose continuity. [/example] The previous theorem says that $C_b(X,Y)$ is closed inside $B(X,Y)$ under uniform limits. The remaining question is whether Cauchy sequences in $C_b(X,Y)$ have limits inside $C_b(X,Y)$; the completeness theorem answers yes when $Y$ is complete. [quotetheorem:8713] This result is the metric-space ancestor of many Banach spaces of functions. For example, when $Y=\mathbb R$ or $Y=\mathbb C$, the space $C_b(X,Y)$ carries the sup norm and becomes a [Banach space](/page/Banach%20Space). ## Comparison with Other Ways to Measure Functions ### Pointwise Control Among the standard elementary comparisons on pointwise functions, the uniform metric is the one that demands direct control at every point. Other modes of convergence may be more flexible, but they often allow the largest error to move through the domain without becoming small. [definition: Pointwise Convergence] Let $S$ be a set and let $(Y,e)$ be a metric space. A sequence $f_k:S\to Y$ converges pointwise to $f:S\to Y$ if for every $s\in S$, \begin{align*} e(f_k(s),f(s))\to0 \end{align*} as $k\to\infty$. [/definition] This definition keeps the inputs independent. It is useful when the domain is large and no uniform control is available, but it is too weak for many limiting operations. Since a uniform estimate holds for all inputs at once, it should imply every fixed-input estimate, but this implication still has to pass through the definitions carefully: a single $N$ must be converted into a valid choice for each fixed $s$. This comparison also identifies the exact point where pointwise convergence is weaker, because the reverse direction would require those fixed-input choices to be made uniformly. [quotetheorem:8714] The converse fails, as the examples $x^k$ on $(0,1)$ and on $[0,1]$ show in different ways. Pointwise convergence loses information about where the largest error occurs. ### Integral Control Integral metrics measure average error rather than worst-case error. This is often the right choice in measure theory, but it is not the same geometry as the uniform metric. [definition: $L^1$ Distance on Integrable Equivalence Classes] Let $(E,\mathcal E,\mu)$ be a [measure space](/page/Measure%20Space). The $L^1$ distance on integrable equivalence classes is the function \begin{align*} d_1:L^1(E,\mathcal E,\mu)\times L^1(E,\mathcal E,\mu)\to[0,\infty) \end{align*} defined, for equivalence classes of integrable functions $f,g:E\to\mathbb R$ with functions equal $\mu$-a.e. identified, by \begin{align*} d_1(f,g)=\int_E |f-g|\,d\mu. \end{align*} [/definition] On actual pointwise functions the same formula is only a pseudometric, since [changing a function on a null set](/theorems/4915) does not change the integral. The $L^1$ distance can be small even when the uniform distance is large, because a tall error supported on a small set may have small integral. [example: Narrow Spikes] Here $\mathcal L^1$ denotes one-dimensional [Lebesgue measure](/page/Lebesgue%20Measure) on $[0,1]$. Let $S=[0,1]$, let $Y=\mathbb R$, and for each $k\ge 1$ define $f_k:[0,1]\to\mathbb R$ by $f_k(x)=1$ when $0\le x\le 1/k$ and $f_k(x)=0$ when $1/k<x\le1$. Since $f_k$ only takes the values $0$ and $1$, we have $|f_k(x)|=f_k(x)$ for every $x\in[0,1]$, so $|f_k|$ is the indicator function of $[0,1/k]$. Hence, by the defining property of the [Lebesgue integral](/page/Lebesgue%20Integral) of an indicator function, \begin{align*} \int_{[0,1]} |f_k|\,d\mathcal L^1=\mathcal L^1([0,1/k])=\frac{1}{k}-0=\frac{1}{k}. \end{align*} Since $1/k\to0$, the $L^1$ distance from $f_k$ to $0$ tends to $0$. The uniform distance records the largest pointwise error instead. For every $x\in[0,1]$, the value $|f_k(x)-0|$ is either $|1-0|=1$ or $|0-0|=0$, so $|f_k(x)-0|\le1$. Also $0\in[0,1/k]$, so \begin{align*} |f_k(0)-0|=|1-0|=1. \end{align*} Thus $1$ is an upper bound for the set of pointwise errors and is actually attained, so \begin{align*} D(f_k,0)=\sup_{x\in[0,1]} |f_k(x)-0|=1. \end{align*} The integral sees that the support $[0,1/k]$ has length $1/k$, while the uniform metric sees that the spike still has height $1$ for every $k$. [/example] This distinction explains why sup norms are natural for uniform approximation and stability estimates, while $L^p$ norms are natural for measure-theoretic averaging and weak formulations of PDEs. ## Approximation and Stability Uniform estimates are valuable because they are stable under composition with uniformly continuous maps. If the output error is small everywhere and a second map does not amplify small errors unpredictably, then the composed functions remain close everywhere. [definition: Uniform Continuity] Let $(Y,e)$ and $(Z,d)$ be metric spaces. A function $\Phi:Y\to Z$ is uniformly continuous if for every $\varepsilon>0$ there exists $\delta>0$ such that for all $y_1,y_2\in Y$, \begin{align*} e(y_1,y_2)<\delta \implies d(\Phi(y_1),\Phi(y_2))<\varepsilon. \end{align*} [/definition] [Uniform continuity](/page/Uniform%20Continuity) is the correct hypothesis because the same $\delta$ must work at every output value attained by the functions. Ordinary continuity can fail to provide such global control on unbounded subsets. The next question is whether this uniform control is enough to pass uniform convergence through composition. [quotetheorem:8715] This theorem is a common tool for passing limits through nonlinear operations. It is especially useful when all functions take values in a compact subset of $Y$, since continuous maps on compact metric spaces are uniformly continuous. [example: Squaring Bounded Uniform Limits] Let $S$ be a set and let $f_k,f:S\to\mathbb R$ be bounded real-valued functions with $D(f_k,f)\to0$. Suppose there exists $M>0$ such that $|f_k(s)|\le M$ and $|f(s)|\le M$ for every $s\in S$ and every $k\in\mathbb N$. For each $s\in S$, factor the difference of squares: \begin{align*} f_k(s)^2-f(s)^2=(f_k(s)-f(s))(f_k(s)+f(s)). \end{align*} Taking absolute values and using $|ab|=|a||b|$ gives \begin{align*} |f_k(s)^2-f(s)^2|=|f_k(s)-f(s)|\,|f_k(s)+f(s)|. \end{align*} By the triangle inequality in $\mathbb R$ and the common bound by $M$, \begin{align*} |f_k(s)+f(s)|\le |f_k(s)|+|f(s)|\le M+M=2M. \end{align*} Therefore, for every $s\in S$, \begin{align*} |f_k(s)^2-f(s)^2|\le 2M|f_k(s)-f(s)|. \end{align*} Also, by the definition of the uniform metric, \begin{align*} |f_k(s)-f(s)|\le \sup_{t\in S}|f_k(t)-f(t)|=D(f_k,f). \end{align*} Combining the last two inequalities gives, for every $s\in S$, \begin{align*} |f_k(s)^2-f(s)^2|\le 2M D(f_k,f). \end{align*} Since $2M D(f_k,f)$ is an upper bound for all pointwise squared errors, their supremum satisfies \begin{align*} D(f_k^2,f^2)=\sup_{s\in S}|f_k(s)^2-f(s)^2|\le 2M D(f_k,f). \end{align*} Because $D(f_k,f)\to0$ and $2M$ is fixed, $2M D(f_k,f)\to0$, so $D(f_k^2,f^2)\to0$. Thus the squared functions $f_k^2$ converge uniformly to $f^2$; the common bound by $M$ prevents the squaring operation from amplifying small uniform errors without control. [/example] The boundedness hypothesis in the example is not decorative. Squaring is not uniformly continuous on all of $\mathbb R$, and uniform convergence alone does not control nonlinear growth unless the relevant values stay in a bounded region. [example: Uniformly Close Inputs but Squared Outputs Not Close] Let $S=\{*\}$, let $Y=\mathbb R$ with the usual metric, and define $f_k,g_k:\{*\}\to\mathbb R$ by $f_k(*)=k$ and $g_k(*)=k+1/k$. For each fixed $k$, the functions $f_k$ and $g_k$ are bounded because their images are the one-point sets $\{k\}$ and $\{k+1/k\}$. Since the domain has only the point $*$, the uniform distance between $f_k$ and $g_k$ is the single pointwise error: \begin{align*} D(f_k,g_k)=\sup_{s\in\{*\}}|f_k(s)-g_k(s)|=|f_k(*)-g_k(*)|. \end{align*} Substituting the definitions gives \begin{align*} |f_k(*)-g_k(*)|=\left|k-\left(k+\frac{1}{k}\right)\right|=\left|-\frac{1}{k}\right|=\frac{1}{k}. \end{align*} Thus $D(f_k,g_k)=1/k$, so the input functions become uniformly close to each other. Now compare their squares at the only point of the domain: \begin{align*} |g_k(*)^2-f_k(*)^2|=\left|\left(k+\frac{1}{k}\right)^2-k^2\right|. \end{align*} Expanding the square, \begin{align*} \left(k+\frac{1}{k}\right)^2=k^2+2k\cdot\frac{1}{k}+\frac{1}{k^2}=k^2+2+\frac{1}{k^2}. \end{align*} Therefore \begin{align*} |g_k(*)^2-f_k(*)^2|=\left|k^2+2+\frac{1}{k^2}-k^2\right|=\left|2+\frac{1}{k^2}\right|=2+\frac{1}{k^2}. \end{align*} Since $2+1/k^2\ge 2$ for every $k$, the squared functions do not become uniformly close. This shows why boundedness of each individual function is not enough for nonlinear stability; the values must stay inside one common bounded region. [/example] ## Beyond and Connected Topics The uniform metric is the entry point to [uniform convergence](/page/Uniform%20Convergence) and to normed spaces of functions. Once the distance $D(f,g)=\|f-g\|_\infty$ is available, questions about sequences of functions become ordinary questions in metric spaces and Banach spaces. A first continuation is [Cambridge IA Analysis Notes](/page/Cambridge%20IA%20Analysis%20Notes), where pointwise estimates, suprema, continuity, and convergence of functions appear in their earliest analytic form. The uniform metric packages these ideas into a single distance. A second continuation is [Cambridge IB Analysis and Topology](/page/Cambridge%20IB%20Analysis%20and%20Topology), where metric spaces and topological consequences of convergence become systematic. The uniform metric is a central example of a metric built from another metric by applying a supremum construction. A third continuation is [Cambridge II Analysis of Functions](/page/Cambridge%20II%20Analysis%20of%20Functions), where function spaces, completeness, and limiting procedures become more structural. The [completeness of bounded function spaces](/theorems/288) is a basic model for later Banach-space arguments. For measure-theoretic function spaces, the contrast with $L^p$ metrics becomes important. In [Geometric Measure Theory III: BV Functions and Sets of Finite Perimeter](/page/Geometric%20Measure%20Theory%20III%3A%20BV%20Functions%20and%20Sets%20of%20Finite%20Perimeter), [convergence in measure](/page/Convergence%20in%20Measure)-based norms and weak notions of derivative replace worst-case pointwise control. The uniform metric remains the comparison point for understanding what those weaker modes of convergence do not control. ## References Androma, [Cambridge IA Analysis Notes](/page/Cambridge%20IA%20Analysis%20Notes). Androma, [Cambridge IB Analysis and Topology](/page/Cambridge%20IB%20Analysis%20and%20Topology). Androma, [Cambridge II Analysis of Functions](/page/Cambridge%20II%20Analysis%20of%20Functions). Androma, [Geometric Measure Theory III: BV Functions and Sets of Finite Perimeter](/page/Geometric%20Measure%20Theory%20III%3A%20BV%20Functions%20and%20Sets%20of%20Finite%20Perimeter). Walter Rudin, *Principles of Mathematical Analysis* (1976). John B. Conway, *A Course in Functional Analysis* (1990).