A function space treats an entire function as one point. The chosen distance then determines what it means for one mathematical model to approximate another. Uniform distance controls every input at once. This makes it strong enough to preserve continuity and weak enough to admit powerful approximation and compactness theorems.
Motivation
A sequence of continuous functions may converge at every input while approaching a discontinuous limit. The functions on converge to zero for and to one at . Each point eventually behaves well, but the required index depends on the point and becomes arbitrarily large near one. Uniform convergence forbids this moving region of error.
Two questions guide this note. Can a continuous function be approximated uniformly by functions from a simpler algebra, such as polynomials? And when does a family of continuous functions contain a uniformly convergent subsequence? Stone–Weierstrass answers the first through algebraic separation of points. Arzelà–Ascoli answers the second through a common modulus of continuity. Neither theorem says that arbitrary pointwise approximation automatically becomes uniform.
Definitions
Definition 2.1.1
(Uniform norm and function algebra)
Let be a nonempty compact metric space. Write for its continuous real-valued functions and set . An algebra is a real vector subspace closed under pointwise multiplication. It separates points if, whenever , some satisfies .
Compactness ensures that the supremum is finite and attained. Uniform convergence means . In particular, the index giving a prescribed error works for all inputs. The uniform closure consists of uniform limits of sequences from ; in a metric space, sequence limits suffice to characterize closure.
Definition 2.1.2
(Equicontinuity and relative compactness)
A family is equicontinuous if for each and there is such that implies for every . It is pointwise bounded if for each . It is relatively compact in the uniform norm if its uniform closure is compact.
Equicontinuity places the quantifier “every function” after a single choice of neighborhood. A separate continuity argument for each function does not give this shared choice. On compact , the definition implies uniform equicontinuity: for every , one works at every pair of points. To see this, take finitely many small neighborhoods on which all functions vary by less than from their center, and use a Lebesgue number for that finite cover.
Concept lensAnalytic
Finite input data control an entire function only with a shared modulus
A finite net of inputs is not already a finite net of functions. Equicontinuity makes the input net useful: closeness of two functions at its centers propagates to every nearby input with an error bound shared by the whole family. Compactness of the domain then reduces infinitely many input conditions to finitely many center comparisons. This is the bridge from pointwise subsequence selection to convergence in the uniform norm.
Reading a common modulus in examples
An estimate of the form for every in the family gives equicontinuity whenever as . The important point is that the same function works for every member. A derivative bound is one way to obtain such an estimate on an interval, but it is not the only way.
For example, put for . For , the inequality gives
Given , choose . Then gives error less than for every , including . No derivative bound common to this family is needed: has unbounded difference quotients at zero. Moreover , so the same estimate gives on . The family is consequently equicontinuous and pointwise bounded, ready for the Ascoli theorem below.
Contrast on . Every is continuous and , but the family is not equicontinuous at . Fix . For any proposed , choose an integer so large that . Then but . The index is allowed to depend on the proposed radius: that is exactly how one disproves a radius claimed to work for all members.
Quick check
(Let on . Give a common modulus, determine the uniform limit, and explain why a common finite derivative bound on is unavailable.)
Solution · Answer
The square-root estimate gives for every . Also . Finally for , which tends to infinity as : even one member cannot have a finite Lipschitz, hence a bounded-derivative, estimate on the whole interval.
Theorem/proposition
Proposition 2.1.3
(Completeness of continuous functions)
The normed space is complete. A uniformly closed subspace is therefore complete as well.
Theorem 2.1.4
(Real Stone–Weierstrass theorem)
If a real subalgebra contains the constant functions and separates points, then .
Theorem 2.1.5
(Arzelà–Ascoli theorem for compact metric domains)
A family is relatively compact in the uniform norm if and only if it is pointwise bounded and equicontinuous. Equivalently, every sequence in the family has a uniformly convergent subsequence whose limit is continuous.
For complex-valued functions, the approximation theorem needs an additional hypothesis: the algebra is closed under complex conjugation. We prove only the real version here. The compactness theorem extends to finite-dimensional real or complex targets by applying the coordinate argument; an arbitrary infinite-dimensional target requires additional pointwise relative compactness.
Proof sketch or proof idea
Proof
(Uniform Cauchy sequences converge uniformly)
For completeness, let be uniformly Cauchy. At each point, is a Cauchy sequence of real numbers; call its limit . Given , choose with for . Letting at each point gives for all . Thus convergence is uniform. Continuity follows from the three-term estimate comparing to , then , then .
From partial sums to a uniform series
An infinite series of functions is defined through its partial sums . Pointwise convergence lets the required depend on . Uniform convergence requires one to control the tail at every input; this distinction is especially useful when a Fourier series is to represent a continuous function.
Theorem 2.1.6
(Uniform Cauchy criterion and the Weierstrass test)
Let be any nonempty set and . The series converges uniformly on if and only if, for every , some makes
In particular, if for all , with and , the series converges uniformly and its sum satisfies
Proof
(Cauchy tails and numerical tails)
Uniform convergence makes and uniformly close to the same limit, so the triangle inequality gives the tail condition. Conversely, that condition makes the partial sums Cauchy at each . Completeness of supplies a pointwise limit . Fix and use the condition with ; taking shows simultaneously for every . This proves uniform convergence without assuming in advance that the summands or the limit are bounded.
For the -test, the triangle inequality bounds each finite tail by the corresponding numerical tail . The latter is small independently of , so the criterion applies. Passing also gives the displayed error estimate. If is a metric space and each is continuous, each is continuous and its uniform limit is continuous by the three-term argument above. Compactness of is not needed for this implication.
Worked example 2.1.7
(Two ways to control a series)
For on , take . The sum is continuous and, for , the integral comparison gives
As a second example, converges uniformly on for every , using ; its tail is at most after the term . On it still converges pointwise to , but not uniformly: the tail is unbounded for each fixed . The domain is part of the assertion.
The -test is sufficient, not necessary. For instance, the constant functions form a uniformly convergent series because its partial sums are the same convergent alternating harmonic sums at every input. Yet any bounds on a nonempty domain satisfy , so their sum diverges. Uniform convergence can use cancellation; absolute numerical bounds discard it.
Quick check
(On let for . Find the pointwise sum and decide whether convergence is uniform.)
Solution · Answer
The finite geometric identity gives , including at . The pointwise sum is for and at . For every , the supremum of the error over is (approach from the left), so convergence is not uniform. Each partial sum is continuous while the limit is discontinuous, giving a second check. Pointwise convergence and bounded summands alone cannot justify interchanging a series limit with continuity.
Polynomial approximation
We next supply the approximation ingredient rather than assuming polynomial approximation. For define the Bernstein polynomial
The weights are nonnegative, sum to one, have mean and variance for the values . Fix . Uniform continuity gives such that when . The total weight at is at most , because each such squared deviation is at least . If , then
The bound is independent of , proving uniform polynomial approximation. An affine change of variable gives the same conclusion on any compact interval. In particular, polynomials approximate the absolute-value function there. No probability theory is needed: the weights, their mean and the squared deviations are all finite algebraic identities.
Proof
(Stone–Weierstrass: from an algebra to a lattice)
Let . Addition and scalar multiplication preserve uniform limits. Multiplication does too, since uniformly convergent sequences are uniformly bounded and . Thus is a closed algebra.
If and , approximate uniformly on by polynomials. Their compositions with belong to and converge uniformly to . Hence is closed under absolute values, and therefore under pointwise maxima and minima through .
Fix and . For each pair , construct with and . For distinct points, rescale and translate a separating function; for equal points use a constant. With fixed, the open sets where cover . Select finitely many and take the maximum of their functions, obtaining . It satisfies everywhere and .
The sets where also cover . Select finitely many and take their minimum . Then everywhere. Since is closed and is arbitrary, . The two finite selections explain exactly why compactness enters the proof; separation of points is used only in the initial two-point interpolation.
Proof
(Ascoli: diagonal selection, then a finite net)
For the sufficient direction of Ascoli, compact metric spaces have a countable dense set . Given a sequence in , pointwise boundedness and the real Bolzano–Weierstrass theorem permit successive subsequence selections converging at , then , and so on. The diagonal subsequence converges at every .
To upgrade this to uniform convergence, choose a common equicontinuity radius for error and a finite net whose centers belong to the dense set. Such a net exists: first cover the compact domain by finitely many balls of half the desired radius, then move each center by less than half that radius to a point of the dense set. The triangle inequality gives the required cover. The diagonal sequence is Cauchy at these finitely many centers. For any , compare the two function values at by traveling through a nearby center. The two continuity errors and one center error total less than . Completeness of now supplies the uniform limit. The same hypotheses pass to the closure, so that closure is sequentially compact and hence compact.
Conversely, compactness of the closure supplies a finite uniform -net whose centers belong to the closure and are therefore continuous functions. They need not belong to the original family. At a fixed point, these finitely many functions are bounded; the approximation error bounds every other function there. Their finitely many continuity neighborhoods also have a common intersection. A three-term estimate transfers their continuity to equicontinuity of the entire family. This proves both necessary conditions without assuming differentiability.
Proof X-ray
Where compactness is used twice in Ascoli
The sufficiency proof uses compactness of twice. First it supplies a countable dense set, so that a diagonal argument can make one subsequence converge at countably many points. Second it covers by finitely many balls of a common equicontinuity radius, so that convergence at finitely many centers controls every input. Equicontinuity is the bridge between the two steps: it compares the value at an arbitrary with the value at a nearby center, with an error independent of the index. Pointwise boundedness enters only in the diagonal step, through Bolzano–Weierstrass, and completeness of turns the uniformly Cauchy subsequence into a continuous limit.
There is a useful order to applying this criterion. First identify the domain and the topology: a derivative bound on a noncompact domain does not supply the missing finite cover. Next obtain a shared continuity estimate, for example a common Lipschitz constant or a common modulus tending to zero. Finally control the values at enough points. On a compact domain, equicontinuity and pointwise boundedness together imply a global uniform bound, by a finite cover and comparison with its centers. This is a consequence of the two hypotheses together, rather than a replacement for checking equicontinuity.
Worked examples
Worked example 2.1.8
(1. Polynomial approximation and a quantitative construction)
On , ordinary real polynomials contain constants and separate points because the function does. Stone–Weierstrass makes them uniformly dense. For , the Bernstein construction gives , so the uniform error is exactly . The construction proves existence without solving for a best approximating polynomial; best approximation is a separate problem.
Worked example 2.1.9
(2. Why separation cannot be omitted)
On , the algebra of even polynomials contains constants but cannot distinguish from . It cannot uniformly approximate . If is even and both endpoint errors were less than one, we would have and , which are incompatible. Every uniform limit of even functions remains even.
Worked example 2.1.10
(3. A relatively compact family produced by derivative control)
Consider continuously differentiable functions on with and . The mean value theorem gives and . Ascoli therefore gives relative compactness in . The uniform limits need only be Lipschitz, not continuously differentiable: smooth approximations to show why the closure may contain a corner.
In contrast, is uniformly bounded by one but fails equicontinuity at one. Given , choose within of one; then , while . No common continuity radius works. Every subsequence has the same discontinuous pointwise limit, so no subsequence converges uniformly.
Counterexample mode
A Lipschitz bound on a noncompact domain
Claim: on any metric domain, a uniformly bounded family with a common Lipschitz constant is relatively compact in the uniform norm. On let . Every takes values in and has Lipschitz constant one, so the family is pointwise bounded and equicontinuous. For , however, and , so and no subsequence is uniformly Cauchy. The functions converge pointwise to zero while the bump escapes to infinity. The repair is the hypothesis that the domain be compact, which supplies the finite cover used in the proof.
Common mistakes
Common mistake
(Boundedness and equicontinuity are independent)
Uniform boundedness controls heights; equicontinuity controls changes over short distances. Neither condition substitutes for the other. The constant functions are equicontinuous but not pointwise bounded. The powers above are bounded but not equicontinuous. When checking the definition, state exactly what may depend on: letting each function choose its own radius does not establish equicontinuity.
Common mistake
(Stone–Weierstrass needs an algebra, and gives no uniform degree)
Stone–Weierstrass concerns an algebra, not a collection containing one separating function. The span of and separates points on an interval but cannot approximate every continuous function; it is not closed under multiplication. Also, the theorem gives arbitrary accuracy for each chosen target. On , no fixed polynomial degree gives error less than one for every function in the unit ball. For a proposed degree , choose a continuous function taking alternating values at successive points. Any polynomial approximating it with error less than one must alternate signs there and hence have at least distinct zeros, which its degree forbids. The required degree can depend on the target and the accuracy. This obstruction is specific to an infinite domain; on a finite domain the whole continuous-function space is finite-dimensional.
Common mistake
(Relative compactness allows limits outside the family)
Finally, “relatively compact” leaves room for limits outside the original family. To obtain compactness of the family itself, check that it is uniformly closed. Derivative bounds often survive as Lipschitz bounds, while differentiability may be lost. The topology of convergence determines which constraints are retained; convergence of function values must not be mistaken for convergence of derivatives.
Summary
The uniform norm converts common error control into a metric. Its completeness retains continuous limits. A separating algebra becomes dense through polynomial approximation and finite maxima and minima. An equicontinuous pointwise bounded family becomes relatively compact through diagonal selection and a finite net of inputs. Approximation and compactness share finite-control arguments but answer different questions.
Exercises
Quick check
(Prove that the set of functions on satisfying and is compact in the uniform norm.)
Solution · Answer
The common Lipschitz bound gives equicontinuity, and gives boundedness. Ascoli yields relative compactness. Uniform limits retain the value at zero and, after passing to the limit in , the Lipschitz inequality. The family is closed, so it equals its compact closure.
Quick check
(Show that polynomials in are uniformly dense in , although they are not dense in .)
Solution · Answer
On , is injective, so the real algebra generated by it contains constants and separates points. Apply Stone–Weierstrass. On all its members are even and fail to separate opposite points, giving the obstruction in the worked example on why separation cannot be omitted. A missing odd power in the formula is not itself an obstruction; failure to distinguish inputs is.
Quick check
(Show that evaluation is continuous for the uniform norm. Explain why convergence at every evaluation does not imply uniform convergence.)
Solution · Answer
The inequality proves continuity. The sequence on converges at every input, but its pointwise limit is discontinuous. A uniform limit of continuous functions is continuous, so this convergence cannot be uniform.
Quick check
(Suppose is an equicontinuous sequence on compact and converges pointwise to . Prove that the convergence is uniform.)
Solution · Answer
Fix . Compactness and equicontinuity give a common radius for error for all . Passing to the pointwise limit gives the corresponding bound for , possibly with equality. Choose a finite net of inputs within that radius. Pointwise convergence gives one sufficiently large index for error less than at all net points. Comparing through a nearby net point bounds by at most , uniformly in . This argument does not require a separate boundedness assumption because the pointwise limit is already supplied.