Complex coefficients make an algebra larger, but they do not automatically make it capable of approximating every complex-valued continuous function. Polynomials in distinguish points of the unit circle, yet they cannot uniformly approximate there. Adding conjugates repairs this obstruction. Restricting the same algebra to a short arc can also change the answer, even without adding a single new polynomial.
The distinction is between the functions available exactly and those available as uniform limits. The complex Stone–Weierstrass theorem gives a useful sufficient algebraic condition. Its proof also reveals a weaker closure condition that is necessary and sufficient, and explains why geometry of the domain matters.
Motivation
The real approximation theorem uses maximum and minimum operations to combine local approximations. Complex numbers do not carry the ordering needed for that proof. We cannot simply repeat it with complex-valued maxima. Instead, we must recover a sufficiently rich algebra of real-valued functions inside the complex algebra.
For , the natural candidates are and . These formulas use the conjugate of . Having and allowing complex scalar multiplication does not guarantee that belongs to the algebra. The extra hypothesis earns exactly this transition from complex separation to real separation.
Definitions
Let be a nonempty compact metric space, and equip with . The real theorem proved in the preceding function-space note will be used: a real subalgebra of containing constants and separating points is uniformly dense.
Definition 2.2.1
(Complex algebra, conjugation and separation)
A complex subalgebra is a complex vector subspace closed under pointwise multiplication. It is unital if it contains the constant function one; complex linearity then supplies every constant. It separates points if every distinct admit an with . It is closed under conjugation if whenever , where .
This last property is also called self-adjointness of the function algebra. Here it means pointwise conjugation, not an assertion about the domain or adjoint of an operator. A bar over an algebra will denote uniform closure; a bar over a function or scalar denotes conjugation. Write when both operations occur in one argument.
Definition 2.2.2
(Real part algebra and vanishing nowhere)
The real-valued part of is , treated as a real algebra. An algebra vanishes nowhere if for every there is some with . Different points may require different functions. This is weaker than asking in advance for one algebra member nonzero everywhere.
Uniform closure preserves the algebra operations. For products, if and uniformly, then is uniformly bounded and tends uniformly to zero. Conjugation is an isometry: .
Theorem/proposition
Theorem 2.2.3
(Complex Stone–Weierstrass theorem)
If a unital complex subalgebra separates points and is closed under conjugation, then .
Proposition 2.2.4
(The exact closure criterion for a separating unital algebra)
Suppose is a unital complex algebra separating points. It is uniformly dense if and only if for every . Thus exact conjugation closure is sufficient, but is not a necessary property of the original algebra itself.
Corollary 2.2.5
(Coordinate polynomials on compact sets)
For any nonempty compact , polynomials with complex coefficients in , restricted to , are uniformly dense in . Equivalently, polynomials in the real coordinate functions are dense. No interior, convexity or smooth-boundary assumption on is needed.
Theorem 2.2.6
(A nonunital version on a compact domain)
If a complex algebra separates points, is closed under conjugation and vanishes nowhere, then it is uniformly dense even if it does not initially contain the constant one.
The hypotheses are used with their stated quantifiers. Separation concerns pairs of distinct inputs; vanishing nowhere concerns one input at a time. A compact-domain theorem does not automatically become a theorem for arbitrary continuous functions on a noncompact space, where the supremum may be infinite.
Proof sketch or proof idea
Recover a real separating algebra
Proof
(Reduce to the real theorem)
Assume first that is unital and conjugation-closed. Both and belong to by the formulas in the motivation. This real-valued part contains real constants and is closed under real linear combinations and products. If , at least one of its real and imaginary parts has different values at . Therefore separates points.
The real Stone–Weierstrass theorem applies. For any and , choose with and . Then and
This proves the complex theorem. It imports the already proved real approximation result, including its lattice argument; it does not pretend complex numbers themselves have a compatible maximum operation.
Proof X-ray
Conjugation is what makes the real reduction stay inside the algebra
Separating points with a complex function gives separation by either its real or imaginary part. That observation becomes useful only when those parts belong to the admissible algebra. Conjugation supplies this membership. Once a real separating algebra is available, the rest of the proof is the real theorem applied twice, with a controlled final complex error.
Exact membership can be replaced by approximate membership
Proof
(Approximate conjugates suffice)
Density clearly implies that for every . Conversely, assume precisely this inclusion. If , choose converging uniformly to . Each belongs to , and the isometry of conjugation shows that they converge uniformly to . Closedness gives .
Thus is a conjugation-closed unital complex algebra and still separates points because it contains . Apply the complex theorem to . It is both dense and closed, so . This proves the closure criterion without assuming the conjugates were exact original algebra members.
Proof
(Coordinates separate and conjugate)
For the coordinate corollary, coordinates distinguish distinct points of . Conjugating a monomial exchanges its coordinate and conjugate-coordinate powers and conjugates its coefficient, so the algebra is conjugation-closed. It contains one. The equivalent real-coordinate description follows from
Construct an approximate unit instead of assuming a unit
Proof
(Build an approximate unit by compactness)
Now suppose satisfies the nonunital theorem's hypotheses. The open sets for cover . Compactness supplies finitely many whose sets still cover it. Consequently
for some positive constants . There is now one positive function in the algebra. Approximate uniformly on by real polynomials , using real polynomial approximation. Then : every term contains a positive power of , so no separate constant function in has been used. Moreover,
If , the choice gives the same conclusion directly. Thus . This closure is a conjugation-closed complex algebra separating points; the unital theorem now makes it all of . Compactness and the no-common-zero condition have supplied the missing unit, rather than hidden it in a polynomial substitution.
The no-common-zero hypothesis cannot be dropped.
Counterexample mode
A common zero blocks the missing unit
Claim: a conjugation-closed complex algebra that separates points is dense, even without a unit. On take the polynomials with . They form a conjugation-closed algebra, and separates points. But every member vanishes at , so for every in the algebra, and the constant one is not approximable. The repair is the hypothesis of the nonunital theorem: the algebra must vanish nowhere. On the same algebra has no common zero, and the worked example below recovers the unit.
Worked examples
Worked example 2.2.7
(1. The circle needs both frequency directions)
On , conjugation gives . Polynomials in are precisely finite Laurent sums . They contain constants, separate points and are conjugation-closed, so they uniformly approximate every continuous function on the circle.
In contrast, let . Direct integration of integer exponentials gives
The absolute value of this average is at most , so that error is at least one. The zero polynomial achieves error one. Thus the exact distance from to analytic polynomials on the circle is one. The same obstruction holds on the closed unit disk by restriction to its boundary. Complex coefficients, a unit and point separation have not removed the missing negative frequency.
Concept lensAnalytic
Density as the absence of an annihilating measure
A subspace is uniformly dense exactly when the only complex Borel measure on with for every is . This combines the Hahn–Banach theorem with the Riesz representation of bounded functionals on . The circle example exhibits such a measure: annihilates every analytic polynomial, since each has positive frequency, yet . Because has total variation one, this single number already gives the distance bound one.
Worked example 2.2.8
(2. A quantitative approximation of a nonanalytic target)
On the square , , the function is continuous and has Lipschitz constant one in Euclidean distance. Polynomials in can approximate it uniformly by the coordinate corollary. For positive integers , a concrete construction is the product Bernstein polynomial
The nonnegative weights sum to one. Their squared deviations in the two coordinates have weighted sums and . The Lipschitz bound and weighted Cauchy–Schwarz therefore give
This is a finite weighted-sum calculation, with no probability prerequisite. In particular has error at most . The first polynomial is . Substituting the coordinate formulas above writes every such approximation as a polynomial in . Holomorphicity of the target has never been assumed.
Worked example 2.2.9
(3. A short arc changes what can be approximated)
Let and let be analytic polynomials restricted to . Put . The polynomials
satisfy . Since on the arc,
The closure of therefore contains and hence the conjugate of every polynomial. The closure criterion proves that is dense in . Yet is not exactly a polynomial on this arc: such an identity would make have infinitely many zeros, forcing it to be the zero polynomial, which fails at zero. This is a dense algebra that is not itself conjugation-closed. It does not contradict the circle obstruction because the domain has changed.
Worked example 2.2.10
(4. A missing constant can be uniformly recovered)
On , take polynomials with zero constant coefficient. They form a conjugation-closed complex algebra, separate points using , and vanish nowhere. They do not contain one exactly: a polynomial equal to one throughout an interval must be the constant polynomial one. Nevertheless,
has zero constant coefficient and . The nonunital theorem applies. On , the same algebra instead vanishes at zero and cannot approximate one with error less than one. Whether an algebra has a common zero matters as well as whether its written formulas contain a constant term.
Common mistakes
Common mistake
(Conjugate the variable, not only the coefficients)
Conjugation of a function is not obtained merely by conjugating the coefficients while leaving its complex coordinate unchanged. The conjugate of is . On the circle this introduces negative powers; on a disk, is not , and the latter is not even defined at zero.
Common mistake
(Sufficient is not necessary)
The theorem's conjugation hypothesis cannot be deleted from a general sufficient criterion, as the circle example proves. It is also incorrect to call exact conjugation closure necessary for every dense algebra: the short arc disproves that stronger claim. What is necessary and sufficient, with a unit and separation, is approximate availability of conjugates in the uniform closure.
Common mistake
(Density is not convergence of a chosen procedure)
Finally, density is an existence statement about suitable approximants, not a conclusion that any familiar coefficient procedure converges uniformly. Laurent-polynomial density does not prove uniform convergence of ordinary Fourier partial sums; the earlier Fejér theorem supplies a particular convergent averaging procedure. Quantitative rates also require additional information about the target, such as the Lipschitz estimate used in the Bernstein construction.
Summary
Conjugation allows complex separation to produce real-valued separating functions, so the real theorem proves the complex one. Uniform closure can supply conjugates or a missing unit even when the original algebra does not contain them exactly. Circle, arc and interval examples show why the algebra and its domain must be checked together. Concrete coordinate polynomials connect the theorem to actual approximation formulas.
Exercises
Quick check
(On , consider polynomials in . Verify conjugation closure and show that the uniform distance from to this algebra is exactly one. Which hypothesis fails?)
Solution · Answer
Conjugation exchanges and and conjugates coefficients, so the algebra is conjugation-closed. Every member has the same value at and , hence does not separate these points. If approximated the coordinate function with error , then . Thus , and achieves equality.
Quick check
(Express as a polynomial in , and compute its conjugate polynomial.)
Solution · Answer
The polynomial is . Its conjugate is . The real-coordinate square is real-valued; the second term equals and changes sign under conjugation.
Quick check
(For an integer , prove that the distance from to analytic polynomials on the circle is exactly one.)
Solution · Answer
Multiply the error by and average over a period. The target contribution is one; every polynomial term has strictly positive integer frequency , so its average is zero. The uniform error is therefore at least one. The zero polynomial attains one on the circle.
Quick check
(In the short-arc example, give a degree bound ensuring error at most , where . Verify that suffices for .)
Solution · Answer
The exact error is , so it suffices to choose , remembering that . For , , so the eleven-term polynomial works. This certificate comes from the geometric construction, not from the abstract density theorem alone.
Quick check
(On , let be polynomials satisfying . Prove that its uniform closure is exactly , not the whole continuous-function space.)
Solution · Answer
Every uniform limit from vanishes at zero, by continuity of evaluation. Conversely, for continuous with , complex polynomial approximation on the interval gives uniformly. Set . Then . Thus the indicated vanishing subspace is exactly the closure. The unit cannot be recovered because all algebra members share a zero.