Proof Architecture for Stone-Weierstrass

A Proof Built from Modules

The Stone-Weierstrass argument is easier to use when its proof is viewed as a sequence of tasks rather than as one long construction. The previous tutorial established the Real Stone-Weierstrass Theorem using local patches, a finite subcover, and a finite maximum. Here we organize that architecture and isolate two reusable tools: a finite patching lemma with asymmetric errors, and a precise rule for allocating the final approximation error.

Let \(K\) be a nonempty compact subset of \(\mathbb{R}\), let \(A\) be a unital subalgebra of \(C(K)\) that separates points, and let \(\overline{A}\) denote its closure in the supremum norm. The results from the previous tutorial are available as established tools: the One-Sided Patching at a Fixed Point Theorem supplies local-to-global upper control for a patch, and the Absolute Values in the Uniform Closure Theorem ensures that \(\overline{A}\) is closed under finite maxima and minima. We will cite those results, not reconstruct them.

The proof architecture separates four jobs. Point separation supports interpolation; the one-sided patching result supplies functions with a global upper estimate; continuity creates neighborhoods where those functions also have a lower estimate; and compactness turns the neighborhoods into a finite family. A finite maximum then assembles the estimates. Finally, membership in \(\overline{A}\) lets us replace the assembled function by an actual member of \(A\).

The value of this decomposition is diagnostic. If a proof has local estimates but no open cover, the continuity step is missing. If it has an open cover but no finite subcover, it has not yet justified a finite lattice operation. If it constructs a function in \(\overline{A}\) but ends by calling it an element of \(A\), it has skipped the final approximation step.

A Finite Patching Lemma

The central assembly step does not need all the hypotheses of Stone-Weierstrass. Once a finite cover and its candidate functions are in hand, the pointwise estimate follows from a simple maximum. The upper and lower tolerances need not be the same, which is useful when different parts of an argument produce different error sizes.

Proof. Fix \(x\in K\). Since every candidate satisfies the global upper estimate, \(h_i(x)\leq f(x)+\alpha\) for each \(i\). Taking the maximum over the finite list gives \(g(x)\leq f(x)+\alpha\). Since the \(U_i\) cover \(K\), there is an index \(j\) with \(x\in U_j\). The lower estimate for that patch gives \(h_j(x)>f(x)-\beta\). Since \(g(x)\geq h_j(x)\), we have \(g(x)>f(x)-\beta\). These establish both bounds. For continuity, the maximum of two continuous real-valued functions is continuous because \(u\vee v=(u+v+|u-v|)/2\). Applying this identity repeatedly shows that the maximum of finitely many continuous functions is continuous. \(\square\)

The lemma exposes why the maximum is the appropriate operation. The global upper bounds survive because every candidate lies below the same ceiling. The local lower bounds survive because, at each point, at least one covering patch lies above the common floor, and the maximum cannot be smaller than that patch. The lemma itself does not require compactness: compactness is what supplies the finite cover in applications where the initial cover may have infinitely many sets.

Keeping the Final Error Budget Honest

Patching usually produces a function in a closure, while the theorem’s conclusion asks for an element of the original algebra. That last change of object is a second approximation, and its error must be added to the patching error. The next result records the exact bookkeeping rule.

Proof. Since \(h\in\overline{A}\), the definition of closure in the norm metric implies that for every \(\rho>0\) there exists \(a\in A\) with \(\|a-h\|<\rho\). By the triangle inequality,

This proves the first assertion. If \(\eta+\rho<\varepsilon\), then the displayed strict inequality gives \(\|a-f\|<\varepsilon\), as required. \(\square\)

The distinction between a weak bound and a strict bound matters. The intermediate construction may only establish \(\|h-f\|\leq\eta\), not a strict inequality. There is no problem: the strict inequality in the final conclusion comes from choosing the closure approximation with error strictly less than \(\rho\). A proof should state where its strictness comes from rather than silently changing \(\leq\) to \(<\).

Why the Finite Cover Is a Real Step

A finite maximum is an operation already available from the lattice identities in the uniform closure. An infinite pointwise supremum is a different operation and is not automatically available. Compactness is precisely what allows the proof to avoid needing that additional closure property: it reduces the open cover to finitely many patches before taking a maximum.

Auditing the Architecture

A reliable proof audit follows the dependencies in order. First check that the local patch theorem applies: its algebraic hypotheses and point-separation condition must be in place. Next check that the proposed neighborhoods really are open and cover every point. Then verify that compactness produces a finite subcover and that the operation used to combine the resulting finite family is available in the space being used. Finally, track every error through the transfer from \(\overline{A}\) back to \(A\).

There is a small but important distinction between pointwise and uniform conclusions. The finite patching lemma gives a pointwise strict lower estimate and a weak upper estimate. These imply \(|g(x)-f(x)|\leq\max\{\alpha,\beta\}\) at every point, hence a supremum-norm bound with \(\leq\). That is sufficient for the two-stage error budget. It is not necessary to claim a strict supremum bound at this stage, and retaining the weak inequality makes the argument robust even when the set under consideration is not compact.

The Real Stone-Weierstrass Theorem from the previous tutorial is the completed result of these modules; its proof is not a black box for the details above. The architecture explains what each hypothesis contributes and makes the argument adaptable: in another approximation problem, one can ask whether local candidates, a finite cover, a stable combining operation, and a compatible error budget are available. If one of these components is absent, the proof needs a new idea at that specific stage.