Stone-Weierstrass Motivation

From Lattice Operations to Approximation

The previous tutorial showed that pointwise maxima and minima preserve continuity. In approximation theory, this matters because a maximum or minimum can combine local approximations into a single function that works across a whole domain. To use this idea, we need an algebra of candidate approximants with enough flexibility to match a target at selected points.

Throughout, let $K$ be a nonempty compact subset of $\mathbb{R}$, and consider real-valued continuous functions on $K$. An algebra of such functions can add, scale, and multiply its members. Two further properties will be central: it contains the constant functions, and it can distinguish any two different points of $K$. These properties are the starting point for the Stone-Weierstrass approximation strategy.

Definition: A subalgebra $A$ of $C(K)$ is a linear subspace closed under pointwise multiplication. It is unital if it contains every constant function. It separates points of $K$ if, whenever $x,y\in K$ and $x\neq y$, some $u\in A$ satisfies $u(x)\neq u(y)$.

Separating points is weaker than approximating an entire function: it only asks for one algebra element to distinguish a particular pair. Yet, together with constants and the algebra operations, it has an immediate consequence. An element that distinguishes two points can be rescaled and shifted to take any prescribed pair of values there.

Theorem (Two-Point Interpolation): Let $A$ be a unital subalgebra of $C(K)$ that separates points. For every $f\in C(K)$ and every $x,y\in K$, there is a $g\in A$ such that $g(x)=f(x)$ and $g(y)=f(y)$.

Proof. If $x=y$, take the constant function $g=f(x)$, which belongs to $A$. Now suppose $x\neq y$. Choose $u\in A$ with $u(x)\neq u(y)$. The denominator in the following expression is therefore nonzero:

g=f(x)\mathbf{1} +\frac{f(y)-f(x)}{u(y)-u(x)}\bigl(u-u(x)\mathbf{1}\bigr),

where $\mathbf{1}$ denotes the constant function with value $1$. Because $A$ is unital and is a linear subspace, $g\in A$. At $x$, the factor $u-u(x)\mathbf{1}$ vanishes, so $g(x)=f(x)$. At $y$, the formula gives

g(y)=f(x)+\frac{f(y)-f(x)}{u(y)-u(x)}\bigl(u(y)-u(x)\bigr)=f(y).

This proves the claim in both cases. $\square$

Worked Example: Interpolating at Two Points with Polynomials

Let $K=[-1,2]$, let $f(t)=t^2+1$, and prescribe the points $x=-1$ and $y=2$. The coordinate function $u(t)=t$ belongs to the polynomial algebra and has different values at the two points. The interpolation formula gives

g(t)=f(-1)+\frac{f(2)-f(-1)}{2-(-1)}(t-(-1)) =2+\frac{5-2}{3}(t+1)=t+3.

Direct substitution verifies both required values: $g(-1)=2=f(-1)$, and $g(2)=5=f(2)$. This linear polynomial matches the quadratic at those two points, but it does not match it throughout the interval; for example, $g(0)=3$ while $f(0)=1$. Two-point interpolation is a local ingredient, not yet a uniform approximation theorem.

Why the Uniform Closure Has Lattice Operations

The functions in $A$ need not themselves be closed under taking maxima and minima. But their uniform limits have more flexibility. Write $\overline{A}$ for the closure of $A$ in the supremum norm on $C(K)$. We first record why this closure remains an algebra.

Proposition: If $A$ is a subalgebra of $C(K)$, then $\overline{A}$ is also a subalgebra of $C(K)$. If $A$ contains the constant functions, so does $\overline{A}$.

Proof. Let $f,g\in\overline{A}$. There are sequences $f_n,g_n\in A$ that converge uniformly to $f,g$, respectively. The sums $f_n+g_n$ and scalar multiples $cf_n$ belong to $A$ and converge uniformly to $f+g$ and $cf$. Hence those limits belong to $\overline{A}$.

For products, uniform convergence of $g_n$ implies that $(g_n)$ is uniformly bounded: there is a finite $M$ such that $\|g_n\|_\infty\leq M$ for every $n$. Using the triangle inequality and the supremum-norm product estimate,

\|f_ng_n-fg\|_\infty \leq \|f_n\|_\infty\|g_n-g\|_\infty +\|f_n-f\|_\infty\|g\|_\infty.

The sequence $(f_n)$ is uniformly bounded as well, because it converges uniformly. Both terms on the right tend to zero, so $f_ng_n$ converges uniformly to $fg$. Since $f_ng_n\in A$, it follows that $fg\in\overline{A}$. Finally, every constant function in $A$ is already in its closure. $\square$

The Weierstrass Approximation Theorem now supplies a way to obtain absolute values in $\overline{A}$, even if $A$ has no operation that directly takes the absolute value of a function. This is the key bridge between algebraic operations and the lattice operations from the previous tutorial.

Theorem (Absolute Values in the Uniform Closure): If $A$ is a unital subalgebra of $C(K)$, then $|h|\in\overline{A}$ for every $h\in\overline{A}$. Consequently, $\overline{A}$ is closed under pointwise maxima and minima.

Proof. Fix $h\in\overline{A}$, and set $M=\|h\|_\infty$. If $M=0$, then $h$ is the zero function and $|h|=h\in\overline{A}$. Suppose $M>0$. The scalar function $t\mapsto |t|$ is continuous on the compact interval $[-M,M]$. By the Weierstrass Approximation Theorem, for every $\eta>0$ there is a polynomial $p$ such that

\sup_{-M\leq t\leq M}\bigl|p(t)-|t|\bigr|<\eta.

By the preceding proposition, $\overline{A}$ is an algebra. Thus $p(h)\in\overline{A}$, since $p(h)$ is a finite sum of scalar multiples of powers of $h$, and the constant functions belong to $\overline{A}$. Because $|h(x)|\leq M$ for every $x\in K$, the polynomial estimate gives

\|p(h)-|h|\|_\infty =\sup_{x\in K}|p(h(x))-|h(x)||<\eta.

This holds for every positive $\eta$, so $|h|$ belongs to the closure of $\overline{A}$. The set $\overline{A}$ is closed, and therefore $|h|\in\overline{A}$. By the lattice formulas from the previous tutorial,

h\vee k=\frac{h+k+|h-k|}{2}, \qquad h\wedge k=\frac{h+k-|h-k|}{2},

both $h\vee k$ and $h\wedge k$ belong to $\overline{A}$ whenever $h,k\in\overline{A}$. $\square$

Worked Example: A Nonsmooth Function in the Polynomial Closure

Take $K=[-1,1]$ and let $A$ be the polynomial algebra. The coordinate function $h(x)=x$ belongs to $A$, so the absolute-value theorem shows that $|x|\in\overline{A}$. The lattice formula then gives

x\vee 0=\frac{x+|x|}{2} = \begin{cases} 0,&-1\leq x\leq0,\\ x,&0\leq x\leq1. \end{cases}

The two expressions agree at $x=0$, and this function is therefore in the uniform closure of the polynomials, even though it has a corner there and is not itself a polynomial. In particular, for every $\varepsilon>0$, some polynomial $p$ satisfies $\|p-(x\vee0)\|_\infty<\varepsilon$. No explicit coefficients are needed for this conclusion: it follows from the approximation theorem for the scalar absolute-value function and algebraic closure.

Finite Covers Turn Local Matching into a Global Bound

Two-point interpolation gives a function that matches a target at a selected pair of points. It does not by itself control that function everywhere else. Compactness and lattice operations address that gap in stages. The next result shows how to combine finitely many two-point interpolants to obtain a one-sided bound across all of $K$.

Theorem (One-Sided Patching at a Fixed Point): Let $A$ be a unital subalgebra of $C(K)$ that separates points. For $f\in C(K)$, $x\in K$, and $\varepsilon>0$, there is an $h_x\in\overline{A}$ such that $$ h_x(x)=f(x) \qquad\text{and}\qquad h_x(z)<f(z)+\varepsilon\quad\text{for every }z\in K. $$

Proof. For each $y\in K$, the Two-Point Interpolation Theorem gives $g_y\in A$ with $g_y(x)=f(x)$ and $g_y(y)=f(y)$. This also applies when $y=x$, or in that case one may take the constant function $g_x=f(x)$. The function $g_y-f$ is continuous and is zero at $y$. Hence the set

U_y=\{z\in K:g_y(z)<f(z)+\varepsilon\}

is open relative to $K$ and contains $y$. The sets $U_y$, as $y$ ranges over $K$, cover $K$. By compactness, finitely many of them, say $U_{y_1},\ldots,U_{y_m}$, still cover $K$. Define

h_x=g_{y_1}\wedge\cdots\wedge g_{y_m}.

Each $g_{y_i}$ lies in $A\subseteq\overline{A}$, and $\overline{A}$ is closed under finite minima. Thus $h_x\in\overline{A}$. Also, $g_{y_i}(x)=f(x)$ for every $i$, so their minimum at $x$ is $f(x)$. For any $z\in K$, choose $i$ with $z\in U_{y_i}$. Then

h_x(z)\leq g_{y_i}(z)<f(z)+\varepsilon.

This proves both claims. $\square$

Worked Example: Why the Minimum Is the Useful Operation

Suppose a finite cover of $K$ has been obtained from three neighborhoods, with corresponding functions $g_1,g_2,g_3\in\overline{A}$. At each point $z$, at least one of the neighborhood conditions gives $g_i(z)<f(z)+\varepsilon$. Taking the minimum ensures the same upper bound at every point:

h(z)=g_1(z)\wedge g_2(z)\wedge g_3(z) =\min\{g_1(z),g_2(z),g_3(z)\} <f(z)+\varepsilon.

For instance, if at a particular point the three values are $4.2$, $3.7$, and $5.1$, then $h(z)=3.7$. If the covering neighborhood there guarantees $g_2(z)<f(z)+\varepsilon$, this minimum inherits that bound. A maximum would not: it could select one of the other values instead. At the fixed point $x$ in the patching theorem, every candidate has value $f(x)$, so taking the minimum does not disturb the required equality.

The Approximation Strategy

The one-sided patching result illustrates the roles of the hypotheses. Point separation gives two-point interpolants. Continuity makes the desired inequality persist on a neighborhood. Compactness reduces the collection of neighborhoods to a finite cover. Closure under minima then combines the corresponding functions while preserving the inequality.

The remaining idea is to vary the fixed point $x$. The patched function $h_x$ equals $f$ at $x$; by continuity, it stays above $f-\varepsilon$ on some neighborhood of $x$. A finite selection of these neighborhoods covers $K$, and a pointwise maximum of the corresponding functions preserves the lower bounds while retaining the upper bounds already established. This is the complementary patching step that turns local matching into a uniform approximation argument.

The discussion also clarifies why a subalgebra needs more than closure under addition and multiplication. The polynomial algebra on an interval has natural candidates, but a general subalgebra may contain too few functions unless it can distinguish points. Conversely, distinguishing points alone is not yet a global estimate: uniform control depends on continuity, compactness, and the ability to combine candidates by lattice operations.

Takeaway: A unital point-separating algebra can interpolate any continuous target at two chosen points. Its uniform closure is closed under absolute values, maxima, and minima; compactness then allows finitely many local approximants to be patched into a one-sided global bound.

Check Your Understanding

Use the definitions and results in this tutorial to answer the following questions.

What does it mean for a subalgebra of $C(K)$ to be unital and to separate points?
How does an element that separates $x$ and $y$ lead to an algebra element matching prescribed values $f(x)$ and $f(y)$?
Why does the uniform closure of an algebra remain closed under multiplication?
How does polynomial approximation of the scalar absolute-value function imply that $|h|\in\overline{A}$?
In the one-sided patching theorem, why is a finite subcover needed, and why is a minimum used?
What additional patching step would turn the one-sided bound into a two-sided uniform approximation?

What You'll Learn