Domain Geometry And Boundary Effects In Approximation

Why the Domain Matters

A neural approximation bound is never just a statement about a function and a network size. It also depends on where the function is defined, which norm measures the error, and what regularity is available near the edge of the domain. A bound proved on a cube can often be used on a smaller domain, but only after checking how the target is represented there and how the error norm changes.

Boundary behavior matters because smoothness inside a domain need not hold uniformly up to its boundary. A function can be smooth at every interior point while its derivatives become unbounded near the edge. It may still have a continuous extension, or even a Sobolev extension, but those are different regularity statements. Approximation results that require bounded derivatives on a surrounding cube cannot be invoked from interior smoothness alone.

Domains, Norms, and Restriction

Suppose \(K\) lies inside a cube \(Q\), \(F\) is a function on \(Q\), and \(f\) is its restriction to \(K\). Any network that approximates \(F\) on \(Q\) also approximates \(f\) on \(K\): restricting the set over which an error is measured cannot increase a supremum. For integral error, there is a separate volume factor. The measurability of the error is essential, not optional.

To verify the bound, measurability makes the integral well-defined, and the pointwise error bound gives \[ |F(x)-g(x)|^p\le\varepsilon^p\quad(x\in K). \] Integrating yields \(\int_K|F-g|^p\,dx\le\varepsilon^p|K|\). Taking the \(p\)-th root proves the result. If \(|K|=0\), the integral and the stated upper bound are both zero. For \(K\) of positive finite measure, the inequality also explains why a uniform error guarantee implies an \(L^p\) guarantee, with a factor depending on the domain’s volume. \(\square\)

Rescaling Cubes and Tracking Geometry

Bounds are often stated on the unit cube \(C=[0,1]^d\), even when the data domain is a different cube \(Q=\prod_{j=1}^d[a_j,b_j]\), where \(b_j>a_j\). The affine map \(T:C\to Q\) given by \(T(u)_j=a_j+(b_j-a_j)u_j\) transfers a target \(f\) on \(Q\) to \(F(u)=f(T(u))\) on \(C\). If \(N\) approximates \(F\), then \(x\mapsto N(T^{-1}(x))\) approximates \(f\) on \(Q\).

For ReLU networks, composing inputs with an affine map changes the first layer’s weights and biases but not the number of layers or units: the first affine transformation can absorb \(T^{-1}\). If an affine output transformation is needed, it can likewise be absorbed into the output layer. The error norms still transform with the domain: uniform error is unchanged, while an integral incurs the Jacobian factor.

Indeed, the Jacobian determinant of \(T\) is the constant \(\prod_{j=1}^d(b_j-a_j)\). Applying the change-of-variables formula to the nonnegative measurable function \(|h|^p\) gives \[ \int_Q|h(x)|^p\,dx =\left(\prod_{j=1}^d(b_j-a_j)\right)\int_C|H(u)|^p\,du. \] Taking \(p\)-th roots proves the \(L^p\) identity. Since \(T\) is a bijection between \(C\) and \(Q\), the two functions attain the same set of absolute values, proving equality of the uniform norms. \(\square\)

Boundary Regularity and Extensions

An approximation theorem on a cube may assume that the target has bounded derivatives throughout that cube. For a target initially specified only on a domain \(\Omega\), such a theorem requires an appropriate extension to the cube. The extension must agree with the target on \(\Omega\) (or almost everywhere for an \(L^p\) statement), and it must have the regularity and norm bounds required by the approximation theorem. Extending by zero is not automatically valid: it can create a jump at the boundary even when the target is well behaved inside.

A standard Sobolev extension theorem states that if \(\Omega\) is a bounded Lipschitz domain, then for integer \(s\ge0\) and \(1\le p<\infty\), such a bounded extension operator exists, with \(C\) depending on the domain, \(s\), and \(p\). A bounded domain with a smooth boundary satisfies the Lipschitz condition. This theorem supplies a route from domain-based Sobolev assumptions to ambient Sobolev assumptions; it does not assert that zero extension works, nor does it by itself prove any particular neural approximation rate. To obtain a neural bound, one must also have an approximation theorem whose hypotheses match the extended function’s regularity.

Interval Approximation and a Boundary-Sensitive Example

The interval makes the distinction between a target’s smoothness and its boundary behavior especially visible. A continuous function on a compact interval can be uniformly approximated by piecewise-linear functions, and a one-dimensional ReLU network can represent such a spline. The approximation rate of a particular spline depends on how the knots are placed and on the target’s regularity up to the endpoints.

Common Pitfall: Treating Interior Smoothness as an Extension

• Wrong reasoning: “The function has derivatives everywhere inside the domain, so a cube approximation theorem for smooth functions applies.” This ignores whether those derivatives remain controlled near the boundary and whether the target has an extension with the required norm.
• Correction: Check the theorem’s exact regularity assumption, including whether it concerns derivatives on the closure, weak derivatives on the domain, or a specified ambient extension. Use an extension result only when its domain and function-space hypotheses hold.
• Wrong reasoning: “A uniform bound automatically gives an \(L^p\) bound, regardless of how the functions are defined.” An \(L^p\) norm requires a measurable error function.
• Correction: State measurability on the integration domain; continuity of both functions is a common sufficient condition. Then the bound follows by integrating the pointwise error bound.