Why does universal approximation say nothing about a network's output far outside its fitted range?

## The quantifier is doing the work Read the theorem's order of quantifiers carefully: *for a compact set `K`, a continuous target `f` and a tolerance `eps`, there exist weights such that the network is within `eps` of `f` **on `K`***. `K` is fixed first. The conclusion is a statement about the maximum error over `K` and only over `K`. There is no clause about points outside, so nothing follows about them — not 'the error is a bit larger', not 'the trend continues'. Literally nothing. This matters in practice because 'universal approximator' gets quoted as if it were a property of the model that travels with it to any input. It is not. It is a property of a hypothesis class relative to a bounded region. ## What actually happens off the region Outside the fitted range the output is decided by the architecture's asymptotics: - **Saturating activations (logistic sigmoid, tanh).** Each hidden unit computes `sigma(w . x + b)`. Push `x` far enough in any direction and every pre-activation is large in magnitude, so every unit sits at its saturation value. The output becomes a fixed combination of saturated constants: a plateau. The model flatlines at a level that has nothing to do with the target out there. - **ReLU and its relatives.** The network is a continuous piecewise-linear function with finitely many breakpoints, all determined by the finitely many hidden units. Beyond the outermost breakpoint in a given direction there are no more kinks, so the function is a single affine map from there to infinity. It extrapolates as a straight line (or hyperplane), with a slope fixed by whichever units are active. Take the concrete case: a one-hidden-layer model fit to a wiggly target on `[0, 10]`, accurate to a hair inside that interval, evaluated at `x = 40`. With tanh units you get a constant. With ReLU units you get a line. Neither has any relationship to the target's true value at 40. The model has not failed to learn; it was never asked about 40, and its class cannot invent structure it was not constrained on. ## Why width does not help A common wrong instinct is to widen the layer. Width buys resolution inside the region — more bumps, more breakpoints, a tighter fit on `K`. It cannot buy behaviour off `K`, because every finite network has finitely many breakpoints or saturation thresholds, and past the last one the asymptotic form is fixed. You can move the boundary out by training on a larger region, but only if you have data there; adding units to a model fit on `[0, 10]` does not tell it anything about 40. ## Enlarging the region If the deployment range is really `[0, 100]`, then set `K = [0, 100]` and the theorem's guarantee covers it — but two costs appear. First, the width needed grows with how much structure the target has across the larger region. Second, and more importantly, the theorem hands you weights assuming full knowledge of the target on `K`; in practice data has to do that job, so you need examples spanning the enlarged region. Declaring a wider domain without data covering it changes nothing about what the model knows. ## The operational consequence Treat the fitted range as part of the model's contract. Record the input range the model was fit over, check incoming values against it at serving time, and route out-of-range inputs to an explicit path — a refusal, a clamp, a fallback rule, or an alert — rather than silently returning a plateau or an extrapolated line that looks like a number and reads like a prediction. In a review, the sentence 'it is a universal approximator, so it will handle that input' should always draw the question: universal over which region? ## What a strong answer sounds like Name the compactness condition, say what the architecture does asymptotically for both saturating and piecewise-linear units, state that width does not help, and finish with the operational move: pin the supported input range and enforce it.