A smoother estimates f(x) from the data near x. Scatter n points uniformly on the cube [−1, 1]d: the small cube [−r, r]d holds on average n rd of them, which collapses as d grows. To keep a fixed fraction q of the data, the neighbourhood must have half-width q1/d, close to the whole cube: it is no longer local. The price shows in the risk: for a twice-differentiable density the best mse falls only like n−4/(4+d), so the sample size for a fixed accuracy grows exponentially in d.
The sample-size panel uses the counting estimator of f(0) with a cube of side h (Exercise 4.6) for the standard Normal density in d dimensions: mse ≈ Ah4 + B/(nhd) with A = (Δf(0))2/576 and B = f(0); at the best h it equals a constant times n−4/(4+d). Accuracy is measured relative to f(0)2, since f(0) = (2π)−d/2 itself shrinks with d. The grey line is the parametric benchmark, where the mse falls like 1/n in every dimension. These are first-order approximations: for large d the best h is not small, and the true sample sizes are larger still. The simulation uses a seeded generator; the upper left panel shows the first two coordinates.
code/ch04_smoothing_concepts.py (ex46, ex46_n_for_relmse).