We observe W = X + U and want the density f of X. The density of W is f blurred by the error density, and in Fourier terms the blur is a product: ψW(t) = ψ(t)ψU(t). So estimate ψW from the data, divide by the known ψU, damp high frequencies with a kernel of bandwidth h, and invert. The division blows up the noise at the frequencies where ψU is small: for Normal errors that happens fast, and the best possible rate is only (log n)−2.
All three estimates are computed the same way: f̂(x) = (1/2π)∫ e−itx ψ̂W(t) ψK(th) m(t) dt, with ψ̂W the empirical characteristic function and ψK(s) = (1 − s2)3 on [−1, 1]. The naive estimate has m = 1, the deconvolving estimate (10.20) m = 1/ψU, and the small-σ correction m = 1 + σu2t2/2, the first two terms of 1/ψU for Normal errors (it needs a finite variance, so it is not shown for Cauchy errors). The risk curves are exact (Parseval), not simulated. The Laplace error has variance σu2; the Cauchy error has scale σu and no variance at all.
code/ch10_other_topics.py (fourier_kde, exact_mise, fig_s2, fig_s3).