A kernel density estimate puts a small bump of width h on every observation and adds them up: f̂(x) = (1/n) Σ Kh(x − xi). A small h follows every observation (low bias, high variance, a wiggly curve); a large h smooths real features away (high bias, low variance). Use the real expression values of four genes in 101 mouse embryo cells (Figure 3.19 of Holmes and Huber 2019), or a simulated mixture whose true density is known, so that the error can be measured. The bandwidth is the standard deviation of the kernel, as in R's density().
Silverman's rule (R's bw.nrd0) is 0.9 min(s, IQR/1.34) n−1/5, Scott's (bw.nrd) uses 1.06 instead of 0.9. The AMISE-optimal bandwidth needs the true density, so it is offered only for the simulation: h* = [R(K) / (R(f″) n)]1/5 for a kernel with unit variance. Simulated samples use a seeded generator, so everyone sees the same samples.
code/ch03_data_visualization.R (Figure 3.19).