Computing contours from discrete points. about d3-contour

[mbostock]

Released in 1.1!

[mbostock]

Seems like geom_density2d defaults to a grid size of 100×100 (n defaults to 100) and by default uses bandwidth.nrd as the Gaussian kernel. I’m guessing that should map pretty closely to a stack blur radius.

[curran]

Random thought - A 1D version of this Gaussian kernel would be a really nice addition to D3, possibly as a sibling of d3-array#histogram. I see this was worked on long ago in d3/Add kernel density estimation. PR #143.

[mbostock]

Here’s a related example (just updated): https://bl.ocks.org/mbostock/4341954

That example computes the kernel density estimation the brute force way: it evaluates the kernel for each data point (n) for each grid point (m) = O(nm). This tends to be prohibitive for estimating two-dimensional density since the number of grid points in two dimensions is much larger (e.g., m = 480×250 = 120,000 instead of m = 100).

One way of making the brute force method faster would be to sort the data, and then keep a moving window of data points whose kernel value is non-zero. Although if your bandwidth is large, then this wouldn’t be a big improvement, since you’d still be evaluating the kernel on a large number of data points for each grid point.

Or you could use the same fast blur technique d3-contour uses for two-dimensional density estimation. This is less accurate because the data points are first rounded to the resolution of the grid before blurring, but often this is acceptable.

[curran]

I really like the idea of blur(histogram(data)).