TypeError: Cannot read property 'circle' of null

[micooz]

I encountered an issue that special input data can crash voronoi.polygons.

Reproduce Demo

var d3 = require("d3");

var width = 400;
var height = 400;

// change width and height to 100 works well
// var width = 100;
// var height = 100;

var x = d3.scaleLinear();
var y = d3.scaleLinear();

x.domain([0, 1]).range([0, width]).nice();
y.domain([0, 1]).range([height, 0]).nice();

var voronoi = d3.voronoi()
  .extent([[-1, -1], [width + 1, height + 1]])
  .x((pt) => x(pt.x))
  .y((pt) => y(pt.y));

// TypeError: Cannot read property 'circle' of null
voronoi.polygons([
  {x: 0.9999994485078688, y: 0},
  {x: 0.9999994512180275, y: 0}
]);

I thought it was my scale range problem, but the following code works:

voronoi.polygons([
  {x: 0.1, y: 0},
  {x: 0.2, y: 0}
]);

https://runkit.com/micooz/595cabe5640c440013382792

[mbostock]

Here’s a reduction:

var voronoi = d3.voronoi();

voronoi([
  [399.999779, 0],
  [399.999780, 0]
]);

In short the problem is the rounding applied during construction isn’t enough to prevent certain numerical instabilities; these two points are approximately, but not exactly, coincident. This is probably not something I can fix since this code is ported from Raymond Hill’s implementation of Steven Fortune’s algorithm. However, it might be fixable if this library is rewritten to use Vladimir Agafonkin’s delaunator, so that it first computes the Delaunay triangulation and then derives the dual Voronoi diagram. I believe this is generally considered the more numerically stable approach, and it’ll probably be faster…

[Fil]

https://beta.observablehq.com/d/2e32a5a801b9ed02 shows this is probably fixed now in d3-delaunay

I'm saying “probably” because d3-delaunay can't yet handle cases with n=2 points, cf https://github.com/d3/d3-delaunay/issues/19

[garrickf]

Does d3-delaunay allow for a mapping from the Voronoi polygons to data points (with the other attributes in our original data), like d3-voronoi does? I wanted to use this because of the ability to make it easier to hover over points in a scatter plot, but I'm also running into the issue of some of my data points being approximately coincident.

[mbostock]

Does d3-delaunay allow for a mapping from the Voronoi polygons to data points

Yes. Cell i in the Voronoi diagram corresponds to input point i.

[garrickf]

Does d3-delaunay allow for a mapping from the Voronoi polygons to data points

Yes. Cell i in the Voronoi diagram corresponds to input point i.

Thanks for the quick reply! I think my issue might be different: the number of cells I get is less than the number of points I pass in, so perhaps some of my points are actually coincident? I found a way to create a mapping for at least some of my points that have cells, which will probably suffice for hovering over information. Thanks for your help!

[Fil]

It will indeed be fixed by https://github.com/d3/d3-delaunay/pull/64