Visualization is a fundamental part of modern data-centered applications: a plot can show you in the blink of an eye if your data has the shape that you’re expecting, but having to retrieve all your samples just to cram them in a graph without enough pixels to show them all is clearly not a good idea.

Downsampling seems the obvious next step, but how to choose which samples to keep and which to throw away? The key idea is to take the samples that make the overall shape of your data as similar to the original one as possible.