John Urschel's Favorite Theorem
blogs.scientificamerican.com
blogs.scientificamerican.com
His favourite theorem is described here https://arxiv.org/abs/0808.0163. It's a theorem in graph theory, and it says that if you have any graph, you can find a subgraph with "not too many edges" that is "like" the original graph, in a specific sense.
"Not too many" means that for some value d which you get to choose, the number of edges is no more than about d times the number of vertices. (In general, the number of edges in a graph can scale like the number of vertices squared.)
"Like the original graph" means that a thing called the "graph Laplacian" of the cut-down graph is a good approximation to the graph Laplacian of the original. Here's one way to think about what the graph Laplacian is. Imagine a sort of "heat flow" process on the graph, where at any given time each vertex has a temperature, and at each time step you move each vertex's temperature towards the average of its neighbours' temperatures. All the possible temperature assignments constitute a finite-dimensional vector space; call it V; the graph Laplacian is a linear map from V to itself that maps a temperature assignment to the resulting change produced by one step of the heat-flow process. It's a sort of discrete version of the second derivative operator. And it turns out that lots of interesting properties of a graph can be computed just from knowing the graph Laplacian, or even just from knowing its eigenvalues.
"Good approximation" means that the relative error is on the order of 1/sqrt(d) where d is that value you got to choose earlier. So the more edges you're allowed to keep, the better the approximation can be made.
(All of the above is a handwavy approximation. If you understood everything I said and want the correct version, you can find it in the paper itself.)