Some aspects of real-world road networks seem to have an approximate scale invariance property, motivating study of mathematical models of random networks whose distributions are exactly invariant under Euclidean scaling. This requires working in the continuum plane, so making a precise definition is not trivial. We introduce an axiomatization of a class of processes we call scale-invariant random spatial networks, whose primitives are routes between each pair of points in the plane. One concrete model, based on minimum-time routes in a binary hierarchy of roads with different speed limits, has been shown to satisfy the axioms, and two other constructions (based on Poisson line processes and on dynamic proximity graphs) are expected also to do so. We initiate study of structure theory and summary statistics for general processes in the class. Many questions arise in this setting via analogies with diverse existing topics, from geodesics in first-passage percolation to transit node-based route-finding algorithms.

Motivated by the shape of transportation networks such as subways, we consider a distribution of points in the plane and ask for the network G of given length L that is optimal in a certain sense. In the general model, the optimality criterion is to minimize the average (over pairs of points chosen independently from the distribution) time to travel between the points, where a travel path consists of any line segments in the plane traversed at slow speed and any route within the subway network traversed at a faster speed. Of major interest is how the shape of the optimal network changes as L increases. We first study the simplest variant of this problem where the optimization criterion is to minimize the average distance from a point to the network, and we provide some general arguments about the optimal networks. As a second variant we consider the optimal network that minimizes the average travel time to a central destination, and we discuss both analytically and numerically some simple shapes such as the star network, the ring, or combinations of both these elements. Finally, we discuss numerically the general model where the network minimizes the average time between all pairs of points. For this case, we propose a scaling form for the average time that we verify numerically. We also show that in the medium-length regime, as L increases, resources go preferentially to radial branches and that there is a sharp transition at a value L_{c} where a loop appears.