The paper works within a framework where probability distributions can be used as "coordinate systems" for doing geometry. Just as you can describe points in space using different coordinate grids, you can describe and compare probability distributions using other distributions as reference frames. Within this setup, you can define a notion of an "average" of distributions, called a probability barycenter, along with related summary statistics called Kolmogorov moments. The challenge is that the space of valid coordinate systems is enormous and infinite-dimensional, making it hard to work with computationally.
To make the framework tractable, the paper develops a way to approximate any valid coordinate system using simpler, finite-dimensional building blocks. The key insight is that switching between two coordinate systems on the same interval can always be described by a strictly increasing distortion function on the unit interval. The paper shows that a natural class of these distortion functions, piecewise-linear ones that only go upward, can be written as weighted combinations of simple "ramp" functions, like mixing a handful of basic shapes to build a more complex one. The mixing weights live in a simplex, meaning they are non-negative and sum to one, which gives a clean geometric picture. The paper also studies a smoother approximation method using Bernstein polynomials, which connect to order statistics from a uniform distribution and have nice mathematical properties.
The practical payoff is that these approximations come with explicit error guarantees. If you approximate the distortion function closely, the resulting barycenters and moments are also close, with quantifiable bounds on how much they can drift. The paper also points out a pleasing connection to quantile estimation: in a specific but natural setting, the barycenter turns out to be a quantile of the data, with the particular quantile level determined by the distortion being used. Numerical examples are included to show both approximation schemes in action and to illustrate how errors in the coordinate system translate into errors in the computed barycenters.