Random Fourier features are a common way to approximate shift-invariant kernels, such as the Gaussian kernel exp(-1/2 ||x-y||^2). The approach uses independent samples from the kernel's spectral measure to construct an unbiased estimator of the kernel. This makes large-scale kernel methods more tractable.
A new arXiv paper, "Random Fourier Features in Near-Linear Time," presents a method that brings this construction into near-linear time. The abstract describes the estimator as built from independent samples from the spectral measure, but does not give further algorithmic details in the available text.
The significance of the work lies in the potential speedup for kernel approximation. However, the abstract does not specify the exact complexity or the range of kernels covered beyond shift-invariant ones. The paper appears to be a direct extension of the classical random Fourier feature framework.