The Nyström approximation is a technique for efficiently approximating large kernel matrices by computing exact kernel evaluations on a small subset of landmark points and using these to reconstruct a low-rank approximation of the full matrix. In the context of WiFi/CSI sensing for occupancy detection and crowd counting, it matters because it reduces the computational cost of kernel-based machine learning methods — such as Gaussian processes or support vector machines — that would otherwise be prohibitively expensive when applied to high-dimensional CSI feature spaces with large numbers of samples. Key variants include uniform random sampling of landmark points, k-means-based selection, and greedy approximation strategies, which differ in how representative landmarks are chosen to balance approximation accuracy against computational efficiency.
Dictionary term