dirksen-2016-subgaussian-jl-generalization

Summary

Dirksen showed that the fast, sparse dimensionality-reduction guarantee doesn't depend on using one particular random-number recipe; it works for a very broad family of random matrices as long as their entries are independent, centered, normalized, and not too heavy-tailed. In practice, this means the system can use many more standard random constructions without losing the theoretical guarantee, making the result far more flexible to apply.