“…Also, as defined here, n is a power of 2, but variants of this construction exist for other values of n.) Importantly, applying the randomized Hadamard transform, i.e., computing the product xDH for any vector x ∈ R n takes O(n log n) time (or even O(n log r) time if only r elements in the transformed vector need to be accessed). Applying such a structured random projection was first proposed in [71,72], it was first applied in the context of randomized matrix algorithms in [80,81], and there has been a great deal of research in recent years on variants of this basic structured random projection that are better in theory or in practice [73,81,82,83,84,85,1,86,87,88,89,90]. For example, one could choose Ω = DHS, where S is a random sampling matrix, as defined above, that represents the operation of uniformly sampling a small number of columns from the randomized Hadamard transform.…”