actually I think the best psu-random generator would be "hierarchical." That is, you use a very fast+simple generator for a while, then reseed it using a slower but more random generator, that in turn would be re-seeded at longer intervals with a slower but still more random generator etc. You aim to get nearly the speed of the fastest but nearly the randomness of the slowest generator. The "feedin" technique is a way to do such reseeding in a "continual slow drip" rather than all at once. Seems to me the hierarchical idea can be backed up with a THEOREM (or pretty close): If "randomness" is measured by expected computational complexity of a statistical test that finds a problem with it, and if achievable randomness grows with generator runtime exponentially, then it is possible to achieve arbitrarily enormous randomness with only a constant factor slower average runtime than the simplest generator.