Summary of Pseudorandom Number Generation by Nonlinear Methods Yu Zhang yz08@fsu.edu In this paper, the author summarized recent research on nonlinear pseudorandom number generators. Many classical methods are widely used in generating random numbers such as linear congruential method. But the sequences produced by these methods have many regular features. So, researchers wanted to use nonlinear methods to improve the performance of the sequence. Discrepancy is defined by DN (t0, t1, …, tN-1)=supJ|FN(J)-V(J)|, J is any subinterval of [0, 1), FN(J) is N-1 times the number of points falling into J, and V(J) is the s-dimension volume of J. It measures the equidistribution of the pseudorandom numbers. The first generator discussed by the author is the general nonlinear congruential generator. Sequences formed by this method have the discrepancy DN(1)=DN(x0, x1, … ,xN-1) which satisfies DN(1)=O (N-1p1/2(log p)2). The discrepancy Dp(s)=Dp(x0, x1, … ,xp-1) where xn=(xn, xn+1, … ,xn+s-1) is used to analyze the statistical independence features of the sequence. For sequence over the full period, the discrepancy satisfies Dp(s) = O (p-1/2(log p)s) for 2 ≤ s ≤ p; and for sequence over parts of the period, the discrepancy DN(s) = DN(x0, x1, … ,xN-1) where xn = (xsn, xsn+1, … ,xsn+s-1) satisfies DN(s) = O(N-1p1/2(log p)s+1) for 2 ≤ s ≤ d-1. The author also stated another method called explicit inversive congruential method. Assume we have a prime number p which is larger than 5, we define Z p={0, 1, …, p-1} as the finite field of order p. For integers a and b belong to Z p, we can get an explicit inversive congruential sequence (yn) by yn = ̅̅̅̅̅̅̅̅̅ an + b ≡ (an + b)p−2 (mod p). Then, we can obtain a sequence (xn) in the interval [0, 1) by xn=yn/p. The discrepancy satisfies DN(1)=O(N-1p1/2(log p)2). Like the analysis of the statistical independence features above, for full period, Dp(s) = O (p-1/2(log p)s), and for independence of parts of the period, DN(s) = O(N-1p1/2(log p)s+1). As we summarized before, the explicit inversive congruential method has a good feature: any hyperplane in Zpd contains at most d points of the set Vd. So, it has favorable equidistribution, statistical independence and structural features. Furthermore, a large number of parallel streams of explicit inversive congruential pseudorandom numbers also make it suitable for parallel computations.