On random variate generation when only moments of Fourier coefficients are known (Q1123530)

Essentially the method is a combination of acception/rejection with use of approximating series such that each succeeding member within the series is needed indeed only with (heavily) decreasing probability, which fact results in (unexpected) low expense in computation in all probability, for each number to be generated. The discussed cases and examples exhibit also easily evaluable rejection criteria. The needed proofs succeed under weak assumptions concerning the distribution and the approximations. Knowledge of the first Fourier coefficient or moments is presupposed; for even functions with convex decreasing Fourier coefficients even no estimate of the series remainder is needed: usually the calculation requires comparison of the remaining rest of the infinite series, which may be delimited from above by an estimation with moderate computation-up to sufficient precision. Besides standard Fourier series other kinds are considered: Legendre and Laguerre/Hermite polynomials, furthermore estimations derived from Lipschitz conditions. The resulting programs are elegantly short, even for usage of so-called \(\theta\) factors with Fourier series (reminding of relaxation coefficients) to improve (``smoothe'') the approximation and with many classes for the calculation of good estimates to the series remainder. The probable, expected numbers of iteration, i.e. mainly the number of needed members in the series, for one random number to be generated are derived. The user should check the variance of these numbers and the effect of the limited precision in the computing system, when the programs are applied.