In bulk quantum computation one can manipulate a large number of indis
tinguishable quantum computers by parallel unitary operations and meas
ure expectation values of certain observables with limited sensitivity
. The initial state of each computer in the ensemble is known but not
pure. Methods for obtaining effective pure input states by a series of
manipulations have been described by Gershenfeld and Chuang (logical
labeling) [Science 275. 350 (1997)] and Cory et al. (spatial averaging
) [Proc. Natl. Acad. Sci. USA 94, 1634 (1997)] for the case of quantum
computation with nuclear magnetic resonance. We gives different techn
ique called temporal averaging. This method is based on classical rand
omization, requires no ancilla quantum bits, and can be implemented in
nuclear magnetic resonance without using gradient fields. We introduc
e several temporal averaging algorithms suitable for both high-tempera
ture and low-temperature bulk quantum computing and analyze the signal
-to-noise behavior of each. Most of these algorithms require only a co
nstant multiple of the number of experiments needed by the other metho
ds for creating effective pure stales.