The joint probability of obtaining, by chance, so many low-probability outcomes is very remote [P < 0.001, based on Fisher's technique for combining probabilities from independent tests of significance (Sokal and Rohlf, 1981)]. Therefore, although we cannot discern any clear trend in the among-population variance for mean fitness, we know it was initially equal to 0 and that significant variation arose early in the experiment and persisted throughout the 10,000 generations.
Nonetheless, we cannot absolutely exclude the possibility that the populations might eventually converge in mean fitness. In fact, when similar analyses were performed after 2000 generations, we suggested that the populations might have diverged in mean fitness only transiently, owing to stochastic variation in the timing of a series of substitutions leading to the same fitness peak (Lenski et al., 1991). The reason for our earlier caution in claiming sustained divergence was not because the among-population genetic variance was much smaller in the first 2000 generations (Figure 7), but rather because the rate of adaptation continued at a much more rapid pace (Figure 6), leaving open the possibility that the less fit populations would soon catch up to their better-adapted cousins. But with an additional 8000 generations, the rate of fitness improvement has become so slow that eventual convergence now seems very unlikely. Whereas the average rate of improvement between generations 1000 and 2000 was 0.108 (±0.020 SEM) per 1000