For over a decade algorithms that predict RNA secondary structure via identification of the minimum free energy (MFE) structure have deservedly been the dominant players in computationally predicting these structures. However, messenger RNAs (mRNA) likely exist in a population of structures, and in fact some message structures have been found to regulate expression through conformational change. To address the utility of MFE structures to efficiently represent this population we compared MFE structures to representative samples from the Boltzman weighted ensemble of structures. We found that human mRNAs are unlikely to occupy the minimum free energy (MFE) state, and that the Boltzman weighted ensemble of secondary structures of these messages often contain multiple classes. Our findings further indicate that MFE structures often ineffectively represent these ensembles, and that one or a few centroid structures more effectively capture the overall character of these ensembles. Surprisingly, we also find that these centroid structures predict the secondary structure of structural RNAs with fewer errors that MFE structures. These findings may well be reflections of the curse of dimensionality associated with such high dimensional problems, and likely indicate the futility of optimum for high dimensional inferences on more general graphs such as inferences on regulatory networks.