Ensemble Averaging and the Curse of Dimensionality

Abstract

When comparing climate models to observations, it is often observed that the mean over many models has smaller errors than most or all of the individual models. This paper will show that a general consequence of the nonintuitive geometric properties of high-dimensional spaces is that the ensemble mean often outperforms the individual ensemble members. This also explains why the ensemble mean often has an error that is 30% smaller than the median error of the individual ensemble members. The only assumption that needs to be made is that the observations and the models are independently drawn from the same distribution. An important and relevant property of high-dimensional spaces is that independent random vectors are almost always orthogonal. Furthermore, while the lengths of random vectors are large and almost equal, the ensemble mean is special, as it is located near the otherwise vacant center. The theory is first explained by an analysis of Gaussian- and uniformly distributed vectors in high-dimensional spaces. A subset of 17 models from the CMIP5 multimodel ensemble is then used to demonstrate the validity and robustness of the theory in realistic settings.