1: \begin{abstract}

2: We introduce a class of dimension reduction estimators based on an

3: ensemble of the minimum average variance estimates of functions that

4: characterize the central subspace, such as the characteristic

5: functions, the Box--Cox transformations and wavelet basis. The ensemble

6: estimators exhaustively estimate the central subspace without imposing

7: restrictive conditions on the predictors, and have the same convergence

8: rate as the minimum average variance estimates. They are flexible and

9: easy to implement, and allow repeated use of the available sample,

10: which enhances accuracy. They are applicable to both univariate and

11: multivariate responses in a unified form. We establish the consistency

12: and convergence rate of these estimators, and the consistency of a

13: cross validation criterion for order determination. We compare the

14: ensemble estimators with other estimators in a wide variety of models,

15: and establish their competent performance.

16: \end{abstract}