1: \begin{abstract}

2: A new methodology is proposed for generating realizations of a random

3: vector with values in a finite-dimensional Euclidean space that are

4: statistically consistent with a data set of observations of this vector.

5: The probability distribution of this random vector, while a-priori not

6: known, is presumed to be concentrated on an unknown subset of the

7: Euclidean space. A random matrix is introduced whose columns are

8: independent copies of the random vector and for which the number of

9: columns is the number of data points in the data set. The approach is

10: based on the use of (i) the multidimensional kernel-density estimation

11: method for estimating the probability distribution of the random

12: matrix, (ii) a MCMC method for  generating realizations for the random

13: matrix, (iii) the diffusion-maps approach  for discovering and

14: characterizing the geometry and the structure of the data set, and

15: (iv) a reduced-order representation of the random matrix, which is

16: constructed using the diffusion-maps vectors associated with the first

17: eigenvalues of the transition matrix relative to the given data

18: set. The convergence aspects of the proposed methodology are

19: analyzed and a numerical validation is explored through three

20: applications of increasing complexity. The proposed method is found to

21: be robust to noise levels and data complexity as well as to the intrinsic

22: dimension of data and the size of experimental data sets.  Both the

23: methodology and the underlying mathematical framework presented in

24: this paper contribute new capabilities and perspectives at the

25: interface of uncertainty quantification, statistical data analysis,

26: stochastic modeling and associated statistical inverse problems.

27: % for boundary value problems, in the design of experiments for random

28: % parameters, and in signal processing and machine learning.

29: \end{abstract}

30: