drsc

Dimensionality Reduction for Spectral Clustering
[IDX,W,FVAL,SUMD,ITER] = DRSC(X, K, SIGMA, VARARGIN)

 [IDX, W, FVAL, SUMD, ITER] = DRSC(X, K, SIGMA) produces a clustering of the
 N-by-D data matrix (X) into (K) clusters, by identifying the optimal
 (K-1)-dimensional linear subspace to project the data. (SIGMA) is the
 bandwidth parameter for the Gaussian kernel used to estimate the kernel
 matrix.

 [IDX, W, FVAL, SUMD, ITER] = DRSC(X, K, SIGMA) returns the cluster assignment,
 (IDX); the projection matrix (W); a vector of values of the projection index
 at each iteration (FVAL); the sum of squared distances to the cluster centres
 in the optimal linear subspace, (SUMD); and finally the iteration at which
 the algorithm terminated, (ITER). If DRSC fails to converge ITER=0

 [IDX, W, FVAL ,SUMD, ITER] = DRSC(X, K, S, 'PARAM1',val1, 'PARAM2',val2, ...)
 specifies optional parameters in the form of Name,Value pairs. 

 'v0' - D-by-Q matrix of initial projection vectors. Q determines dimensionality of
    projection subspace
    (default: (K-1) first principal components)

 'maxit' - Number of DRSC iterations (default: 50)

 'maxitdim' - Number of gradient descent iterations for each dimension (default: 50)

 'ftol' - Stopping criterion for change in objective function value over consecutive iterations
    (default: 1.e-4, suggested in Niu et al. AISTATS 2011)

 'verb' - Verbosity. Values greater than 0 enables progress monitoring during execution
    (default: 0)

Reference:
D. Niu, J.G. Dy and M.I. Jordan. Dimensionality Reduction for Spectral Clustering.
Proceedings of the 14th International Conference on Artificial Intelligence and Statistics,
volume 15 of JMLR W&CP, pages 552-560, 2011.