restrdeter

restrdeter computes determinant restriction

Syntax

  • out=restrdeter(eigenvalues, niini, restr)example
  • out=restrdeter(eigenvalues, niini, restr, tol)example
  • out=restrdeter(eigenvalues, niini, restr, tol, userepmat)example

Description

restrdeter restricts the determinant according to the constraint specified in scalar restr. This function is called in every concentration step of function tclust in case determinant restriction is needed

example

out =restrdeter(eigenvalues, niini, restr) Example using all default options.

example

out =restrdeter(eigenvalues, niini, restr, tol) Determinant restriction when an eigenvalue is 0.

example

out =restrdeter(eigenvalues, niini, restr, tol, userepmat) An example using option arguments tol and repmat.

Examples

expand all

  • Example using all default options.
  • Suppose v=3 and k=4 so the matrix containing the eigenvalues is 3-by-4 First column of matrix eigenvalues contains the eigenvalues of the first group Second column of matrix eigenvalues contains the eigenvalues of the second group Thrid column of matrix eigenvalues contains the eigenvalues of the third group Fourth column of matrix eigenvalues contains the eigenvalues of the fourth group

    rng(10,'twister')
    eigenvalues=abs(10*randn(3,4));
    % niini is the column vector containing the sizes of the 4 groups
    niini=[30;40;20;10];
    out=restrdeter(eigenvalues,niini,1.1)
    disp('Input matrix of unrestricted eigenvalues')
    disp(eigenvalues)
    disp('Output matrix of restricted eigenvalues which satisfy determinant constraint')
    disp(out)
    disp('Ratio between largest and smallest determinant')
    disp(max(prod(eigenvalues))/min(prod(eigenvalues)))
    disp('Ratio between largest and smallest restricted determinants')
    disp(max(prod(out))/min(prod(out)))
    out =
    
        4.7173   50.0736    7.6327    3.7940
       10.7021    0.2257    3.8281    3.6133
        2.8139   11.4263    4.8620   10.3628
    
    Input matrix of unrestricted eigenvalues
        6.4581   11.1831   11.6942    7.3428
       14.6513    0.0504    5.8652    6.9930
        3.8523    2.5519    7.4491   20.0559
    
    Output matrix of restricted eigenvalues which satisfy determinant constraint
        4.7173   50.0736    7.6327    3.7940
       10.7021    0.2257    3.8281    3.6133
        2.8139   11.4263    4.8620   10.3628
    
    Ratio between largest and smallest determinant
      715.8613
    
    Ratio between largest and smallest restricted determinants
        1.1000
    
    

  • Determinant restriction when an eigenvalue is 0.
  • Suppose 5 variables and six groups

    av=abs(randn(5,6));
    % The third eigenvalue of the second groups is set to 0
    av(3,2)=0;
    % Maximum ratio among determinants must be equal to 1.6.
    restr=1.6;
    % Group sizes
    niini=[30;40;20;10;50;100];
    disp('Original values of the determinants')
    disp(prod(av))
    % Apply the restriction
    a=restrdeter(av,niini,restr);
    disp('Restricted eigenvalues which satisfy determinant constraint')
    disp(a)
    disp('Values of restricted determinants')
    disp(prod(a))
    disp('Maximum value of ratio among determinants')
    disp(max(prod(a))/min(prod(a)))
    Original values of the determinants
        0.0055         0    0.0010    0.0000    0.1825    0.0432
    
    Restricted eigenvalues which satisfy determinant constraint
        1.3868   53.0992    0.6626    0.0566    0.6163    1.2339
        0.0670   61.2269    0.5191    0.6177    0.3601    1.0498
        0.6905    0.0000    0.4029    1.3884    0.6211    0.1682
        0.1217    7.6070    0.1578    0.7623    0.3348    0.1529
        1.1867   61.2269    0.4240    0.2507    0.3214    0.4454
    
    Values of restricted determinants
        0.0093    0.0093    0.0093    0.0093    0.0148    0.0148
    
    Maximum value of ratio among determinants
        1.6000
    
    

  • An example using option arguments tol and repmat.
  • Suppose 3 variables and six groups

    av=abs(randn(3,6));
    % Maximum ratio among determinants must be equal to 1.6.
    restr=1.6;
    % Group sizes
    niini=[30;40;20;10;50;100];
    % Apply the restriction using a tolerance of 1e-12 and use MATLAB
    % function repmat for the computations
    tol=1e-12;
    repm=1;
    a=restrdeter(av,niini,restr,tol,repm);

    Related Examples

    expand all

  • Determinant restriction when all eigenvalues of a group are 0.
  • Two variables and five groups.

    av=abs(randn(2,5));
    restr=1.6;
    niini=[30;40;20;10;50];
    av(:,2)=0;
    a=restrdeter(av,niini,restr);
    disp('Maximum value of ratio among determinants')
    disp(max(prod(a))/min(prod(a)))

  • Determinant restriction when all eigenvalues of two groups are 0.
  • niini=[30;40;20;10;50];
    av=abs(randn(2,5));
    av(:,2:3)=0;
    a=restrdeter(av,niini,restr);
    disp('Maximum value of ratio among determinants')
    disp(max(prod(a))/min(prod(a)))

    Input Arguments

    expand all

    eigenvalues — Eigenvalues. Matrix.

    v x k matrix containing the eigenvalues of the covariance matrices of the k groups.

    v is the number of variables of the dataset which has to be clustered.

    Data Types: single| double

    niini — Cluster size. Column vector.

    k x 1 vector containing the size of the k clusters

    Data Types: single| double

    restr — Restriction factor. Scalar (default) or vector of length 2.

    If restr is a scalar it defines the maximum ratio of the determinants which is allowed. In other words, we impose the constraint on the covariance matrices:

    \[ \frac{\max_{j=1,...,k} |\Sigma_j|}{\min_{j=1,...,k} |\Sigma_j|} \leq restr \] where $restr \geq 1$. In this case the "shape" constraint (as defined below) applied to each group is fixed to $c_{shape}=10^{10}$, to ensure the procedure is (virtually) affine equivariant. In other words, the decomposition or the $j$-th scatter matrix $\Sigma_j$ is \[ \Sigma_j=\lambda_j^{1/v} \Omega_j \Gamma_j \Omega_j' \]

    where $\Omega_j$ is an orthogonal matrix of eigenvectors, $\Gamma_j$ is a diagonal matrix with $|\Gamma_j|=1$ and with elements $\{\gamma_{j1},...,\gamma_{jv}\}$ in its diagonal (proportional to the eigenvalues of the $\Sigma_j$ matrix) and $|\Sigma_j|=\lambda_j$. The $\Gamma_j$ matrices are commonly known as "shape" matrices, because they determine the shape of the fitted cluster components. The following $k$ constraints are then imposed on the shape matrices:

    \[ \frac{\max_{l=1,...,v} \gamma_{jl}}{\min_{l=1,...,v} \gamma_{jl}}\leq c_{shape}, \text{ for } j=1,...,k, \]

    The particular case $restr=1$ forces all determinants of the scatter matrices to be equal i.e. $|\Sigma_1|=...= |\Sigma_k|$.

    If $restr$ is a vector of length 2 the second element refers to $c_{shape}$ of the previous equation. In other words, for example if $restr=[3, 10]$ we impose the $k+1$ constraints

    \[ \frac{\max_{j=1,...,k} |\Sigma_j|}{\min_{j=1,...,k} |\Sigma_j|} \leq restr(1)=3 \] and \[ \frac{\max_{l=1,...,v} \gamma_{jl}}{\min_{l=1,...,v} \gamma_{jl}} \leq restr(2)=10, \text{ for } j=1,...,k, \]

    Different constrained clustering problems can be defined when varying $restr(1)$ and $restr(2)$. In particular, we are ideally searching for spherical clusters when $restr(2)=1$.

    Models with variable volume and spherical clusters are handled with $1<restr(1)<\infty$ and $restr(2)=1$. The $restr(1)=restr(2)=1$ case often yields a very constrained parametrization because it implies spherical clusters with equal volumes.

    Data Types: single| double

    Optional Arguments

    tol — tolerance. Scalar defining the tolerance of the procedure.

    The default value is 1e-8

    Example: 'tol',[1e-18]

    Data Types: double

    userepmat — use builtin repmat. Scalar.

    If userepmat is true function repmat is used instead of bsxfun inside the procedure. Remark: repmat is built in from MATLAB 2013b so it is faster to use repmat if the current version of MATLAB is >2013a

    Example: 'userepmat',1

    Data Types: double

    Output Arguments

    expand all

    out —Restricted eigenvalues which satisfy the determinant constraint. Matrix

    v-by-k matrix containing restricted eigenvalues.

    The ratio between the determinants (that is the product of the columns of matrix out) is not greater than restr

    References

    Fritz H., Garcia-Escudero, L.A. and Mayo-Iscar, A. (2013), A fast algorithm for robust constrained clustering, "Computational Satistics and Data Analysis", Vol. 61, pp. 124-136.

    [Available at http://www.eio.uva.es/infor/personas/tclust_algorithm.pdf]

    This page has been automatically generated by our routine publishFS