How to use DynStatCov

DynStatCov comes with a simple interface for the following work-flow:

1. Compute an initial co-variance matrix from a base set of observations
2. Add a new observation to update the co-variance matrix
3. Get the current co-variance matrix as numpy array
4. Go to step 2.

See the library reference for a description of the complete functionality.

Examples

Example with initially three observations of four features each, which is updated once.

>>> import numpy as np
>>> from dynstatcov import Dynstatcov
>>> X = np.array([[3, 4, 1, 2],
[4, 4, 0, 3],
[1, 3, 4, 2]], dtype=numpy.float64)

>>> dsc = Dynstatcov(X)
>>> dsc.get_cov()
array([[ 2.33333397, 0.83333397, -3.16666698, 0.66666603],
[ 0.83333397, 0.33333206, -1.16666603, 0.16666603],
[-3.16666698, -1.16666603, 4.33333302, -0.83333397],
[ 0.66666603, 0.16666603, -0.83333397, 0.33333397]], dtype=float64)

>>> dsc.get_n_samples()
3

>>> y = np.array([3, 4, 1, 2], dtype=numpy.float64)
>>> dsc.update(y)
>>> dsc.get_cov()
array([[ 1.58333337, 0.58333337, -2.16666675, 0.41666669],
[ 0.58333337, 0.25 , -0.83333337, 0.08333334],
[-2.16666675, -0.83333337, 3. , -0.5 ],
[ 0.41666669, 0.08333334, -0.5 , 0.25 ]], dtype=float64)

Internally, only the upper part of the symmetrical co-variance matrix is stored and the full matrix reconstructed on each call to get_cov(). Alternatively, you can also request the upper triangular matrix, which is slightly faster.

>>> dsc.get_cov_tri()
array([ 1.58333337, 0.58333337, -2.16666675, 0.41666669, 0.25 ,
-0.83333337, 0.08333334, 3. , -0.5 , 0.25 ], dtype=float64)

We can also remove a sample setting the optional second argument of update() non-zero.

>>> dsc.update(y, 1)
>>> dsc.get_cov()
array([[ 2.33333397,  0.83333397, -3.16666698,  0.66666603],
       [ 0.83333397,  0.33333206, -1.16666603,  0.16666603],
       [-3.16666698, -1.16666603,  4.33333302, -0.83333397],
       [ 0.66666603,  0.16666603, -0.83333397,  0.33333397]], dtype=float64)

Note on precision

DynStatCov can be compiled with different precisions (e.g numpy.float32 or numpy.float64). You might have to test, which version you are running. By default, its double precision. If required, you can always change typedef of DTYPE_t in the cython code and re-compile, as described in Configuration of DynStatCov.

Note on update() method

The sample will not get checked for its length. You are yourself responsible, that it fits the length of the observations passed on class initialization.