[SciPy-User] struggling and fighting with linalg

josef.pktd at gmail.com josef.pktd at gmail.com
Thu Oct 21 12:03:39 EDT 2010


There are a lot of recommendations in the mailing list and recipes in
various packages how to use linalg more efficiently.
However, yesterday I wasted several hours how to do the multivariate
normal distribution with cholesky factor and solve.

The other problem we have in some of the code is that svd or similar
are calculated several times when individual numpy/scipy functions are
called (e.g. pinv, logdet). Dag has oomatrix, and there are some
packages on the matlab fileexchange (I didn't look at those yet).

As a collection of recipes and reminder for myself, which is which and
what needs to be transposed in s v d, I started to write some linalg
array classes (in attachment). It's essentially just a holder for the
decomposition and operations defined with it, using arrays not
matrices and without any operator overloading.

Still incomplete and only partially tested, and I haven't yet gone
back to doing some statistics with it.

Initially, I was looking for a cholesky decomposition of
(block-)Toeplitz matrices. But those and sparse cholesky seem to be
unavailable in the BSD-compatible landscape, but are for GPL.

If anyone can provide code snippets or (short) references for missing
parts for the linalg, it would reduce my trial-and-error time. I
haven't worked my way through all the suggestions on the mailing list
yet. As for my initial intention, I would also like to get a
multivariate normal class with batteries included, so I don't have to
figure it out and test each time I need a piece.

Josef
-------------- next part --------------
'''Recipes for more efficient work with linalg using classes


intended for use for multivariate normal and linear regression 
calculations

x  is the data (nobs, nvars)
m  is the moment matrix (x'x) or a covariance matrix Sigma

examples:
x'sigma^{-1}x
z = Px  where P=Sigma^{-1/2}  or P=Sigma^{1/2}

Initially assume positive definite, then add spectral cutoff and
regularization of moment matrix, and extend to PCA


Author: josef-pktd
Created on 2010-10-20
'''


import numpy as np
from scipy import linalg


#this has been copied from nitime a long time ago
#TODO: ceck whether class has changed in nitime
class OneTimeProperty(object):

     
    """A descriptor to make special properties that become normal attributes.
 
    This is meant to be used mostly by the auto_attr decorator in this module.
    Author: Fernando Perez, copied from nitime
    """
    def __init__(self,func):
        
        """Create a OneTimeProperty instance.
 
         Parameters
         ----------
           func : method
           
             The method that will be called the first time to compute a value.
             Afterwards, the method's name will be a standard attribute holding
             the value of this computation.
             """
        self.getter = func
        self.name = func.func_name
 
    def __get__(self,obj,type=None):
        """This will be called on attribute access on the class or instance. """
 
        if obj is None:
            # Being called on the class, return the original function. This way,
            # introspection works on the class.
            #return func
            print 'class access'
            return self.getter
 
        val = self.getter(obj)
        #print "** auto_attr - loading '%s'" % self.name  # dbg
        setattr(obj, self.name, val)
        return val


class PlainMatrixArray(object):
    '''Class that defines linalg operation on an array
    
    simplest version as benchmark
    
    linear algebra recipes for multivariate normal and linear
    regression calculations    
    
    '''
    def __init__(self, data=None, sym=None):
        if not data is None:
            if sym is None:
                self.x = np.asarray(data)
                self.m = np.dot(self.x.T, self.x)
            else: 
                raise ValueError('data and sym cannot be both given')
        elif not sym is None:
            self.m = np.asarray(sym)
            self.x = np.eye(*self.m.shape) #default
            
        else:
            raise ValueError('either data or sym need to be given')
        
    @OneTimeProperty
    def minv(self):
        return np.linalg.inv(self.m)
    
    @OneTimeProperty    
    def m_y(self, y):
        return np.dot(self.m, y)
        
    def minv_y(self, y):
        return np.dot(self.minv, y)

    @OneTimeProperty        
    def mpinv(self):
        return linalg.pinv(self.m)

    @OneTimeProperty        
    def xpinv(self):
        return linalg.pinv(self.x)
    
    def yt_m_y(self, y):
        return np.dot(y.T, np.dot(self.m, y))
        
    def yt_minv_y(self, y):
        return np.dot(y.T, np.dot(self.minv, y))
    
    #next two are redundant
    def y_m_yt(self, y):
        return np.dot(y, np.dot(self.m, y.T))
        
    def y_minv_yt(self, y):
        return np.dot(y, np.dot(self.minv, y.T))

    @OneTimeProperty        
    def mdet(self):
        return linalg.det(self.m)
    
    @OneTimeProperty    
    def mlogdet(self):
        return np.log(linalg.det(self.m))
    
    @OneTimeProperty
    def meigh(self):
        evals, evecs = linalg.eigh(self.m)
        sortind = np.argsort(evals)[::-1]
        return evals[sortind], evecs[:,sortind]

    @OneTimeProperty        
    def mhalf(self):
        evals, evecs = self.meigh
        return np.dot(np.diag(evals**0.5), evecs.T)
        #return np.dot(evecs, np.dot(np.diag(evals**0.5), evecs.T))
        #return np.dot(evecs, 1./np.sqrt(evals) * evecs.T))        

    @OneTimeProperty    
    def minvhalf(self):
        evals, evecs = self.meigh
        return np.dot(evecs, 1./np.sqrt(evals) * evecs.T)    


  
class SvdArray(PlainMatrixArray):
    '''Class that defines linalg operation on an array
    
    svd version, where svd is taken on original data array, if
    or when it matters
    
    no spectral cutoff in first version
    '''
    
    def __init__(self, data=None, sym=None):
        super(SvdArray, self).__init__(data=data, sym=sym)
            
        u, s, v = np.linalg.svd(self.x, full_matrices=1)
        self.u, self.s, self.v = u, s, v
        self.sdiag = linalg.diagsvd(s, *x.shape)
        self.sinvdiag = linalg.diagsvd(1./s, *x.shape)

    def _sdiagpow(self, p):
        return linalg.diagsvd(np.power(self.s, p), *x.shape)

    @OneTimeProperty
    def minv(self):
        sinvv = np.dot(self.sinvdiag, self.v)
        return np.dot(sinvv.T, sinvv)
        

    @OneTimeProperty
    def meigh(self):
        evecs = self.v.T
        evals = self.s**2 
        return evals, evecs
    
    @OneTimeProperty    
    def mdet(self):
        return self.meigh[0].prod()
        
    @OneTimeProperty    
    def mlogdet(self):
        return np.log(self.meigh[0]).sum()
    
    @OneTimeProperty    
    def mhalf(self):
        return np.dot(np.diag(self.s), self.v)
     
    @OneTimeProperty
    def xxthalf(self):
        return np.dot(self.u, self.sdiag)
        
    @OneTimeProperty
    def xxtinvhalf(self):
        return np.dot(self.u, self.sinvdiag)
        

class CholArray(PlainMatrixArray):
    '''Class that defines linalg operation on an array
    
    cholesky version, where svd is taken on original data array, if
    or when it matters
    
    plan: use cholesky factor and cholesky solve
    nothing implemented yet
    '''
    
    def __init__(self, data=None, sym=None):
        super(SvdArray, self).__init__(data=data, sym=sym)



def testcompare(m1, m2):
    from numpy.testing import assert_almost_equal, assert_approx_equal
    decimal = 12
    
    #inv
    assert_almost_equal(m1.minv, m2.minv, decimal=decimal)
    
    #matrix half and invhalf
    #fix sign in test, should this be standardized
    s1 = np.sign(m1.mhalf.sum(1))[:,None]
    s2 = np.sign(m2.mhalf.sum(1))[:,None]
    scorr = s1/s2
    assert_almost_equal(m1.mhalf, m2.mhalf * scorr, decimal=decimal)
    assert_almost_equal(m1.minvhalf, m2.minvhalf, decimal=decimal)
    
    #eigenvalues, eigenvectors    
    evals1, evecs1 = m1.meigh
    evals2, evecs2 = m2.meigh
    assert_almost_equal(evals1, evals2, decimal=decimal)
    #normalization can be different: evecs in columns
    s1 = np.sign(evecs1.sum(0))
    s2 = np.sign(evecs2.sum(0))
    scorr = s1/s2
    assert_almost_equal(evecs1, evecs2 * scorr, decimal=decimal)
    
    #determinant
    assert_approx_equal(m1.mdet, m2.mdet, significant=13)
    assert_approx_equal(m1.mlogdet, m2.mlogdet, significant=13)

####### helper function for interactive work
def tiny2zero(x, eps = 1e-15):
    '''replace abs values smaller than eps by zero, makes copy
    '''
    mask = np.abs(x.copy()) <  eps
    x[mask] = 0
    return x
    
def maxabs(x):
    return np.max(np.abs(x))


#if __name__ == '__main__':

        
n = 5
y = np.arange(n)
x = np.random.randn(100,n)
autocov = 2*0.8**np.arange(n) +0.01 * np.random.randn(n)
sigma = linalg.toeplitz(autocov)

mat = PlainMatrixArray(sym=sigma)
print tiny2zero(mat.mhalf)
mih = mat.minvhalf
print tiny2zero(mih) #for nicer printing

mat2 = PlainMatrixArray(data=x)
print maxabs(mat2.yt_minv_y(np.dot(x.T, x)) - mat2.m)
print tiny2zero(mat2.minv_y(mat2.m))

mat3 = SvdArray(data=x)
print mat3.meigh[0]
print mat2.meigh[0]

testcompare(mat2, mat3)

'''
m = np.dot(x.T, x)

u,s,v = np.linalg.svd(x, full_matrices=1)
Sig = linalg.diagsvd(s,*x.shape)

>>> np.max(np.abs(np.dot(u, np.dot(Sig, v)) - x))
3.1086244689504383e-015
>>> np.max(np.abs(np.dot(u.T, u) - np.eye(100)))
3.3306690738754696e-016
>>> np.max(np.abs(np.dot(v.T, v) - np.eye(5)))
6.6613381477509392e-016
>>> np.max(np.abs(np.dot(Sig.T, Sig) - np.diag(s**2)))
5.6843418860808015e-014

>>> evals,evecs = linalg.eigh(np.dot(x.T, x))
>>> evals[::-1]
array([ 123.36404464,  112.17036442,  102.04198468,   76.60832278,
         74.70484487])

>>> s**2
array([ 123.36404464,  112.17036442,  102.04198468,   76.60832278,
         74.70484487])

>>> np.max(np.abs(np.dot(v.T, np.dot(np.diag(s**2), v)) - m))
1.1368683772161603e-013

>>> us = np.dot(u, Sig)
>>> np.max(np.abs(np.dot(us, us.T) - np.dot(x, x.T)))
1.0658141036401503e-014

>>> sv = np.dot(Sig, v)
>>> np.max(np.abs(np.dot(sv.T, sv) - np.dot(x.T, x)))
1.1368683772161603e-013


'''


More information about the SciPy-User mailing list