Conditioning the Covariances

Conditioning the covariances is accomplished without explicitly computing $ {\bf\Sigma}_i$ as well. As mentioned in Table 13.1, step 5, there are two methods, the BIAS and CONSTRAINT methods. The BIAS method is simpler. On the other hand, the CONSTRAINT method delivers a better PDF estimate because the covariances are not biased and appears to converge faster. But, it may interfere with the monotonic increasing property of the E-M algorithm, i.e. that the total log-likelihood always goes up, but this is still an unresolved issue. Both methods are based on the idea of independent measurement error in the elements of $ {\bf z}$. Let $ {\bf D}$ be a diagonal covariance matrix with $ {\bf D}_{n,n}=\rho^2_n$. The two methods differ in how they regard $ {\bf D}$. The BIAS method assumes $ {\bf D}$ is an a priori estimate of $ {\bf\Sigma}$, while the CONSTRAINT method assumes $ {\bf D}$ is a measurement error covariance.

The BIAS method is implemented by adding $ {\bf D}$ to the newly formed covariance estimate. But, because we do not work with $ {\bf\Sigma}$ directly, it is necessary to implement the conditioning as follows: Let $ {\bf X}^\prime = {\bf QR}$. The upper triangular matrix $ {\bf R}$ is retained and $ {\bf Q}$ is discarded. Next, we form the matrix as shown below for the case $ P=3$:

$\displaystyle {\bf R}^* =
\left[ \begin{array}{c} {\bf R} \cdots  {\rm dia...
...
\rho_1 & 0 & 0 \\
0 & \rho_2 & 0 \\
0 & 0 & \rho_3
\end{array} \right]
$

It may be verified that $ {\bf R}^{*\prime} {\bf R}^*$ is the same as $ {\bf\Sigma}_i$ with the diagonal adjustments. Next, the QR-decomposition of $ {\bf R}^*$ is computed and the upper triangular part is stored.

The CONSTRAINT method assumes that $ {\bf\Sigma}={\bf\Sigma}_0 + {\bf D}$ where $ {\bf\Sigma}_0$ is an arbitrary covariance. Let the eigendecomposition of $ {\bf\Sigma}$ be $ {\bf\Sigma}={\bf V} {\bf S}^2 {\bf V}^\prime$. Clearly, then

$\displaystyle {\bf S}^2= {\bf V}^\prime {\bf\Sigma}_0 {\bf V} +
{\bf V}^\prime {\bf D} {\bf V}.
$

Thus, the diagonal elements of $ {\bf S}$ can be no smaller than the square root of the diagonal elements of $ {\bf V}^\prime {\bf D} {\bf V}$. Note that $ {\bf V}$ and $ {\bf S}$ may be obtained from the SVD of the Cholesky factor of $ {\bf\Sigma}$:

$\displaystyle {\bf\Sigma}= {\bf R}^\prime {\bf R},
$

and

$\displaystyle {\bf U S V}^\prime = {\bf R}.
$

It is implemented in this way in software/gmix_step.m(tmpvar corresponds to $ {\bf R}$):
       [U,S,V]=svd(tmpvar,0);
        S = diag(S);
        S = max(S,sqrt(diag( V' * diag(minvar) * V )));
        tmpvar = U * diag(S) * V';
        [q,tmpvar] = qr(tmpvar,0);
where the last two steps re-construct $ {\bf R}$, then force it to be upper triangular.

Consider the following example. Data was created using a mixture of 2 Gaussians using the code segment below:

     %
     % produce data that is from two Gaussian populations
     %
     fprintf('Creating data : ');
     N=4096;
     mean1=[2 3]';
     cov1= [2 -1.6; -1.6 2];
     mean2=[1.3 1.3]';
     cov2= [.005 0; 0 .005];
     x1 = chol(cov1)' * randn(DIM,N);
     x1=x1+repmat(mean1,1,N);
     x2 = chol(cov2)' * randn(DIM,N);
     x2=x2+repmat(mean2,1,N);

     data1 = [x1 x2];
Next, a GM parameter set was initialized with 2 modes with random starting means. Next, software/gmix_step.mwas iterated 50 times using the BIAS and the CONSTRAINT method. This experiment was repeated 9 times. In each trial, the same starting point was used for both methods. The results are plotted in Figures 13.2 and 13.3.
Figure 13.2: Convergence performance of the BIAS and CONSTRAINT methods. The CONSTRAINT method is consistently faster and achieves a higher log-likelihood every time.
\includegraphics[scale=0.66, clip]{cb.eps}
Figure: Typical results of training using the BIAS (left) and CONSTRAINT (right) methods. Each method used $ \rho _n=0.5$. Note that for the BIAS method, the covariance of the large mode is too fat, but for the CONSTRAINT method it is correct. For the small mode, the mode size is much smaller than $ \rho _n$ and therefore both methods produce similar results, as would be expected.
\includegraphics[width=3.2in,height=3.2in, clip]{br.eps} \includegraphics[width=3.2in,height=3.2in, clip]{cr.eps}
Note that the BIAS method has covariances that are biased and appear somewhat larger than necessary. In every case, the CONSTRAINT method converged faster and achieved a higher log-likelihood.

Baggenstoss 2017-05-19