I have the following problem : I'm calculating the sample covariance matrix in the frequency domain ( $y_{k}$ is the FFT of a time domain $k_{th}$ symbol vector signal , basically a simulated received signal) as follows:


Well , the next step on my algorithm is to solve a optimization process, as an essential part of it I need to compute the eigenvalues by doing SVD ,method of powers, etc... in MATLAB of the following equation:


To not get into too many details because I believe my issue comes from a numerical problem and many insights of the actual algortihm / context aren't needed .Let's assume, $A$ is simply a predefined matrix that I compute.

The REAL ISSUE appears now because $\mathbf R$ is ill-conditioned as MATLAB tells me so. So the inverse procedure seems to be failing and the eigenvalues I'm obtaining are incredibly small due to this issue (In fact I only get 1 eigenvalue different than zero). The dimensions of $\mathbf R$ are typically large ( since they are compressed depends on the actual compression ratio I'm using but let's say $32\times 32$ for example).

One approach to solve this problem I found is to use diagonal loading as:

$$(\left(\mathbf R+\sigma\mathbf I\right)^{-1}\mathbf A)\quad\text{with}\quad\sigma > 0$$

This seems to be solving the problem, the eigenvalues are now scaled due to this background "noise". My question is how can I obtain the truly well scaled eigenvalues because , later on, in my algorithm these eigenvalues will serve as weights since I'm considering them as an actual power estimate.

Note : I've been playing with the cond() function in MATLAB for $\sigma= 0.05$ ,cond(R+sigma*I)= 2 , which is not bad I believe.

Feel free to ask more questions about the problem. But I think my question relates to a purely numerical issue involving eigenvalues and covariances matrices.


You can write

$$ R=YY^H $$ where $Y$ is a matrix of size $N\times N_f$ and $N$ is the dimension of $y_k$. $Y$ contains all the measured $y_k$ as its columns.

Then, the rank of $R$ is upper bounded by $N_f$. In particular, if $N_f<N$, $R$ will always be a singular matrix. So, if you have too few measurements, you will likely run into the problem you described.

For your second part, how do the Eigenvalues of $R$ and $R+\sigma I$ relate: Consider the eigenvalue decomposition of $R$ by $$ R=PVP^{-1}$$ where $P$ contains the eigenvectors and V is the diagonal matrix with the eigenvalues (they can be zero). Then you have

$$ R+\sigma I=P V P^{-1} + \sigma I = P(V+\sigma I)P^{-1} $$

This means that the eigenvalues are all increased by $\sigma$.

  • $\begingroup$ In the first part, I made sure Nf ( Number of frames ) is larger than the dimension of the vector . In the second part , the scaling factor of the eigenvalues as a function of sigma I was looking for is the relation between $(R)^{-1}*A$ and $(R+\sigma*I)^{-1}*A$. Sorry if I din't make that clear.Thank you very much for answering. $\endgroup$ – Ricardo García Nov 8 '16 at 22:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.