I've a question regarding to eigenvalues, sinve I am not very familiar with the concept. Suppose I've a matrix $X'X$ in the case of an OLS regession. And lets assume that the regarding eigenvalues are very small. This imlpies strongly multicollinearity and a high variance of the OLS estimators. Is there any relationship between the eigenvalues of $X'X$ and the ones of $(X'X)^{-1}$ ?

The following point should be correct: The smaller the eigenvalues of $X'X$ or the larger the eigenvalues of $(X'X)^{-1}$ the nearer we are to multicollinearity and $Var(\beta)$ will be large.


Let's say $\lambda$ is one of the eigen values of $X'X$.

For eiegen vector $v$, this equation is hold : $X'X v= \lambda v$

And this can be rewritten like this : $(X'X)^{-1} v= \lambda^{-1} v$

which implies that the eiegen values of the inverse matrix is reciprocal of those of the original matrix.

I hope this is helpful for you.

I am not a native English speaker. Any improvement is welcome.Thank you in advance.

  • $\begingroup$ Thank you. This would underline my last sentence. And the larger the eigenvalues of the inverse are the large the variance of the coefficients, or am I wrong ? $\endgroup$ – Leo96 Aug 29 '18 at 9:19
  • $\begingroup$ How about using spectral decomposition? That allows us to write like this : $(X'X)^{-1}=A' \Lambda A=\sum_{i=1}^n \lambda_i v_i v_i'$ $\endgroup$ – KDG Aug 29 '18 at 9:29
  • $\begingroup$ and the variance covariance matrix of beta vector is $\sigma^2 (X'X)^{-1}$ $\endgroup$ – KDG Aug 29 '18 at 9:30
  • 1
    $\begingroup$ I think this would show that your statement is true clearly. $\endgroup$ – KDG Aug 29 '18 at 9:31
  • $\begingroup$ I am not familiar with specral decomposition, still a beginner in statistics^^. But it seems as larger the the eigenvalues of the inverse as larger the variance, via spectral decompositoin $\endgroup$ – Leo96 Aug 29 '18 at 9:31

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.