23
$\begingroup$

I have a number of multivariate observations and would like to evaluate the probability density across all variables. It is assumed that the data is normally distributed. At low numbers of variables everything works as I would expect, but moving to greater numbers results in the covariance matrix becoming non positive definite.

I have reduced the problem in Matlab to:

load raw_data.mat; % matrix number-of-values x number of variables
Sigma = cov(data);
[R,err] = cholcov(Sigma, 0); % Test for pos-def done in mvnpdf.

If err>0 then Sigma is not positive definite.

Is there anything that I can do in order to evaluate my experimental data at higher dimensions? Does it tell me anything useful about my data?

I'm somewhat of a beginner in this area so apologies if I've missed out something obvious.

$\endgroup$
  • $\begingroup$ It sounds like your data are too sparse for the high-dimension representations. Are you planning on running regression models with this data? $\endgroup$ – Jonathan Thiele Jun 14 '12 at 16:58
34
$\begingroup$

The covariance matrix is not positive definite because it is singular. That means that at least one of your variables can be expressed as a linear combination of the others. You do not need all the variables as the value of at least one can be determined from a subset of the others. I would suggest adding variables sequentially and checking the covariance matrix at each step. If a new variable creates a singularity drop it and go on the the next one. Eventually you should have a subset of variables with a postive definite covariance matrix.

| cite | improve this answer | |
$\endgroup$
  • 23
    $\begingroup$ +1. It's also worth noting that all covariance matrices are positive definite and all positive definite matrices are the covariance matrix of some multivariate distribution. Therefore, saying "non-positive definite covariance matrix" is a bit of an oxymoron. It appears the OP was really just saying that the sample covariance matrix was singular which can happen from exactly collinearity (as you've said) or when the number of observations is less than the number of variables. $\endgroup$ – Macro Jun 14 '12 at 17:23
  • 4
    $\begingroup$ Some stats software can be induced to correct this problem automatically. E.g., Stata's regress command will automatically drop extra variables when some are collinear (and its output can be saved in a form that identifies these variables and marks a non-collinear subset for future use). A likely complication, though, is that the variables might not necessarily be collinear, but they may be close enough that propagation of floating point error in the Cholesky decomposition produces negative eigenvalue estimates, making the variables collinear for all practical purposes. $\endgroup$ – whuber Jun 14 '12 at 17:55
  • 1
    $\begingroup$ @whuber, there is similar functionality in R as well - regression models automatically drop variables from the linear predictor if there is exact collinearity. $\endgroup$ – Macro Jun 14 '12 at 18:35
  • 2
    $\begingroup$ @whuber, it's a bit hacky but you can do a similar trick. If g is your linear model, then colnames(model.matrix(g))[-which(is.na(coef(g))==TRUE)][-1] should return the names of the predictors entered into the model that are not exactly collinear. This works by checking which coefficients were NA (that's how R indicates a variable was dropped), and finding the corresponding column names of the model matrix (deleting the intercept column). By the way, that won't work if there are no collinear terms so an if statement to check that which(is.na(coef(g))==TRUE) isn't empty is required :) $\endgroup$ – Macro Jun 14 '12 at 18:59
  • 8
    $\begingroup$ @Macro All covariance matrices are positive semi-definite. When they are singular they are not positive definite because x'Ax>0 for all vectors x for the matrix A to be positive definite. In the singular case x'Ax=0 occurs for some x. $\endgroup$ – Michael R. Chernick Jun 14 '12 at 19:22
0
$\begingroup$

It's a valid result. The estimate for that component of the covariance matrix is zero, which may very well be true! It can lead to computation difficulties, but some algorithms in R (I don't know about Matlab) can handle this. I don't understand why people get upset about this and insist on fitting more parsimonious models.

| cite | improve this answer | |
$\endgroup$
0
$\begingroup$

One point that I don't think is addressed above is that it IS possible to calculate a non-positive definite covariance matrix from empirical data even if your variables are not perfectly linearly related. If you don't have sufficient data (particularly if you are trying to construct a high-dimensional covariance matrix from a bunch of pairwise comparisons) or if your data don't follow a multivariate normal distribution, then you can end up with paradoxical relationships among variables, such as cov(A,B)>0; cov(A,C)>0; cov(B,C)<0.

In such a case, one cannot fit a multivariate normal PDF, as there is no multivariate normal distribution that meets these criteria - cov(A,B)>0 and cov(A,C)>0 necessarily implies that cov(B,C)>0.

All this is to say, a non-positive definite matrix does not always mean that you are including collinear variables. It could also suggest that you are trying to model a relationship which is impossible given the parametric structure that you have chosen.

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ Your answer is wrong on so many levels. Anyhow, consider a covariance matrix with 1's on the diagonal, and 1/2 for cov(1st and 2nd components), 1/2 for cov(1st and 3rd components), and -1/2 for cov(2nd and 3d components). The covariance matrix has eigenvalues approximately 0.15, 1.35, 1.50, providing a counterexample to the assertion in your 2nd paragraph. $\endgroup$ – Mark L. Stone May 5 '16 at 16:31
  • $\begingroup$ @MarkL.Stone, you're right about the 2nd paragraph, but I wonder if some of this is ambiguous & could be saved under a generous interpretation. Eg, I wonder if, in paragraph 1, "don't have sufficient data... trying to construct a high-dimensional covariance matrix from a bunch of pairwise comparisons" refers to having a lot of missing data & using the pairwise complete observations to compute each element in the covariance matrix. $\endgroup$ – gung - Reinstate Monica May 5 '16 at 17:04

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.