risk - What is the best way to "fix" a covariance matrix that is not positive semi-definite?

I have a sample covariance matrix of S&P 500 security returns where the smallest k-th eigenvalues are negative and quite small (reflecting noise and some high correlations in the matrix).

I am performing some operations on the covariance matrix and this matrix must be positive definite. What is the best way to "fix" the covariance matrix? (For what it's worth, I intend to take the inverse of the covariance matrix.)

One approach proposed by Rebonato (1999) is to decompose the covariance matrix into its eigenvectors and eigenvalues, set the negative eigenvalues to 0 or (0+epsilon), and then rebuild the covariance matrix. The issue I have with this method is that:

1. 
   

   the trace of the original matrix is not preserved, and

   
   


2. 
   

   the method ignores the idea of level repulsion in random matrices (i.e. that eigenvalues are not close to each other).

   
   

Higham (2001) uses an optimization procedure to find the nearest correlation matrix that is positive semi-definite. Grubisic and Pietersz (2003) have a geometric method they claim outperforms the Higham technique. Incidentally, some more recent twists on Rebonato's paper are Kercheval (2009) and Rapisardo (2006) who build off of Rebonato with a geometric approach.

A critical point is that the resulting matrix may not be singular (which can be the case when using optimization methods).

What is the best way to transform a covariance matrix into a positive definite covariance matrix?

UPDATE: Perhaps another angle of attack is to test whether a security is linearly dependent on a combination of securities and removing the offender.

Answer

Nick Higham's specialty is algorithms to find the nearest correlation matrix. His older work involved increased performance (in order-of-convergence terms) of techniques that successively projected a nearly-positive-semi-definite matrix onto the positive semidefinite space.

Perhaps even more interesting, from the practitioner point of view, is his extension to the case of correlation matrices with factor model structures. The best place to look for this work is probably the PhD thesis paper by his doctoral student Ruediger Borsdorf.

Higham's blog entry covers his work up to 2013 pretty well.