Ravi Varadhan
2014-Dec-21 17:32 UTC
[R] Subject: short-sale constraint with nonpositive-definite matrix in portfolio optimization
Hi, You can try a projected gradient approach, which is implemented in the spg() function in the "BB" package. You have to provide a projection function which will take an infeasible matrix as input and will give a feasible (i.e. positive-definite) matrix as output. This kind of matrix projection is quite easy to do, and in fact, there are functions in R to do this (e.g., see posdefify() function in "sfsmisc" package). There is a recent review article on spectral projected gradient algorithm in J Statistical Software. http://www.jstatsoft.org/v60/i03 Hope this is helpful, Ravi [[alternative HTML version deleted]]