On some systems, one may be told that the model-implied covariance matrix is not positive-definite when inverting it during the fit process.
The theory was that this is a numerical issue - another possibility is that this occurs when the estimation results in under-identified models for some values of lambda.
My suggestion is to calculate pseudo-inverses instead. If the covariance matrix is positive definite, the pseudo-inverse is equivalent to the actual inverse. In the border cases where the matrix is just a bit off an invertible matrix, the difference should be negligible. In cases where it is way off, one should actually notice this as really non-identified models tend to yield unreasonable estimates.
On some systems, one may be told that the model-implied covariance matrix is not positive-definite when inverting it during the fit process.
The theory was that this is a numerical issue - another possibility is that this occurs when the estimation results in under-identified models for some values of lambda.
My suggestion is to calculate pseudo-inverses instead. If the covariance matrix is positive definite, the pseudo-inverse is equivalent to the actual inverse. In the border cases where the matrix is just a bit off an invertible matrix, the difference should be negligible. In cases where it is way off, one should actually notice this as really non-identified models tend to yield unreasonable estimates.