L-BFGS implementation as a general optimization pattern (http://en.wikipedia.org/wiki/Limited-memory_BFGS)
Logistic Regression implemented against L-BFGS primitive
Unit / Regression tests leveraging Logistic Regression data generator
Would be interested to see your implementation against jblas.
We're trying to be consistent with the linear algebra libraries we use and not add additional dependencies to the project unless we really think we need to. Is there anything that's missing in jblas that you'd need for your implementation? Jblas is pretty good from a performance standpoint and does a good job at avoiding unnecessary object creation.
The unbounded version is there in mallet but I have the c++ version of both unbounded and bounded bfgs optimizer derived from fortran code and tested on large scale problems. I think it will be very useful for mllib optimization package. My plan was to bring first version using bridj due to efficiency. Another option is to jblas it...Any thoughts ?