Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.

## Requests for new functionality

Beginner
336 Views
I'm not sure what the best way to communicate requests for new functionality in MKL is, so I'll post here, hoping that the post will get to the right person at Intel. I know that with limited resources, it may not be possible to create this functionality. But I'm also sure that I'm not the only one who would take advantage of it. So here are some things that I really wish MKL could do.

1. A digamma and polygamma function. These functions come up when working with the log densities of probability distributions that include beta and gamma functions. The digamma function is the first derivative of the log gamma function, and knowing that first derivative is helpful when computing the mode of certain distributions. These functions are available in GSL, but not MKL.
2. A symmetric rank-1 and rank-k update for the Cholesky decomposition of a matrix. These functions would be equivalent to the _syr and _syrk BLAS functions, but would take Cholesky decompositions of the A matrix as the input, and return the Cholesky of A as the output. Why is this useful? The fastest way to compute the density of, or draw from, a multivariate normal distribution is by using the Cholesky of the precision matrix to convert to/from a standard MVN. When this matrix changes repeatedly, I need to do the update, and then take the Cholesky decomposition over and over. This seems inefficient to me. I know that there are some algorithms out there that can do this, but it would be quite nice if MKL included that functionality.
3. Trust region algorithms that are more amenable to general nonlinear optimization, as opposed to solving a nonlinear least squares algorithm. All that would be necessary here would be some efficient code to solve the trust region subproblem, given a current estimate of a gradient, Hessian and trust radius. An example would be the Steihaug conjugate gradient algorithm.
Please let me know if requests like this should go elsewhere. Or, please let me know if there are existing solutions to these problems that I don't know about. But it seems to me that these three things would be of general interest, and could take advantage of optimizations of the Intel architecture in a way that most users would not be able to implement on their own.
Thanks,
Michael
3 Replies
Employee
336 Views
HI Michael,
Could you describe more deeply why current version of Trust region solver in MKL is not convenient for you?
With best regards,
Moderator
336 Views
Michael,
All is OK. This is the right place to give us your feedbacks and requests regarding MKL.
Item 1 we have the similar request and we just added your notes.
Regarding item 2 the new Feature Request has been submitted to the internal database and we will back to you as soon as this functionality will be available.