Does jaxopt.LBFGS required double precision?

[joacorapela]

The PyTorch implementation of LBFGS requires double precision (please refer to this issue).

jaxopt.ScipyMinimize uses double precision by default.

Does jaxopt.LBFGS require double precision?

[mblondel]

No it uses the same precision as the parameters.

[joacorapela]

Thanks @mblondel .

jaxopt.ScipyMinimize says:

Note that some methods relying on FORTRAN code, such as the L-BFGS-B solver for scipy.optimize.minimize, require casting to float64.

Can I assume that this does not hold for jaxopt.LBFGS and that there should not be very large differences in the result of a jaxopt.LBFGS optimization if I use simple or double precision parameters?

[mblondel]

• jaxopt.ScipyMinimize is a wrapper around SciPy, which uses a FORTRAN implementation for LBFGS-B. The FORTRAN code uses double precision.
• jaxopt.LBFGS is a pure implementation in JAX, which supports any dtype.

[joacorapela]

Thanks @mblondel