Re-starting or starting over?

[normvcr]

Occasionally, during the running of the optimization, I will need to change the variables (x) that are passed in the interface of the evaluate() routine. This "change of variables" does not affect the value of objective function (F), only how F is parametrized. For example, if F involves a unit vector (x,y), I may use just x to describe the unit vector; but, should x become too large, I will start using y, instead. One way to do this, is to restart L-BFGS with the new parametrization, but this is not the most efficient approach. How much work is involved to have a routine that efficiently updates the L-BFGS data structures for a new x supplied by the user?