Trust Region Full Newton not scale invariant ?
When using very high thresholded gradient, the radius square scales ~ linearly with the threshold, but the "xx prediction" scales inversely, which makes the behavior of Full Newton TR very dependant on the scale, for a constant mu. Probably linked to the inappropriate norm being used ?
If _stabilization is O(n), alpha is O(n), xx is O(n) too but the radius² is ~1/n gradient norm squared is ~1/n too. (So mu * radius = grad norm holds)
Edited by Boris Martin