1: \begin{abstract}

2:   Optimization algorithms can see their local convergence rates deteriorate when

3:   the Hessian at the optimum is singular.

4:   These singularities are inescapable when the optima are non-isolated.

5:   Yet, under the right circumstances, several algorithms preserve their

6:   favorable rates even when optima form a continuum (e.g., due to

7:   over-parameterization).

8:   This has been explained under various structural assumptions, including the

9:   Polyak--{\L}ojasiewicz inequality, Quadratic Growth and the Error Bound.

10:   We show that, for cost functions which are twice continuously differentiable

11:   ($\mathrm{C}^2$), those three (local) properties are equivalent.

12:   Moreover, we show they are equivalent to the Morse--Bott property, that is,

13:   local minima form differentiable submanifolds, and the Hessian of the cost

14:   function is positive definite along its normal directions.

15:   We leverage this insight to improve local convergence guarantees for

16:   safe-guarded Newton-type methods under any (hence all) of the above

17:   assumptions.

18:   First, for adaptive cubic regularization, we secure quadratic convergence even

19:   with inexact subproblem solvers.

20:   Second, for trust-region methods, we argue convergence can fail with an exact

21:   subproblem solver, then proceed to show linear convergence with an inexact one

22:   (Cauchy steps).

23: \end{abstract}

24: