1: \begin{abstract}

2: 	Zero-One Composite Optimization (0/1-COP) is a prototype of nonsmooth, nonconvex optimization problems and it has attracted much attention recently.

3: 	The augmented Lagrangian Method (ALM) has stood out as a leading methodology for such problems. 

4: 	The main purpose of this paper is to extend the classical theory of ALM from

5: 	smooth problems to 0/1-COP.

6: 	We propose, for the first time, second-order optimality conditions for 0/1-COP.

7: 	In particular, under a second-order sufficient condition (SOSC), we prove R-linear convergence rate of the proposed ALM. 

8: 	In order to identify the subspace used in SOSC, we employ the proximal operator of

9: 	the 0/1-loss function, leading to an active-set identification technique.

10: 	Built around this identification process, we design practical stopping criteria for 

11: 	any algorithm to be used for the subproblem of ALM. 

12: 	We justify that Newton's method is an ideal candidate for the subproblem and

13: 	it enjoys both global and local quadratic convergence.

14: 	Those considerations result in an inexact Newton ALM (iNALM).

15: 	The method of iNALM is unique in the sense that it is active-set based, it is inexact (hence more practical), and SOSC 

16: 	%instead of widely assumed 	Kurdyka-$\L$ojasiewicz (K$\L$) properties 

17: 	plays an important role in its R-linear convergence analysis. 

18: 	The numerical results on both simulated and real datasets show the fast running speed and high accuracy of iNALM when compared with several leading solvers.

19: 	

20: 	\noindent{\bf \textbf{Keywords}:}

21: 	zero-one composite optimization problem, 

22: 	inexact augmented Lagrangian method,

23: 	second-order optimality conditions,

24: 	P-stationary point, 

25: 	convergence rate, 

26: 	subspace Newton method.

27: 	% \PACS{PACS code1 \and PACS code2 \and more}

28: %	\subclass{90C26 \and 90C30 \and 90C90}

29: \end{abstract}