“…The 0 norm has been widely used as the sparsity constraint to construct a sparse tracking portfolio (Benidis, Feng, and Palomar 2018). However, imposing the 0 constraint makes the regularised regression problem NP-hard and requires search heuristics, such as genetic algorithms (Ni and Wang 2013;Li, Sun, and Bao 2011;García, Guijarro, and Oliver 2017), Tabu search (García, Guijarro, and Oliver 2017), simulated annealing (Chang et al 2000;Woodside-Oriakhi, Lucas, and Beasley 2011) and transformation (Coleman, Li, and Henniger 2006;Wang et al 2012). These algorithms are not guaranteed to find the optimal solution, and in many situations the search space grows super-linearly.…”