Exploring photometric redshifts as an optimization problem: an ensemble MCMC and simulated annealing-driven template-fitting approach

Using a 4D grid of ∼2 million model parameters (Δz = 0.005) adapted from Cosmological Origins Survey photometric redshift (photo-z) searches, we investigate the general properties of template-based photo-z likelihood surfaces. We find these surfaces are filled with numerous local minima and large degeneracies that generally confound simplistic gradient-descent optimization schemes. We combine ensemble Markov Chain Monte Carlo sampling with simulated annealing to robustly and efficiently explore these surfaces in approximately constant time. Using a mock catalogue of 384 662 objects, we show our approach samples ∼40 times more efficiently compared to a 'brute-force' counterpart while maintaining similar levels of accuracy. Our results represent first steps towards designing template-fitting photo-z approaches limited mainly by memory constraints rather than computation time.