Learning to Concentrate: Multi-tracer Forecasts on Local Primordial Non-Gaussianity with Machine-Learned Bias

Local primordial non-Gaussianity (LPNG) is predicted by many non-minimal models of inflation, and creates a scale-dependent contribution to the power spectrum of large-scale structure (LSS) tracers, whose amplitude is characterized by bϕ . Knowledge of bϕ for the observed tracer population is therefore crucial for learning about inflation from LSS. Recently, it has been shown that the relationship between linear bias b1 and bϕ for simulated halos exhibits significant secondary dependence on halo concentration. We leverage this fact to forecast multi-tracer constraints on f loc NL. We train a machine learning model on observable properties of simulated IllustrisTNG galaxies to predict bϕ for samples constructed to approximate DESI emission line galaxies (ELGs) and luminous red galaxies (LRGs). We find σ(f loc NL) = 2.3, and σ(f loc NL = 3.7, respectively. These forecasted errors are roughly factors of 3, and 35% improvements over the single-tracer case for each sample, respectively. When considering both ELGs and LRGs in their overlap region, we forecast σ(f loc NL) = 1.5 is attainable with our learned model, more than a factor of 3 improvement over the single-tracer case, while the ideal split by bϕ could reach σ(f loc NL) < 1. We also perform multi-tracer forecasts for upcoming spectroscopic surveys targeting LPNG (MegaMapper, SPHEREx) and show that splitting tracer samples by bϕ can lead to an order-of-magnitude reduction in projected σ(f loc NL for these surveys.