Recovering stellar population parameters via two full-spectrum fitting algorithms in the absence of model uncertainties

TL;DR

This study compares two full-spectrum fitting algorithms, pPXF and STARLIGHT, analyzing their biases and scatter in recovering stellar population parameters from mock spectra, highlighting differences in accuracy, bias, and computational speed.

Contribution

It provides a detailed comparison of pPXF and STARLIGHT algorithms, focusing on biases, accuracy, and computational efficiency in stellar population analysis.

Findings

pPXF shows predictable bias trends with age and metallicity.

STARLIGHT biases increase with higher dust extinction and error spectrum shape.

pPXF is significantly faster than STARLIGHT in processing spectra.

Abstract

Using mock spectra based on Vazdekis/MILES library fitted within the wavelength region 3600-7350\AA, we analyze the bias and scatter on the resulting physical parameters induced by the choice of fitting algorithms and observational uncertainties, but avoid effects of those model uncertainties. We consider two full-spectrum fitting codes: pPXF and STARLIGHT, in fitting for stellar population age, metallicity, mass-to-light ratio, and dust extinction. With pPXF we find that both the bias in the population parameters and the scatter in the recovered logarithmic values follows the expected trend. The bias increases for younger ages and systematically makes recovered ages older, $M_*/L_r$ larger and metallicities lower than the true values. For reference, at S/N=30, and for the worst case ($t=10^8$yr), the bias is 0.06 dex in $M_*/L_r$, 0.03 dex in both age and [M/H]. There is no significant dependence on either E(B-V) or the shape of the error spectrum. Moreover, the results are consistent for both our 1-SSP and 2-SSP tests. With the STARLIGHT algorithm, we find trends similar to pPXF, when the input E(B-V)<0.2 mag. However, with larger input E(B-V), the biases of the output parameter do not converge to zero even at the highest S/N and are strongly affected by the shape of the error spectra. This effect is particularly dramatic for youngest age, for which all population parameters can be strongly different from the input values, with significantly underestimated dust extinction and [M/H], and larger ages and $M_*/L_r$. Results degrade when moving from our 1-SSP to the 2-SSP tests. The STARLIGHT convergence to the true values can be improved by increasing Markov Chains and annealing loops to the "slow mode". For the same input spectrum, pPXF is about two order of magnitudes faster than STARLIGHT's "default mode" and about three order of magnitude faster than STARLIGHT's "slow mode".