Inferring Galaxy Star Formation Histories from Integrated Spectral Energy Distributions: Target-Specific Representation Selection and Robust Certification
Recovering galaxy star formation histories (SFHs) from integrated spectral energy distributions (SEDs) is an ill-posed inverse problem because of temporal integration, physical degeneracies, limited wavelength coverage, and forward-model uncertainty. We develop a target-specific framework that evaluates the minimum worst-case root-mean-square error for each candidate representation under an error tolerance and stability cap, then selects the least complex certified candidate. The procedure distinguishes failure within the candidate family, failure of the unrestricted linear estimator class, and unresolved cases. Forward-operator mismatch and nonlinear remainders are incorporated through robust envelopes, while risk certification and post-selection coverage are treated separately. Four benchmarks test the theory, convex implementation, risk brackets, false certification, and coverage. Using the same data for selection and inference reduces nominal 95~{\%} coverage to 90.25~{\%}. Under operator mismatch, the nominal procedure falsely certifies its returned fixed weights, whereas robust certification selects a richer representation whose envelope risk remains below tolerance. In an age-resolved SED proxy, nominal certification yields a library-relative false certification in all six scenarios. Robust certification reports only for full wavelength coverage at signal-to-noise ratios of 30 and 60; the returned weights satisfy the tolerance over the mismatch library and its convex hull. These results support internal validity and computational reproducibility, but not external validity for production stellar population synthesis libraries or observations. Accordingly, the framework reframes the choice of reported SFH resolution as a certifiable, target-specific decision: it returns the least complex representation justified by the stated scientific, observational, and modeling conditions, and otherwise withholds the report.