Parametrization dependence and Delta Chi-squared in parton distribution fitting

Parton distributions functions (PDFs), which are essential to the interpretation of data from high energy colliders, are measured by representing them as functional forms containing many parameters. Those parameters are determined by fitting a wide variety of experimental data. The best-fit PDF set is obtained by minimizing the standard $\chi^2$ measure of fit quality. The uncertainty range is estimated in the Hessian method by regarding as acceptable, all fits for which $\chi^2$ lies within $\Delta\chi^2$ of its minimum. The appropriate value of $\Delta\chi^2$ for this purpose has been estimated by a variety of arguments to be approximately 50 - 100 for a 90% confidence limit. This paper resolves the long-standing paradox of why that empirical value is so much larger than the $\Delta\chi^2=2.7$ for 90% confidence that would be expected on the basis of standard Gaussian statistics.