All references I have consulted have been extremely sketchy about this point. The (2 dimensional) nonlinear sigma model in some Riemannian manifold $M$ with metric $g_{\mu\nu}$ has action $$S = \frac{1}{2} \int_M d^2 \sigma~ g_{\mu \nu} \partial_\alpha X^\mu \partial_\alpha X^\nu$$ Here, $\alpha$ is the "world sheet" index describing the embedding, and $\mu$ is the target space index. When we define the path integral, it should be $$\int \mathcal{D} X e^{-S}$$ Standard techniques to compute the renormalization of the sigma model use the background field expansion, suitably adapted to preserve manifest target space coordinate invariance. BUT, a crucial assumption going into the background field expansion and the whole apparatus of perturbative quantum field theory is that the path integrals are over linear field spaces; otherwise the Gaussian integral formulas which give us the determinants and Feynman diagrams break down even at the formal level. However, the NLSM is defined as an integral over maps to a curved manifold. Putting it differently, the usual discretized measure is not generally covariant. How does the "textbook approach" account for this point?
The quickest fix would be to discretize and define $$\mathcal{D}X = \sqrt{\det g(X)} \prod_i dX_i,$$ but even this is not very satisfactory since it renders the integrals non-Gaussian. Any insight would be appreciated.