A quadratic estimator view of the transfer function correction in intensity mapping surveys

In single dish neutral hydrogen (HI) intensity mapping, signal separation methods such as principal component analysis (PCA) are used to clean the astrophysical foregrounds. PCA induces a signal loss in the estimated power spectrum, which can be corrected by a transfer function (TF). By injecting mock signals of HI into the data and performing the PCA cleaning, we can use the cleaned mock HI signal to cross-correlate with the original mock, and estimate the signal loss as a TF, ${T}(\vec{k})$. As expected, a correction of ${T} (\vec{k})^{-1}$ restores the cross-power between the HI and optical galaxies. However, contrary to intuition, the HI autopower also requires a ${T}(\vec{k})^{-1}$ correction, not ${T}(\vec{k})^{-2}$. The ${T}(\vec{k})^{-1}$ correction is only known empirically through simulations. In this Letter, we show that the ${T}(\vec{k})^{-1}$ correction in autopower is universal, and can be analytically proven using the quadratic estimator formalism through window function normalization. The normalization can also be used to determine the TF correction for any type of linear process. Using the window function, we demonstrate that PCA induces mode-mixing in the power spectrum estimation, which may lead to biases in the model inference.