Normalization is a critical step in data processing that influences downstream analyses. Normalization aims to adjust for technical variations in data acquisition, facilitating accurate comparisons across heterogeneous datasets. In this paper, we identify key challenges in scRNA-seq normalization, including the simplex nature of reads, compositional bias from the mRNA population, technical and biological outliers, and non-linear relationships between the input and output. We introduce a new framework to address these challenges by modeling the measurement function and robust learning of parameters. Empirical validations on real datasets demonstrate the effectiveness of the proposed normalization method, RFNorm, in preserving lower-dimensional mathematical structures crucial for cell type and state analysis. This is assessed through the invariance of k-nearest neighbor graphs comparing the performance of RFNorm against established methods.