Libraries are sequenced to different depths, so each sample gets a size factor sj with EKij = sj qij. The column sum is a poor estimate because a few genes can dominate it; the median of ratios to the geometric-mean reference is robust as long as most genes (or balanced numbers of genes) do not change. Once counts are normalised, their variance across replicates grows like the Poisson mean at low counts and like αμ2 at high counts: the gamma-Poisson law.
code/ch08_count_data_glm.R (sf_median_ratio, figures 8.1 to 8.3).