RNA-Seq count size factors are defined in formula 5 of Anders & Huber (2010)

With pre-specified geometric means, are size factors supposed to be the same for identical samples regardless of total count matrix context?

That is, if I calculate the size factor for a single sample or if I extract that size factor for that sample from a larger context, shouldn't they be identical if the geometric mean was fixed?

For example:

library(DESeq2)

set.seed(353567)

ddsRaw <- makeExampleDESeqDataSet(n=1000, m=40)
gm <- exp(rowMeans(log(counts(ddsRaw))))

dds <- estimateSizeFactors(ddsRaw, geoMeans=gm)

ddsSubset <- estimateSizeFactors(ddsRaw[, 10:20], geoMeans=gm)

all.equal(sizeFactors(dds)[10:20], sizeFactors(ddsSubset))  # Size factors are not equal

I think the code below from estimateSizeFactorsForMatrix() appears to be responsible for the dataset-dependent size factors, but I do not understand how it relates to formula 5, because it is now no longer solely dependent on the reference geometric means.

if (incomingGeoMeans) {
  sf <- sf/exp(mean(log(sf)))
}

Thanks!

The last chunk there is just to set size factors to have geometric mean of 1 for any particular dataset, regardless of their relation to geoMeans. I implemented it this way intentionally. So the normalization of a new dataset will be identical up to a single scaling factor regardless of the samples. What would your desired behavior be? I don't think it makes sense to have the size factors far from 1.