Cuffdiff Sample v/s replicates FPKM

[vinay052003]

I ran cuffdiff on two multi-replicate samples. Cuffdiff reports FPKM for each sample as well as FPKM for each replicate.
I was wondering how does it calculate the sample FPKM from its replicate's FPKMs.
I am sure it's not simple average across the replicates because I've done that and the numbers don't match.