Precomputed or new PCA on subsetted data

[dravichandar]

I am trying to get an understanding of how PCA neighbor graph affects the computed velocity field. I used the dynamic model to infer velocities from a subsetted seurat object (focusing only on 2/6 clusters that show expansion in disease). I get some marked differences in results with (a) using the original PCA that was computed for all 6 clusters and then just subsetted to cells in the 2 vs (b) recomputing the PCA on just the 2 clusters (whether via Seurat or scvelo itself for pca)
The former seems to make better biological sense. Just curious if this is expected behavior