PCA looses to Truncated SVD in sustaining proximity of vectors in lower dimensional space - why?

Question

I have high dimensional vectors (1024-dim) which represent textual information (output of a NLP Bert-model). I want to apply dimension reduction and experimented with LSA (t