Time-Continuity of PCA/SVD

Question

I have a daily set of observations (Jacard similarity matrix) that I want to embed in a much lower dimensional space.

Thus I run PCA or SVD for each day and plot the projection on the top 2 dimensions (they usually explain ~90% of variance, but this is beyond the point).

My problem is that when I animate the embedding, I see that it flips along the 2nd coordinate every now and then, i.e., the embedding is not time-continuous.

I wonder what can be done to ensure time-continuity.

E.g., using a "running average" of, say, 3 days of data might improve continuity, but is unlikely to ensure it.