I had a similar objection reading the article. If you take a look at [0] they go into a bit more detail. Briefly, you can distribute your decrypt function over

I had a similar objection reading the article. If you take a look at [0] they go into a bit more detail. Briefly, you can distribute your decrypt function over the additions and multiplications by homomorphism and then bound the error, then you can judiciously choose your weighting to compensate. [0] https://courses.csail.mit.edu/6.857/2015/files/yu-lai-payor.pdf https://courses.csail.mit.edu/6.857/2015/files/yu-lai-payor....