I think you might be interpreting error in a manner different than intended. An estimate is virtually always wrong, but can often be very close to correct. Suppose you measure the length of a carrot with a ruler that is only graduated down to 1/8 inches. If one end of the carrot is lined up with the zero mark, the other end is extremely unlikely to lie exactly on a n/8 mark. The exact length remains unknown, but can be estimated with reasonable accuracy. The difference between the actual length and the estimated length is an error, but it might not be (probably isn't, at least in this example) consequential.
I guess I had always assumed that the loss occurred at the encoding end, where some of the signal was selectively removed to reduce the file size. I guess that was kind of naive