Earth Mountain almost there, page-450

You definitely do not have to know that something works in order to patent it.

My favourite example is a data compression algorithm that was guaranteed to always reduce the original data in size by at least one bit. The logical problem arises - you could repeatedly compress any length of data, again and again, and ultimately reduce any amount of input into just a single 1 or 0. Obviously that is impossible

Still, a patent costs a lot of time and money - and you can only patent something that can be accurately described. Things that are patented have either very high potential commercial value, or a high likelihood of working, otherwise they are a waste of money to the patentholder.