You can get very good lossless compression with floating point numbers, Facebook's Gorilla paper comes to mind. I usually use it for delta-of-delta encoding whi

You can get very good lossless compression with floating point numbers, Facebook's Gorilla paper comes to mind. I usually use it for delta-of-delta encoding which provides very high compression for time series. While that won't really help in your case, their floating point encoding could help compressing matrices quite efficiently. http://www.vldb.org/pvldb/vol8/p1816-teller.pdf http://www.vldb.org/pvldb/vol8/p1816-teller.pdf [page 5]