You can't do a bitmap in general b/c random integers are really, really big.
If there is a compression algorithm that works on 10% of data sets, and does massive harm on the other 90%, you can use or not use it, at the cost of a small amount of header information and a bunch of CPU time, and it doesn't matter that in general, on average, it's quite bad. All that matters is whether the times it's good beat the header info cost. I think. I'm not sure this is helpful but it doesn't require a compression algorithm being a net gain on average.
If there is a compression algorithm that works on 10% of data sets, and does massive harm on the other 90%, you can use or not use it, at the cost of a small amount of header information and a bunch of CPU time, and it doesn't matter that in general, on average, it's quite bad. All that matters is whether the times it's good beat the header info cost. I think. I'm not sure this is helpful but it doesn't require a compression algorithm being a net gain on average.