1) It splits a region (it causes a string of zeros to have a one in them) (adds 8 bytes)
2) It coalesces a region (it removes the last remaining zero between two regions) (subtracts 8 bytes)
I'm sure there are other solutions, and as was pointed out it doesn't really deal with duplicates. I suspect a multi-value bit field might cover that.
I simulated it with a Python program and you get about 10^4 coalesces. That doesn't even begin to make a dent in 10^6, and storing this will still take ~7 megabytes.
Right.
and you'll end up with 8 bytes x 10^6 = 7.6 megabytes.
Not if you encode cleverly. The successive differences will be on the scale of 10^8/10^6 = 100, which is a very small number. It takes 7 bits to store, or at least 8 bits in a variable-width encoding.
An argument that you need context info: if you don't use context info, each bit pattern needs to be rounded up to the nearest number of whole bits. That waste's about N/2 bits. According to my calculations, that would raise the bound from 0.96MiB to 1.02MiB, i.e. too much.
But really, it's even worse. Let's assume you can indeed define such a compression and you manage to use the context sufficiently to avoid wasting too many bits. Further, you manage to do so without using much temporary space at all.
That means you'll need to be updating your sorted list in place. Each update is likely to be a non-trivial task, since you need to find the place to update (with variable-width encoding and no space for luxuries like an index, that's probably a linear scan), and you need to take into account some amount of context - and then, because you've changed some data, recode possibly unbounded amounts (so N) of following context. That makes this sorting algorithm quadratic; and it'll have a high scaling factor too since the operations we're counting here aren't cheap: we're not moving bits around, we're decoding and recoding them all the time in a context-sensitive way and that means doing lots and lots of branches.
Assuming the 1MB limit is reasonably inspired because the chip neads to be tiny, you're probably not dealing with the fastest chip in the world. As an estimate, I'd guess this algorithm (if we could make it) would be about 1000000 times slower than a normal sort (assuming a 13 times higher constant factor), and you'd be running it on something very very slow. On a modern processor, a plain sort takes about 75ms, making this thing 75 ks, or approximately 1 day. On that very limited 1MB proc...
So even you manage to pull it off (very hard), it's almost certainly going to be very slow.
I think the parent's solution works. The typical separation between numbers is about 2^7: that works out to 8 bits per number -- 7 for the length of the '00...0', plus one length-one '1'. 8 bits * 10^6 is 1 MiB exactly. You need a variable-length encoding, so you can fit 2^7 in 7 bits while still allowing larger numbers to be encoded.