It's a neat little trick to efficiently squeeze four 5-bit values into a single 16-bit value when we don't care about the order of those 5-bit values and we want to get real close to that Shannon limit.
This is good for storing things like a series of order independent options in an entry or entry identifiers in a very large hash table. Note that even for values > 5 bits, we can still apply this trick if we're willing to move the rest of the to match the sorted order of those 5 bits but obviously you wouldn't want to shuffle around hundreds of bytes just to save 4 bits.
This trick works by trading control over the ordering of the numbers for extra precision in the four numbers.
There are multichoose(24, 4) = 3876 possible unique values for a multiset of 4 4-bit values, whereas there are 65536 possible unique values for a 16 bit integer. log2(multichoose(24, 4)) = 11.92 so if we can map all possible values to a 12 bit representation, it leaves us with 4 bits to play around with and that's enough to store the extra bit in our four 5-bit numbers. Theoretically we have a bit more than 4 bits to play around with as log2(216 / 3876) = 4.08 or from 16 - 11.92 but in practice that fraction is a bit harder to extract without a larger or more indirect lookup table.
More generally, a multiset of m n-bit numbers cannot hold more than m * n bits of information and it is possible to leverage that difference to store m' n'-bit numbers in the same m * n bits as long as log2(multichoose(m', n')) <= m * n. log2(multichoose(25, 4)) = 15.67 bits < 16 bits.
- More information on multisets: Wikipedia
- More general methods: Compressing Sets and Multisets of Sequences
- Practical applications: Cuckoo Filters, Compressing Genomes