Can We Overcome the n log n Barrier for Oblivious Sorting?

It is well-known that non-comparison-based techniques can allow us to sort n elements in o(n log n) time on a Random-Access Machine (RAM). On the other hand, it is a long-standing open question whether (non-comparison-based) circuits can sort n elements from the domain [1..2k] with o(kn log n) boolean gates. We consider weakened forms of this question: first, we consider a restricted class of sorting where the number of distinct keys is much smaller than the input length; and second, we explore Oblivious RAMs and probabilistic circuit families, i.e., computational models that are somewhat more powerful than circuits but much weaker than RAM. We show that Oblivious RAMs and probabilistic circuit families can sort o(log n)-bit keys in o(n log n) time or o(kn log n) circuit complexity. Our algorithms work in the indivisible model, i.e., not only can they sort an array of numerical keys --- if each key additionally carries an opaque ball, our algorithms can also move the balls into the correct order. We further show that in such an indivisible model, it is impossible to sort Ω(log n)-bit keys in o(n log n) time, and thus the o(log n)-bit-key assumption is necessary for overcoming the n log n barrier. Finally, after optimizing the IO efficiency, we show that even the 1-bit special case can solve open questions: our oblivious algorithms solve tight compaction and selection with optimal IO efficiency for the first time.