Another nontrivial example comes from the area of average case complexity. Rainer Schuler proves interesting properties of the class he calls PP−comp, see [1].
PP−comp is the class of languages that are accepted in polynomial time on μ-average for every polynomial time computable (P-computable) distribution μ. Naturally, P⊆PP−comp holds, since the existence of a deterministic polytime algorithm implies that it remains efficient on the average, no matter the what the input distribution is. However, the condition of running in average polynomial time for every P-computable input distribution appears strong enough to suspect PP−comp=P.
Surprisingly, Schuler proves that there is a language L∈PP−comp, which is Turing-complete for E, that is,
E⊆PPP−comp(∗)
This implies the unconditional separation
PP−comp≠P. While the latter also uses the fact
E≠P, which follows from the Time Hierarchy Theorem, the novel part (*) builds on different tools: beyond diagonalization, it employs resource bounded measure and Kolmogorov complexity.
Reference:
[1] R. Schuler, "Truth-table closure and Turing closure of average polynomial time have different measures in EXP," CCC 1996, pdf