When Good Theorems Have Bad Proofs

Here is one of my favorite examples of a bad proof for what turns out to be a correct theorem.

Theorem: If NP is in BPP then the whole polynomial-time hierarchy is in BPP.

Let's focus on simply showing Σ₂ is in BPP if NP is in BPP. The rest is straightforward induction. Here is our first proof:

Σ₂=NP^NP⊆ NP^BPP⊆BPP^BPP=BPP. Do you see the problem with this proof?

To get a correct proof (first due to Zachos) we need to use Arthur Merlin games. Consider a Σ₂ language L as an ∃∀ expression. Since NP is in BPP, we can replace the ∀ with a probabilistic test. This gives us what is known as MA or a Merlin-Arthur game where the powerful Merlin sends a message that Arthur can probabilistically verify. A now classic result shows that MA is contained in AM, where Arthur provides a random string to Merlin who must then provide a proof based on that string. Once again we apply the NP in BPP assumption to allow Arthur to simulate Merlin probabilistically and now we have a BPP algorithm for L.

The problem in the first proof is in the second "⊆". The assumption NP in BPP does not imply NP^A in BPP^A for all A.