What is the difference between AM and IP

Question

Intro

I am trying to understand how those two models of interactive proof are different. I understand that

• $\text{AM}$ relies on public coins (the prover knows the random bits used by the verifier)
• $\text{IP}$ relies on private coins (the prover does not)

And here are some common results in those classes

• Goldwasser and Sipser showed that if we allow a polynomial number of rounds $Q$, we have that $\text{IP}[Q] \subseteq \text{AM}[Q+2]$.

• It has also been shown that if we allow a constant number of rounds $k$ in $\text{AM}$, we have that $\text{AM}[k] \subseteq \text{AM}[2]$

This seems to imply that if we limit $\text{IP}$ to a constant number of round $k$, we also have that $\text{IP}[k] \subseteq \text{AM}[k+2] \subseteq \text{AM}[2]$, and that these classes are very similar.

---

The thing that bugs me

But then, why do we have the huge difference that $\text{IP} = \text{PSPACE}$, while $\text{AM} \subseteq \prod_2^{\text{P}}$?

(see the Complexity Zoo on AM and IP)

---

Question(s)

1. Why the difference? My assumption is that there is a convention to use $\text{AM}$ to speak about $\text{AM}[k]$, private coin, constant $k$ rounds protocols, and $\text{IP}$ to speak about $\text{IP}[\text{poly}]$ public coin, polynomial rounds protocols. Is it correct?

2. Do the results holds if we swap the classes? $\text{IP}[k] \subseteq \prod_2^{\text{P}}$ does, but what about $\text{AM}[\text{poly}] = \text{PSPACE}$?

3. As an additional question, we do know that a constant number of messages $k$ can be condensed into only two messages. Is there a similar result for a polynomial number of messages?

Answer 1

$AM$ is defined to be $AM[2]$, i.e. languages with 2-rounds Arthur-Merlin protocol, which is equal to $AM[k]$ for any constant $k$.

Two extra rounds are required to overcome the fact that the coins are public, i.e.

$IP[k]\subseteq AM[k+2]$ for any $k$ , constant or not, as stated in complexity zoo.

This means that $IP[poly]=AM[poly]=PSPACE$ (which answers your second question).

If $IP$ can be limited to a constant number of rounds (rather than polynomial), then we would have $PSPACE\subseteq IP[k]\subseteq \Pi^p_2$, and the hierarchy collapses to the second level, since $PH\subseteq PSPACE$.