It is known that adding bounded-error randomization to PSPACE doesn't add power. That is, BPPSAPCE=PSPACE.

It is famously unknown whether P=BPP, but it is known that $BPP\subseteq \Sigma_2\cap \Pi_2$.

Thus, it is possible (while conjectured to be false) that adding probability to P adds expressive power.

My question is whether we know (or have evidence of) the border between P and PSPACE where adding randomization no longer adds power.


Are there any problems that are known to be in $BP\Sigma_i$ (resp. $BP\Pi_i$) that are not known to be in $\Sigma_i$ (resp. $\Pi_i$)? And similarly for $BPPH$ v.s. $PH$?

  • 6
    $\begingroup$ BPPH=PH. xxxxxxxxxxxxx $\endgroup$ – Emil Jeřábek Apr 30 '15 at 10:49
  • $\begingroup$ @EmilJeřábek - thanks, do you have a reference for this result? $\endgroup$ – Shaull Apr 30 '15 at 11:28
  • 7
    $\begingroup$ This is just a relativization of the Gács–Sipser–Lautemann theorem. $\endgroup$ – Emil Jeřábek Apr 30 '15 at 14:11
  • 4
    $\begingroup$ Though for a tighter bound, it’s better to relativize the inclusion $\mathrm{AM}\subseteq\Pi^P_2$, which gives $\mathrm{BP}\Sigma^P_i\subseteq\Pi^P_{i+1}$ (for $i\ge1$), and dually $\mathrm{BP}\Pi^P_i\subseteq\Sigma^P_{i+1}$. $\endgroup$ – Emil Jeřábek Apr 30 '15 at 15:06

There is a difficulty with the premise of your question — "when does randomization stops helping within $\mathrm{PSPACE}$ — because it suggests that the computational classes $\mathrm{X}$ such that $\mathrm{P \subseteq X \subseteq PSPACE}$ form some sort of linear hierarchy when this is not evident.

We can illustrate this by comparisons between the polynomial hierarchy and counting classes. As Emil Jeřábek indicates in the comments, $$\begin{align*} \mathrm{BP\cdot \Sigma_i^p \subseteq \Pi_{i+1}^p} &&\text{and}&& \mathrm{BP\cdot \Pi_i^p \subseteq \Sigma_{i+1}^p} \end{align*}$$ by relativisation of $\mathrm{AM \subseteq \Pi_2^p}\,$; and therefore $\mathrm{BP \cdot PH = PH}$. On the other hand, Toda's Theorem shows that $$\mathrm{PH \subseteq BP\cdot\oplus P}.$$ If you suppose that "randomization has stopped adding power by the time you ascend to $\mathrm{PH}$", then you will be tempted to suspect that because $\mathrm{PH \subseteq BP \cdot \oplus P}$, perhaps in fact $\mathrm{BP \cdot \oplus P = \oplus P}$. But I don't know that anyone conjectures this, or even that $\mathrm{PH \subseteq \oplus P}$ (which would be a necessary consequence); I think that any result of this sort would be considered a major breakthrough.

Of course, if you only care about the polynomial hierarchy, and more generally (to scale up to $\mathrm{PSPACE}$) quantified boolean formulas, then you can extract some sort of linear answer to your question — in which case Emil's comments are about as complete an answer as you are likely to get.

  • $\begingroup$ Thanks! I was indeed thinking more of the polynomial hierarchy than other classes. In fact, this question stems from studying restrictions of temporal logics, so there is some sort of hierarchy among them, and counting classes are less relevant. $\endgroup$ – Shaull May 1 '15 at 5:40
  • 1
    $\begingroup$ You may want to find a more pointed version of your question then, and try again. :-) $\endgroup$ – Niel de Beaudrap May 1 '15 at 6:18
  • 3
    $\begingroup$ Concerning the "randomization has stopped adding power" point: we also have $\mathrm{BP\cdot BPP=BPP}$, but this doesn't imply $\mathrm{BP}\cdot\mathcal C=\mathcal C$ for all classes $\mathcal C\supseteq\mathrm{BPP}$. $\endgroup$ – Emil Jeřábek May 1 '15 at 10:47
  • $\begingroup$ @Emil: sure enough, though a fair complaint may be that there's already randomness there. This raises the question of whether (for any class, however specified) one can tell whether it already 'contains randomness', but that's a much more complicated kettle of fish. $\endgroup$ – Niel de Beaudrap May 2 '15 at 4:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.