$$ \newcommand{\P}{\text{P}} \newcommand{\BQP}{\text{BQP}} \newcommand{\BPP}{\text{BPP}} \newcommand{\PSPACE}{\text{PSPACE}} \newcommand{\SP}{\text{#P}} $$ $$ \newcommand{\ket}[1]{\lvert #1 \rangle} \newcommand{\bra}[1]{\langle #1 \rvert} \newcommand{\coloneqq}{\mathrel{:=}} \newcommand{\dim}{\text{dim}} $$

One of the stupidest things about being an undergrad is that I never have enough time for research—and I am only halfway though!! Anyway, without further ado, here is a question I have thought a little bit about but never solved. I apologize in advance for any errors, I haven’t seriously thought about this problem in almost a year.

Question: Is QSZK/qpoly contained in EXP/poly?

My interest in this question came from trying to understand the limits on the power of quantum interactive proofs with advice. Aaronson (2005) showed that BQP/qpoly is contained in PP/poly and Raz (2009) showed that QIP(2)/qpoly = ALL. Due Watrous’ (2002) we know that QSZK lies between BQP and QIP(2). Moreover, it has a nice complete problem so a natural and seemingly easy question to bound the power of QSZK/qpoly. This question is also left as an open problem in Aaronson (2018) where they showed that PDQP/qpoly = ALL.

My conjecture is that QSZK/qpoly is contained in EXP/poly. My intuition is that a modification of Aaronson’s (2003) original proof that BQP/qpoly is contained in EXP/poly works here.

Sketch of Aaronson’s proof

(Adapted from Aaronson’s slides)

Let $A$ be a BQP/qpoly algorithm. Fix an input length $n$ and an advice state $\ket{\psi}$. We can make the error of $A$ exponentially small by taking polynomially many copies $\ket{\psi}^{\otimes p(n)}$ of the advice.

Define $S_0$ to be the set of advice states that cause the algorithm to output $0$ with probability $1-\epsilon$, and $S_1$ to be the set of advice states that cause the algorithm to output $1$ with probability $1-\epsilon$.

Claim 1. There exist orthogonal subspaces $H_0, H_1$ such that $S_0$ is exponentially close to $H_0$ and $S_1$ is exponentially close to $H_1$.

Proof. Given an advice state $\ket{\varphi}$, the acceptance probability of a BQP/qpoly algorithm is given by \begin{equation}\label{acceptance} \bra{\varphi} \rho \ket{\varphi}. \end{equation} where $\rho$ is a quantum state (positive semidefinite trace one matrix). Define \begin{align} H_0:& \;\text{Subspace spanned by eigenvectors of $\rho$ with eigenvalues in $[0,1/3]$}\\ H_1:& \;\text{Subspace spanned by eigenvectors of $\rho$ with eigenvalues in $[2/3,1]$} \end{align} It is known (see this proof on Math StackExchange) that eigenvectors of distinct eigenvalues of quantum states are orthogonal so these subspaces are orthogonal as claimed.

For each length $n$, define the classical advice for the EXP/poly algorithm to be a polynomial number of strings $z_i$, each of length $n$ encoding a positive integer $z_i \leq 2^n$. Call this set of strings $B$.

We will now describe an EXP/poly algorithm for $A$, that takes advice of the form mentioned above.

Fix an input $x \in \{0,1\}^n$ and loop through all $y \leq x$ in lexicographic order. (We can do this in EXP.)

Define $T_0$ to be the entire Hilbert space and we will iteratively define $T_y$ be the subspace of advice states $\ket{\psi}$ compatible with the inputs $1$ to $y$ as follows. First, define \begin{align} \Pi_0:& \;\text{Projection of $T_{y-1}$ onto $H_0$}\\ \Pi_1:& \;\text{Projection of $T_{y-1}$ onto $H_1$} \end{align} and then for each $y$, do the following:

  1. If $y \notin B$, choose the larger subspace; that is, $T_y \coloneqq \Pi_0$ if $\dim(\Pi_0) \geq \dim(\Pi_1)$, and $T_y \coloneqq \Pi_1$ otherwise.
  2. If $y \in B$, choose the smaller subspace; that is, $T_y \coloneqq \Pi_1$ if $\dim(\Pi_0) \geq \dim(\Pi_1)$, and $T_y \coloneqq \Pi_0$ otherwise.

Notice that each time we pick the smaller subspace, $\dim(T_y)$ is at least halved (because $\Pi_0$ and $\Pi_1$ are projections onto orthogonal subspaces) so we only need polynomially many bits of advice to get exponentially close to $\ket{\psi}^{\otimes p(n)}$. If we assume that this works (exercise for the interested reader), then by equation \eqref{acceptance}, we get the acceptance probability of $A$.

The QSZK case

The naïve approach is to use the above procedure to learn about $\ket{\psi}$ and then use the ideas of Kitaev and Watrous (2000) to simulate the quantum interactive proof system in EXP. But that does not work in general because of the aforementioned result of Raz (2009) who showed that QIP(2)/qpoly = ALL.

So, to prove this, one needs to take advantage of the statistical zero-knowledge restriction. It is easy to see that an analogue of Watrous’ (2002) Quantum State Distinguishability (QSD) where the circuits take as input $\ket{\psi} \otimes \ket{0}^{\otimes n}$, which I will call Advice Quantum State Distinguishability (AQSD), is complete for QSZK/qpoly. I imagine that an analog of Watrous’ proof that QSD is in PSPACE works for AQSD, which would put QSZK/qpoly in EXP/poly. (We still need EXP because the above procedure to learn $\ket{\psi}$ takes exponential time.)

Another Question. If QSZK/qpoly is strictly less powerful than QIP/qpoly, I wonder if QSZK itself is less powerful than QIP. Right now, the best upper bound we have on QSZK is PSPACE and due is Watrous (2002). It is natural to think that QSZK is contained in PP but Bouland et al. (2016), resolving a question of Watrous, showed that a proof that even SZK is in PP would require non-relativising techniques. So, what about PPPP? or the Counting Hierarchy?

Acknowledgements. I thank Andrew Drucker and John Watrous for helpful discussions. (and sorry for giving up!)