Bounds on approximating Max $k$XOR with quantum and classical local algorithms

Talk

Kunal gives a talk here.

Background

kXOR is a constraint satisfaction problem where every clause involves $$k$$ Boolean variables. Their combined parity must be either even or odd.

The threshold algorithm is competitive with QAOA on MaxCut, which is why it was studied here.

Main idea

There are four parts:

• A formula for the performance of the depth-1 QAOA on regular, triangle-free instances of kXOR. It doesn’t depend on the number of variables.
• A formula for the performance of the (classical) threshold algorithm on regular, triangle-free instances of kXOR.

With these two parts, you can compare the performances. The QAOA is better than the threshold algorithm for $$k > 4$$.

• The best possible performance given a large random regular kXOR instance. This value is calculated from a connection to spin glass theory. The code is available here and could be used for more types of CSPs.

With this part, you can compare the best possible performance with that of the algorithms above. The best possible performance scales as $$1/2 + \Theta(\sqrt{k/D})$$ but the other algorithms scale much more poorly with $$k$$.

• A proof that some fully-satisfiable Max 3XOR instances cannot be solved with QAOA until the depth is logarithmic with the number of variables. This is proven through a weak version of the NLTS conjecture. It extends a main result in Bravyi et al.

The paper has two main contributions, a comparison of depth-1 QAOA with classical threshold algorithms, and an obstruction result for log-depth QAOA on Max3XOR that follows from showing of the existence of models with a partial Z2 symmetry that satisfy the NLTS conjecture.

First, the authors show through numerics that for triangle-free instances of Max-kXOR, depth-1 QAOA performance asymptotes to a larger value in k than depth-1 local classical algorithms such as the threshold algorithms by Hirvonen et al. and by Hastings. Specifically, QAOA1 beats the classical algorithms beyond k = 4.

Then, the authors use techniques from spin glass physics to bound the best possible performance on random, bounded-degree max kXOR instances, and show numerical evidence to suggest that both QAOA1 and the threshold algorithm are far from achieving this performance. While QAOA with p>1 may be hard to analyze, it would be interesting to see whether higher-depth versions of the threshold algorithm (see, e.g., doi:10.1103/PhysRevA.103.052413) approach the Parisi bound and how fast.

Finally, the obstruction result follows from a straightforward extension of the original result of Bravyi et al. for total Z2 symmetry: one adds a small node set and includes one representative from the set in each clause in the original instance. This is a nice observation and suggests that the same hammer could be useful to give obstruction results for QAOA on other problems.

tags: QAOA  kXOR  Parisi