First, the world thought that D-Wave hadn't built a quantum computer; then, it thought there was a quantum computer in the box; next, there was disappointment that the D-Wave machine didn't speed things up (but might still be quantum); and now, it starts to look like it's not quantum after all.
In the latest instalment in the duelling papers battle at Arxiv, researchers from IBM and UC Berkeley (including quantum computing pioneer Umesh Vazirani) tip-toe the fine line between academic debunking and professional courtesy: rather than saying outright “this is not a quantum computer” (which is how the results are being headlined in some quarters), they state with more restraint that “classical models for the D-Wave machine are not ruled out.”
To put the latest experiment in context: last year, a paper authored by Sergio Boixo of the University of Southern California, Matthias Troyer of ETH Zurich and others concluded the D-Wave computer performs quantum annealing – in other words, in a limited sense, it is a quantum computer.
They reached this conclusion by correlating the D-Wave machine's input-output behaviour with a quantum model called simulated quantum annealing, and comparing that to the predictions of two classical computing models, simulated annealing and classical spin dynamics. Since there was a poor correlation between the D-Wave and the classical models, and a good correlation with the predictions of the quantum models, the Boixo/Troyer paper decided that “the device performs quantum annealing.”
If The Register understands the new paper correctly, there's a problem with this conclusion: what if there are other classical models that could predict the D-Wave machine's behaviour, without resorting to quantum-level explanations.
That's what the UC Berkeley / IBM team (lead author Seung Woo Shin) sought to explore, starting with the basis of the Boixo/Troyer work: “the core of the argument [that D-Wave has built a quantum computer – El Reg] is based on the finding that the D-Wave machine and simulated quantum annealing generally found the same set of instances to be 'hard' and the same set of instances to be 'easy.'”
The new experiment puts forward a different classical model as the basis of comparison: it “replaces each spin in the D-Wave machine with a magnet pointing in some direction in the XZ plane. Each magnet is subject to both an external magnetic field as well as dipole-dipole couplings due to interactions with nearest neighbours on the so-called Chimera graph,” the paper states.
“The external magnetic field is attenuated at the same rate as in [the Boixo/Troyer paper], while the couplings between magnets also follow the schedule [from the same paper. We consider the input-output behaviour of this model on the same set of one thousand inputs.”
Their conclusion is that their classical model achieves “as good or better correlation with the D-Wave machine's input-output behaviour than simulated quantum annealing does”.
The Register notes that Matthias Troyer provided a suggestion to the Berkeley / IBM team, which they followed: a direct comparison between their model and simulated quantum annealing, which also revealed a strong correlation (and thus a demonstration of the validity of the Berkely / IBM model.
“One way to view this result is that our model is the classical analogue of a mean-field approximation to simulated quantum annealing, and that for the set of problems solved by D-Wave One, this approximation is very accurate,” they write.
Of course, the problem with all academic analyses of the D-Wave computer is that all any researcher can test is the input-output behaviour of the machine, since the company keeps its internals a closely-guarded secret. ®