Skip to content
Quantum.world

The Signal

Paper claims learning advantage for shallow quantum circuits

Researchers say they have proved a distribution-learning advantage for constant-depth quantum circuits over a restricted class of classical circuits. The theoretical result uses highly structured data and does not establish an advantage over deeper classical circuits.

Paper claims learning advantage for shallow quantum circuits

Researchers have reported a theoretical learning advantage for shallow quantum circuits over a restricted class of classical circuits, in a paper published by npj Quantum Information on November 10. The result concerns a task in which a generator learns to reproduce an unknown probability distribution from examples.

The comparison uses constant-depth quantum circuits with one- and two-qubit gates and constant-depth classical circuits with bounded fan-in and fan-out. The authors describe the advantage as unconditional.

What the learner produces

The task is set in the probably approximately correct, or PAC, framework. A learner receives samples from an unknown distribution and must produce a generator whose output distribution is close to the target.

The paper compares the kinds of generators the learner is allowed to output, rather than claiming that every quantum learning algorithm outperforms every classical one. Its central result is that, for the constructed distribution class, a shallow quantum generator can be learned to arbitrary accuracy, while the restricted classical generators incur a non-negligible error.

Closeness is assessed using total variation distance. The paper also sets a probability of failure for the learning procedure, so the accuracy requirement is paired with a requirement on how reliably the learner meets it.

From sampling to learning

The paper builds on earlier sampling results by Bene Watts and Parham. Its authors introduce a hyperplane-learning problem into a function used in that work, turning the sampling construction into a distribution-learning task.

In the construction, the learner must recover a parameter from examples. Once that parameter is known, the procedure can specify a shallow quantum circuit that generates the target distribution.

The authors describe an efficient learning algorithm and say that, under their stated conditions, it can identify the parameter with the required probability and output the circuit that generated the target. Their proof compares that circuit's output with the performance attainable by the restricted classical circuit class.

Correlations and limits

The authors identify non-local correlations in states that can be prepared by constant-depth circuits as the source of the advantage.

The result is a mathematical comparison, not a reported hardware experiment. Its authors are N. Pirnay, S. Jerbi, J.-P. Seifert and J. Eisert; the journal published the peer-reviewed accepted research ahead of its final version of record.

The authors say the data used in the problem remain highly structured, even though the task is formulated as a machine-learning problem. They also identify proving advantages over deeper classical circuits and moving toward unstructured data as further work.

Sources

  1. npj Quantum Information
  2. arXiv

Related stories

Explore

More articles