Skip to content
Quantum.world

The Signal

Preprint tests tunneling search on a photonic chip

A preprint reports tests of a tunneling-based search method on a photonic waveguide chip, using one- and two-dimensional nonconvex landscapes. Its supplement says the experiments measured classical-light intensity, and that the data do not establish an end-to-end speedup over stochastic gradient descent.

Preprint tests tunneling search on a photonic chip

A photonic waveguide chip was used to test a quantum-tunneling search across one- and two-dimensional nonconvex landscapes, with the researchers reporting that a changing Hamiltonian helped distinguish the global minimum. The preprint, posted to arXiv on Nov. 12, describes quantum tunneling search, or QTS, as a method combining a quantum tunneling walk with local annealing.

Encoding the landscape in waveguides

For the tests, the team encoded objective-function values in the refractive-index profile of femtosecond-laser-written waveguides. Each waveguide represented a point in the landscape: lower objective values corresponded to lower on-site potentials, while coupling between neighboring guides allowed light to move through the array.

The researchers tested 29-by-1 and 21-by-9 waveguide arrays, with waveguides spaced 10 micrometers apart. The one-dimensional arrays were used for landscapes with multiple wells; the larger arrays represented two-dimensional landscapes. The chip’s waveguide couplings and refractive-index profiles were fixed by fabrication rather than electronically reprogrammed during measurement.

In QTS, an incomplete local-annealing process prepares an input state in part of the landscape. The state then evolves across the broader landscape, and the method uses an average of its distribution over propagation to identify minima. The researchers also varied the Hamiltonian during propagation, increasing the potential’s relative strength to concentrate the output around lower-energy locations.

Tests across multiple wells

In a two-dimensional double-well test, the preprint reports that an average over propagation lengths up to 90 millimeters highlighted both local minima, even though the input was not the exact ground state of the starting well. The team also tested one-dimensional three- and five-well landscapes. For the five-well case, it says the available waveguide length was insufficient for the distribution to converge, and presents a varying-Hamiltonian approach as an alternative.

For a randomly shaped five-well landscape, the researchers report that both constant- and varying-Hamiltonian experiments highlighted the local minima over the measured propagation range. They say the varying Hamiltonian gave a more uniform distribution and explored distant regions more strongly. In a separate test with uneven minima, measured intensity peaked at the location labeled x=14, identified in the paper as the global minimum, after 90 millimeters of propagation.

Speedup claim has an important limit

The paper compares the optical results with stochastic gradient descent, or SGD, a classical optimization method. Its abstract describes the work as demonstrating quantum speedups, but the supplement draws a boundary around that comparison: the photonic curves are measured output-intensity fractions, while the SGD curves are distributions from software trajectories, and their horizontal axes represent different quantities.

The authors say a factor of 60 was used only to make the plotted spatial spread of the SGD results resemble that of the photonic search. They state that the data do not support an end-to-end speedup estimate. Such an estimate would need to account for factors including objective-function calls, trajectories, execution time, optical input and output, and the one-time costs of preparing and calibrating the waveguide arrays.

The supplement also clarifies that the main experiments used classical-light intensity measurements, not single-photon counting; its two-photon coincidence maps are simulations. The authors identify a two-photon experiment as possible future work, alongside extending the search to higher-dimensional landscapes and encoding different objective functions on a reconfigurable chip.

Sources

  1. Non-Convex Optimization via Quantum Tunneling Search on a Photonic Chip
  2. arXiv (full text)

Related stories

Explore

More articles