Photon loss
and error budgets

1. Fixed loss removes the scaling case for random boson sampling

Boson sampling sends photons through a network of beam splitters and asks where they land, and it has been a leading candidate for showing a quantum computer doing something a classical one cannot. Armando Angrisani and Hela Mhiri, from the University of Copenhagen, EPFL and Sorbonne Université, show that under any fixed nonzero loss rate a classical computer can sample efficiently, in polynomial time, with high probability over Haar random networks. That covers single photon and fixed-squeezing Gaussian inputs, fermion sampling with magic inputs, and, for the photonic cases, thermal as well as pure loss.

Its authors note that the runtime bounds grow exponentially as loss weakens, so a polynomial time result does not mean today's experiments can be simulated in practice.

Youngrong Lim at Chungbuk National University reached a matching result in earlier work, posted on Zenodo on October 21 and listed on arXiv this week (arXiv 2611.01434), and notes that it does not by itself show that any particular device can be spoofed.

Both are average case, asymptotic proofs about sampling experiments used as advantage claims, not about error corrected photonic designs. The question for every future sampling claim: how does loss scale as the experiment grows?

arXiv listing November 6 · Theory

2. Longer bit flip times and better error models

Cat qubits store information in two opposite states of a superconducting oscillator, so bit flips are suppressed exponentially with photon number while phase flips grow only linearly. Kerr cats have been stuck at bit flip times of milliseconds. A Weizmann Institute and QuamCore team used the same nonlinear interaction that confines the cat to pass excitations into a lossy readout mode. From repeated readout records they inferred a bit flip time of 448 ± 57 seconds, which the paper puts more than five orders of magnitude beyond previous Kerr cat results.

Phase flip times stay below a microsecond, about 113 nanoseconds at the highest photon number where both were measured, so the protection is strong on one side only. Dissipative cat qubits had already passed 100 seconds; this brings Kerr cats into that range.

Riverlane went after the errors nobody wrote down (arXiv 2611.06605). Its method learns correlated error mechanisms from detector statistics and hands them to the decoder. In simulation with injected correlated errors, logical error fell 75% at code distance 15. On published data from a Google superconducting device, with no Google authors, Lambda, the error suppression per step in code size, rose 9% between distances 5 and 7.

Both approaches target errors that error correction has to handle. Will the phase flip side of cat qubits improve too?

arXiv listing November 4 · Experiment

3. The plumbing of scale: colder surroundings, steadier ions

Heat radiation from room temperature surroundings knocks Rydberg atoms out of their states and has long been suspected of setting off avalanches that remove many atoms at once. JILA, NIST and CU Boulder put an atom array behind a 45 kelvin radiation shield. Single atom decay driven by that radiation fell 30% to 50%, and avalanche rates fell by about 18% to 22% in arrays of up to 70 atoms, which a model with no free parameters matched once other measured triggers were included.

Trapped ions face a different problem when they move. A team at China's National University of Defense Technology showed single qubit gates that hold up while an ion is shuttled between zones of a chip trap under a deliberately applied magnetic field gradient (arXiv 2611.06136), beating conventional gates under shuttling. Two qubit entangling gates are the stated next step.

The shield targets correlated loss that a single gate fidelity does not describe. How much shielding will commercial atom machines need, and at what cost?

arXiv listing November 5 · Experiment