Tracing photon noise from atomic scattering to quantum algorithms

From scattering to an error model

A quantum-computing model often begins by assigning probabilities to different errors. This study instead starts with the interaction that produces them: a single photon passing an ensemble of two-level atoms in a chiral waveguide, modeled as supporting propagation in one direction.

The photon carries a qubit in its polarization. Its frequency, however, is not perfectly sharp: the model gives the photon a Gaussian spectrum with bandwidth characterized by σ. Different frequencies experience different polarization transformations as the photon encounters the atoms.

Hong derives those transformations from the microscopic interaction Hamiltonian, the mathematical description of the photon, atoms and their coupling. He then averages over the frequency information that is not retained when describing the polarization qubit. This produces an effective quantum channel—a rule describing how transmission changes the qubit.

The distinction matters: the modeled photon need not disappear for its polarization to suffer noise. Even lossless scattering can correlate polarization with frequency, leaving the polarization alone less faithfully preserved.

When extra protection helps

The study feeds this physically derived channel into simulations of a rotated planar code, a form of surface-code quantum error correction. The code spreads one logical qubit across multiple physical qubits. Here, each physical qubit is a photon traveling through its own waveguide.

The simulations examine code distances 3, 5, 7 and 9. A distance-d code uses d² physical qubits, so larger distances mean greater encoding overhead. The question is whether that overhead actually reduces the chance of an incorrect logical result.

For identical, lossless atoms with equal, real coupling to both polarization modes, and a photon spectrum centered on atomic resonance, the simulations show threshold behavior. Below a boundary in atom number and coupling strength, increasing code distance suppresses logical errors. Increasing either the number of scattering atoms or the coupling strength pushes the system toward worse performance in the investigated regime.

Six numerical threshold points are fitted approximately by NΓ/σ = 0.028. Here, N is the number of atoms and Γ characterizes atom–waveguide coupling. An analytical approximation for small NΓ/σ explains the scaling: the physical bit-flip probability grows approximately as 2√π times NΓ/σ. At the fitted boundary, that corresponds to an error probability of roughly 10%.

This is a threshold for the particular transmission-noise model and ideal correction procedure, not a general statement that quantum hardware can tolerate 10% errors in every operation. Each simulated data point uses 20,000 shots, and the boundary is inferred from finite code sizes.

Designing around the dominant error

Under the symmetric, resonant assumptions, the effective physical noise contains only Pauli X errors: quantum bit flips that exchange the two computational-basis states. Other conditions need not preserve this special structure. In this case, a commonly used simplification called Pauli twirling reproduces the original channel exactly, rather than merely approximating it.

Hong then carries the resulting logical error probabilities into quantum phase estimation, an algorithm for estimating the phase associated with an operator’s eigenstate. These simulations use distance-9 error correction and examine atom numbers below the correction threshold.

For target phases 7/8 and 115/128, the average estimated phase moves farther from the target as atom-induced noise increases. Ancilla registers of 4, 6 and 8 qubits show different resolution effects: all can represent 7/8 exactly, whereas only the 8-qubit register can represent 115/128 exactly.

The error bias also suggests a redesign. Hong expresses the ancilla operations and final measurements in the Hadamard, or X, basis. The redesigned and standard circuits are equivalent without noise, but respond differently to errors at the modeled transmission locations.

An X error immediately before an X-basis measurement does not change its outcome probabilities. Earlier errors can still matter, so the redesign does not eliminate the noise. Nevertheless, the paper reports improved simulated performance at Γ/σ = 0.00125. The supplied text does not establish a numerical improvement factor.

A modeling bridge, not a finished device

This is a theoretical and computational proof of concept, presented in an arXiv preprint, version 1, posted on 9 November 2026. It is not an experimental demonstration of a working error-corrected computer.

Preparation, gates and measurements are idealized, and correction uses a single round of syndrome extraction. The model excludes photon loss and atom-to-atom disorder. It also treats each controlled power in phase estimation as one ideal local operation, rather than many separate transmissions.

Those assumptions limit device-level conclusions. The useful result is narrower: microscopic knowledge of noise can guide both error correction and the choice between otherwise equivalent algorithm implementations.