A three-user experiment, not a deployed network
On September 3, 2026, Reem Mandil and colleagues posted the first version of a preprint describing a laboratory twin-field quantum key distribution network built as a 127-kilometer fiber ring. Three user stations, named Alice, Bob and Danny, were placed around four spools of SMF-28 ultra-low-loss fiber. Any two users could attempt key generation while the third remained inactive. This is a three-user network with three possible pairings, although the paper’s main results table reports trials for only Alice-Bob and Alice-Danny. It is not a citywide field deployment or an operating commercial service.
The experiment matters because long-distance twin-field QKD usually demands exact optical phase control. Quantum key distribution uses measurements of carefully prepared light to let two parties create matching secret key material and expose disturbances that could indicate interception. Twin-field protocols send very weak optical pulses toward a central measurement station, where their interference supplies the useful detection events. Loss rises sharply with fiber distance. Twin-field methods are designed to improve the rate-distance relationship, but only when the arriving fields remain coherent enough to interfere reliably.
How the Sagnac ring works
The Toronto-led setup used one continuous-wave distributed-feedback laser at the central station, Charlie. An optical switch turned that light into bursts, and the pulses entered the ring through a 50:50 beam splitter. Copies traveled clockwise and counterclockwise through the same loop. Each active user changed the intensity and phase of only its assigned passing pulses. After completing the loop, the two directions returned to Charlie and interfered at the same beam splitter. Two single-photon avalanche detectors recorded which output clicked. The detector efficiency was 10 percent, and the paper reports a dark-count probability of about 3 × 10^-7 per pulse.
That shared route is the core engineering choice. In many twin-field arrangements, distant users operate separate lasers and send light through separate fiber paths. Frequency drift in the lasers and changes in fiber length or refractive index move their relative phase. Systems may respond with frequency locking, strong reference pulses, feedback electronics or later computational correction. Those techniques can work, but they add equipment and consume transmission time. In this Sagnac ring, both interfering pulse directions originated from the same laser and traversed the same physical loop. Common disturbances therefore tend to affect both directions similarly.
What eliminating phase stabilization actually saves
The authors report that the arrangement operated without active phase stabilization and without phase postcompensation. That does not mean the apparatus was passive or self-maintaining. It removed one difficult control problem, the need to continually estimate and correct relative optical phase. The potential benefit is practical: one conventional laser replaces multiple sources that would otherwise need tight frequency coordination, and the experiment used avalanche photodiodes rather than superconducting nanowire detectors. Whether that produces lower total cost in a deployable network remains unmeasured because the paper provides no complete bill of materials, operating-cost study or field-maintenance record.
Polarization still required active correction. Light traveling through fiber can have its polarization rotated differently as temperature, stress and other environmental conditions alter the fiber’s birefringence. Poor alignment reduces interference and can also impair the lithium-niobate modulators used by each participant. The team tapped a small fraction of the light at four points, counted photons, and drove four electronic polarization controllers. A custom Arduino-based counter sampled detector counts every 500 milliseconds. A thresholded gradient-ascent routine adjusted a controller only when its monitored rate fell below a target. This is active feedback, but it corrects polarization rather than phase.
Measured stability and backscatter control
With that polarization system running, the researchers recorded interference visibility for one hour in 60-second integration windows. Visibility describes how sharply the output changes between constructive and destructive interference. They report 93 ± 1 percent. Monitoring intensities at the four control detectors had relative standard deviations of 1.2, 8.9, 4.5 and 1.4 percent. The paper notes occasional dips, early drift at one monitor and a small systematic trend at another. These details make the result more useful than a single peak number because they show both stability and remaining variation.
Rayleigh backscattering posed a separate problem. Tiny density variations in fiber scatter some launched light backward, and a bidirectional ring can direct that noise into the same detectors looking for extremely weak quantum signals. Polarizing beam splitters filtered roughly half of this randomly polarized background. The optical switch also kept pulses on for 280 microseconds and off for 654 microseconds, a 30 percent duty cycle, so scattered light could decay before detection. The resulting total noise was about 1 × 10^-6 per pulse. The authors estimate backscatter at roughly three times the detector dark-count rate, enough to remain a meaningful part of the error budget.
The distances and rates require separate readings
The distance figures describe different parts of the apparatus and should not be blended. The complete ring measured 127 kilometers. Charlie-to-Alice was 50 kilometers and Charlie-to-Danny was 27 kilometers. Bob’s effective distance to Charlie depended on the active pairing: 52 kilometers for Alice-Bob and 75 kilometers for Danny-Bob. The headline performance came from Alice-Bob, whose two arms totaled 102 kilometers and 45 decibels of loss, including the modeled effect of 10 percent detector efficiency. The average quantum bit error rate was 6.49 percent, from 7.50 percent at one detector and 5.49 percent at the other.
For Alice-Bob, the reported asymptotic rate was 1.398 × 10^-5 secret bits per emitted pulse. “Asymptotic” assumes an effectively unlimited data set. At the paper’s finite-data size of 10^10 trials, that pair produced a listed rate of zero. Alice-Danny covered 77 kilometers and achieved 7.190 × 10^-5 bits per pulse asymptotically and 4.408 × 10^-6 bits per pulse in the finite-data calculation. Its average error rate was 5.02 percent. The paper is internally inconsistent about its loss: Table 1 lists 19 plus 16 decibels, totaling 35, while the accompanying prose says 34 decibels. That discrepancy needs author clarification.
What was simulated and what remains unknown
An even more consequential qualification concerns what was physical and what was modeled. The researchers experimentally implemented the signal basis, the setting whose error rate depends on interference visibility. The decoy-state gains used to estimate phase errors were generated by an optimized channel simulation using the measured channel losses. Sending probabilities and decoy intensities were optimized as well. The resulting quantities are protocol-level secure-key-rate calculations, not evidence that a complete hardened QKD product exchanged production keys under every required setting. The paper also says extra taps, photodetectors and bandpass filters would be needed at user stations to defend against strong injected light and other side channels in a fully secure implementation.
The demonstrated progress is therefore specific but worthwhile. A long fiber loop preserved strong interference without a phase-locking system, while automated polarization control and timed transmission handled two problems the geometry did not remove. This division of labor is the useful result: passive common-path cancellation for phase disturbances, active feedback for polarization, and deliberate quiet periods for backscatter. The design may simplify metropolitan quantum-network experiments with unequal paths and several users. Larger networks, ordinary fiber, higher loss and security-hardened stations could erase part of that advantage. Field reliability, sustained throughput, total cost and performance beyond three users remain unknown.