A laboratory experiment in Padova has reported secret-key generation across a deployed channel that joined optical fibre to a 620-m atmospheric path. The study tested continuous-variable quantum key distribution, or CV-QKD, a quantum-key method based on measurements of light signals. Its reported asymptotic secret-key rates were 0.42 Mbit/s for coherent states and 0.93 Mbit/s for squeezed states under their respective channel conditions. The squeezed-state result demonstrates squeezed-state CV-QKD over the deployed atmospheric channel, but the two rates are not a like-for-like head-to-head comparison.
That qualification matters because the headline figures come from asymptotic analysis rather than a positive finite-data security result. In the highest-transmittance combined cluster, 2.15 × 10^8 symbols were still insufficient for positive composable key extraction under the reported parameters, even with epsilon optimization and high reconciliation efficiency. Composable security is the stricter test of whether a finite block of data can support a usable security guarantee. The paper therefore reports promising measured-channel rates, but not a completed finite-size demonstration.
A channel made of fibre and free space
The research question was whether a locally generated local-oscillator CV-QKD system could operate across a concatenated fibre-free-space channel while using adaptive post-processing. The paper describes this kind of mixed path as largely unexplored. Its authors interpret the results as a possible route toward heterogeneous networks that combine fibre, terrestrial free space and satellite links, although the experiment itself concerns the deployed testbed and its measured conditions.
The short configuration had approximately 10 to 18 dB of total loss. The long configuration used 2 km of fibre and had approximately 13 to 21 dB of total loss. The Results section presents four representative data sets: coherent-state measurements over short and long channels, a squeezed-state short-channel measurement, and four coherent-state short-channel measurements processed jointly. Together, these cases allowed the researchers to examine different quantum-state sources and several ways of handling the same broad hybrid-link problem.
The coherent and squeezed protocols did not use entirely separate system architectures. Both shared a locally generated local oscillator, a common optical reference, along with modulation, synchronization and digital signal processing. They differed in the source that prepared the quantum state, the detection front end and the security analysis. For coherent states, the analysis used Gaussian modulation without switching between measurement settings and treated the detectors as trusted. The squeezed-state analysis additionally estimated the VS and VAS variance terms from measured data and likewise treated the detectors as trusted.
Processing followed the channel
The central processing choice was to adapt the data treatment to the channel instead of imposing one fixed partition. The procedure grouped measurements by transmittance, the amount of signal making it through the link, and optimized the cluster boundaries for their weighted secret-key contribution. For each candidate grouping, it evaluated covariance-matrix averaging and de-fading, then selected cluster sizes using local channel statistics and the amount of residual fading.
In one reported comparison data set, adaptive clustering had a higher weighted secret-key rate than each conventional choice tested. Its rate was 27% higher than the best threshold approach, 31% higher than the best linear-scale binning and 41% higher than the best decibel-scale binning. The adaptive method retained 78.6% of the acquired data, compared with 60.5% for the threshold optimum, 65.0% for linear binning and 49.6% for dB binning.
Those percentages describe a processing comparison, not a universal ranking of methods. The comparisons were not randomized, so they do not provide a randomized causal estimate that adaptive clustering will always outperform fixed threshold or binning strategies. They show what happened in the reported data set under the tested settings, while leaving open how the result would transfer to other channel conditions.
The channel itself was not static. Measured coherence times, an estimate based on how long the transmission conditions remained correlated, were 21.1 ms for the coherent-state short-channel measurement, 8.9 ms for the coherent-state long-channel measurement and 16.9 ms for the squeezed-state measurement. Processing frames were 1.25 × 10^4 symbols, or 100 microseconds, for coherent data and 1.25 × 10^5 symbols, or 1 ms, for squeezed data.
That difference in frame design is part of the study's engineering trade-off. The squeezed-state data were handled in longer blocks, and the supplied analysis identifies those longer frames as a limitation because they reduce the resolution available for partitioning the transmittance distribution. The result is therefore a test of processing choices tied to measured channel dynamics, not evidence that the same cluster boundaries will transfer unchanged to every modulation setting or link.
Blind reconciliation added another gain
Adaptive clustering was only part of the processing pipeline. The researchers also used rate-adaptive blind reconciliation, a classical decoding stage that can make additional attempts at different rates instead of committing to one information-reconciliation pass. The paper reports up to 19% additional key from blind reconciliation alone within the unified post-processing framework. Using the complete adaptive chain, the authors extracted up to 318 kbit of asymptotic secret key from a single measurement.
Compared with one-shot reconciliation, blind reconciliation was associated with total secret-key-rate gains of 12.06% in the long-channel measurement, 16.77% in the coherent-state short-channel measurement and 18.79% in the squeezed-state measurement. The corresponding extracted-key-length gains were 11.68%, 16.50% and 17.72%. These were the three reported measurements, so the figures describe those cases rather than a typical gain across all possible links.
The remaining test is finite-size security
The strongest caveat is the gap between an asymptotic calculation and a finite-data demonstration. The highest-transmittance combined cluster contained 2.15 × 10^8 symbols, yet the analysis still found no positive composable key under the present parameters. That does not invalidate the reported asymptotic rates, but it changes what they establish: performance in the measured configurations under asymptotic security bounds, not a deployable finite-size guarantee.
The evidence also remains limited to the Padova deployment and its tested conditions. Further work is needed under stronger turbulence, longer or higher-loss channels and satellite configurations. The supplied analysis also identifies the computational cost of real-time reconciliation as an open implementation question; blind reconciliation requires additional decoding attempts and classical communication, and real-time operation was not demonstrated.
The full acquisition count and a formal sample-size rationale were not reported, and uncertainty intervals for the rates and gains were not provided in the supplied analysis. The main comparison also used reported data sets rather than a randomized design. Taken together, those limits leave the finite-size security, transferability and real-time practicality of the framework open, even as the experiment demonstrates its operation on the measured hybrid link.
The document is arXiv:2608.20088v2, dated 23 August 2026, and the supplied metadata lists no journal or DOI. Its acknowledgments report support from European Union programmes, the Danish National Research Foundation, Innovation Fund Denmark, MEYS and the Czech Science Foundation. The paper says the underlying data are available in reference [49] and reports no conflicts of interest.
Paper data and sources
Original title: Squeezed- and coherent-state quantum key distribution over a deployed hybrid fibre-free-space channel
Authors: Dnan A. E. Hajomer, Huy Q. Nguyen, Ivan Derkach et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text