Can measurement-device-independent QKD remain secure with flawed state preparation over 300 km?

A 303.37 km key with three source flaws modeled

The paper asks whether measurement-device-independent (MDI) quantum key distribution can keep a provable secret key when the states Alice and Bob actually send differ from the states they intended. MDI-QKD removes assumptions about the receiver's detectors by having both users send signals to an untrusted relay, Charlie, which performs an interference measurement and announces outcomes. The remaining weak point is the source. Here the authors model three source imperfections at once and still report secure key distribution over 303.37 km of standard single-mode fiber, with a secret key rate of 6.26E-11 at that distance.

The three flaws are encoding uncertainty, intensity fluctuation, and imperfect vacuum states. Encoding uncertainty means the phase or bit encoding deviates from its nominal value; intensity fluctuation means the decoy and signal pulse energies vary; an imperfect vacuum state means the nominally empty pulse still carries a small residual intensity because the extinction ratio is finite. The paper treats these as state-preparation uncertainty (SPU), meaning the prepared state is not exactly the one the security proof assumes.

The distance claim is comparative, not absolute. The authors state that earlier MDI-QKD experiments addressing only encoding SPU reached a maximum of 170 km, and that their result extends the distance by 83.56 km in the finite-size case while covering more SPU types. That comparison is the paper's own framing of what is new.

Operator dominance converts measured states into bounds

The security proof needs two quantities for single-photon events: how often they occur, and how often they carry phase errors that leak information to an eavesdropper. Neither is measured directly. The authors instead build operator inequalities, called operator dominance conditions, that bound the unknown single-photon state from above and below using states the experiment can actually prepare and count.

Concretely, the condition in the paper's Eq. (9) writes a target operator as a real-coefficient combination of the four single-photon states in the Z and X bases, and Eq. (12) gives analogous lower and upper bounds built from decoy intensities. Once such an inequality holds, the counting rate of the target state can be bounded by the measured counting rates of the test states. The authors describe this as giving a tight estimate of the secret key rate rather than a loose worst-case one.

The vacuum-state problem needs extra handling because time-bin encoding relies on the vacuum as a signal state, yet a finite extinction ratio makes it imperfect. The authors introduce an auxiliary weak decoy at intensity omega to stand in for the vacuum and derive the first condition in Eq. (12) from inequalities on the unnormalized states. They argue the estimate stays tight because the term q0,o + q1,o - 1 is close to zero.

Key rate falls steeply as fiber length grows

The experiment reports results at five fiber lengths. The secret key rate R_MDI is 2.06E-6 at 101.15 km, 2.80E-7 at 152.40 km, 5.51E-8 at 202.31 km, 3.92E-9 at 253.56 km, and 6.26E-11 at 303.37 km. These are rates per the paper's reported units, and they span roughly five orders of magnitude across the range.

Other reported quantities move in the same direction. The gain Q[mu,mu] falls from 1.44E-5 at 101.15 km to 1.83E-9 at 303.37 km. The single-photon counting rate s11 falls from 2.83E-4 to 4.80E-8 over the same span. The phase error rate ex is not monotonic: it is 0.0939, 0.0845, 0.1880, 0.2331, and 0.2703 across the five lengths, so the longest link also carries the highest reported phase error rate.

These are different measures and should not be substituted for one another. Q[mu,mu] is a detection gain for the signal-intensity pair, s11 is a single-photon counting rate, ex is a phase error rate, and R_MDI is the final secret key rate. The paper's security claim rests on the combination, not on any single number.

Gaussian fits turn measured drift into model input

The security analysis needs a description of how far prepared states stray. The authors characterize the phase distributions with Gaussian fits, reporting means and standard deviations for the four phase settings on each side; for example, the standard deviations range from 0.0143 to 0.0245. They do the same for intensity distributions at each of the five distances, listing means and standard deviations for the signal, decoy, and weaker decoy intensities plus the vacuum setting.

The intensity means are close to their nominal values but not identical, and the standard deviations are small relative to the means. At 303.57 km, for instance, Alice's signal intensity has mean 0.199829 and standard deviation 0.000847. The paper also reports the probabilities of preparing each intensity pair, such as 0.7574 for the (mu, o) pair on Alice's side.

A separate measurement checks laser stability. The authors record the frequency difference of two independent free-running lasers every 0.1 seconds over 30 minutes using a beat setup with a 50:50 beam splitter and classical photodetectors, and report the resulting frequency difference in the range shown in their Fig. 5. This characterizes the source, not the security proof itself.

The proof assumes uncorrelated pulses and bounded drift

The operator dominance conditions in Eq. (12) rely on the absence of correlations between pulses, such as those from the pattern effect. The authors note that pattern sifting with alternate key distillation, or a reference technique for setting-choice correlations, could remove such effects, and they state that here the independent and identically distributed assumption is taken as guaranteed.

The sufficient conditions in Eq. (34) require relations among the intensity sets, such as mu > o and mu + o > nu + o'. Under intensity fluctuation these sets become intervals rather than fixed values. The authors argue the conditions remain reasonable because only a small fraction of states, those far from nominal intensity under the Gaussian model, would violate them, and they suggest optimizing intensities in advance if the sign analysis of the coefficients fails.

The discussion also names what is not covered: encoding correlated with other degrees of freedom, which could create side channels and break the qubit assumption, and correlations between nearby pulses from the pattern effect, setting choices, or phase. The authors present these as directions for future work rather than as solved problems.

Security margin matters more than the distance record

The paper's most decision-relevant comparison is not the 303.37 km figure alone but what it costs to keep security under more source flaws. The authors report that at roughly 250 km their system maintains a secret key rate similar to long-distance MDI-QKD experiments that do not account for SPU. In other words, the added modeling of encoding, intensity, and vacuum imperfections did not collapse the rate at that distance in their comparison.

That comparison is the paper's own, and it is a statement about similar rates, not identical ones. The reported rate at 253.56 km is 3.92E-9, and readers should treat the near-250 km claim as the authors' cross-experiment comparison rather than as a controlled head-to-head measurement.

The practical reading is that the operator dominance framework lets an implementation keep a provable key while admitting measured, Gaussian-characterized source drift, at the price of a rate that falls steeply with distance and a phase error rate that rises to 0.2703 at the longest link. The relevant threshold for a deployer is therefore not whether 300 km is reachable, but whether the measured drift at their own link stays inside the modeled distributions that the proof consumes.

Sources used in this answer