Quantum key distribution (QKD) lets two parties grow a shared secret key while bounding how much an eavesdropper can learn. Practical systems send dim laser pulses instead of true single photons. A few of those pulses carry two or more photons. An eavesdropper can quietly peel off one of the extra photons and read it later, and this leaves no trace in the error rate. This is the photon-number-splitting attack. The standard defense, the decoy-state method, varies the laser brightness at random and checks that detection statistics stay consistent across brightness levels. That check is only as good as the statistics. With a finite number of pulses, and a laser whose brightness drifts from pulse to pulse, the error bars on the inferred single-photon behavior can become so wide that no key survives.
We propose to treat decoy-state parameter estimation as what it physically is: a rare-event counting experiment. At long distance only about one pulse in a million is detected, so the right statistical yardstick is the square root of the number of detections, not the number of pulses sent. The best rigorous bound for drifting, correlated sources, Azuma's martingale inequality, uses the second yardstick. The core claim is that a variance-aware martingale bound, inverted so that it can be solved from the observed count, removes most of this penalty without giving up security against adaptive attacks. Pairing it with per-pulse intensity monitoring, in place of a worst-case drift interval, should recover positive keys in the low-data, high-drift regime where current analyses collapse.
Core Claim
Two papers set the stage, and each fixes half of the problem. Improved key-rate bounds for practical decoy-state quantum-key-distribution systems inverts the Chernoff bound. Instead of asking how far an observed count can stray from its mean, it bounds the unknown mean given the observed count. This is tight for rare counts and gains "up to 7 km" of secure distance over Chernoff plus Hoeffding at $N=10^{10}$. It assumes independent, identically distributed pulses with perfectly stable intensities, so it does not apply when the laser drifts.
Finite-key analysis for a practical decoy-state twin-field quantum key distribution takes the other half. It drops the independence assumption by using Azuma's inequality and models intensity fluctuations of size $\delta_\mu$ explicitly. It beats the repeaterless PLOB bound (Pirandola–Laurenza–Ottaviani–Banchi, the best possible rate without quantum repeaters) with $10^{12}$ signal pulses at dark-count rate $10^{-7}$. It also reports that at $\delta_\mu = 0.3$ the protocol can fail to beat that bound.
Our diagnosis is that much of that failure is a statistical artifact rather than a physical limit, and it has two separable sources. The first is that Azuma charges every pulse the full range of a detection indicator, whether or not detection was likely. Freedman-type martingale inequalities instead charge only the accumulated conditional variance $V_N=\sum_i p_i(1-p_i)$, where $p_i$ is the probability, given the past, that pulse $i$ is detected. With probability at least $1-\epsilon$, the deviation of the observed count from its conditional mean is roughly
$$ \text{Azuma: } \sqrt{2N\ln(1/\epsilon)} \qquad \text{versus} \qquad \text{Freedman: } \sqrt{2V_N\ln(1/\epsilon)}+\tfrac{2}{3}\ln(1/\epsilon). $$
Because $V_N \le \sum_i p_i$, which is the expected detection count $\approx N\eta$ for overall detection probability $\eta$, the ratio between the two widths scales as $\sqrt{\eta}$. This is the same $1/\sqrt{\text{count}}$ scaling that governs any Poisson-limited measurement. At $\eta\sim10^{-6}$ the improvement is about three orders of magnitude in the fluctuation term.
The obstacle is that $V_N$ depends on the very mean being estimated. We therefore invert the Freedman inequality in the same way the inverse-Chernoff paper inverted Chernoff, solving for the largest and smallest conditional means consistent with the observed count. The result is a bound that is rigorous under adversarial, adaptive eavesdropping, including photon-number-splitting attacks that choose which pulses to block based on history. It also needs no knowledge of the true variance.
The second source of looseness is the worst-case treatment of drift. Every pulse is assumed to sit anywhere in $[\mu(1-\delta_\mu),\mu(1+\delta_\mu)]$. A tap-off intensity monitor at Alice gives a per-pulse record instead. We bin pulses by monitored intensity and give the decoy linear program a separate set of yield constraints for each bin. Only the monitor's calibrated error then enters as a worst-case term.
Research Program
The theoretical work has three parts. The first is the inverted Freedman bound for sums of detection indicators. It comes with a closed-form or one-dimensional root-finding solution, and we prove that it is never looser than inverse Chernoff in the independent case, up to a constant term in $\ln(1/\epsilon)$.
The second part is a composable parameter-estimation step for decoy-state BB84 and for the two-decoy twin-field protocol. In twin-field QKD, two senders send pulses to a central untrusted station that interferes them. The estimation step uses the martingale bound per intensity bin. It feeds lower bounds on the single-photon yield $Y_1$ and upper bounds on the single-photon error rate $e_1$ into the standard key-rate formula.
The third part is an explicit statement of the assumption that the monitor record is private to Alice and is correlated with true intensity within a calibrated error. We also analyze how the key rate degrades if that error is mis-specified.
The empirical work is simulation, because the file records no public hardware QKD datasets. Lacuna dataset and code searches returned nothing domain-specific. We reproduce the channel models of the cited papers: 0.2 dB/km fiber loss, dark counts of $10^{-7}$ and $10^{-8}$, and security parameter $\epsilon\approx10^{-10}$. We vary block size from $10^{8}$ to $10^{14}$ pulses. Laser drift is modeled as a correlated autoregressive process whose marginal spread matches $\delta_\mu\in\{0,0.1,0.2,0.3\}$. As a secondary check, the bound is applied to non-Poissonian sources: the modified coherent states of Measurement-device-independent QKD with Modified Coherent State, whose finite-key results rest on Gaussian approximations. The question is whether their reported 10–40 km gains survive a rigorous treatment.
Baselines
The naive reference is the Gaussian fluctuation model, a $\pm k\sigma$ error bar, which is how the modified-coherent-state paper handles finite data. It gives optimistic key rates but no composable guarantee. We report it as an upper envelope and measure how often its intervals fail to cover the true yield.
The rigorous independent-pulse baseline is the inverse Chernoff analysis of Improved key-rate bounds for practical decoy-state quantum-key-distribution systems. Its reported figure is a secure-distance gain of up to 7 km over Chernoff plus Hoeffding at $N=10^{10}$. We run it on identical channel parameters with intensities and sending probabilities optimized by the same optimizer and iteration budget as our method. When $\delta_\mu>0$, its independence assumption is violated. We still report its key rates there, flagged as unsound, to show what tightness costs.
The strongest sound competitor for drifting sources is the Azuma-based analysis of Finite-key analysis for a practical decoy-state twin-field quantum key distribution. Its reported results are a PLOB crossing with $10^{12}$ signal pulses at $P_d=10^{-7}$, tolerable loss "exceeding 80 dB" at $N=10^{14}$ with stable sources, and possible failure to beat PLOB at fluctuation magnitude 0.3. We reimplement it with the same drift model and the same optimizer. Beating it means a sound key at $\delta_\mu=0.3$ that exceeds PLOB at some distance for $N\le10^{12}$. It also means reaching a given key rate with at least a tenfold smaller block size at moderate drift.
Two ablations isolate the ingredients. One applies the inverted Freedman bound with the worst-case drift interval and no monitor. The other applies Azuma with monitor binning. If almost all the gain comes from the variance-aware bound, monitoring is unnecessary hardware. If it comes from binning, the statistical contribution is minor. We also run a Kato-type martingale inequality as a further sound comparator. Such variance-sensitive bounds exist in the broader QKD literature, but we found no reported result on Lacuna for them in this setting.
Why not just…?
Why not just use the Gaussian approximation, which every experimental physics analysis already uses? The inverse-Chernoff paper argues it is not information-theoretically rigorous because it ignores how the failure probability depends on data size. At $\epsilon\sim10^{-10}$ with a few hundred detections, Poisson tails are far from Gaussian. Our coverage experiment measures empirical failure rates directly, run at a larger $\epsilon$ so that failures are observable, under adaptive photon-number-splitting strategies. It will show whether the Gaussian model fails in this regime.
Why not just apply the inverse Chernoff bound to drifting lasers and accept a small error? Correlated drift lets the conditional detection probability wander in ways an adversary can align with her blocking decisions. Independence-based bounds then have no guarantee, which is why the twin-field paper moved to Azuma. We quantify the unsoundness rather than argue it, through coverage tests under autoregressive drift with an attacker who correlates blocking with intensity.
Why not just use an existing variance-sensitive martingale inequality such as Kato's? This is the real threat. If such a bound already closes most of the gap, our statistical contribution shrinks to the inversion and its proof of dominance over inverse Chernoff. The remaining contribution would then be the monitored-intensity binning and the source-agnostic extension. The head-to-head comparison in the ablations settles this, and we will report the outcome either way.
Why not just collect more pulses or buy a more stable laser? The mode-pairing analysis notes that $10^{13}$ pulses take hours of uninterrupted operation, and satellite or drone links have short windows. More stable sources help, but our $\delta_\mu$ sweep shows how much drift a given block size tolerates. That converts the hardware question into a quantitative specification.
Proposed Experiments
The first decisive check is soundness. For each method we simulate $10^5$ sessions at $\epsilon=10^{-3}$ under three attackers: no attack, a photon-number-splitting attacker who blocks single-photon pulses, and a photon-number-splitting attacker whose blocking adapts to past intensities. We count how often the true $Y_1$ falls below the reported lower bound. Our method must stay at or below $\epsilon$.
The second check is tightness. We plot key rate against distance at $N\in\{10^{9},10^{10},10^{12}\}$ and $\delta_\mu\in\{0,0.3\}$. We also report the minimum block size that gives any positive key at 100 km and the minimum block size that beats PLOB for twin-field QKD. The prediction, from the $\sqrt{\eta}$ scaling, is that the gain over Azuma grows with distance.
The third check is the ablation decomposition. It attributes the gain between variance awareness and monitoring, and includes the monitor-miscalibration sweep that shows how gracefully the bound fails.
The proposal is falsified if the inverted Freedman bound gains less than a factor of two in minimum block size over Azuma at $\delta_\mu=0.3$. It is also falsified if a Kato-type bound matches it within noise without the inversion.