Tracking a Drifting Blockade: Memory-Based Drive Control for Single-Photon Sources

A photon blockade is a condition where an optical cavity, once it holds one photon, refuses to admit a second. The device then emits photons one at a time, which is what single-photon sources, optical switches and photonic quantum processors need. The condition is fragile. It holds only while the driving laser's frequency and phase sit at a narrow sweet spot, and in real solid-state devices that sweet spot keeps moving. Quantum dots "blink," going dark for stretches of time. Their emission frequency wanders with charge noise. In the interference-based variant of blockade, a small error in phase or detuning turns clean one-at-a-time emission into photon bunching, where photons arrive in clumps. Today these devices are characterized at a fixed operating point. There is no treatment of how to keep them at the sweet spot over hours.

This proposal argues that keeping a blockade alive is a visual-tracking problem in disguise, and that the machine-learning literature on temporal appearance drift has already worked out the right structure for it. That literature covers trackers that must follow an object whose appearance changes, and which must not be fooled by occlusions or look-alike distractors. In our setting the "object" is the device's spectral and photon-statistics signature. Its "appearance" drifts as the emitter wanders. Blinking plays the role of occlusion, and neighbouring resonances play the role of distractors. We propose a controller that keeps a small key-value memory of past (drive setting, measured signature) pairs, borrowed from space-time memory trackers. It adds a write rule that refuses to store observations taken while the emitter is dark or ambiguous. After blinking or recurring spectral jumps, the controller re-locks immediately by retrieving the matching past state instead of re-scanning.

Core Claim

The claim is narrow and testable. Under realistic drift, blinking, and a finite photon-counting budget, a write-gated memory controller will keep a quantum-dot or interference-based blockade source below a target antibunching level for a larger fraction of wall-clock time than two alternatives: a standard dither lock, and a recursive Bayesian filter with no memory of past states. Antibunching is measured by $g^{(2)}(0)$, the normalized rate at which two photons arrive together; values below 1 mean photons arrive one at a time, and 0 is a perfect single-photon source. The advantage should come specifically from two things. The first is recurrence: charge configurations of a quantum dot tend to revisit earlier states, much as a rotating object returns to earlier views. The second is refusing to update during dark periods. If the advantage does not survive an ablation that removes the memory, the idea is wrong.

The connection between the fields runs in one direction. Tracking supplies the method and photon blockade supplies the application. We are not proposing photonic hardware for video tracking. The obvious bridge ideas in that direction, such as blockade as a hardware gate for video features or quantum interference as a distractor filter, have no physical path to handling megapixel classical video. The closest existing work, Universal Logical Quantum Photonic Neural Network Processor via Cavity-Assisted Interactions, trains a cavity-based photonic network by backpropagation for static state preparation (>99.9% fidelity on 2-photon Haar-random and 4-photon N00N states). It treats fabrication error as a fixed perturbation that deeper networks absorb. It never treats the operating point as something that drifts during operation, which is our entire subject.

Research Program

The physics motivates the problem in three ways. Probing the ladder of dressed states and nonclassical light generation in quantum-dot–cavity QED maps the Jaynes-Cummings dressed states of an InAs dot in a photonic-crystal cavity. The Jaynes-Cummings model describes a single emitter coupled to a single cavity mode, and its dressed states are the mixed light-matter energy levels this coupling produces. The paper reports that the dot is active only about 65% of the time. That is an occlusion rate far worse than most tracking benchmarks. For interference-based ("unconventional") blockade, Phase-Controlled Tunable Unconventional Photon Blockade in a Single-Atom-Cavity System shows that the drive phase controls the destructive interference. Strong photon antibunching in weakly nonlinear two-dimensional exciton-polaritons predicts $g^{(2)}(0)\sim10^{-3}$, but only at a finely tuned nanodot radius of about 42 nm. Both results point to a very sharp optimum, which is exactly the regime where slow drift is most damaging.

The tracking literature supplies the method. Video Object Segmentation Using Space-Time Memory Networks introduced reading the current state from a bank of past keys and values by attention. It updates the tracker without test-time optimization. STMTrack: Template-Free Visual Tracking With Space-Time Memory Networks showed that a small pool of six frames (first, most recent, and four intermediates) is enough to follow deforming objects. The tracking file flags a weakness in such small pools: heuristic eviction can discard early states that later recur. We take that weakness seriously, because recurrence is the dominant structure in dot spectral wandering. AiATrack: Attention in Attention for Transformer Visual Tracking admits only high-overlap frames into its short-term memory. State-Aware Anti-Drift Object Tracking documents how unguarded updates poison a tracker. These two papers motivate our write gate, which is the analogue of AiATrack's overlap threshold for a photon-counting device.

Concretely, we build a simulator and a controller. The simulator integrates the driven Jaynes-Cummings master equation, and a two-mode interference model for unconventional blockade. The emitter frequency follows a random telegraph process between a few charge configurations plus slow diffusion. A dark state is occupied roughly 35% of the time, matching the reported duty cycle. Detection is finite-count: each control step yields a transmission count and a noisy coincidence estimate of $g^{(2)}(0)$.

The controller works in four steps. It encodes a short window of counts at a few probe detunings into a key. It retrieves from memory by softmax attention over stored keys, in the form $W(p,q)$ used by space-time memory networks. It outputs a drive detuning and phase as the attention-weighted value, and then refines them with a few local probes. Finally, it writes the new (key, drive setting) pair only if two conditions hold. The emitter must be judged bright, which means the transmission count exceeds a threshold set by the dark-state count distribution. And the measured $g^{(2)}(0)$ must be below a target, so that the stored setting is a good one.

The memory is managed by merging near-duplicates rather than evicting old entries, so that rare recurring states survive. As an optional third platform, Rydberg-blockade drive-detuning drift can be simulated with Pulser: An open-source package for the design of pulse sequences in programmable neutral-atom arrays. Theory for that platform comes from Photon-Photon Interactions via Rydberg Blockade.

Baselines

The naive reference is a fixed drive setting. It is optimized once at the start of the run and never changed, the way blockade is characterized in the papers above. It tells us how fast drift destroys blockade when nothing is done. The second naive reference is periodic full re-scan: every $T$ seconds the controller sweeps detuning (and phase, for unconventional blockade), then re-centers. It is run under the same total photon budget as our method, so that time spent scanning counts as time without blockade.

The strongest classical competitor is a dither lock. The drive frequency is modulated slightly, and the error signal is fed back to hold the lower-polariton transmission peak, or the $g^{(2)}(0)$ minimum in the unconventional case. This is what an experimentalist would actually build first. Its modulation depth and loop gain get the same tuning budget, a fixed number of simulated hours of hyperparameter search, as our controller.

The strongest estimation competitor is a switching Bayesian filter. It is a particle filter over emitter frequency with an explicit dark state and discrete charge-configuration modes, and it is given the true generative model of the simulator. This is deliberately generous, because it is the model-based ideal.

Two ablations isolate our key ingredients. One removes memory retrieval, which leaves a purely recursive local-probe controller. The other removes the write gate, so that every observation is stored. There is no reported result found on Lacuna for any feedback-stabilized photon-blockade source, so no external number exists to beat. The static targets we refer to are the predicted $g^{(2)}(0)\sim10^{-3}$ optimum and the 65% active fraction. Beating the baselines means two things at matched photon budget and matched latency: a higher fraction of wall-clock time with $g^{(2)}(0)<0.5$, and a shorter median re-lock time after dark periods. The margins must hold across five or more drift seeds and exceed their variability.

Why not just…?

Why not just use a dither lock, as for any laser-cavity system? This is a real threat for conventional blockade when there are no blinking or discrete jumps. There a smooth peak exists and the lock should do well, and we expect it to tie our method. It fails where the tracking analogy bites. During dark periods the error signal vanishes and the lock wanders. After a charge jump it has to re-find the peak from scratch. In unconventional blockade the target is a minimum of $g^{(2)}(0)$ rather than an intensity peak, and $g^{(2)}(0)$ can only be estimated from slowly accumulating coincidence counts, which makes the dither signal noisy and slow. The telegraph-drift and blinking-rate sweeps below show exactly where the lock stops being competitive. If it never stops, we report that.

Why not just run a particle filter with the right noise model? If the drift model is known and low-dimensional, a switching filter is close to optimal and may match or beat us. We give it that advantage deliberately. Our bet rests on model mismatch. Real dots have an unknown number of charge configurations, and each configuration has its own optimal phase in unconventional blockade. A non-parametric memory of settings that actually worked avoids needing a correct generative model. The decisive check is a mismatch sweep: the filter is given a wrong number of modes or wrong transition rates. If memory does not help under mismatch, the contribution reduces to "use a Bayesian filter," and we will say so.

Why not just build better devices? Charge stabilization and better fabrication will reduce drift, and nothing here argues against them. But the blinking and edge-defect problems are documented features of current platforms, and the sensitivity of interference-based blockade to phase and detuning is intrinsic. That sensitivity is the price of working at weak coupling. Any source operated for long periods will face some residual drift, and the controller's value can be read directly from how its advantage shrinks as simulated drift amplitude is reduced.

Proposed Experiments

The main study is a grid over three settings: drift type (slow diffusion, telegraph jumps, and both), dark-state fraction (0%, 35%, 60%), and blockade mechanism (conventional in the strong-coupling regime, with the higher-manifold feature near detuning $0.5g$ as a known distractor peak; unconventional with phase control). In each cell we run every controller for long simulated durations. We report the fraction of time with $g^{(2)}(0)<0.5$ and $<0.1$, delivered single-photon rate, re-lock time after dark periods, and photons spent on probing.

The decisive checks are three. First, memory must beat the no-memory ablation specifically in telegraph-jump settings and not in pure diffusion. That would show the gain comes from recurrence, as the space-time memory analogy predicts. Second, the ungated-write ablation must degrade as the dark fraction grows, reproducing in physics the memory-poisoning failure seen in tracking. Third, under model mismatch the memory controller must beat the switching filter. If the first check fails, the tracking analogy is wrong for this problem.

A small portability check closes the loop in the other direction. The bright-state write gate becomes a gate on STMTrack's six-frame pool driven by the tracker's confidence, and merge-instead-of-evict becomes the pool's management policy. We evaluate on the full-occlusion sequences of LaSOT: A High-Quality Benchmark for Large-Scale Single Object Tracking, comparing against unmodified STMTrack under identical training and inference. The file gives no LaSOT number for STMTrack; for reference, AiATrack reports 69.0% AUC on LaSOT. We claim no gain beyond the modified STMTrack itself. This check only tests whether the device-derived memory rule is neutral or helpful for video.

If a real quantum-dot device becomes available through collaboration, the same controller would run on hardware with Hanbury-Brown and Twiss coincidence counting. The comparison against the dither lock would be repeated at matched integration time.