Quantum networks cannot amplify their signals. The no-cloning theorem forbids copying an unknown quantum state, so photons carrying entanglement fade exponentially with distance. Networks work around this by chaining short links through repeaters, which are nodes that store entanglement in a quantum memory and stitch neighbouring links together, and by routing over satellite and aircraft links whose loss is lower than fiber. That workaround only pays off if the network knows, a few milliseconds ahead, which links will be good. Free-space links drift with cloud, haze, turbulence and satellite elevation, and stored qubits decohere quickly. A router that acts on stale link estimates wastes scarce entangled pairs. We propose a link-loss forecaster that learns continually as weather and orbits change, and that never forgets a condition it has seen before. When fog returns next week, the forecaster should already know fog.
The mechanism takes one idea from exemplar-free incremental learning, meaning methods that learn a sequence of tasks without storing past raw data. Analytic continual learning keeps only a small matrix summarizing past inputs, and its recursive closed-form updates give exactly the same weights as retraining on all past data. Its known weakness is that it needs a frozen feature extractor, which breaks when the input domain shifts. For link loss we can use a feature map that does not go stale: the physics of attenuation. Log-transmissivity is close to linear in a small set of physically derived quantities. Regime changes then alter the coefficients, not the features. Our claim is that a bank of regime-specific analytic regressors on physics features, with an analytic rule that decides which regime is active and when a new one appears, forecasts link quality better right after a regime recurs than the obvious adaptive filter does. It also updates well within memory coherence times, and it raises end-to-end entanglement fidelity when its forecasts drive a standard router.
Core Claim
In fiber, transmissivity is $\eta = e^{-\alpha L}$, where $\alpha$ is the attenuation coefficient and $L$ the length, so $\log \eta$ is linear in $L$. Free-space links add terms for diffraction, pointing error, atmospheric extinction (which scales with the air mass along the path, and so with elevation angle) and turbulence-induced fading. Each term is approximately additive in log-transmissivity. We write $\log \eta_t \approx w_r^\top \phi(x_t)$. Here $x_t$ is the link telemetry at time $t$ (geometry, elevation, pointing jitter, local weather sensors), $\phi$ is a fixed physics-derived feature map with a few dozen dimensions, and $w_r$ is the coefficient vector of regime $r$, for example clear night, daytime turbulence, thin cloud, or haze.
Each regime keeps only an autocorrelation matrix and a cross-correlation vector. Its weights are the ridge solution
$$ w_r = \Big(\lambda I + \sum_{t \in \mathcal{T}_r} \phi(x_t)\phi(x_t)^\top\Big)^{-1} \sum_{t \in \mathcal{T}_r} \phi(x_t)\log\eta_t, $$
where $\mathcal{T}_r$ is the set of past time steps assigned to regime $r$ and $\lambda$ is a regularizer. The inverse is updated by rank-one recursions in $O(d^2)$ time for feature dimension $d$. As in ACIL, which proved that recursive analytic updates equal joint training on all past data, each $w_r$ is exactly what one would get by storing every sample of that regime and refitting.
Regimes recur without announcement. That matches the blurry, recurring-class setting that GACL handles by splitting classes into seen and unseen and updating a shared autocorrelation memory. We borrow that split for regimes. An analytic classifier on weather telemetry assigns each step to a known regime. A residual test, which checks whether the forecast error is inconsistent with every stored regime, creates a new regime. What is new is not recursive least squares, which is decades old. It is the combination of three things: a physics feature map that removes the frozen-backbone objection, regime-partitioned exact memory so that recurring conditions are recalled rather than relearned, and evaluation by the downstream quantity that matters, which is entanglement delivered before memories decohere.
Relation to Prior Work
The closest bridges between quantum systems and continual learning concern local circuits, not channels. CL-QAS uses elastic weight consolidation (EWC), a penalty that discourages changes to weights judged important for earlier tasks. It applies EWC to a reinforcement learning agent that designs variational circuits for sequential ECG and finance tasks. Ye and Chen reuse a library of past policies to synthesize Bell states as gate noise drifts. Both motivate the idea that drift in quantum hardware calls for continual learning. Neither models photon loss, repeaters or routing, and both use penalty-based or caching mechanisms rather than exact memory.
On the networking side, SPARQ trains a deep Q-network router on space-air-ground graphs, ranking paths by transmissivity rather than distance. Its router is trained offline and fixed thereafter, which motivates our focus on post-deployment drift. SPARQ also supplies our simulation setting and our main end-to-end comparison. The optimization approaches, Benders decomposition for SATQN routing and two-stage stochastic programming with its case study, assume known weather distributions. They constrain our design because they already handle planning well, so we target short-horizon forecasting instead of replacing them. Pant et al. show that multi-path routing on grids escapes exponential rate decay when links succeed often enough. This is why accurate per-link estimates matter: path choice on a mesh is only as good as the link weights fed to it.
On the learning side, GPM restricts gradient updates to directions orthogonal to those used by past tasks and reports zero backward transfer on 20-split miniImageNet. GPM also admits that this protected space fills up over long sequences, which is a real concern for an endless stream of weather. DOLFIN brings orthogonal low-rank adapters to federated settings. We use both as neural baselines rather than as our method. Our regimes are few and recurring, so they do not saturate.
Research Program
We will first build the feature map and validate it. For each link type in a SPARQ-style simulator (fiber, ground to high-altitude platform, platform to low-orbit satellite), we derive $\phi$ from standard loss terms. We then check that a single regime's linear fit explains held-out log-transmissivity within that regime. If it does not, the whole premise fails early and cheaply.
Second, we will build a nonstationary channel generator. It will produce months of simulated telemetry in which regimes switch and recur, with realistic dwell times, gradual device aging layered on top, and occasional regimes never seen before. No public free-space quantum link telemetry dataset appears in our sources, so this generator is a contribution in itself, and its assumptions will be stated explicitly.
Third, we will implement the regime-bank forecaster. It includes the analytic regime classifier and the residual test for creating new regimes. We will also give a slow drift term such as aging a small exponential forgetting factor inside each regime. That treatment is separate from regime switching.
Fourth, we will couple the forecasts to routing in two ways. One is maximum-product-transmissivity routing, which is shortest path on $-\log \hat\eta$. The other feeds forecasts as inputs to a SPARQ-style Q-network. We will measure delivered fidelity and resolved requests under a memory coherence time.
Baselines
The naive reference is persistence: forecast each link's next transmissivity as its last measured value and route on that. It costs nothing and is surprisingly strong when conditions are stable. We will also report SPARQ's shortest-distance router as the no-learning floor.
The strongest simple competitor is recursive least squares on the same physics features with a single global model and an exponential forgetting factor. This is the textbook adaptive filter. It uses identical features, data and update cost, and it differs from our method only in having no regime memory. We will tune its forgetting factor on a validation stream with the same budget we give our residual threshold and regularizer.
The strongest existing learned system is the SPARQ router. Its paper reports that it resolves 39% more teleportation requests than shortest-path routing and finds a path in about 3.94 ms, against 90 to 93 ms for traditional algorithms (SPARQ). We will reimplement it in our simulator, train it on the same early portion of the stream, and let it fine-tune online. We will run it with and without GPM projection and with and without EWC, all on equal wall-clock budgets. We found no reported result on Lacuna for link-transmissivity forecasting under recurring regimes, so every forecasting number will be produced here.
We add an unrestricted replay baseline: a neural forecaster retrained on all stored telemetry. It is an upper-bound reference, not a privacy-constrained competitor.
Two ablations isolate our key ingredients. The first replaces the physics feature map with ACIL-style random feature expansion of raw telemetry, which tests whether the physics features matter. The second collapses all regimes into one analytic model with no forgetting, which is exact joint training over everything. It should fail precisely because old regimes contaminate current forecasts.
Beating the baselines means lower forecast error in the first minutes after a known regime recurs than both forgetting-factor least squares and fine-tuned GPM. It also means end-to-end fidelity at or above the online SPARQ router, with per-step updates far under the coherence time.
Why not just…?
Why not just use recursive least squares with a forgetting factor? It is the real threat, because it shares our features and update cost. Its weakness is structural. To track a fast switch it must forget quickly, and then it has discarded the previous occurrence of the regime it is returning to. It therefore relearns fog every time fog comes back. Our decisive check measures error and delivered fidelity as a function of time since a regime recurred. If a well-tuned forgetting factor matches us there, the regime bank adds nothing and we will report that.
Why not just keep a replay buffer, since telemetry is not private? For link data, privacy is not the argument, and we do not make it. The argument is that exact per-regime memory already equals refitting on every stored sample of that regime, at $O(d^2)$ memory per regime with $d$ a few dozen. That suits payload computers on satellites and aircraft. The replay baseline tells us whether a flexible neural model that sees everything beats a linear physics model. If it does by a wide margin, the conclusion is that the feature map is too crude, not that replay is required.
Why not just route on live measured transmissivity? When measurements are fresh and free, this is nearly optimal, and forecasting helps little. In practice, measuring a quantum link consumes photons and time, and entanglement attempts must be scheduled ahead across links whose stored qubits are decaying. We will sweep the forecast horizon relative to memory coherence time. We expect the proposal's advantage to vanish at zero horizon and to grow as the horizon approaches the coherence time. That boundary is a result worth reporting either way.
Why not just train one policy on a diverse mix of regimes offline? SPARQ found that training on dynamic graphs beats a static snapshot by 15% in fidelity. So domain randomization clearly helps within the training distribution. It cannot cover regimes that first appear after deployment, or slow aging. Our generator deliberately holds out regimes to test exactly that case.
Proposed Experiments
The decisive experiments run on the simulated stream at three scales: a small fiber-plus-platform chain, a SPARQ-sized space-air-ground network, and a 2D mesh in the style of Pant et al.
Forecasting is scored by mean squared error in log-transmissivity, split into three periods: steady state, the first window after a known regime recurs, and the first window after a novel regime appears. Networking is scored by average delivered fidelity and resolved teleportation requests, with memory coherence time swept over an order of magnitude. Cost is scored by per-update latency and memory on a single embedded-class CPU.
The main result we expect is a clear gap over forgetting-factor least squares and over fine-tuned GPM in the recurrence window. We expect parity in steady state, and a gap in delivered fidelity that grows with forecast horizon.
Two outcomes would falsify the idea. The first is that the physics features explain little within-regime variance. The second is that the recurrence-window advantage disappears once the forgetting factor is tuned. A third check guards regime creation: on streams with no novel regimes, the residual test should create almost no spurious regimes, and on streams with many regimes the bank's memory should grow only with the number of distinct regimes, not with time.
References
- ACIL: Analytic Class-Incremental Learning with Absolute Memorization and Privacy Protection
- GACL: Exemplar-Free Generalized Analytic Continual Learning
- Continual Quantum Architecture Search with Tensor-Train Encoding: Theory and Applications to Signal Processing
- Quantum Architecture Search via Continual Reinforcement Learning
- SPARQ: Efficient Entanglement Distribution and Routing in Space–Air–Ground Quantum Networks
- Efficient Entanglement Routing for Satellite-Aerial-Terrestrial Quantum Networks
- Resource Allocation in Quantum Key Distribution (QKD) for Space-Air-Ground Integrated Networks
- Quantum-Secured Space-Air-Ground Integrated Networks: Concept, Framework, and Case Study
- Routing entanglement in the quantum internet
- Gradient Projection Memory for Continual Learning
- DOLFIN: Balancing Stability and Plasticity in Federated Continual Learning