Several recent papers propose that small quantum circuits could make better scheduling decisions for satellite networks than classical neural networks. The decisions in question include which server handles a user's computation, which satellite takes which task, and how bandwidth is split. These papers compare quantum and classical agents by counting environment samples or model parameters. That comparison leaves out what a quantum circuit actually costs to run. A variational quantum circuit (a short sequence of tunable quantum gates whose settings are learned, like weights in a network) does not return a number when you evaluate it. It returns one random measurement outcome, so estimating each Q-value (the agent's predicted payoff for an action) takes hundreds to thousands of repeated runs, called shots. Each shot is also corrupted by the hardware noise typical of today's noisy intermediate-scale quantum (NISQ) processors. A satellite network, meanwhile, gives each decision a hard time limit: the scheduling epoch, the few minutes of a ground contact pass, the latency deadline of a user's request.
This note proposes to re-measure the claimed quantum advantage in orbital edge scheduling using shots and wall-clock time as the currency, under realistic noise. It also proposes one constructive fix: an action-selection rule that keeps sampling a circuit only until the best action is statistically separated from the runner-up, and then stops. The expected outcome is a map of where, if anywhere, a quantum decision agent fits inside an orbital decision deadline and still beats a classical network of equal size. The fix should widen that region.
Core Claim
The published quantum-for-satellite results are measured in the wrong units. Edge Intelligence in Satellite-Terrestrial Networks with Hybrid Quantum Computing pairs a 16-qubit circuit with a classical network inside a double deep Q-learning agent to choose offloading servers across three LEO satellites. It reports better sample efficiency than classical double deep Q-learning, but in noiseless simulation, and it names hardware noise as an open requirement. Quantum Reinforcement Learning for Coordinated Satellite Systems reads out action probabilities over a $2^{16}$-sized action space from circuit measurements and reports the highest normalized reward against multi-agent RL, independent Q-learning, deep Q-learning and Monte Carlo baselines. It likewise leaves qubit limits and physical noise unmodeled. Neither paper counts shots.
We claim two things. First, once each action decision is charged its shot cost and noise is modeled, these advantages will shrink or reverse at typical orbital decision deadlines. Second, most of the shot cost is wasted on decisions where the best action is obvious, so an adaptive stopping rule can recover much of the lost ground.
The stopping rule rests on a standard concentration bound. To tell apart two actions whose true Q-values differ by a gap $\Delta$, with per-shot outcome variance $\sigma^2$ and error probability at most $\delta$, a Gaussian or Hoeffding argument needs roughly
$$ n \approx \frac{2\sigma^2 \log(1/\delta)}{\Delta^2} $$
shots. The decision then has to fit inside the epoch, which requires $n\, t_{\text{shot}} + t_{\text{overhead}} \le \tau$. Here $t_{\text{shot}}$ is the time per shot, $t_{\text{overhead}}$ covers queueing, compilation and link round trip, and $\tau$ is the decision deadline. Because $n$ grows as $1/\Delta^2$, a fixed shot budget is too large for easy decisions and too small for close ones. Hardware noise makes this worse because it shrinks measured gaps.
What is new relative to prior work is threefold. We evaluate these agents under a deadline-and-shot accounting that none of the cited quantum-satellite papers uses. We add stronger classical baselines. We apply sequential, per-decision shot allocation at inference time, which differs from shot allocation during training.
Research Program
The work draws on three roles of prior results. The satellite papers above motivate the study, since they supply the claims under test and admit they were never checked against hardware noise.
Variational Quantum Circuits for Deep Reinforcement Learning constrains the design. On a Frozen-Lake grid world, its quantum Q-learning agent used 28 parameters where tabular Q-learning used 64, and runs on IBM's ibmq-valencia device closely tracked noiseless simulation. That suggests noise is survivable when the classical optimizer only needs the ranking of Q-values to be preserved. We therefore expect the binding constraint to be shot count rather than noise alone.
An Adaptive Optimizer for Measurement-Frugal Variational Algorithms supplies the method family we build on. Its iCANS optimizer adapts shots per parameter during gradient training and outperformed Adam and SPSA per shot on chemistry-style and circuit-compiling tasks. We keep iCANS for training, but the decision-time problem is different: choosing one action per step, repeatedly, against a clock. That calls for a best-arm identification rule, which samples candidate actions until a confidence bound separates the leader from the rest.
The program has three pieces. The first is rebuilding the two environments from their published specifications. The offloading environment has four user devices, three LEO satellites at 600 km, millimeter-wave access links, inter-satellite backhaul, energy harvesting, and bandwidth scenarios from 50 to 110 MHz. The coordinated scheduling environment has a factored action space up to $2^{16}$. We rebuild them from the papers' descriptions because we have no evidence the original code is available.
The second piece is an evaluation harness. It simulates quantum circuits with a device-style noise model: relaxation and dephasing times, and one- and two-qubit gate errors, swept over a range rather than fixed to any single device. It charges every decision its shots, times each shot, and adds a round-trip overhead. That overhead matters because any plausible NISQ processor for this use sits on the ground behind the satellite-terrestrial link, not in orbit, given cryogenic and power limits.
The third piece is the adaptive action selector and its analysis. We will measure the distribution of Q-value gaps these agents actually face, which tells us how much adaptive stopping can save before we build anything elaborate.
Baselines
The naive reference points come first. One is a greedy or random server-selection heuristic, which fixes the floor of the reward scale. The other is the quantum agents exactly as published: noiseless, with implicit infinite shots. That second reference is the upper bound the prior papers report against, and every other number is read as a fraction of it.
The strongest existing quantum methods are the hybrid quantum double deep Q-learning agent of Edge Intelligence in Satellite-Terrestrial Networks with Hybrid Quantum Computing and the basis-measurement agent of Quantum Reinforcement Learning for Coordinated Satellite Systems. Each is run with a fixed shot count per decision, swept over several values, under the noise sweep. No reported numerical result for reward, convergence or energy was found on Lacuna for either environment. Both papers report only qualitative superiority ("higher rewards with fewer data points", "highest normalized reward"). Our reimplementation therefore has to re-establish their noiseless advantage before any shot-aware comparison is meaningful. If we cannot reproduce that advantage, we will report the failure as a result.
The strongest classical competitor is a classical double deep Q-network whose parameter count matches the quantum agent's trainable parameters. We also run a second classical network with a factored action head: one output per sub-decision rather than one per joint action. This is the standard way classical RL handles a $2^{16}$-sized action space, and it is the natural rebuttal to the "curse of dimensionality" claim.
All agents receive the same environment seeds, the same number of environment interactions, and the same hyperparameter search budget, measured in trials. Every classical agent is also charged its real inference time, so the comparison is on equal footing.
The key ablation isolates the stopping rule. We run the same trained quantum agent with fixed shots, with adaptive stopping, and with infinite shots.
The proposal succeeds if adaptive stopping reaches reward within a small margin of the infinite-shot agent while meeting the deadline on nearly all decisions, in a regime where fixed-shot agents miss deadlines or lose reward. The quantum agent counts as beating the classical baselines only if it exceeds both parameter-matched classical networks on reward at equal wall-clock time, including shots.
Why not just…?
Why not just use a small classical network? This is the most serious threat, and we expect it may win. A parameter-matched classical network gives an exact Q-value in microseconds with no shot noise. The only evidence in the file favoring the quantum side is parameter efficiency: 28 versus 64 parameters in Variational Quantum Circuits for Deep Reinforcement Learning. But a 64-parameter classical model is not a memory problem on any flight or ground computer. The equal-wall-clock comparison settles this directly. If the classical network dominates across the whole noise and deadline sweep, the finding is that quantum schedulers are not justified for these tasks, and we will say so plainly.
Why not just use enough shots, fixed at a large number? Under a loose deadline this works, and our sweep will show where. The $1/\Delta^2$ scaling means any fixed budget either wastes time on easy decisions or misjudges close ones. The fixed-versus-adaptive ablation, together with the measured gap distribution, shows how much a single well-tuned constant loses. If the gaps turn out to be nearly uniform, adaptivity buys little, and the fixed budget is the honest recommendation.
Why not just factor the classical action space? The $2^{16}$ action space in Quantum Reinforcement Learning for Coordinated Satellite Systems was compared against classical baselines that appear to treat actions jointly. A factored-head classical agent may dissolve the dimensionality problem without any quantum circuit. We include it as a first-class baseline. If it matches the quantum agent even in noiseless simulation, the published advantage came from weak baselines rather than from quantum representation.
Proposed Experiments
The first experiment reproduces both published agents noiselessly against the classical and factored baselines. The second repeats every comparison with shot charging and the noise sweep, across deadlines that range from a fraction of a second, for per-request offloading, to minutes, for pass-level scheduling. We report reward, the environment's energy objective, deadline-miss rate, total shots, and wall-clock time per decision.
The third experiment measures the empirical distribution of top-two Q-value gaps under noise and uses it to predict, from the displayed bound, how many shots adaptive stopping should save. It then checks whether the observed savings match that prediction. A match indicates the mechanism is understood; a mismatch suggests noise-induced bias rather than variance is the limiting factor.
Finally, for the smallest configuration only, we run the trained circuit's inference on real cloud quantum hardware, if access permits. The purpose is to test whether simulated shot times and noise-induced gap shrinkage match reality, as Variational Quantum Circuits for Deep Reinforcement Learning found for its grid world.
The decisive check is a single plot per environment: reward against wall-clock time per decision, with one curve per agent. It shows whether any region exists where the adaptive quantum agent lies above both classical curves. If no such region exists, the conclusion is that current quantum proposals for satellite scheduling do not survive realistic accounting.