A reconfigurable intelligent surface is a flat panel of tunable elements that redirects a wireless or optical beam around an obstacle. In classical radio, surfaces with built-in amplifiers ("active" surfaces) beat purely reflecting ("passive") ones by tens of decibels, because they lift the signal before the second hop's loss and the receiver's own noise can bury it. Quantum links that use surfaces to route entangled light to blocked users have so far used only passive surfaces, on the intuition that amplifiers wreck fragile quantum states. This note asks when that intuition is right. We model the active surface element as a quantum amplifier, a device that must add a minimum amount of noise, and we measure what it does to the classical post-processing cost of recovering the quantities a quantum user actually needs, such as entanglement fidelity or state purity.
The claim is a sharp, testable rule rather than a yes-or-no verdict. An active element placed mid-link lowers the recovery cost only when the noise downstream of the surface, referred back to the surface, is larger than the amplifier's own added noise. That downstream noise is the second-hop loss plus detector and background counts. The rule predicts a crossover that existing passive-only designs cannot see. It also says how much thermal noise, a direct consequence of the power and heat overhead of active hardware, the amplifier can tolerate before it stops paying for itself.
Core Claim
Two results frame the question. On the wireless side, Active RIS vs. Passive RIS: Which Will Prevail in 6G? built a 2.36 GHz element with more than 25 dB of reflection gain. It modeled both the noise the amplifier injects and amplifies ("dynamic noise") and the noise of the surrounding circuitry ("static noise"), and reported roughly 40 dB higher SNR than a passive surface at 256 elements. On the quantum side, Information recoverability of noisy quantum states and Retrieving Nonlinear Features from Noisy Quantum States show that nonlinear features of a state can be recovered from a noisy copy if and only if the noise channel is invertible as a linear map. They also show that the recovery cost, the multiplicative blow-up in the number of measurement shots, is the quantity that matters in practice. The one paper that puts a surface inside a quantum link, Reconfigurable Intelligent Surface (RIS)-Assisted Entanglement Distribution in FSO Quantum Networks, uses a passive surface and a single-path link model. It reports that weather can cut network rate by 45% and that rate-maximizing placement can cost 83% of fidelity.
The research file frames the open question as whether active noise makes the composite channel singular, meaning non-invertible. For the continuous-variable model we analyze first, that framing is too coarse. A chain of loss and phase-insensitive amplification with nonzero end-to-end transmission is an invertible map. What changes is the cost of inverting it. The question to answer is therefore when the inversion cost falls rather than rises.
The mechanism is easiest to see in one quadrature, one of the two field components that a homodyne detector measures. Let the first hop have transmissivity $\eta_1$ and the second hop $\eta_2$. The surface amplifier has gain $G$ and thermal excess $n_{ex}$, measured in photons. The detector adds $n_d$ photons of noise. Variances are written in vacuum (shot-noise) units. Referring everything back to the input, the variance the receiver effectively sees is
$$\chi(G) = V_{in} + \frac{1-\eta_1}{\eta_1} + \frac{(G-1)(1+2n_{ex})}{\eta_1 G} + \frac{(1-\eta_2) + 2n_d}{\eta_1 \eta_2 G},$$
where $V_{in}$ is the input variance. The third term is the amplifier's own cost. It is bounded above by $(1+2n_{ex})/\eta_1$, a bound that follows from the minimum noise any phase-insensitive amplifier must add. The fourth term is everything after the surface, and it is divided by $G$. This is the quantum analogue of the classical rule that amplifiers belong as early in a receive chain as possible.
Comparing $G \to \infty$ with $G = 1$ gives the crossover. Amplification helps exactly when
$$1 + 2n_{ex} < \frac{(1-\eta_2) + 2n_d}{\eta_2}.$$
Because recovery cost grows with $\chi$, the same inequality should mark where active surfaces reduce the sampling overhead of retrieving fidelity and purity. Verifying that, or finding where it fails, is the project. Two places where it could fail are non-Gaussian, single-photon encodings and multi-mode receivers.
Research Program
The first component is a channel model for the active quantum surface. It reuses the dynamic-noise and static-noise decomposition of Zhang et al. and maps it onto quantum quantities. Dynamic noise becomes the amplifier's added noise with thermal excess $n_{ex}$. Static noise becomes background and detector counts $n_d$.
This mapping is where the overhead of active components becomes physics rather than an engineering footnote. Zhang et al. (2023, Active RISs: Signal Modeling, Asymptotic Analysis, and Beamforming Design) and Basar (2023, Reconfigurable Intelligent Surfaces for 6G: Emerging Hardware Architectures, Applications, and Open Challenges) identify heat dissipation as the barrier to dense active arrays. In our model, heat raises $n_{ex}$, which shrinks the region where amplification helps. The program therefore reports a maximum tolerable excess noise as a function of link geometry. That gives hardware designers a concrete thermal target instead of a qualitative warning.
The second component computes recovery cost. For continuous-variable entanglement, the state is a two-mode squeezed vacuum (a standard source of entangled light) truncated to a few photons per mode. We compute the inverse-map cost exactly with the semidefinite program, a standard convex optimization, used in the observable shift method. Retrieving Nonlinear Features from Noisy Quantum States notes that this program becomes intractable beyond about five qubits, so we keep truncations small and pair them with analytic bounds from the Gaussian formula above.
For the single-photon polarization encoding that Chehimi et al. use, an amplifier does not act cleanly on a qubit. We model its effect inside the one-photon subspace as extra multi-photon contamination after post-selection. This regime is where the clean crossover rule is most likely to break, and where true near-singularity could appear, for example from detector saturation. We treat it as the main risk to the claim.
The third component relaxes the single-antenna assumption. Every cross-direction paper in the file, from Chehimi et al.'s single-path quantum link to the multi-user, single-antenna-user analyses of Active RIS vs. Passive RIS and Beyond Diagonal RIS: Passive Maximum Ratio Transmission and Interference Nulling Enabler, treats each user as a single receiving mode. With a two-aperture or two-spatial-mode receiver, one amplifying element feeds both output modes. Its added noise is then correlated across modes. Per-mode analysis can then mislead in either direction, overstating the cost if the receiver can cancel correlated noise or understating it if it cannot. We repeat the crossover analysis with a two-mode receiver and test whether the single-mode rule still predicts the multi-mode outcome.
Adsorption-coupled reduction plays only a narrow role, and we say so plainly. Its core idea is to concentrate a dilute contaminant at a surface before an irreversible conversion step. The materials work in Two-Step Hydrothermal Reaction Enhances Removal of Cr(VI)... and New Functions of Polyaniline shows how concentration ahead of conversion beats dilute-phase reaction, which is the same structure as amplifying before a noisy detector. The file's Gap 3 suggests using reversible redox materials as low-power surface tuning, but the file contains no radio-frequency or optical measurements of such materials. We therefore do not build on them. Their only appearance is as a hypothetical zero-bias-power phase-tuning entry in the power accounting, clearly labeled as unmeasured.
Baselines
The naive reference point is a passive surface with no mitigation, evaluated in the free-space optical simulation of Chehimi et al.: 1550 nm, atmospheric turbulence and pointing errors modeled with Meijer G-functions, and phase damping. We report raw fidelity and purity error at the receiver. For context, that paper reports that rate-maximizing placement can drop fidelity by 83%. Any proposed configuration must improve on the raw passive link at equal source rate and equal number of measurement shots.
The strongest existing method on this exact question is Chehimi et al.'s optimized passive surface. It places the surface in three dimensions and allocates per-user generation rates by simulated annealing, and it reports a 64% fairness improvement and a solution within 6% of exhaustive search. We rerun their placement optimizer on our scenarios and add observable-shift mitigation on top, so the passive baseline gets both its best geometry and the best known post-processing. For observable shift itself the file gives no reported scalar accuracy ("no reported result found on Lacuna for observable-shift Tr[ρ²] error"), so we measure it directly. Beating this baseline means a lower shot count to reach a target fidelity-estimation error, at equal surface area, with the active surface's placement re-optimized under the same annealing budget.
A learned mitigation baseline addresses the worry that a data-driven corrector could make hardware changes unnecessary. Beyond Circuit Connections: A Non-Message Passing Graph Transformer Approach for Quantum Error Mitigation, a graph-transformer corrector trained without ideal labels, reports a mean absolute error of 0.0246, against 0.0421 for a graph neural network and 0.1933 for zero-noise extrapolation, on 15-qubit Trotterized circuits. We adapt its self-supervised training to the link setting with the same number of training samples per configuration. Because learned post-processing cannot exceed the invertibility limit, the comparison tests whether practical learned correction already closes the gap that active hardware would close.
Two ablations isolate the key ingredient. The first sets $n_{ex}$ to zero, giving a quantum-limited amplifier with ideal cooling, and then sweeps it upward. This separates the benefit of gain from the cost of heat. The second moves the same amplifier from the surface to the receiver, which isolates the value of amplifying mid-link.
Why not just…?
Why not just make the passive surface bigger? Classically, Zhang et al. show that a passive surface needs roughly 2.5 million elements to match an active one. That figure comes from radio-frequency multiplicative fading, though, and does not transfer automatically to optics, where apertures are cheap in element count but limited by pointing error. The experiment answers this directly. We sweep passive aperture against active gain at matched total area and report the passive size needed to match each active configuration's recovery cost. If modest passive enlargement matches the active results across most of the weather and geometry range, the active proposal is confined to a narrow regime, and we will report that.
Why not just amplify at the receiver? This is the most serious alternative. A receiver-side amplifier avoids putting power and heat on a remotely mounted surface. The formula above predicts that a receiver amplifier suppresses detector noise $n_d$ but not second-hop loss $1-\eta_2$, because the signal has already lost that photon flux. The surface position should therefore win only when the second hop is lossy, as with long or turbulent paths, and tie when it is clean. The relocation ablation tests this, and we expect a substantial region where the receiver wins. Mapping that region is part of the result, not a threat to hide.
Why not just fix it in post-processing? Information recoverability results show that post-processing can only invert the channel, at a cost that grows with $\chi$, and can never restore what the data processing inequality removes. Learned correctors such as the graph-transformer method still have to pay this inverse cost through more samples or more training data. They also need retraining as noise drifts, a limitation the graph-transformer paper itself notes. The learned baseline above measures how much of the active surface's benefit software alone recovers at equal shot budgets.
Proposed Experiments
The first experiment validates the crossover in the Gaussian regime. Over a grid of $(\eta_1, \eta_2, n_{ex}, n_d)$ values drawn from Chehimi et al.'s turbulence and weather settings, we compute the exact inverse-map cost on truncated two-mode squeezed states. We then check whether the sign change in recovery cost matches the analytic inequality. The decisive outcome is agreement within truncation error. A systematic shift would mean recovery cost does not track $\chi$ monotonically, which would itself be a reportable finding.
The second experiment is the single-photon stress test. We repeat the comparison with polarization-entangled pairs under the contamination model and look for configurations where gain makes the effective qubit channel nearly singular, so that its inverse cost diverges. This is the regime where the "amplifiers destroy quantum states" intuition would be vindicated.
The third experiment is the multi-mode test. With two receiver modes and a shared surface amplifier, we test whether the single-mode crossover still predicts the outcome, with and without a receiver that jointly processes both modes.
The fourth experiment is the overhead ledger. For every configuration that beats the strongest passive baseline, we report the required gain, the maximum tolerable $n_{ex}$, and the implied cooling and power requirement, using Zhang et al.'s amplifier model. The result is a curve of recovery-cost reduction per watt of surface power. The proposal succeeds if there is a nonempty, physically plausible region where an active surface beats both the optimized passive surface with mitigation and the receiver-side amplifier, and if the simple inequality predicts where that region lies.
References
- Active RIS vs. Passive RIS: Which Will Prevail in 6G?
- Information recoverability of noisy quantum states
- Retrieving Nonlinear Features from Noisy Quantum States
- Reconfigurable Intelligent Surface (RIS)-Assisted Entanglement Distribution in FSO Quantum Networks
- Active RISs: Signal Modeling, Asymptotic Analysis, and Beamforming Design
- Active RISs: Signal Modeling, Asymptotic Analysis, and Beamforming Design
- Reconfigurable Intelligent Surfaces for 6G: Emerging Hardware Architectures, Applications, and Open Challenges
- Reconfigurable Intelligent Surfaces for 6G: Emerging Hardware Architectures, Applications, and Open Challenges
- Beyond Diagonal RIS: Passive Maximum Ratio Transmission and Interference Nulling Enabler
- Two-Step Hydrothermal Reaction Enhances Removal of Cr(VI) from Wastewater Using Nitrogen-Doped Starch-Based Hydrothermal Carbon
- New Functions of Polyaniline
- Beyond Circuit Connections: A Non-Message Passing Graph Transformer Approach for Quantum Error Mitigation