The Entanglement Price of Coherent Dark Matter Sensor Networks

Ultralight dark matter, such as axions with masses near a nano-electronvolt, would behave less like particles and more like a weak classical wave that rotates every quantum sensor inside a region hundreds of meters to kilometers across in the same way. Eliminating Incoherent Noise shows how to exploit this. It applies a binary tree of two-sensor quantum gates, with depth $\log_2 N$, that funnels the tiny common rotation of $N$ sensors into the excited-state population of a single sensor before anything is measured. Noise is then paid once rather than $N$ times. In their simulations the signal power grows as $N^2$ against noise growing as $N$. That scheme assumes every sensor can interact with its neighbors through high-fidelity gates, which in practice means one cryostat. Networks that span separate cryostats or sites, which is where large arrays would actually live, have no such interconnect. The gates would have to run over entanglement shared across the sites, and nobody has asked how much entanglement that takes.

This note proposes to answer that question with the tools of asymptotic entanglement theory. The core claim is that the entanglement cost of coherent signal combining is far below the naive one entangled qubit pair per sensor per shot. The reason is that both the signal and the thermal noise live almost entirely in a small subspace: the states with at most one excitation among the sensors involved. If the combining gate is restricted to that subspace, and the sensor states from many shots are compressed within one coherence time of the dark matter field, the cost per sensor per shot should fall to roughly the entropy of a single sensor's state, which is small. We will bound this cost from below with semidefinite programs, match it from above with explicit protocols, and turn the result into a curve of sensitivity against entanglement budget. Experimentalists can then use that curve to decide whether a distributed array is worth building.

Core Claim

Model each sensor, after one integration window, as a qubit holding excitation probability $p = \sin^2\theta + p_{\rm th}$. Here $\theta$ is the small rotation induced by the dark matter field and $p_{\rm th}$ is the thermal excitation probability. One layer of the tree acts on two sensors held by different parties. The quantity we care about is the entanglement cost $E_C$ of simulating that two-sensor gate between separate sites with free local operations and classical communication: the minimum asymptotic rate of shared Bell pairs consumed per use. Our prediction is

$$ E_C^{\rm gate}(p) \;\lesssim\; h(p) + O(p), \qquad h(p) = -p\log_2 p - (1-p)\log_2(1-p). $$

The bound comes from compressing the incoming sensor state into its typical subspace and teleporting only that, as in Schumacher-style compression. It requires blocking $n$ shots together. This is physically allowed because the axion field stays phase-coherent for roughly $10^6$ oscillation periods, so many shots share the same $\theta$ and are close to identical copies. For ABRACADABRA-scale frequencies of 75 kHz to 2 MHz (ABRACADABRA-10 cm design), that leaves plenty of shots per coherence time. When $p \ll 1$, $h(p)$ is far below one, which is the cost of naive per-shot teleportation. The open questions are how tight this is and where thermal noise erases the savings.

Research Program

The work has three parts.

The first part is lower bounds. Computable and Faithful Lower Bound on Entanglement Cost introduces a hierarchy of positive-partial-transpose (PPT) state sets and an SDP-computable quantity, the logarithmic fidelity of $k$-negativity. That quantity lower-bounds entanglement cost and was extended to bipartite channels. Partial transposition is a matrix test for entanglement, and "PPT" states pass it. This paper supplies our method. We apply it to the noisy two-sensor combining channel, including the $T_1$ and $T_2$ noise and $10^{-4}$ gate error used by Shu, Xu and Xu, as a function of $p$, $p_{\rm th}$ and the number of shots blocked. Its channel results were demonstrated on generic examples such as the Werner-Holevo channel, never on a sensing channel driven by a weak external field.

The second part is achievable protocols that are restricted to a subspace. Ideas from subspace-constrained state synthesis enter here as the design principle. We build combining gates that act nontrivially only on the span of $|00\rangle, |01\rangle, |10\rangle$ and teleport only the typical low-excitation subspace of blocked shots. Their consumed-ebit rate is then compared to the SDP bound. Irreversibility under PPT operations constrains this comparison: cost and distillable entanglement genuinely differ for mixed states even with generous free operations. A gap between our lower and upper bounds may therefore be real rather than an artifact of a loose bound. Reversible entanglement beyond quantum operations shows that the gap closes to the logarithmic negativity only under maps that cannot be physically implemented. We use its logarithmic negativity value as a reference line, never as an achievable target.

The third part is a bound-entanglement diagnostic. Thermal noise could push the intermediate two-node states of a partially executed tree into the PPT-but-entangled region. That region is studied in Robustness of PPT entangled states and Non-full-rank bound entangled states, which construct such states from unextendible product bases and show that they persist under small perturbations. In that region, correlations generated at one tree level cannot be redistilled to repair a later level. We map the boundary in the $(p_{\rm th}, T_2)$ plane, which gives a noise budget beyond which distributed trees cannot be patched with entanglement purification. Not All Entangled States are Useful for AAQPT finds that bound entangled states make poor probes for full tomography of an unknown process. That result supports treating this region as one to avoid, not exploit.

The deliverable is a sensitivity-versus-ebit-rate frontier for networks of $N = 2$ to $32$ sensors, split across $2$ to $N$ sites.

Baselines

The naive reference point is independent local readout. Each sensor is measured on its own and the populations are summed classically, giving the standard-quantum-limit scaling $N^{-1/2}$ reported by Shu, Xu and Xu. It is run under the same noise model, shot budget and integration time. Any distributed protocol must beat it at a stated ebit rate, or it is pointless.

The strongest existing method is the co-located tree of Shu, Xu and Xu, with ideal local gates at $10^{-4}$ infidelity. They report sensitivity scaling of $N^{-3/2}$ against $N^{-1/2}$, validated in simulations of up to 32 sensors. We treat it as the ceiling. A distributed protocol "beats the question" if it recovers at least 90% of the co-located signal-to-noise ratio at an ebit rate below $0.1$ per sensor per shot. We reproduce their simulation setup so that the comparison uses identical noise parameters.

The naive distributed baseline is full gate teleportation: one Bell pair consumed per qubit per gate, with no subspace restriction or blocking. It achieves the co-located performance at maximal cost.

The key ablation uses the same blocked compression without the restriction to the low-excitation subspace, and separately the restriction without blocking. This isolates which ingredient produces the savings.

We also report the classical post-processing reference from TIDMAD. There, a fully connected network reaches a denoising score of 6.43 against a no-denoising score of 1.0 on ABRACADABRA time series. We treat it as complementary, applied after readout in every arm. No reported result was found on Lacuna for the entanglement cost of any sensing protocol, so the cost axis has no prior number.

Why not just…?

Why not just measure each sensor locally and combine with AI denoising? Denoising acts after readout. By then, each of the $N$ readouts has already added its own thermal and amplifier noise, so it cannot recover the pre-measurement $N^2$-versus-$N$ advantage. The TIDMAD authors themselves note that their limits did not surpass earlier ABRACADABRA runs, partly because of hardware coupling. Our local-readout-plus-denoising arm quantifies this directly. If it matches the distributed quantum arm at realistic $p_{\rm th}$, our conclusion is that distributed quantum links are not worth building. That is a useful negative result.

Why not just put all sensors in one cryostat? When that is possible, it should be done, and our work does not apply. The question matters only when array size or engineering forces several sites, while the axion coherence length at nano-electronvolt masses still covers them. We state this scope explicitly rather than claim generality.

Why not just teleport every sensor state? It works and is our naive distributed baseline. The proposal stands or falls on whether the subspace-restricted protocol is substantially cheaper. If thermal population makes $h(p)$ close to one, the savings vanish, and the frontier will show it.

Why not use bound entangled or subspace-engineered initial states as noise-robust sensors? The signal is a common classical field acting on unentangled sensors. Shu, Xu and Xu get their advantage without initial entanglement, and Lu et al. suggest that bound entangled probes are poor at characterizing unknown processes. For these reasons we use subspace structure in the readout and in the noise diagnostic, not in the probe state.

Decisive Checks

The first check is whether the SDP lower bound on the two-sensor channel's cost, evaluated at $p_{\rm th}$ values set by ABRACADABRA-band temperatures, falls well below one ebit. If it does not, the core claim fails, whatever protocols we build. The second check is whether our explicit protocol comes within a constant factor of that bound across $N$ up to 32 in Qiskit-style noisy simulation. The third check is whether the PPT-entangled region appears inside realistic noise parameters, which would cap the number of tree levels that can span sites. Together these settle whether coherent dark matter sensing can be distributed, and at what entanglement price.