Which Multipartite Entanglement Measure Predicts the Sensing Gain of a Noisy Dark-Matter Sensor Array?

Searches for wave-like dark matter, such as ultralight axions that behave like a classical field oscillating at a fixed frequency, increasingly use arrays of quantum sensors such as superconducting qubits or microwave cavities. The field pushes all the sensors in step, while thermal and dephasing noise hits each one independently. Eliminating Incoherent Noise uses this asymmetry. Its circuit, a tree of two-qubit gates, gathers the shared signal into one sensor so that signal power grows as $N^2$ while noise grows as $N$. Sensitivity then improves as $N^{-3/2}$ instead of the $N^{-1/2}$ of independent sensors summed classically, and this holds in simulation up to $N=32$ sensors with gate errors near $10^{-4}$. The authors suggest that pre-sharing entangled states could push further, toward the Heisenberg limit of $1/N$. One example is a GHZ state, an $N$-qubit superposition of "all zero" and "all one" that responds $N$ times faster to a common phase. Real arrays, however, never hold a clean GHZ state. They hold a decohered mixture, and nobody has said how much multipartite entanglement must survive for the entangled array to beat the unentangled coherent protocol.

This proposal asks one question. As noise grows, the entangled array's advantage over the coherent unentangled protocol of Shu et al. vanishes at some noise level. Which computable entanglement quantity, if any, crosses its own threshold at that same point? The candidates are genuine multipartite entanglement (GME, where all $N$ parties are jointly entangled and no split into separable groups explains the state), the three-tangle, and plain bipartite entanglement across cuts. They disagree sharply on the same noisy GHZ family, so the question has a sharp empirical answer. If one of them tracks the crossover, it becomes a cheap check, run from a few measurements before each data-taking run, of whether the array is in an advantageous regime.

Where the two directions actually meet

The multipartite-entanglement literature in the file offers a range of tractable tools, all developed with no sensing task in mind.

  • Exact three-tangle. Efficient numerical method to calculate the three-tangle of mixed states computes the three-tangle of mixed states with replica-exchange Monte Carlo, meaning many copies of a stochastic search run at different "temperatures" that swap configurations to escape local minima. Its results match analytical values to within $10^{-7}$, and it reports that for a GHZ state mixed with white noise the three-tangle stays exactly zero until the GHZ weight reaches about $p\approx0.7$.
  • CHSH-subspace GME test. Characterizing multipartite entanglement by violation of CHSH inequalities writes concurrence as a sum of two-qubit Bell-inequality violations over subspaces. This gives a GME test, $X+Y+Z>8$, that works down to purity about 0.56 in local dimension 4, where earlier methods needed 0.82.
  • Hierarchical concurrence bound. A lower bound of concurrence for multipartite quantum systems bounds four-party concurrence through its three- and two-party pieces. It detects entanglement of the noisy four-qubit state whenever $t>1/9$.

These are the measuring instruments. Their thresholds on the same state family differ by a factor of several. That disagreement is the reason "which one predicts sensing gain" is a real question rather than a formality.

The indirect-detection literature plays a smaller and more honest role here. Its signal-to-noise picture says that independent backgrounds can only be beaten by $\sqrt{N}$ averaging. Combined searches for dark matter in dwarf spheroidal galaxies stacks 354.3 hours over four galaxies to reach about $10^{-24}\,\mathrm{cm^3/s}$ at 1 TeV, and remains limited by astrophysical systematics. The one mechanism in the file that breaks the $\sqrt{N}$ barrier, exploiting a signal that is coherent across detectors against noise that is not, lives in laboratory sensor arrays rather than in astrophysical telescopes. Cherenkov-telescope data carry no quantum state to entangle, so we do not pretend to apply entanglement measures to MAGIC or LHAASO photons. Indirect detection supplies the signal-to-noise accounting and the classical reference point. The quantum work happens in the wave-dark-matter sensor setting that Eliminating Incoherent Noise established.

Core Claim

A simple scaling argument shows why the answer is not obvious. Under global white noise, a GHZ state with weight $p$ reads out a common phase $\phi$ with contrast $p$, and its phase is amplified to $N\phi$. Its per-shot sensitivity is then roughly $1/(pN)$, compared with $1/\sqrt{N}$ for unentangled sensors, so its edge over independent sensors survives down to $p^\ast\sim N^{-1/2}$. The relevant baseline, though, is Shu et al.'s coherent protocol, which already gains beyond $\sqrt{N}$ through gate-based concentration. The true crossover therefore depends on how the tree circuit's own error budget compares with GHZ decoherence.

The entanglement thresholds behave differently. The three-tangle threshold sits near 0.7 for three qubits. GME witness thresholds depend on the witness used. Under local dephasing, which is more physical than global white noise, GHZ contrast decays as $e^{-N\gamma t}$ while the coherent product-state protocol degrades far more gently.

We hypothesize that no single-copy GME witness predicts the crossover under local dephasing. Instead, the quantity that tracks it is the strength of the full $N$-body correlation, which the CHSH-subspace overlaps and the three-tangle capture better than biseparability tests. Confirming or refuting this, across noise models and array sizes, is the contribution.

Research Program

Step 1: simulator. We build a noisy-circuit simulator in Qiskit that reproduces the setting of Eliminating Incoherent Noise. It includes $T_1$ relaxation, $T_2$ dephasing, and two-qubit gate errors swept around $10^{-4}$. A weak common phase rotation stands in for the dark-matter field. On this simulator we compare three schemes at matched total sensing time and matched gate counts:

  • independent sensors read out separately;
  • the tree-circuit coherent protocol;
  • a GHZ-initialized array read out by parity.

The figure of merit is the minimal detectable phase per unit time.

Step 2: entanglement quantities on the actual states. At each noise setting we compute the entanglement quantities on the actual pre-signal state of the array:

  • the three-tangle by replica-exchange Monte Carlo, for $N=3$;
  • the CHSH-subspace GME criterion and its tight lower bound;
  • the hierarchical concurrence lower bound, for $N=4$;
  • bipartite PPT negativity across all cuts.

For larger $N$, where convex-roof quantities become intractable, we restrict to the CHSH-subspace and concurrence bounds. These need only correlation measurements, not full tomography.

Step 3: matching thresholds to the crossover. We locate, for each noise model, the noise level where the GHZ array stops beating the coherent protocol. We then ask which quantity's threshold, or which monotone value, best matches it across $N=3$ to $32$.

Step 4: multi-copy activation. Activation of genuine multipartite entanglement shows that two copies of an isotropic GHZ state can be genuinely multipartite entangled even when one copy is not, and that this holds even from states with positive partial transpose across every cut. Such states are undistillable, so the effect is distinct from ordinary entanglement purification. We test whether two such copies, combined by the local Hadamard-product map from that paper, give any sensing gain over one copy used twice. A null result would be informative: it would show that activated GME is not a sensing resource and would sharpen what "resource" means for this application.

Baselines

Naive reference: independent sensors. The do-nothing reference is $N$ independent sensors read separately and summed classically, the laboratory analogue of the likelihood stacking in Combined searches for dark matter in dwarf spheroidal galaxies. Its sensitivity scales as $N^{-1/2}$. It runs on the same simulator, with the same noise parameters and the same total time.

Strongest competitor: the coherent tree-circuit protocol. The strongest existing method is the protocol of Eliminating Incoherent Noise. It is the right competitor because it already beats the standard quantum limit without any pre-shared entanglement, so any entangled scheme must clear its bar rather than the $\sqrt{N}$ bar. The paper reports sensitivity scaling as $N^{-3/2}$, signal power as $N^2$, and noise as $N$ up to $N=32$ at gate infidelity around $10^{-4}$. No reported absolute sensitivity value was found on Lacuna for this protocol, so we will reproduce its scaling first as a validation gate.

Fairness and what counts as beating it. Fairness means identical $T_1$, $T_2$, and gate error per two-qubit gate. It also means charging the GHZ preparation circuit's gates to the entangled scheme, and an equal shot budget. The entangled array "beats" the coherent protocol only where its sensitivity is lower at matched resources. The entanglement-quantity study "succeeds" if one quantity's threshold predicts the crossover noise level within the sweep resolution across all $N$ tested and both noise models.

Ablation. The ablation that isolates the key ingredient, entanglement as opposed to state preparation overhead or readout choice, swaps the GHZ state for a product state run through the same preparation-depth idle noise and the same parity readout.

Entanglement-quantity baselines. Here the baseline is bipartite negativity across cuts. Each multipartite quantity must out-predict it. We anchor these quantities against the reported benchmarks:

Our implementations must reproduce these numbers before being trusted.

Why not just…?

Why not just stack independent sensors and accept $\sqrt{N}$? Classical stacking is the reference, and it loses by construction whenever the signal is coherent across sensors. Eliminating Incoherent Noise already shows an $N^{-3/2}$ protocol beating it. The interesting comparison is therefore not with stacking but with that coherent protocol. Stacking stays in the study only as the floor.

Why not just use bipartite entanglement across cuts? Bipartite tests are cheap, but they cannot tell an array where all $N$ sensors share one correlation from one where only pairs do. Activation of genuine multipartite entanglement shows that states can look biseparable or PPT on every cut and still carry multipartite structure. Sensing gain from GHZ states comes from the full $N$-body correlation, so we expect bipartite negativity to mispredict the crossover. Including it as a predictor tests this directly. If it predicts as well as the multipartite quantities, that is a publishable negative result about the necessity of multipartite measures here.

Why not skip entanglement measures and calibrate with an injected test tone? This is the serious threat. Injecting a known signal measures the actual gain directly, with no theory. Our answer is partial. A tone calibration tells you the gain at one operating point but not why it falls off, nor how it extrapolates to larger $N$ or new noise conditions. A predictive entanglement quantity does both, and it can be estimated from correlation measurements already needed for state characterization. If the experiments show that tone calibration costs fewer shots than the best entanglement estimate for equal predictive accuracy, we will say the practical value is nil and the contribution is explanatory. We measure this shot-cost comparison explicitly.

Why not use plain simulated annealing for the three-tangle? Efficient numerical method to calculate the three-tangle of mixed states reports that constrained simulated annealing gets trapped on high-rank noisy GHZ states, while replica exchange reaches $10^{-7}$ agreement. We therefore adopt replica exchange and use annealing only as a sanity check.

Decisive Checks

Check 1: matching thresholds. For $N=3$ and $N=4$ under both global white noise and local dephasing, we plot the GHZ-versus-coherent-protocol crossover against each entanglement quantity's threshold on the same states. If the three-tangle's $p\approx0.7$ threshold or the CHSH-subspace GME threshold sits within one sweep step of the crossover under both noise models, while bipartite negativity does not, the hypothesis that full-body correlation strength is the operative resource is supported. If no quantity tracks the crossover under local dephasing, the hypothesis fails. In that case we report that entanglement certificates are the wrong pre-run check for these arrays.

Check 2: scaling in $N$. Extending to $N=8$ to $32$ with the measurement-only bounds tests whether the predictor's agreement survives growth in $N$. The scaling argument predicts that the white-noise crossover drifts as $N^{-1/2}$, while witness thresholds need not.

Check 3: two-copy activation. The two-copy activation experiment either finds sensing gain from activated GME or finds none. Either outcome directly answers whether the single-copy boundary is the one that matters for signal-to-noise.