Certification Time for Weak-Force-Mediated Entanglement between Levitated Particles

The paper in detail

Two identical particles move along one transverse direction, with dimensionless quadratures obeying \([\hat q_s,\hat p_{s'}]=2i\delta_{ss'}\) and an interaction \(\hat V=\tfrac{\hbar g}{2}(\hat q_{\rm A}-\hat q_{\rm B})^2\) obtained by expanding a central potential. For Gaussian states the covariance matrix \(\mathbb V\) carries everything, and the state is entangled exactly when the smallest symplectic eigenvalue of the partially transposed covariance matrix is below one:

\[\tilde\nu_-=\sqrt{\tfrac12\Bigl[\tilde\Delta-\sqrt{\tilde\Delta^2-4\det\mathbb V}\Bigr]}<1,\qquad \tilde\Delta=\det\mathbb A+\det\mathbb B-2\det\mathbb C.\] Each run yields a joint noisy reading of both quadratures with outcome covariance \(\mathbb\Sigma=\mathbb V+\mathbb V_{\rm m}\), and the estimate \(\check{\mathbb V}\) averages \(N\) such readings. Propagating the fluctuations of \(\check{\mathbb V}\) through the gradient \(\mathbb G=\partial\tilde\nu_-/\partial\mathbb V\) gives \(\mathrm{Std}(\check{\tilde\nu}_-)\simeq\alpha/\sqrt N\) with \(\alpha^2=2\,\mathrm{tr}(\mathbb G\mathbb\Sigma\mathbb G\mathbb\Sigma)\). Requiring the gap \(1-\tilde\nu_-\) to exceed \(z\) standard deviations gives \(N_\star\). With \(\mathbb V_{\rm m}=\xi\mathbb I\), the bound \(\alpha\ge\tilde\nu_-+\xi\) holds for every state, so near separability even the best state needs \(N_\star\propto(1-\tilde\nu_-)^{-2}\) runs. Moving \(\tilde\nu_-\) from 0.999 to 0.990 cuts the cost a hundredfold.

In release–recapture, the traps are switched off, the particles evolve under their mutual force, and they are recaptured for readout. The common mode spreads freely while the differential mode oscillates at \(\omega_-=\sqrt{4g\Omega}\), and the mismatch creates correlations. Holding the interaction phase \(\theta=\omega_-t\) and the ratio \(r=\Gamma_{\rm d}/g\) of dark momentum diffusion to coupling fixed while \(\epsilon=\sqrt{4g/\Omega}\to0\), the witness tends to \(\tilde\nu_-\simeq rF(\theta)\), independent of \(\epsilon\), but the fluctuation coefficient grows as \(\alpha\simeq\xi H(\theta)/\epsilon\), with \(F(\pi)=\pi/\sqrt6\) and \(H(\pi)=\sqrt3/2\). Including the release time in each cycle,

\[T_\star(\theta)\simeq\frac{z^2\xi^2H(\theta)^2}{\epsilon^2\,[1-rF(\theta)]^2}\Bigl(\frac{\theta}{\epsilon\Omega}+t_{\rm rest}\Bigr).\] Entanglement appears only for \(g>g_{\rm ent}\simeq\pi\Gamma_{\rm d}/\sqrt6\) at \(\theta=\pi\). The release time that minimises \(T_\star\) lies beyond the most entangling one, because the drop in \(\alpha\) outweighs the smaller gap and longer cycle, which gives a severalfold saving at weak coupling. In continuous monitoring, \(T_\star=(z\beta/(1-\tilde\nu_-))^2\tau_{\rm c}\), where \(\tau_{\rm c}\) counts how long the record takes to supply one independent sample. There the certifiable coupling stays close to the boundary for generating steady-state entanglement at all, while release–recapture certifies at much weaker couplings. At low diffusion its limit is the time budget rather than the noise. Optimising the release time is left out of that comparison, and calibration uncertainty in \(\mathbb V_{\rm m}\) remains to be included.

Transferable methods

1. Delta-method error bars for a covariance witness

   The witness is a smooth function of an estimated covariance matrix, so its statistical error to leading order is the gradient contracted with the covariance of the estimator. Two identities for the gradient of \(\tilde\nu_-\) turn this into a closed formula and a state-independent lower bound.

   For zero-mean Gaussian samples with covariance \(\mathbb\Sigma\), the sample second moments satisfy \(N\,\mathrm{Cov}(\check{\mathbb V}_{ij},\check{\mathbb V}_{kl})=\mathbb\Sigma_{ik}\mathbb\Sigma_{jl}+\mathbb\Sigma_{il}\mathbb\Sigma_{jk}\) (Isserlis). Linearising \(\check W\simeq W+\mathrm{tr}[\mathbb G(\check{\mathbb V}-\mathbb V)]\) with symmetric \(\mathbb G\) gives \(N\,\mathrm{Var}(\check W)\simeq2\,\mathrm{tr}(\mathbb G\mathbb\Sigma\mathbb G\mathbb\Sigma)\). For \(W=\tilde\nu_-\), symplectic diagonalisation gives \(\mathbb G=\mathbb W^{\mathsf T}\mathbb W/2\), positive and of rank two, with \(\mathbb G\mathbb V\mathbb G=\tfrac{\tilde\nu_-}{2}\mathbb G\); and since \(\tilde\nu_-\) is homogeneous of degree one, \(\mathrm{tr}(\mathbb G\mathbb V)=\tilde\nu_-\). Expanding with \(\mathbb\Sigma=\mathbb V+\xi\mathbb I\),

   \[\alpha^2=\tilde\nu_-^2+2\xi\tilde\nu_-\,\mathrm{tr}\,\mathbb G+2\xi^2\,\mathrm{tr}\,\mathbb G^2=\tfrac12\bigl[(\tilde\nu_-+2\xi g_1)^2+(\tilde\nu_-+2\xi g_2)^2\bigr],\] where \(g_{1,2}\) are the nonzero eigenvalues of \(\mathbb G\). The symplectic condition on the rows \(v_1,v_2\) of \(\mathbb S^{-1}\) that select the \(\tilde\nu_-\) mode gives \(|v_1\mathbb Jv_2^{\mathsf T}|=1\), so \(\mathrm{tr}\,\mathbb G=(\lVert v_1\rVert^2+\lVert v_2\rVert^2)/2\ge\lVert v_1\rVert\lVert v_2\rVert\ge1\). The minimum under \(g_1+g_2\ge1\) is at \(g_1=g_2=\tfrac12\), giving \(\alpha\ge\tilde\nu_-+\xi\). Opposite local squeezings \(u\) of a two-mode squeezed vacuum give \(\mathrm{tr}\,\mathbb G=\cosh2u\) and \(\alpha=\tilde\nu_-+\xi\cosh2u\).

   Editorial check: with a gradient computed by symmetric finite differences, a symmetric thermal two-mode squeezed state with \(\tilde\nu_-=0.449\) and \(\xi=1\) gives \(\mathrm{tr}\,\mathbb G=1.000\) and \(\alpha=1.449=\tilde\nu_-+\xi\), and locally squeezed states with \(u=0.5\) and \(u=1\) give \(\alpha=1.992\) and \(4.366\), equal to \(\tilde\nu_-+\xi\cosh2u\). A Monte Carlo of the estimator with \(N=4000\) samples and 2000 repetitions gives \(\sqrt N\,\mathrm{Std}=1.466\), 2.056 and 4.58 for the same states; the excess over the asymptotic values shrinks as \(N\) grows.

   How to reuse it

   Any entanglement or nonclassicality witness computed from an estimated covariance or density matrix inherits a sample cost \(N_\star\propto(\text{gap})^{-2}\) with a prefactor set by its gradient. Before choosing a target state, compute \(\alpha\) along with the witness: a local Gaussian unitary applied before readout (here, undoing local squeezing) can reduce \(\alpha\) without changing the entanglement. The identity \(\mathbb G\mathbb V\mathbb G=\tfrac{\tilde\nu_-}{2}\mathbb G\) is specific to symplectic eigenvalues, but Euler’s theorem for homogeneous witnesses applies generally.

2. Weak-coupling limit at a fixed noise-to-coupling ratio

   To ask how weak a force can be and still be certified, the paper scales the diffusion rate with the coupling, \(\Gamma_{\rm d}=rg\), and the release time with \(1/\epsilon\). This separates what decides whether entanglement exists (the ratio \(r\)) from what decides how long certification takes (the size \(\epsilon\)).

   The exact second moments of the common and differential modes are expanded in \(\epsilon=\sqrt{4g/\Omega}\) at fixed \(\mu=2\bar n+1\), \(r\) and \(\theta\). The invariants scale as \(\det\mathbb V\sim\mu^2r^2\theta^3(\theta-\sin\theta\cos\theta)/(6\epsilon^2)\) and \(\tilde\Delta\sim\mu^2(\sin\theta-\theta\cos\theta)^2/\epsilon^2\), so \(\det\mathbb V/\tilde\Delta^2=O(\epsilon^2)\) and \(\tilde\nu_-^2\simeq\det\mathbb V/\tilde\Delta\). The factors of \(\mu\) and \(\epsilon\) cancel:

   \[\tilde\nu_-\simeq r\,\frac{\sqrt{\theta^3(\theta-\sin\theta\cos\theta)}}{\sqrt6\,\lvert\theta\cos\theta-\sin\theta\rvert}=rF(\theta),\qquad F(\pi)=\frac{\pi}{\sqrt6}.\] Setting \(rF(\pi)=1\) gives the entanglement threshold \(g_{\rm ent}=\pi\Gamma_{\rm d}/\sqrt6\). In contrast, the momentum entries of \(\mathbb G\) grow as \(1/\epsilon\), so \(\alpha\simeq\xi H(\theta)/\epsilon\) and \(N_\star\propto\epsilon^{-2}\propto1/g\). With the release time \(\theta/(\epsilon\Omega)\) dominating each cycle, \(T_\star\propto\epsilon^{-3}\propto g^{-3/2}\).

   Editorial check: evaluating the exact moments (the paper’s Eqs. S42–S47, with \(\bar n=0.65\)) at \(\theta=\pi\) gives \(\tilde\nu_-=0.3313\), 0.3217 and 0.3207 for \(r=0.25\) at \(\epsilon=10^{-2},10^{-3},10^{-4}\), converging to \(rF(\pi)=0.3206\); for \(r=0.5\) at \(\theta=2.5\) the values converge to \(rF=0.5354\). The reference couplings of the paper’s Table 1 follow from \(g/\Omega=-nV_{\rm int}(d)(q_{\rm zp}/d)^2/(\hbar\Omega)\): we obtain \(3.17\times10^{-5}\) (Coulomb), \(3.48\times10^{-9}\) (Casimir–Polder) and \(7.79\times10^{-22}\) (gravity), as listed.

   How to reuse it

   When extrapolating any decoherence-limited protocol to weak interactions, scale the noise with the coupling rather than holding it fixed. The limit then splits into a generation criterion depending only on the ratio and a sampling cost carrying all the dependence on interaction strength. This makes clear which improvements (isolation, readout noise, time budget) move which threshold.

Source: Main text Eqs. 1–18; End Matter Appendix A (Eqs. A1–A8) and Appendix B; Supplemental Material S2–S4 (Eqs. S9–S57). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It replaces “maximise the entanglement” by the quantity an experiment actually pays for, the time to certify it. The bound depends on the readout noise and the entanglement gap alone, so it is a floor for any covariance-based test, and it shows that a less entangled state can be certified sooner. At low decoherence, the minimum certifiable coupling is set by the time budget, so better isolation alone buys little.

Limitations. The analysis is asymptotic in the number of measurements (delta method), assumes perfectly calibrated measurement noise, and treats one transverse mode with the interaction expanded to quadratic order. Continuous monitoring is evaluated for one LQG feedback strategy. The weak-coupling formulas are not uniform as the diffusion ratio or the interaction phase goes to zero. The paper’s reference couplings put gravity at \(g/\Omega\approx7.8\times10^{-22}\), about twelve orders of magnitude below Casimir–Polder, and it does not claim that a gravitational test is within reach.

Description based on the full text of the linked arXiv version, read 2026-10-29; equations follow the paper’s notation.