Foundations of quantum mechanics · Theory

Linear monitoring-feedback model for linearized gravity

Daisuke Miki, Alfred Li, Yiqiu Ma, Yubao Liu, Yanbei Chen

arXiv:2611.17642v1 ↗PDF ↗Submitted 13 November 2026

Question
One way to make gravity classical while matter stays quantum is to let an auxiliary observer continuously measure each mass and feed the classical record back as the gravitational force. Such models reproduce Newtonian attraction but add noise, and different proposals assume different observers. Which features of the observers' inputs decide whether the extra noise has a floor, whether the masses can become entangled, and whether the scheme can imitate a genuinely quantum gravitational field?
Result
Working in linear input–output theory, the paper sorts models by the correlations in the observers' input fields. With independent, uncorrelated (vacuum) inputs, each mass suffers a force noise of at least \(2\hbar M\omega_{\mathrm{SN}}^2\), where the Schrödinger–Newton frequency \(\omega_{\mathrm{SN}}\) is set by how tightly nuclei sit at their lattice sites. Squeezed but local inputs can remove the single-mass floor, yet the two-mass noise matrix still obeys \(S_{AA}S_{BB}-\lvert S_{AB}\rvert^2\ge4\hbar^2K_{AB}^2\), the known no-entanglement bound. With inputs correlated across the masses, the noise can be squeezed jointly, the masses can become entangled, and with a retarded response the scheme reproduces the commutators and fluctuations of linearized quantum gravity.

The paper in detail

Whether gravity must be quantum is usually tested by asking whether it can entangle two masses. A complementary idea is that a classical gravitational field has to be sourced by classical information about where the mass is, and acquiring that information disturbs the mass. Kafri, Taylor and Milburn made this concrete: an auxiliary observer measures each mass continuously and a feedback force built from the record supplies gravity. The price is extra force noise from the measurement's back action and imprecision.

This paper writes the whole construction in the Heisenberg picture. Each mass \(x\) is coupled to an observer's input quadratures \(a_1,a_2\), the record is the output quadrature, and the feedback force is a chosen linear kernel of the record. The experimenter's own readout is a second, independent channel. The deterministic response to forces then matches the target gravitational interaction exactly, and every difference between models sits in the noise the observers inject, its correlations and its commutators.

Self-gravity is the first surprise. If the field is sourced by a classical record \(x_c\) of the body's position, the body's quantum mass distribution feels a restoring force \(-M\omega_{\mathrm{SN}}^2(x-x_c)\), with \(\omega_{\mathrm{SN}}^2=Gm/(6\sqrt\pi\,x_{\mathrm{zp}}^3)\) set by nuclei of mass \(m\) localized to \(x_{\mathrm{zp}}\sim10^{-12}\) m. Because \(x-x_c\) is the record's imprecision, this term injects noise, and balancing it against back action gives the floor \(2\hbar M\omega_{\mathrm{SN}}^2\), the free-mass standard quantum limit at \(\omega_{\mathrm{SN}}\). The Diósi–Penrose collapse model has the same structure, with \(\omega_{\mathrm{DP}}=\omega_{\mathrm{SN}}(x_{\mathrm{zp}}/\sigma_{\mathrm{DP}})^{3/2}\).

The regimes then follow from which input correlations are allowed. Local squeezing can hide the self-gravity noise of one mass, but the mutual-gravity feedback couples each mass's force to the other observer's imprecision, so the joint noise keeps a determinant floor of \(4\hbar^2K_{AB}^2\), where \(K_{AB}=2GM_AM_B/d^3\). That floor is the condition under which a local classical channel cannot entangle. Inputs entangled across the two observers remove it, much as shared entanglement enables teleportation. For a retarded response, the force commutators of the feedback scheme equal those of a quantum field, so a suitable input state reproduces its fluctuations as well.

The experimental message is a hierarchy. Single-mass force-noise measurements at the Schrödinger–Newton scale, between \(2\pi\times8\) and \(2\pi\times78\) mHz depending on the material, are the most accessible test of the simplest models. Mutual-gravity noise sits about \((\omega_{\mathrm{SN}}/\omega_G)^2\sim10^4\) lower, and only an entanglement test can separate local-record models from nonlocal mediators. Open questions include whether tying the monitoring rate to microscopic constituents restores a single-mass floor even with local squeezing.

Transferable methods

  1. Trading back action against imprecision

    Each observer's noise has two parts: back action, which grows with the measurement strength \(\alpha\), and imprecision of the record, which shrinks with it. When the record is fed back as a force, both parts become force noise, and the paper's single-mass floor and two-mass determinant bound are both statements about this trade-off.

    For one mass, the record is \(x_c=\hbar b_2/\alpha=x+\hbar a_2/\alpha\), so the self-gravity feedback \(-M\omega_{\mathrm{SN}}^2(x-x_c)\) adds \((\hbar M\omega_{\mathrm{SN}}^2/\alpha)\,a_2\) to the back action \(\alpha a_1\). For vacuum inputs, \[S_{\mathscr F}=\alpha^2+\frac{\hbar^2M^2\omega_{\mathrm{SN}}^4}{\alpha^2}\ \ge\ 2\hbar M\omega_{\mathrm{SN}}^2,\] with equality at \(\alpha^2=\hbar M\omega_{\mathrm{SN}}^2\) by the inequality of arithmetic and geometric means.

    With two masses and vacuum inputs in the rescaled quadratures (Regime IIa), \[S_{\mathscr F_A}=\hbar\Bigl[2M_A\omega_{\mathrm{CG}}^2+\frac{(M_A+M_B)\,\omega_G^4}{2\omega_{\mathrm{CG}}^2}\Bigr]\ge2\hbar\sqrt{M_A(M_A+M_B)}\,\omega_G^2.\] Editorial correction: the minimum is reached where \(2M_A\omega_{\mathrm{CG}}^2=(M_A+M_B)\omega_G^4/(2\omega_{\mathrm{CG}}^2)\), that is \[\omega_{\mathrm{CG}}^4=\tfrac14\bigl(1+M_B/M_A\bigr)\,\omega_G^4,\] a factor of 4 below the value printed after Eq. (49); the minimum itself, checked numerically, is as stated.

    For general local inputs the forces reduce to \(\mathscr F_A=\sqrt{\hbar K_{AB}}\,(w_{1A}-w_{2B})\) and \(\mathscr F_B=\sqrt{\hbar K_{AB}}\,(w_{1B}-w_{2A})\). Writing each observer's spectra as \(x_j,y_j,c_j\) with \(x_jy_j-c_j^2\ge1\), \[\frac{S_{AA}S_{BB}-\lvert S_{AB}\rvert^2}{\hbar^2K_{AB}^2}=(x_A+y_B)(x_B+y_A)-(c_A+c_B)^2\ \ge\ 4.\] Editorial check: \(2\times10^5\) random local inputs never went below 4, and vacuum inputs saturate it.

    How to reuse it

    Any scheme that implements a force by measurement and feedback inherits a standard-quantum-limit-like noise floor, which local squeezing can redistribute but not remove from the joint noise matrix. The determinant of that matrix is the quantity to bound when asking whether a classical channel could have produced observed correlations.

  2. Matching a quantum field's fluctuations with correlated inputs

    To show that the same architecture can imitate linearized quantum gravity, the paper needs feedback forces with the field's commutators and an input state with the field's noise. The commutators follow from the retarded kernel alone; the noise is a choice of joint Gaussian input state, which the paper writes down explicitly.

    With force \(\mathscr F_I=a_1^I+\sum_JG^{\mathrm{ret}}_{IJ}*a_2^J\) (unit sensing strength), the commutators are \([\mathscr F_I(t_1),\mathscr F_J(t_2)]=-iG^{\mathrm{ret}}_{IJ}(t_1-t_2)+iG^{\mathrm{ret}}_{JI}(t_2-t_1)\), those of the field. Splitting \(\mathbf G^{\mathrm{ret}}=\mathbf A+i\mathbf K\) into Hermitian parts and choosing \[\mathbf S_{22}=\mathbf K^{-1},\quad\mathbf S_{12}=-\mathbf A\mathbf K^{-1},\quad\mathbf S_{11}=\mathbf A\mathbf K^{-1}\mathbf A+\mathbf K\] gives the force spectrum \[\mathbf S^{\mathscr F}=\mathbf S_{11}+\mathbf S_{12}\mathbf G^{\dagger}+\mathbf G\mathbf S_{21}+\mathbf G\mathbf S_{22}\mathbf G^{\dagger}=2\mathbf K,\] the zero-point noise of the field, while the input matrix satisfies the uncertainty condition \(\mathbf S+i\mathbf J\succeq0\).

    Editorial check: for random Hermitian \(\mathbf A\) and positive \(\mathbf K\) of size \(3\times3\), \(\mathbf S^{\mathscr F}-2\mathbf K\) vanished to \(10^{-14}\) and the smallest eigenvalues of \(\mathbf S\pm i\mathbf J\) were zero to rounding, so the state is a minimum-uncertainty one. This requires \(\mathbf K\) invertible, that is, a dissipative (radiating) response at the frequency in question.

    How to reuse it

    A linear quantum channel between position variables can be emulated by measurement, classical feedback and a pre-shared correlated input state. The recipe above gives that state for any target retarded response, which is useful for designing analogue simulations of field-mediated interactions.

Source: Sec. III (Eqs. (4)–(17)); Sec. IV.3 (Eqs. (39)–(60)); Sec. VI.2.1 (Eqs. (82)–(89)). 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 gives one framework for classical-channel gravity, collapse models and linearized quantum gravity, and ranks experimental targets. The single-mass floor at the Schrödinger–Newton scale is far larger than the two-mass scale of entanglement experiments: about 30 times below LISA Pathfinder's measured acceleration noise in amplitude. It also shows that correlated force noise between two masses is not evidence of a quantum mediator.

Limitations. Everything is linear and Gaussian unless stated, and the equivalence with quantum gravity covers linear response and two-point functions only. The single-mass floor depends on treating the internal mass distribution as part of the classical record and on vacuum inputs; local squeezing removes it, so a null result constrains only that class. The Regime IIa optimum is misprinted: minimizing the paper's own Eq. (45) gives \(\omega_{\mathrm{CG}}^4=(1+M_B/M_A)\,\omega_G^4/4\), not \((1+M_B/M_A)\,\omega_G^4\), although the minimum value it quotes is correct. The silicon \(\omega_{\mathrm{SN}}\) is given as \(0.036\ \mathrm{s^{-1}}\) in one place and \(2\pi\times8\) mHz \(\approx0.05\ \mathrm{s^{-1}}\) in another.

Description based on the full arXiv HTML of v1, read 2026-11-16: Sections I–VII, Table 1 and the opening of Appendix A; the determinant-bound proof in Appendix A was not read in full. Editorial numerical checks: the Regime IIa minimum and its location, the determinant bound over \(2\times10^5\) random local Gaussian inputs (smallest value 4.12 in units of \(\hbar^2K_{AB}^2\)), the Regime III input construction (Eqs. (87)–(89)) for random \(3\times3\) responses, the LISA Pathfinder figure \(2.9\times10^{-3}\ \mathrm{fm\,s^{-2}/\sqrt{Hz}}\), the damping requirement \(5\times10^{-12}\ \mathrm{s^{-1}}\) and \(\omega_G\approx4\times10^{-4}\ \mathrm{s^{-1}}\) for silicon.