Coherent Feedback Cooling in Optomechanics

• Coherent Feedback Cooling is a measurement-free technique that processes the optical field coherently to preserve quantum correlations and reduce mechanical motion.
• The protocol leverages controlled phase shifts and delay networks to generate configurable damping or stiffness modifications in various optomechanical platforms.
• Experimental implementations across levitated nanoparticles, cavity membranes, and phononic-crystal systems demonstrate significant cooling performance with low effective phonon numbers.

Coherent Feedback Cooling (CFC) is a class of feedback protocols in which the field carrying information about a target degree of freedom is coherently processed and fed back without any photodetection, so the loop remains measurement-free and the correlations between the system motion and the feedback signal are preserved (Ernzer et al., 2022, Melo et al., 26 Jun 2025). In optomechanics and related hybrid platforms, the feedback phase, delay, and transmission determine whether the returned signal acts primarily as a stiffness shift, a viscous damping channel, or a more general engineered reservoir. Recent realizations span levitated nanoparticles, cavity membranes, ultracoherent phononic-crystal membranes, and spin-mechanical hybrids, with reported cooling to $n\approx344$ phonons at $3\times10^{-7}\,\mathrm{mbar}$ without cryogenics, $n_f=166\pm7$ at room temperature, and $\bar n_m=4.89\pm0.14$ phonons in a $20\,\mathrm{K}$ environment (Melo et al., 26 Jun 2025, Filho et al., 20 May 2026, Ernzer et al., 2022).

1. Conceptual basis and historical placement

CFC differs from standard feedback in that it does not first convert the system state into a classical measurement record. Instead, the output field is routed through a coherent controller or delay network and then reinjected, so both non-commuting quadratures can participate in the loop without the irreversible state modification associated with measurement (Ernzer et al., 2022, Hamerly et al., 2012). In the Linear Quadratic Gaussian setting, early analyses showed that coherent control schemes, in which the resonator is embedded in an interferometer to achieve all-optical feedback, can outperform optimal measurement-based feedback control schemes in the quantum regime of low steady-state excitation number (Hamerly et al., 2012, Hamerly et al., 2012).

Within quantum optomechanics, resolved-sideband cooling was explicitly identified as an example of coherent feedback control, and analytical comparisons with optimal linear measurement-based feedback were carried out on an equal footing in the ground-state-cooling regime (Jacobs et al., 2014). Subsequent work broadened the notion of CFC from static sideband-cooling architectures to traveling-wave loops with explicit phase and delay control, two-pass cavity interactions, ancilla-assisted collision models, and transfer-function-based controller synthesis (Ernzer et al., 2022, Harwood et al., 2022, Fujimoto et al., 2 Apr 2026).

Experimentally, the field has progressed from theoretical demonstrations of performance gains in the quantum regime to direct implementations in several platforms. These implementations share a common logic: information about motion is encoded into an optical or ancilla mode, rotated in phase space by a coherent transformation or propagation delay, and returned so as to preferentially enhance anti-Stokes-like processes or generate a momentum-proportional force (Schmid et al., 2021, Melo et al., 26 Jun 2025, Filho et al., 20 May 2026).

2. Dynamical mechanism and canonical equations

In the levitated-nanoparticle realization, a single silica nanoparticle is held in an optical tweezer and its center-of-mass motion $z_p(t)$ along the beam axis is imprinted as a small phase modulation $\phi_p(t)=B z_p(t)$ on the field it scatters (Melo et al., 26 Jun 2025). After a fixed loop delay $\tau$, the scattered field returns with phase $\phi(t)=\phi_p(t-\tau)+\phi_0$, and its interference with the trapping field shifts the trap equilibrium according to

$z_{\rm eq}(t)=\beta z_p(t-\tau),$

with $3\times10^{-7}\,\mathrm{mbar}$0 set by the collection and re-injection efficiency, field amplitudes, and spatial mode overlap. For a harmonic oscillator of natural frequency $3\times10^{-7}\,\mathrm{mbar}$1, choosing $3\times10^{-7}\,\mathrm{mbar}$2 makes the equilibrium shift track the velocity, producing an effective viscous force

$3\times10^{-7}\,\mathrm{mbar}$3

so the damping is generated without ever measuring $3\times10^{-7}\,\mathrm{mbar}$4 (Melo et al., 26 Jun 2025).

The same mechanism can be written as a delayed self-interaction. For the nanoparticle center-of-mass mode,

$3\times10^{-7}\,\mathrm{mbar}$5

and, to first order in $3\times10^{-7}\,\mathrm{mbar}$6,

$3\times10^{-7}\,\mathrm{mbar}$7

Using

$3\times10^{-7}\,\mathrm{mbar}$8

one obtains both a stiffness modification proportional to $3\times10^{-7}\,\mathrm{mbar}$9 and a damping term proportional to $n_f=166\pm7$0 (Melo et al., 26 Jun 2025).

In cavity optomechanical implementations, the same physics is often expressed through a closed-loop susceptibility. For a linearized cavity mode $n_f=166\pm7$1 and mechanical mode $n_f=166\pm7$2,

$n_f=166\pm7$3

and the coherent-feedback loop modifies the mechanical response as

$n_f=166\pm7$4

with

$n_f=166\pm7$5

The loop phase $n_f=166\pm7$6, delay $n_f=166\pm7$7, and cavity susceptibility $n_f=166\pm7$8 therefore enter the feedback as a complex self-energy whose imaginary part produces damping and whose real part produces reactive shifts (Filho et al., 20 May 2026).

A closely related two-pass cavity treatment yields explicit damping and spring-shift expressions. In the unresolved-sideband membrane experiment,

$n_f=166\pm7$9

so $\bar n_m=4.89\pm0.14$0 and $\bar n_m=4.89\pm0.14$1 determine whether the feedback is dominantly dissipative or dispersive. The optimal-cooling choice is

$\bar n_m=4.89\pm0.14$2

for which $\bar n_m=4.89\pm0.14$3 is maximized and $\bar n_m=4.89\pm0.14$4 (Ernzer et al., 2022).

3. Experimental realizations across platforms

The experimental literature now contains several distinct CFC architectures. Some use the same optical field interacting twice with the same mechanical mode through different cavity modes; others use a traveling-wave delay line and direct reinjection; still others use a remote ancilla oscillator as the coherent controller. In all cases, no classical estimate of the mechanical state is used inside the feedback loop (Ernzer et al., 2022, Melo et al., 26 Jun 2025, Schmid et al., 2021).

Platform	Coherent loop architecture	Reported performance
Levitated nanoparticle (Melo et al., 26 Jun 2025)	Back-scattered light delayed in a $\bar n_m=4.89\pm0.14$5 single-mode fiber and refocused onto the particle; no photodetection	$\bar n_m=4.89\pm0.14$6, $\bar n_m=4.89\pm0.14$7, $\bar n_m=4.89\pm0.14$8 up to $\bar n_m=4.89\pm0.14$9
Cavity membrane (Ernzer et al., 2022)	Same laser interacts twice through two orthogonal-polarization cavity modes with auxiliary LO, tunable phase $20\,\mathrm{K}$0, and delay $20\,\mathrm{K}$1	$20\,\mathrm{K}$2 phonons at $20\,\mathrm{K}$3
Density phononic-crystal membrane (Filho et al., 20 May 2026)	Reflected field displaced, delayed, and re-injected into the cavity input in orthogonal polarization; combined with strong DBC	$20\,\mathrm{K}$4, cooling factor $20\,\mathrm{K}$5, $20\,\mathrm{K}$6
Membrane with atomic spins (Schmid et al., 2021)	Light-mediated spin-mechanics loop with state swaps and stroboscopic spin pumping	$20\,\mathrm{K}$7, $20\,\mathrm{K}$8, $20\,\mathrm{K}$9

These realizations occupy different operating regimes. The cavity membrane experiment was performed in the unresolved-sideband regime $z_p(t)$0 and cooled below the theoretical limit of cavity dynamical backaction cooling in that regime (Ernzer et al., 2022). The ultracoherent phononic-crystal membrane combined CFC with strong dynamical backaction cooling in a relatively narrow cavity and achieved a cooling factor of $z_p(t)$1 at room temperature (Filho et al., 20 May 2026). The levitated-nanoparticle experiment emphasized a purely all-optical, measurement-free loop that preserves correlations between motion and the feedback field (Melo et al., 26 Jun 2025). The spin-membrane platform used coherent swaps with an ancilla that could be tuned from strong coupling to an overdamped regime, thereby interpolating between state-transfer and reservoir-cooling behavior (Schmid et al., 2021).

4. Optimization, limiting noise, and stability

For the levitated nanoparticle, the center-of-mass spectrum $z_p(t)$2 yields an effective mode temperature through equipartition, and the resulting steady-state expression contains both ordinary cooling and a feedback-noise contribution (Melo et al., 26 Jun 2025). In the weak-gain, low-noise regime one recovers the cold-damping form

$z_p(t)$3

but for stronger $z_p(t)$4 the phase-noise term grows as $z_p(t)$5 and eventually limits cooling. Minimizing the full expression with respect to $z_p(t)$6 gives

$z_p(t)$7

This minimum scales linearly in $z_p(t)$8 and $z_p(t)$9, and inversely in the interference contrast $\phi_p(t)=B z_p(t)$0 and $\phi_p(t)=B z_p(t)$1. In the $\phi_p(t)=B z_p(t)$2 fiber loop, the dominant limit is phase noise, with $\phi_p(t)=B z_p(t)$3, and the onset of noise squashing in the in-loop spectrum is the tell-tale signature that phase noise, not particle motion, dominates the loop (Melo et al., 26 Jun 2025).

For the room-temperature membrane, the final occupancy is written as

$\phi_p(t)=B z_p(t)$4

so the role of CFC is separated from thermal loading, dynamical-backaction added noise, and loop-added noise (Filho et al., 20 May 2026). The full model, which includes cavity dynamics, loop delay and loss, and measured technical noise, quantitatively reproduces the measured spectra and the extracted $\phi_p(t)=B z_p(t)$5. The principal experimental limitations are finite loop delay, optical loss, imperfect mode matching, and technical noise, with cavity-mirror vibrations and laser noise contributing $\phi_p(t)=B z_p(t)$6 to the final occupation (Filho et al., 20 May 2026).

Delay is a recurring control parameter rather than a small correction. In the room-temperature membrane experiment, finite loop delay sets $\phi_p(t)=B z_p(t)$7 for optimal cooling (Filho et al., 20 May 2026). In the spin-membrane system, the propagation delay of about $\phi_p(t)=B z_p(t)$8 shifts the optimum damping condition, makes the steady-state $\phi_p(t)=B z_p(t)$9 asymmetric in detuning, and produces self-oscillation regions for negative detuning (Schmid et al., 2021). In a passive interferometric model of cavity optomechanics, moderate delays of order $\tau$0 degrade the cooling only slightly, whereas very long delays effectively remove the benefit of feedback (Harwood et al., 2020). This suggests that delay is simultaneously a resource for quadrature rotation and a constraint through added phase and stability sensitivity.

5. Relation to measurement-based feedback and standard sideband cooling

Measurement-based feedback relies on detecting the mechanical motion, processing the resulting classical signal, and then applying an electrical or optical force. In the levitated-particle comparison, this was stated to incur fundamental back action linked to detection efficiency, electronic noise in amplifiers and delays, and limited loop bandwidth with extra delay (Melo et al., 26 Jun 2025). CFC avoids those particular penalties because no photodetection feeds the loop, and in all-optical implementations it preserves the motion-feedback correlations that measurement would otherwise discard (Melo et al., 26 Jun 2025, Ernzer et al., 2022).

The comparison is especially sharp in the sideband-unresolved regime. Dynamical backaction cooling alone requires $\tau$1 for strong cooling, but coherent feedback does not rely on resolving motional sidebands and can still produce a large $\tau$2 when $\tau$3 and $\tau$4 are chosen to yield negative-momentum feedback (Filho et al., 20 May 2026). In the unresolved-sideband membrane experiment, standard dynamical backaction cooling had a lower bound of about $\tau$5 phonons for the stated parameters, whereas coherent feedback reached $\tau$6 phonons and did so with roughly $\tau$7 of the optical power needed for cavity cooling to the same $\tau$8 in the bad-cavity regime (Ernzer et al., 2022).

Analytical comparisons sharpen the origin of this advantage. In a perturbative treatment of resolved-sideband cooling and optimal linear measurement-based feedback, the maximal coherent-feedback cooling was found to scale as

$\tau$9

whereas the optimal measurement-based scaling was

$\phi(t)=\phi_p(t-\tau)+\phi_0$0

The performance gap was attributed not to the back-action noise of the measurement but to projection noise, with coherent feedback making better use of the same linear interaction (Jacobs et al., 2014).

A common oversimplification is that coherent feedback is always superior. A unified collision-model treatment gives a more conditional result: measurement-based feedback is typically superior in cooling, whilst coherent feedback is better at assisting quantum operations (Harwood et al., 2022). The same framework also shows that when a pure ancilla is available, coherent feedback can purify the system to the pure state $\phi(t)=\phi_p(t-\tau)+\phi_0$1 for any nontrivial coupling angle, whereas if the ancilla is maximally mixed, coherent feedback alone cannot reduce the system entropy at all and a projective measurement-based loop yields a lower-entropy steady state (Harwood et al., 2022). This suggests that the comparative ranking depends on the controller resource and on what is meant by “cooling.”

6. Generalizations, controller synthesis, and prospects

Several recent works recast CFC as a systematic controller-design problem. In the loop-shaping approach, the controller is specified by a transfer function $\phi(t)=\phi_p(t-\tau)+\phi_0$2, the closed-loop optical susceptibility becomes

$\phi(t)=\phi_p(t-\tau)+\phi_0$3

and the controller is designed so that the Stokes process is suppressed while the anti-Stokes process is preserved or enhanced (Fujimoto et al., 2 Apr 2026). For a notch-filter configuration, the shaped rates satisfy

$\phi(t)=\phi_p(t-\tau)+\phi_0$4

so ground-state cooling becomes possible even in the unresolved-sideband regime. The same framework emphasizes standard loop-stability criteria, including Nyquist and gain/phase-margin conditions, and extends naturally to dissipative entanglement generation and stabilization of nonclassical states (Fujimoto et al., 2 Apr 2026).

Passive-feedback proposals further show that CFC can make a broad-band cavity behave as though it were sideband resolved. One proposal uses a narrow-linewidth passive element so that an optomechanical system in the deeply sideband-unresolved regime exhibits dynamics similar to a sideband-resolved system; for representative chip-scale parameters, the optimized final occupation was reported as $\phi(t)=\phi_p(t-\tau)+\phi_0$5–$\phi(t)=\phi_p(t-\tau)+\phi_0$6 for $\phi(t)=\phi_p(t-\tau)+\phi_0$7–$\phi(t)=\phi_p(t-\tau)+\phi_0$8 (Guo et al., 2022). A related Fano-mirror analysis found that a single-sided coherent-feedback loop can achieve ground-state cooling even when the total optical loss is more than seven orders of magnitude larger than the mechanical frequency and the feedback efficiency is relatively low, whereas a more standard double-sided feedback scheme is not appropriate for the highly asymmetric two-mirror Fano cavity (Du et al., 2024).

The outlook toward the motional ground state is now platform specific but technically concrete. For levitated nanoparticles, the stated route is to shorten the delay line, reduce phase noise by better cancellation methods such as Pound–Drever–Hall feed-forward and fiber noise cancellation, work at $\phi(t)=\phi_p(t-\tau)+\phi_0$9, reduce photon recoil with smaller particles, and improve collection and mode matching; a realistic extrapolation combining these was reported to yield $z_{\rm eq}(t)=\beta z_p(t-\tau),$0 corresponding to $z_{\rm eq}(t)=\beta z_p(t-\tau),$1 phonon in an all-optical, measurement-free loop (Melo et al., 26 Jun 2025). For the room-temperature phononic-crystal membrane, increasing $z_{\rm eq}(t)=\beta z_p(t-\tau),$2 by $z_{\rm eq}(t)=\beta z_p(t-\tau),$3, suppressing cavity-frequency noise by more than $z_{\rm eq}(t)=\beta z_p(t-\tau),$4, and improving loop efficiency toward $z_{\rm eq}(t)=\beta z_p(t-\tau),$5 leads in the model to $z_{\rm eq}(t)=\beta z_p(t-\tau),$6, and to $z_{\rm eq}(t)=\beta z_p(t-\tau),$7 in the quantum-noise-limited limit (Filho et al., 20 May 2026).

Beyond cooling, the same coherent loops have been proposed or demonstrated as tools for non-reciprocal interactions, reservoir-engineered entanglement between remote particles, spin-mechanics interfaces, photon-phonon entanglement generation, steady-state squeezing, and transiently assisted state transfer (Melo et al., 26 Jun 2025, Guo et al., 2022, Harwood et al., 2020). CFC is therefore not merely a cooling technique but a broader architecture for quantum control in which dissipation, delay, and interference are engineered coherently rather than through measurement and classical actuation.