Optical Feedback Cooling Overview

• Optical feedback cooling is a control strategy that uses light for both sensing and actuation to reduce the motional energy of systems ranging from single atoms to macroscopic mirrors.
• It employs both measurement-based and coherent/parametric feedback methods to modulate forces like radiation pressure or optical torque, achieving efficient damping and energy extraction.
• Experimental realizations in cavity-QED, levitated nanoparticles, and torsional nanofibers demonstrate the method’s versatility in overcoming noise and back-action limits for high-precision applications.

Optical feedback cooling is the active suppression of motional energy by closing a feedback loop around an optical transducer. In measurement-based implementations, light measures position, cavity transmission, polarization rotation, or scattered-field phase, and the resulting signal is fed back as trap-depth modulation, radiation-pressure or gradient force, optical torque, or electrical/electrostatic actuation. In coherent implementations, the optical field itself is routed through a delay or interferometric controller so that the returning field produces damping without photodetection. The subject therefore spans cavity-QED cooling of single atoms, levitated nanoparticles, torsional nanofibers, gram-scale mirrors, and linearized optomechanical cavities (Koch et al., 2010, Sames et al., 2018, Vovrosh et al., 2016, 0705.1018, Su et al., 2023, Melo et al., 26 Jun 2025).

1. Scope and architectural variants

Optical feedback cooling is not a single protocol but a family of feedback architectures unified by the central role of light in sensing, actuation, or both. Several experiments explicitly use optical readout together with non-optical actuation: the tapered-optical-fiber torsional-mode experiment measures torsion through polarization rotation of transmitted light and feeds back a torque with electrodes, while charged levitated-nanoparticle experiments use optical position detection and synchronized electric fields for cold damping (Tebbenjohanns et al., 2023, Iwasaki et al., 2018, Dania et al., 2020). Other platforms are purely optical on both sides of the loop, including parametric modulation of optical traps, polarization-controlled optical torque on nanofibers, and interference-generated optical forces in levitated tweezers (Sames et al., 2018, Su et al., 2023, Ezzo et al., 12 Mar 2026).

A useful structural distinction is between measurement-based and coherent feedback. In measurement-based feedback, an optical signal is detected, digitized or filtered, phase shifted, and reapplied as a force or stiffness modulation. In coherent feedback, the optical field is not measured to generate the control action; instead, it is processed optically and returned to the plant. This distinction is explicit in measurement-free coherent cooling of a levitated nanoparticle, where delayed scattered light interferes with the trapping field, and in LQG analyses of all-optical controllers for quantum oscillators (Melo et al., 26 Jun 2025, Hamerly et al., 2012, Hamerly et al., 2012).

A second distinction is between direct force feedback and parametric feedback. Direct force feedback implements a viscous-like force or torque proportional to velocity. Parametric feedback modulates the trap stiffness or optical spring near twice the oscillator frequency, so that the timing of the modulation determines whether energy is extracted or injected. Both mechanisms are widely used, but they have different stability, bandwidth, and noise properties (Sames et al., 2018, Vovrosh et al., 2016, Zhong et al., 2017).

2. Core physical mechanisms

Parametric feedback cooling operates by modulating the confining potential near the principal parametric resonance. In the single-atom cavity-QED experiment, the trap depth is modulated as

$$U(t)\propto \bigl[1+\epsilon \cos(2\omega t + \phi_{\mathrm{pfb}})\bigr],$$

with the modulation phase adjusted from repeated weak position measurements so that the trap is weakened near turning points and strengthened near the center, thereby removing kinetic energy. The same basic strategy appears in levitated optomechanics, where the optical spring constant is modulated at twice the oscillation frequency in all three motional directions (Sames et al., 2018, Vovrosh et al., 2016).

Cold damping implements an additional viscous channel. In the tapered-optical-fiber torsional-mode experiment, the measured torsional signal is delayed and amplified so that the externally applied torque becomes

$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$

with the delay tuned so that the feedback is approximately $90^\circ$ out of phase with displacement at resonance and therefore proportional to $\dot{\phi}(t)$. The same operational principle underlies levitated-nanoparticle cold damping, for which the feedback force is approximated by $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$ (Tebbenjohanns et al., 2023, Ezzo et al., 12 Mar 2026).

Coherent delayed feedback replaces measurement and electronic actuation with a purely optical loop. In measurement-free coherent cooling of a levitated nanoparticle, the particle motion imprints a phase $\phi_p(t)\propto z_p(t)$ on scattered light, the field is delayed by $\tau$, and the returned field shifts the equilibrium position of the trap so that the particle feels a force proportional to its own past displacement. For a harmonic oscillator, the ideal delay for viscous cooling is

$\tau=\frac{\pi}{2\Omega},$

and the coherent feedback damping is

$\Gamma_c=\beta\Omega\sin(\Omega\tau).$

This makes the damping sign explicitly phase controllable: the same loop can cool or heat depending on delay and phase (Melo et al., 26 Jun 2025).

3. Experimental realizations and reported performance

The experimental record spans atomic, levitated, torsional, and macroscopic optomechanical systems.

Platform	Feedback implementation	Representative outcome
Single $^{85}$Rb atom in Fabry–Perot cavity	Continuous parametric modulation from cavity-transmission readout	$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$0 storage time; factor of 60 radial improvement; axial improvement by more than 30% (Sames et al., 2018)
Single neutral rubidium atom in high-finesse cavity	Discrete transmission-based trap-depth feedback	$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$1; storage time about $$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$2 (Koch et al., 2010)
Levitated silica nanoparticle in parabolic mirror trap	One-laser, one-photodiode parametric feedback in 3D	mK range; lowest temperatures around $$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$3 (Vovrosh et al., 2016)
Single nanoparticle in optical lattice	Direct feedback with reduced laser phase noise	occupation number about 3 (Kamba et al., 2020)
Levitated nanoparticle with coherent optical loop	Measurement-free delayed optical feedback	$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$4; $$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$5 phonons (Melo et al., 26 Jun 2025)
Tapered optical fiber torsional mode	Optical readout with active feedback torque	$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$6 (Tebbenjohanns et al., 2023)
5 mm-long optical nanofiber torsional mode	Purely optical derivative feedback	about $$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$7; three orders of magnitude cooling (Su et al., 2023)
Gram-scale suspended mirror	Optical spring, optical dilution, and cold damping	$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$8 (0705.1018)
Levitated nanoparticle with interference-based 3D optical damping	Co-propagating auxiliary field in a single beam path	$$M_{\mathrm{ext}}(t)=g\, \gamma_0\, \Omega\, I_{\mathrm{eff}}\, \phi(t-d),$$9, $90^\circ$0, and $90^\circ$1 along $90^\circ$2 (Ezzo et al., 12 Mar 2026)

The single-atom experiments established optical feedback cooling in a regime where standard multibeam laser-cooling geometries are difficult. The 2010 cavity experiment used a discrete two-window rule on cavity transmission and reached $90^\circ$3 with storage times in the 1-second regime, whereas the 2018 experiment replaced this with continuous phase-sensitive parametric feedback and extended a $90^\circ$4 kHz radial mode from $90^\circ$5 to more than $90^\circ$6, while also cooling a much faster $90^\circ$7 kHz axial mode on microsecond timescales (Koch et al., 2010, Sames et al., 2018).

Levitated systems have provided the broadest performance range. The parabolic-mirror trap realized a compact one-laser, one-photodiode geometry and cooled all three center-of-mass modes from room temperature to a few mK, with the best discussion reaching roughly $90^\circ$8 phonons. In a one-dimensional optical lattice, reducing laser phase noise near the oscillation frequency suppressed excess heating so that the occupation number fell to about three, with the remaining limit identified as photon recoil heating (Vovrosh et al., 2016, Kamba et al., 2020). More recently, coherent measurement-free feedback achieved $90^\circ$9 and $\dot{\phi}(t)$0 phonons without photodetection in the control path, and interference-based optical cold damping in a single beam path produced simultaneous 3D cooling of a neutral particle (Melo et al., 26 Jun 2025, Ezzo et al., 12 Mar 2026).

Rotational and torsional degrees of freedom have also become accessible. The tapered-optical-fiber torsional mode combines optical polarimetric readout with active feedback and was cooled from room temperature to $\dot{\phi}(t)$1, while a 5 mm-long nanofiber torsional resonator uses a weak probe to measure rotation and a polarization-controlled drive laser to apply optical torque, reaching a mode temperature ratio as low as $\dot{\phi}(t)$2 in time-domain statistics and about $\dot{\phi}(t)$3 in the reported conclusion (Tebbenjohanns et al., 2023, Su et al., 2023).

At the macroscopic end, a 1-gram suspended mirror was cooled by combining the optical spring effect, optical dilution, and cold damping. The optical spring raised the effective resonance from about $\dot{\phi}(t)$4 to about $\dot{\phi}(t)$5, and the mode temperature reached $\dot{\phi}(t)$6, about $\dot{\phi}(t)$7 below ambient (0705.1018).

4. Noise, back action, and stability limits

The central limitation in many measurement-based schemes is that the measurement itself perturbs the motion. In continuous parametric cooling of a single atom, the resonant probe causes spontaneous-emission momentum kicks, so the atomic oscillation loses phase coherence after only a few oscillations. For the radial mode, the scattering rate is about $\dot{\phi}(t)$8 and the effective quality factor is only $\dot{\phi}(t)$9. Earlier discrete atom feedback reached a temperature floor consistent with signal-to-noise limitations: a $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$0 integration time and weak probe power imply a signal-to-noise ratio of about 1 at an energy of about $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$1, leading to an estimated floor $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$2, close to the observed $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$3 (Sames et al., 2018, Koch et al., 2010).

In levitated nanoparticles, technical phase noise becomes dominant once gas damping is reduced. In the optical-lattice experiment, laser phase noise near the oscillation frequency $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$4 shakes the standing wave and drives resonant heating, and when that phase noise is decreased by orders of magnitude the heating rate drops until the occupation number is limited only by photon recoil, at about three quanta (Kamba et al., 2020). In measurement-free coherent feedback, the dominant limitation is also phase noise, now in the delayed optical loop; the in-loop noise floor at resonance is consistent with

$$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$5

and the minimum temperature is set by the competition between coherent damping and phase-noise-induced force noise (Melo et al., 26 Jun 2025).

Collisional and detector-noise limits remain important in room-temperature levitation. In the parabolic-mirror trap, Allan-deviation analysis identifies background-gas collisions as the dominant noise source over the explored regime, while the supplement estimates a single-photon-recoil limit of about 24 phonons once gas damping is reduced. In interference-based 3D optical cold damping, the effective temperature

$$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$6

shows directly that measurement imprecision sets the cooling floor; this is why the $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$7 axis, with a lower detection floor, cools much more strongly than $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$8 and $$F_{\mathrm{fb}}(t)\approx -m \Gamma_{\mathrm{fb}} \dot q(t)$$9 (Vovrosh et al., 2016, Ezzo et al., 12 Mar 2026).

High feedback gain can also make diagnostics deceptive. Noise squashing is explicitly reported in the torsional-fiber cold-damping experiment, in coherent-feedback levitation, in optical nanofiber torsional cooling, and in Paul-trap feedback studies, so out-of-loop detectors are used whenever possible to avoid mistaking in-loop suppression for actual motional cooling (Tebbenjohanns et al., 2023, Melo et al., 26 Jun 2025, Su et al., 2023, Dania et al., 2020). In single-laser optomechanical feedback, where the same field probes and actuates, stability requires

$\phi_p(t)\propto z_p(t)$0

because in-loop interference can otherwise diverge and drive the system unstable (Kumar et al., 2022). In macroscopic optical-spring cooling, by contrast, the dominant limitation is laser frequency noise at the optical spring resonance rather than measurement back action in the usual cavity-QED sense (0705.1018).

5. Control theory, coherent feedback, and dissipation engineering

The quantum-control literature formulates feedback cooling as a steady-state LQG problem for open quantum systems. In this treatment, the cost is the steady-state occupation number, such as $\phi_p(t)\propto z_p(t)$1 for an optical mode or $\phi_p(t)\propto z_p(t)$2 for a mechanical mode, with the covariance matrix obtained from the corresponding Lyapunov equation. Within this framework, coherent controllers can outperform optimal measurement-based controllers in the low-excitation regime because they process both non-commuting output-field quadratures without measurement-induced loss of fidelity (Hamerly et al., 2012, Hamerly et al., 2012).

A concrete realization of this idea is passive interferometric coherent feedback in cavity optomechanics. For the linearized red-sideband problem, the loop changes the effective optical loss to

$\phi_p(t)\propto z_p(t)$3

and in the weak-coupling resolved-sideband regime the optimal passive coherent-feedback condition is

$\phi_p(t)\propto z_p(t)$4

This reduces the steady-state phonon occupancy and also accelerates the cooling transient. Nonzero delay degrades the performance, but moderate delays do not eliminate it; in the example given, the passive loop changes $\phi_p(t)\propto z_p(t)$5 from $\phi_p(t)\propto z_p(t)$6 at $\phi_p(t)\propto z_p(t)$7 to $\phi_p(t)\propto z_p(t)$8 at $\phi_p(t)\propto z_p(t)$9, $\tau$0 at $\tau$1, and $\tau$2 at $\tau$3 (Harwood et al., 2020).

Hybrid optical feedback can also be implemented by coupling an optomechanical cavity to a second cavity containing an ultracold atomic ensemble. In that model, the atomic cavity acts as an optical feedback filter that changes the effective detuning, optical spring shift, and especially the optomechanical damping rate. At the red sideband in the good-cavity regime, the feedback-enhanced damping is increased by nearly four orders of magnitude relative to the uncoupled case, and the architecture remains useful in the bad-cavity regime where optimal cooling shifts to $\tau$4 (Sarma et al., 2015).

Single-laser feedback cooling occupies an intermediate position between standard measurement-based and coherent-control viewpoints. The same beam is used for readout and radiation-pressure actuation, so the measured homodyne spectrum acquires a feedback-dependent transduction factor. The experiment and model show that these interference effects are not merely parasitic: when the loop is kept within the stable region, they can enhance cooling beyond the conventional auxiliary-laser architecture, with more than two orders of magnitude cooling demonstrated in a silicon photonic crystal nanobeam resonator (Kumar et al., 2022).

6. Significance, recurrent misconceptions, and outlook

One recurrent misconception is that “optical feedback cooling” must mean purely optical actuation. The literature uses the term more broadly. In the tapered-fiber torsional experiment, optical transmitted light supplies the displacement information while electrodes provide the compensating torque; in electric feedback cooling of charged nanoparticles, the motion is measured optically and the actuator is electrical; in Paul-trap cooling, optical and electrical feedback are compared directly and found to have similar cooling efficiencies (Tebbenjohanns et al., 2023, Iwasaki et al., 2018, Dania et al., 2020). Optical feedback cooling is therefore best understood as a control paradigm organized around optical transduction, not a single-force mechanism.

A second misconception is that modulation at $\tau$5 is intrinsically cooling. Parametric modulation near twice the oscillation frequency is only a resonance condition; the sign of the energy flow is set by phase. In the single-atom cavity experiment, out-of-phase modulation amplifies motion and heats the atom, whereas the correct phase relation damps the motion (Sames et al., 2018). More generally, quantum calculations for shot-noise-dominant levitated nanoparticles find that force feedback reaches a lower minimum occupation number than parametric feedback at the same measurement efficiency, with ground-state cooling thresholds of about 10% efficiency for force feedback and about 40% for parametric feedback (Zhong et al., 2017).

The applications are correspondingly broad. Single-atom cavity-QED work emphasizes restricted optical access, reduced optical pumping, and compatibility with strong-coupling geometries such as tight resonators, atom chips, bottle resonators, and microtoroids (Koch et al., 2010, Sames et al., 2018). Torsional-fiber systems emphasize sensing and hybrid quantum interfaces: the tapered optical fiber supports $\tau$6, up to $\tau$7, with $\tau$8, while the optical nanofiber torsional resonator reports a torque sensitivity of about $\tau$9 (Tebbenjohanns et al., 2023, Su et al., 2023). Interference-based 3D optical cold damping is explicitly compatible with neutral particles and is presented as suitable for microcavities, near-surface trapping architectures, and scalable multiparticle configurations; for the axial mode, the route to $\tau=\frac{\pi}{2\Omega},$0 is associated with pressures around $\tau=\frac{\pi}{2\Omega},$1 (Ezzo et al., 12 Mar 2026).

Theoretical work has added an information-thermodynamic perspective. The relation

$\tau=\frac{\pi}{2\Omega},$2

states that the controller’s learning rate exceeds the entropy pumping required for cooling by a positive excess information flow. The same work derives a precision trade-off,

$\tau=\frac{\pi}{2\Omega},$3

and a lower bound on the energetic cost of optical feedback cooling. It also finds that, in a concrete coherent-light-scattering model, the minimum kinetic temperature scales approximately as $\tau=\frac{\pi}{2\Omega},$4 in the small-noise regime, making measurement precision the major factor determining attainable temperature (Dechant et al., 18 Aug 2025).

Taken together, the reported results indicate that optical feedback cooling is no longer a niche alternative to laser cooling or passive dynamical back-action cooling. It is a general control methodology whose experimentally relevant limits are set by phase fidelity, measurement imprecision, recoil, collisional damping, delay, and loop stability, and whose strongest advantages appear in geometries with limited optical access, in non-resolved-sideband settings, and in platforms where coherent optical processing can preserve correlations that measurement-based loops necessarily discard (Sames et al., 2018, Harwood et al., 2020, Hamerly et al., 2012, Melo et al., 26 Jun 2025).