Coherent Feedback

Coherent feedback routes a quantum signal through another quantum system, optical element, microwave circuit, waveguide, cavity, or auxiliary mode without first converting that signal into a classical measurement record. The feedback path can change phases, mix modes, add dynamics, create interference, or return an emitted field to the original system. Since no intermediate measurement result is read out, the signal can retain quantum coherence and correlations.

The short contrast is:

measurement-based feedback: quantum output -> classical record -> actuator
coherent feedback:          quantum output -> quantum controller -> quantum input

This page is the control-design entry point for coherent feedback. The port normalization and boundary relation are developed in Input–Output Theory. The record-based alternative is Measurement-Based Feedback. Autonomous dissipative stabilization is treated in Reservoir Engineering.

A Quantum Signal in the Feedback Path

A coherent feedback loop has three ingredients:

• a plant, the quantum system whose behavior is being shaped;
• a quantum signal channel, usually a traveling field, waveguide mode, cavity mode, phonon mode, or auxiliary system;
• a coherent controller, which processes the signal through unitary scattering, Hamiltonian evolution, or dissipative coupling without measuring it into a classical record.

The controller may be as simple as a mirror, phase shifter, beamsplitter, circulator, or delay line. It may also be a nonlinear quantum system such as a cavity, qubit, mechanical oscillator, Josephson circuit, or atomic ensemble.

The defining feature is not that no dissipation occurs. Traveling fields can carry energy away, cavities can be lossy, and auxiliary modes can be damped. The defining feature is that the feedback signal is not first reduced to a classical measurement outcome used by a classical controller.

Port Language

For a Markovian port with coupling operator $L$, input–output theory writes, in one common convention,

$$b_{\mathrm{out}}(t) = b_{\mathrm{in}}(t) + L(t).$$

Coherent feedback begins when part of $b_{\mathrm{out}}$ is not simply measured or discarded, but is routed into another quantum input. A passive return path with field return amplitude $0\le r<1$, phase $\phi$, and delay $\tau$ can be modeled by a lossless beamsplitter boundary

$$b_{\mathrm{in}}(t) = r e^{i\phi} b_{\mathrm{out}}(t-\tau) + \sqrt{1-r^2}\,b_{\mathrm{vac}}(t),$$

where $b_{\mathrm{vac}}$ is the independent field entering the unused beamsplitter port. The coefficients satisfy $r^2+(1-r^2)=1$, as required by a unitary two-port mixer. The complementary output port and any additional physical losses must also be included. Adding an independent vacuum field with unit coefficient to a unit-amplitude returned field would not preserve the field normalization.

If $\tau$ is small compared with the system timescales, a Markovian network reduction may be possible. If $\tau$ is comparable to the dynamics, the loop has memory and belongs closer to delayed-feedback or non-Markovian modeling.

Cascaded Systems

A useful building block, which by itself is feedforward rather than a closed feedback loop, is a cascaded connection: the output of one system drives another system in one direction. Suppose systems 1 and 2 have coupling operators $L_1$ and $L_2$, and the output of system 1 is fed into the input of system 2 with negligible delay. For vacuum incident on the first system and unit-efficiency one-way propagation, the joint master equation in the boundary convention above is

$$\dot\rho = - \frac{i}{\hbar} [H_{\mathrm{casc}},\rho] + \mathcal D[L_1+L_2]\rho,$$

with

$$H_{\mathrm{casc}} = H_1+H_2 + \frac{\hbar}{2i} \left( L_2^\dagger L_1 - L_1^\dagger L_2 \right).$$

This compact formula is worth reading slowly. The dissipator $\mathcal D[L_1+L_2]$ contains interference between indistinguishable output paths. The Hamiltonian term is directional: it represents the phase-sensitive drive of system 2 by the field emitted from system 1. The equation is not obtained by simply adding two independent dissipators.

Different phase conventions and propagation phases move factors of $e^{i\phi}$ between $L_1$, $L_2$, and the Hamiltonian term. A trustworthy calculation states the convention.

What the Controller Can Do

A coherent controller can shape dynamics through several mechanisms.

Mechanism	Physical action	Typical use
phase-sensitive interference	recombine emitted and incident fields	change effective damping or reflection
mode conversion	scatter one traveling mode into another	route photons or microwave signals
auxiliary dynamics	store and re-emit the signal	add filtering, memory, or nonlinearity
engineered loss	damp the controller mode	turn coherent routing into a designed reservoir
coherent drive from another quantum system	source output feeds a target input	cascaded state transfer or emission shaping
delayed return	field leaves and later re-enters	non-Markovian feedback, revivals, stabilization limits

These mechanisms can suppress a noise channel, enhance a useful transition, stabilize a state, reshape emitted wave packets, cool a mode, or generate entanglement. They can also destabilize the system, add unwanted modes, create long memory, or make a Markovian model invalid.

Relation to Measurement-Based Feedback

Measurement-based feedback can amplify a weak signal, digitize it, apply nonlinear logic, and use classical computation. It also necessarily exposes the signal to detector inefficiency, added noise, finite bandwidth, and record storage. The controller acts on a classical estimate.

Coherent feedback avoids the intermediate measurement. The controller acts on the quantum field or auxiliary system itself. This can preserve phase information and avoid measurement imprecision, but it also removes the easy option of copying, thresholding, or digitally processing the signal. Quantum no-cloning and added-noise constraints still apply: a coherent controller is a physical quantum device, not a perfect information conduit.

The operational question is:

Is the signal converted to a classical record before the control action?

If yes, use the measurement-based feedback language. If no, use coherent network or enlarged-system language.

Relation to Reservoir Engineering

Coherent feedback and reservoir engineering often overlap. A lossy auxiliary controller can take a quantum signal, process it coherently, and then dump entropy into an output line. If the auxiliary relaxes quickly, eliminating it may produce an effective Lindblad generator for the plant. From the plant’s reduced perspective, that looks like engineered dissipation.

The distinction is one of modeling level. If the auxiliary and routed fields are retained explicitly, the description is a coherent-feedback network. If those degrees of freedom are traced out after a controlled Markov approximation, the reduced description may be an engineered reservoir.

This hierarchy is useful:

plant + coherent controller + fields
        -> eliminate fast lossy controller
        -> effective engineered dissipator

The elimination step must be justified by timescale separation, weak coupling where required, and a clear treatment of delay and loss.

A Passive Cavity Loop

Take $L=\sqrt\kappa a$ and use the return boundary above with negligible delay. Set $\zeta=re^{i\phi}$. Solving the boundary condition before substituting it into the Langevin equation gives

$$b_{\mathrm{in}}= \frac{\zeta\sqrt\kappa}{1-\zeta}a +\frac{\sqrt{1-r^2}}{1-\zeta}b_{\mathrm{vac}}.$$

Consequently,

$$\dot a=-\left[i(\omega_c+\delta\omega)+\frac{\kappa_{\mathrm{eff}}}{2}\right]a -\frac{\sqrt{\kappa(1-r^2)}}{1-\zeta}b_{\mathrm{vac}},$$

with

$$\kappa_{\mathrm{eff}}=\frac{\kappa(1-r^2)}{|1-re^{i\phi}|^2},\qquad \delta\omega=\frac{\kappa r\sin\phi}{|1-re^{i\phi}|^2}.$$

The squared noise amplitude is exactly $\kappa_{\mathrm{eff}}$, preserving the oscillator commutator. This passive loop can enhance or reduce damping but cannot make it negative. At $\phi=0$, $\kappa_{\mathrm{eff}}=\kappa(1+r)/(1-r)$; at $\phi=\pi$, it is $\kappa(1-r)/(1+r)$. The zero-delay approximation must be checked against the modified rates. In particular, its apparent divergence as $r\to1$ at zero phase is a sign that a finite delay or bandwidth can no longer be ignored.

Quantum Controller Example

A more flexible coherent controller is another quantum system. For instance, a cavity mode can receive an output field from a qubit-cavity device, filter it through its own resonance, and feed a field back into the plant. The controller’s susceptibility then shapes the frequency dependence of the feedback.

If the controller mode is retained, the joint state evolves on the tensor product Hilbert space:

$$\rho_{\mathrm{tot}} \in \mathcal B(\mathcal H_{\mathrm{plant}}\otimes\mathcal H_{\mathrm{ctrl}}).$$

The plant alone may show memory because the controller stores information and returns it later. Tracing out the controller prematurely can create an inaccurate Markovian master equation.

If the controller damps much faster than the plant and remains close to a fixed state, it may be eliminated. Then the controller appears as an effective plant dissipator, Lamb shift, or noise spectrum. This is the same logic behind many auxiliary-loss constructions in Reservoir Engineering.

Delay and Stability

Delay is not a minor detail in coherent feedback. A loop with propagation time $\tau$ has memory. A boundary condition such as

$$b_{\mathrm{in}}(t) \propto b_{\mathrm{out}}(t-\tau)$$

means the present derivative can depend on system operators at earlier times. If $\tau$ is comparable to a decay time, Rabi period, cavity lifetime, or controller response time, the Markovian approximation can fail.

Networks containing gain, pumping, or driven nonlinear components can also become unstable. Feedback can then produce runaway amplification or self-oscillation. A strictly passive linear network has no energy source for runaway growth, although it can have long-lived or undamped modes and strong phase sensitivity. Loss and finite bandwidth may stabilize the loop, but they add noise and reduce coherence. A design should therefore report:

• loop phase and propagation time;
• internal loss and unused-port noise;
• controller bandwidth and nonlinearity;
• stability margins or eigenvalue checks in the linearized model;
• whether the effective reduced equation remains completely positive.

The non-Markovian warning signs are discussed more generally in What Non-Markovian Means.

Why Coherent Feedback Is Not Free

Coherent feedback avoids measurement imprecision, but it does not evade quantum noise. Unused ports inject vacuum noise. Loss turns coherent routing into unobserved dissipation. Amplification adds noise unless it is phase-sensitive and used within its limits. A nonlinear controller can leak information into uncontrolled modes. A long loop can return old noise with a phase that either helps or hurts.

A coherent controller must be included as part of the physical system. It has a Hamiltonian, coupling operators, initial state, bandwidth, loss channels, and approximations. Treating it as an ideal algebraic operation is usually the first sign that the model is hiding the hard part.

Exercises

Classical or coherent?

An outgoing microwave field is amplified, digitized, processed by an FPGA, and used to set a later drive amplitude. Is this coherent feedback?

Solution

No. The signal is converted into a classical measurement record before the control action. This is measurement-based feedback. The amplifier and digitizer may be quantum-limited or carefully modeled, but the feedback signal used by the controller is classical.

Check the direction of a cascade

For the joint cascaded generator, let $X_1$ act only on system 1, so it commutes with $L_2$. Show that the coupling terms do not change $\dot X_1$.

Solution

The cross terms from the adjoint dissipator are

$$\frac12\left(L_2^\dagger[X_1,L_1]+[L_1^\dagger,X_1]L_2\right).$$

The commutator with $H_{\mathrm{casc}}-H_1-H_2$ is exactly their negative. The remaining equation is the original local generator for system 1. For an operator of system 2 the terms instead reinforce each other. This cancellation explains why both the common-field dissipator and the directional Hamiltonian are required.

Delay limit

A loop has propagation delay $\tau$ comparable to the cavity lifetime $1/\kappa$. Why is a zero-delay Markovian feedback model suspect?

Solution

The cavity can evolve substantially before its emitted field returns. The returning field therefore carries information about an earlier system state, so the present dynamics depends on the past. A zero-delay Markovian model erases that memory and can predict the wrong stability, damping, phase response, and noise spectrum.

Coherent feedback and reservoir engineering

When can a coherent controller be replaced by an engineered dissipator for the plant?

Solution

Only after a controlled elimination. The controller should relax fast compared with the plant, remain near a known reference state, have negligible unresolved delay or memory, and couple in a regime where the reduced generator is valid and completely positive. If the controller stores information on the plant timescale, it should remain part of the system model.

References

• H. J. Carmichael, An Open Systems Approach to Quantum Optics, Springer (1993).
• C. W. Gardiner and M. J. Collett, “Input and output in damped quantum systems: Quantum stochastic differential equations and the master equation,” Physical Review A 31, 3761–3774 (1985).
• C. W. Gardiner, “Driving a quantum system with the output field from another driven quantum system,” Physical Review Letters 70, 2269–2272 (1993).
• J. Gough and M. R. James, “The series product and its application to quantum feedforward and feedback networks,” IEEE Transactions on Automatic Control 54, 2530–2544 (2009). arXiv:0707.0048.
• J. E. Gough, M. R. James, H. I. Nurdin, and J. Combes, “Quantum filtering for systems driven by fields in single-photon states or superposition of coherent states,” Physical Review A 86, 043819 (2012).
• M. R. James, H. I. Nurdin, and I. R. Petersen, “H-infinity control of linear quantum stochastic systems,” IEEE Transactions on Automatic Control 53, 1787–1803 (2008).
• H. Mabuchi, “Coherent-feedback quantum control with a dynamic compensator,” Physical Review A 78, 032323 (2008).
• H. M. Wiseman and G. J. Milburn, Quantum Measurement and Control, Cambridge University Press (2010).