Feedback from Measurement Records

Measurement-based feedback uses a measurement record to choose later operations on the system. In continuous measurement, the loop is conceptually simple:

measurement record -> state estimate -> control action -> new dynamics

The details are not simple. The record is noisy, the measurement has backaction, the controller has delay and finite bandwidth, and the actuator may add its own noise. A trustworthy feedback model must say what record is available, how it is filtered, what information is used causally, and how the control changes the Hamiltonian or open-system generator.

This page is a bridge from continuous trajectories to feedback control. It explains the operational structure needed before using feedback formulas or designing a controller.

Causal Information

Let $\mathcal Y_t$ denote the classical record available through time $t$. A real-time feedback control must be adapted to the record: at time $t$, it can depend only on information already available.

With zero idealized latency, a control amplitude may be written

$$u_t = f_t(\mathcal Y_t).$$

With a delay $\tau$, the causal form is instead

$$u_t = f_t(\mathcal Y_{t-\tau}), \qquad \tau\ge0.$$

This condition is more than bookkeeping. A controller that uses future data is not a real-time feedback controller. It is an offline smoothing or postselection procedure.

The causal state estimate is the filtered state

$$\rho_c(t) = \rho(t\mid\mathcal Y_t),$$

not a smoothed estimate using later data. For the filtering viewpoint, see Quantum Filtering.

From Record to Control

A practical record-based feedback loop has several stages:

• a detector produces a raw voltage, count stream, or quadrature record;
• calibration converts raw units into an idealized record model;
• a filter or estimator converts the record into $\rho_c(t)$ or a smaller set of state variables;
• a control law chooses amplitudes, phases, pulses, thresholds, or dissipative settings;
• an actuator applies a Hamiltonian drive, changes a coupling, opens a loss channel, or triggers a conditional operation.

In equations, a Hamiltonian feedback model often has the form

$$H(t) = H_0 + \sum_a u_a(t)F_a,$$

where the amplitudes $u_a(t)$ depend on the record or the filtered state. If the control depends on the filtered state,

$$u_a(t) = u_a\!\left(\rho_c(t)\right),$$

the controller is using the best state estimate supplied by the measurement model.

The record itself is not the state. For example, a homodyne current contains a conditional signal plus shot noise. Feeding the raw current directly into a drive is possible, but it is a different controller from one that first estimates the state.

Conditioned Closed-Loop Dynamics

Consider a diffusive measurement of one channel $c$ with efficiency $\eta$. Without writing a specific controller, the conditioned closed-loop equation has the schematic form

$$d\rho_c = \mathcal L_{u_t}\rho_c\,dt + \sqrt{\eta}\, \mathcal H[c]\rho_c\,dW_t,$$

where $\mathcal L_{u_t}$ is the generator with the chosen control applied. For Hamiltonian feedback,

$$\mathcal L_{u_t}\rho = - \frac{i}{\hbar} [H_0+\sum_a u_a(t)F_a,\rho] + \cdots ,$$

where the ellipsis denotes the dissipative terms retained in the model.

This equation is closed only after the control law is specified. If $u_t$ depends on $\rho_c(t)$, the conditioned equation is generally nonlinear. If $u_t$ depends directly on a noisy current, care is needed because white-noise currents are distributions rather than ordinary functions.

Ensemble Dynamics with Feedback

The unconditional closed-loop state is the average over records:

$$\rho_{\mathrm{cl}}(t) = \mathbb E[\rho_c(t)].$$

However, it is usually not obtained by simply taking the old master equation and adding a deterministic control Hamiltonian. The control amplitude itself depends on the record, so the averaging must include correlations between the state, the measurement noise, and the applied control.

For Markovian current feedback, special formulas can combine the measurement and feedback into an unconditional master equation with modified Hamiltonian and dissipative terms. Those formulas depend on the record normalization, detection efficiency, and feedback convention. The important structural point is:

$$\text{measure, ignore record} \ne \text{measure, feed record back, then average}.$$

Feedback changes the physical dynamics, not only the observer’s description.

Hamiltonian Feedback

Hamiltonian feedback changes the system through controlled drives. Examples include:

• applying a qubit rotation based on a continuous readout record;
• changing a local oscillator phase adaptively during homodyne detection;
• applying a force to an oscillator from an estimated position or momentum;
• triggering a microwave or laser pulse after a detector click.

The feedback Hamiltonian is often written

$$H_{\mathrm{fb}}(t) = u_tF,$$

where $F$ is the actuator generator. If $u_t$ is a smooth function of a filtered estimate, ordinary stochastic master equation methods are usually adequate. If $u_t$ is proportional to an ideal white-noise current, the product must be interpreted through a consistent stochastic calculus or replaced by a finite-bandwidth controller model.

Hamiltonian feedback can stabilize a state, track a desired trajectory, cool a mode, undo predictable rotations, or compensate for a known measurement backaction. It cannot recover information lost to unmonitored channels.

Event-Based Feedback

For photon counting and other event records, the feedback action may be triggered by clicks. A simple event-based control law has the form

$$\text{if }dN_t=1,\quad \text{apply }U_{\mathrm{click}}.$$

In density-operator language, a click followed by a unitary pulse has the selective update

$$\rho_c \longmapsto \frac{ U_{\mathrm{click}}c\rho_c c^\dagger U_{\mathrm{click}}^\dagger }{ \operatorname{Tr}(c^\dagger c\rho_c) }.$$

No-click intervals can also drive feedback, because absence of events carries information about the state. A threshold rule that reacts only to clicks may ignore useful information contained in the waiting-time distribution.

For the counting record and trajectory dynamics, see Photon Counting and Quantum Jump Trajectories.

Diffusive Feedback

For homodyne, heterodyne, and weak voltage records, feedback is often continuous. A diffusive record has the form

$$dY_t = \mu_t\,dt + dW_t.$$

A controller can use the record increment $dY_t$, the innovation $dY_t-\mu_tdt$, or a filtered estimate such as $\langle A\rangle_c$. These choices are not equivalent:

• record feedback reacts to the noisy measured signal;
• innovation feedback reacts only to the part not predicted by the current estimate;
• state-estimate feedback reacts to a filtered belief about the system.

Finite bandwidth matters. A realistic controller often uses a filtered current

$$J(t) = \int_{-\infty}^t h(t-s)\,dY_s,$$

where $h$ is a causal response function. If the response time is comparable to the system timescale, the controller memory should be included in the model rather than hidden inside a Markovian feedback term.

For the detector models, see Homodyne Detection and Heterodyne Detection.

Feedback and Backaction

Measurement feedback does not make measurement backaction disappear. It uses the record to respond to the backaction that is known to the observer. If the detector is inefficient, some backaction is unobserved. If an environmental channel is unmonitored, its noise cannot be corrected using the measurement record.

This distinction is visible in the information balance:

$$\text{observed information} \le \text{information carried into the environment}.$$

Feedback can be powerful when the relevant disturbance is correlated with the observed record. It is much less effective against noise that leaves no record accessible to the controller.

For the general measurement-disturbance discussion, see Measurement Backaction.

Stabilization and Cooling Examples

A common qubit example is continuous measurement of $\sigma_z$ together with drives around transverse axes. The record estimates the Bloch vector, and the controller chooses a drive that steers the estimated state toward a target. The feedback must be designed with the stochastic nature of the estimate in mind; using the instantaneous noisy current as if it were $\langle\sigma_z\rangle_c$ gives a different and often unstable loop.

For an oscillator, position measurement can be combined with a force approximately proportional to the estimated momentum. In the right regime this damps motion and realizes feedback cooling. The noise floor depends on imprecision, backaction, efficiency, and actuator noise.

In optical and microwave cavities, photon detections or quadrature records can trigger pulses, phase shifts, or adaptive measurements. Circuit-QED experiments use similar ideas with digitized microwave records and real-time electronics.

These examples share the same structure: the record is not the control objective by itself. The objective is a dynamical property of the system, and the record is the information channel through which the controller infers how to act.

Feedback Versus Reservoir Engineering

Measurement-based feedback and reservoir engineering can produce similar-looking stabilized states, but the mechanisms differ.

In measurement-based feedback, a classical record is acquired, processed, and used to choose later actions. The controller must be modeled as part of the experiment if timing, memory, noise, or thermodynamic costs matter.

In Reservoir Engineering, the dissipation itself is designed so that the target state or subspace is attractive without reading out a classical record. The stabilization is autonomous once the engineered couplings are in place.

The distinction is operational:

$$\text{measurement-based feedback: monitored record plus controller;}$$ $$\text{reservoir engineering: designed open-system generator.}$$

Hybrid schemes exist. A monitored auxiliary mode can supply a feedback signal, while an unmonitored auxiliary mode can act as an engineered reservoir. The model should state which degrees of freedom are measured, which are traced out, and which are retained as controllers.

Thermodynamic Bookkeeping

Feedback can change energy flows, entropy production, and apparent cooling limits. The record, controller memory, actuator, and reset procedure are physical resources. Claims about work extraction or cooling by feedback are incomplete unless they specify where information is stored and how the controller is maintained.

For open-system thermodynamic bookkeeping, see Energy, Heat, and Work and Fluctuation Theorems.

Common Mistakes

• Using a smoothed estimate that depends on future data and calling it real-time feedback.
• Ignoring detector latency, finite bandwidth, digitization, or actuator delay.
• Treating the raw noisy current as the system expectation value.
• Forgetting that no-click intervals and innovation residuals are part of the feedback-relevant record.
• Assuming feedback can correct noise carried only by unmonitored channels.
• Averaging a feedback trajectory as if the control were independent of the measurement record.
• Confusing measurement-based feedback with autonomous reservoir engineering.
• Claiming feedback cooling or work extraction without including the measurement record, memory, controller, and reset costs.

Cross-Links

• Measurement Records for record types, calibration, and finite bandwidth.
• Quantum Filtering for causal state estimation from records.
• Bayesian Quantum Measurement for likelihoods and conditional updates.
• Stochastic Master Equations for conditioned dynamics driven by records.
• Diffusive Trajectories and Quantum Jump Trajectories for the two main record classes.
• Measurement-Based Feedback for control objectives, loop architecture, stabilization, and cooling.
• Reservoir Engineering for autonomous dissipative stabilization.
• Dynamical Decoupling for open-loop pulse-based noise filtering.
• Approximation Checklist for Markov, efficiency, positivity, and timescale checks.

References

• H. M. Wiseman, “Quantum theory of continuous feedback,” Physical Review A 49, 2133–2150, 1994.
• A. C. Doherty and K. Jacobs, “Feedback control of quantum systems using continuous state estimation,” Physical Review A 60, 2700–2711, 1999.
• H. M. Wiseman and G. J. Milburn, Quantum Measurement and Control, Cambridge University Press, 2010.
• K. Jacobs, Quantum Measurement Theory and its Applications, Cambridge University Press, 2014.
• A. Barchielli and M. Gregoratti, Quantum Trajectories and Measurements in Continuous Time, Springer, 2009.
• L. Bouten, R. van Handel, and M. R. James, “An introduction to quantum filtering,” SIAM Journal on Control and Optimization 46, 2199–2241, 2007.
• V. P. Belavkin, “Quantum stochastic calculus and quantum nonlinear filtering,” Journal of Multivariate Analysis 42, 171–201, 1992.

Exercises

1. A proposed controller sets $u_t=f(\mathcal Y_{t+\Delta t})$ with $\Delta t\gt0$. Why is this not a real-time feedback law?

Solution

At time $t$, the record through $t+\Delta t$ is not yet available. The law uses future data, so it is acausal as a real-time controller. It could describe offline smoothing, retrospective analysis, or a delayed implementation after time $t+\Delta t$, but it is not feedback available at time $t$.

2. In a diffusive loop, compare feedback based on $dY_t$ with feedback based on the innovation $dY_t-\mu_tdt$.

Solution

The record increment $dY_t$ contains both the predicted signal $\mu_tdt$ and the stochastic innovation. Feedback based directly on $dY_t$ reacts to the full measured signal, including parts already expected from the current state estimate. Feedback based on $dY_t-\mu_tdt$ reacts only to the newly surprising part of the record. These controllers generally produce different closed-loop dynamics.

3. A photon-counting controller applies a pulse after each click but ignores intervals with no clicks. Name one situation where the no-click information matters.

Solution

If the current conditional state predicts a high click rate, a long no-click interval is evidence that the system may have moved into a dark or low-emission state. Ignoring the interval discards information that should change the state estimate. This is especially important in quantum-jump trajectories, shelving dynamics, and feedback designed to stabilize dark states.

4. Why does finite detection efficiency limit feedback performance?

Solution

Finite efficiency means that only part of the information carried away by the monitored channel reaches the controller. The unobserved part still causes decoherence or damping, but the controller cannot condition on its detailed outcome. Feedback can respond to observed backaction and observed signals, not to information that was lost before reaching the record.