Measurement-Based Feedback

Measurement-based feedback uses information acquired during a run to choose later controls. A detector produces a classical record, a filter or estimator converts that record into usable information, and a controller applies a Hamiltonian drive, pulse, dissipative change, or conditional operation. The loop is useful only to the extent that the record contains information about the disturbance or state variable one wants to control.

The control viewpoint is:

monitored output -> causal estimate -> control law -> changed dynamics

This page treats measurement-based feedback as a design strategy inside quantum control. The detailed trajectory language lives in Feedback from Measurement Records, Quantum Filtering, and Stochastic Master Equations. Here the main questions are what the loop is meant to accomplish, what information it can actually use, and how its performance should be checked.

What Makes Feedback Quantum?

Classical feedback uses a measured signal to control a system. Quantum feedback adds three constraints.

First, the measurement changes the state. The same record that supplies information also represents backaction on the system.

Second, the state estimate is conditional. The controller normally acts on $\rho_c(t)$ , the state conditioned on the available record, not on the ensemble state $\rho(t)$ .

Third, some disturbances are unobserved. Detector inefficiency, loss, thermal noise, dephasing, and unmonitored decay produce entropy that no real-time controller can infer from the retained record.

Thus the promise of feedback is not “measurement removes noise.” The promise is more modest and more precise:

if the useful disturbance is correlated with an available record,
a causal controller can use that record to improve the dynamics

This distinction prevents a common overclaim. Feedback can undo known backaction, stabilize conditional states, cool a mode, or correct detected jumps. It cannot reverse information that has already escaped into an unmonitored environment.

Causality and Information State

Let $\mathcal Y_t$ be the record available to the controller through time $t$ . A real-time feedback law must be adapted to that record:

u(t) = f_t(\mathcal Y_t).

With a delay $\tau$ , the implementable law is instead

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

The filtered state is

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

and a state-estimate controller may be written

u(t) = g_t(\rho_c(t)).

This notation is compact, but it encodes a physical requirement. The controller may use only information that has reached the controller and has been processed within its latency and bandwidth. A smoothed estimate $\rho(t\mid\mathcal Y_T)$ with $T\gt t$ is often excellent for offline analysis, but it is not available for control at time $t$ .

Loop Architecture

A feedback design should specify each layer of the loop.

Layer	Typical object	Questions to state
system model	$H$ , $L_\alpha$ , monitored channels	What is the controlled Hilbert space, and what channels are observed?
record model	$dY_t$ , $dN_t$ , digitized voltages	What detector normalization, efficiency, bandwidth, and noise are assumed?
estimator	$\rho_c(t)$ , Bloch-vector estimate, Kalman-like variables	Is the controller using raw records, innovations, or a filtered state?
control law	$u(t)$ , pulse trigger, feedback gain	What function of the available information is applied?
actuator	Hamiltonian drive, force, phase shift, coupling switch	What physical operation is actually delivered?
validation	fidelity, occupation, stability, robustness	What baseline and uncertainty tests show improvement?

Leaving out any layer can hide the actual bottleneck. For example, a formally stabilizing feedback law may fail because the measured quadrature is not the one carrying the relevant information, or because the actuator cannot respond on the required timescale.

Conditioned Closed-Loop Equation

For a diffusive monitored channel $c$ with efficiency $\eta$ , a schematic stochastic master equation is

d\rho_c = \mathcal L_{u(t)}\rho_c\,dt + \sqrt{\eta}\, \mathcal H[c]\rho_c\,dW_t.

Here $\mathcal L_{u(t)}$ includes the Hamiltonian and dissipative terms with the current control applied, and $\mathcal H[c]$ is the measurement innovation superoperator. If the control is Hamiltonian,

H(t) = H_0+u(t)F,

then the generator contains

\mathcal L_{u(t)}\rho \supset - \frac{i}{\hbar} [H_0+u(t)F,\rho].

The equation becomes a closed-loop equation only after $u(t)$ is defined as a causal functional of the record or of $\rho_c(t)$ . If $u(t)=g(\rho_c(t))$ , the conditioned evolution is generally nonlinear. If $u(t)$ is proportional to an idealized white-noise current, the model must state the stochastic convention or replace the ideal law by a finite-bandwidth controller.

For counting records, a feedback action may be event based:

dN_t=1 \quad\Longrightarrow\quad \rho_c \mapsto \frac{ U_{\mathrm{fb}}c\rho_c c^\dagger U_{\mathrm{fb}}^\dagger }{ \operatorname{Tr}(c^\dagger c\rho_c) }.

The no-click intervals also matter. They update the conditional state because the absence of an expected event is information.

Markovian Current Feedback

An idealized but important limit is instantaneous feedback from a continuous current. Suppose a monitored output with collapse operator $c$ produces a normalized diffusive record, and the current is fed into a Hamiltonian actuator $F$ . In a common Markovian-feedback convention, the unconditional master equation can be written schematically as

\dot\rho = - \frac{i}{\hbar}[H,\rho] + \mathcal D[c-iF]\rho + \frac{1-\eta}{\eta} \mathcal D[F]\rho,

up to convention-dependent factors absorbed into $F$ and the record normalization.

This formula is useful because it shows two structural facts. Feedback changes the average physical dynamics, and inefficiency adds extra noise. In the ideal $\eta=1$ limit the monitored current can be used most effectively; when $\eta\lt1$ , some backaction and environmental information are lost, and feedback through $F$ introduces unavoidable excess diffusion.

The exact coefficients depend on how the current is normalized and how $F$ is defined. A serious calculation should therefore cite the convention and derive the equation from the chosen stochastic master equation rather than transplanting a formula across notations.

Stabilizing a Qubit State

Consider a qubit monitored along $Z$ while driven by feedback around $X$ or $Y$ . The measurement record provides information about the $Z$ coordinate of the Bloch vector, while the drive changes the orientation of the state. A simple feedback goal might be to keep the state near a target point on the Bloch sphere.

The controller can use different information:

raw current feedback reacts quickly but injects measurement shot noise into the drive;
innovation feedback reacts to the part of the record not predicted by the current state estimate;
state-estimate feedback uses the filtered Bloch vector and can incorporate known Hamiltonian drift.

These are not equivalent controllers. If the target is a pure state, inefficiency and unmonitored dephasing make perfect stabilization impossible. The achievable steady fidelity is set by a competition among measurement rate, feedback bandwidth, actuator strength, natural decoherence, and delay.

A useful sanity check is the fast-measurement limit. Stronger measurement gives information faster but also increases measurement-induced diffusion in the conjugate degrees of freedom. Feedback performance improves only if the controller can use the additional information before the added backaction and delay dominate.

Feedback Cooling

Feedback cooling uses a record to infer motion and applies a force that damps it. For an oscillator, a common idealization is continuous position measurement together with a feedback force approximately proportional to the estimated momentum:

F_{\mathrm{fb}}(t) \approx - \Gamma_{\mathrm{fb}} \, \hat p_{\mathrm{est}}(t).

At the level of mean motion this resembles viscous damping. At the level of quantum fluctuations, the limit is set by measurement imprecision, backaction, detector efficiency, delay, and force noise. Increasing measurement strength reduces imprecision but increases backaction; an inefficient detector worsens the tradeoff because some backaction is not observed.

This is why feedback cooling should report more than a damping rate. It should state the final occupation, the measurement efficiency, the actuator noise, and the comparison with passive or reservoir-based cooling. Without those details, “cooling by feedback” can hide heating in another part of the loop.

Continuous Measurement Feedback

In continuous measurement, the controller acts while the record is still being generated. The loop therefore has two simultaneous roles:

estimation: infer the conditional state from a noisy stream;
intervention: change future dynamics based on that estimate.

For a homodyne current

dY_t = \mu_t\,dt+dW_t,

the innovation is

dW_t = dY_t-\mu_tdt.

Using the innovation rather than the raw current can make the controller respond to surprise rather than to expected signal. In a well-calibrated filter, innovation residuals have conditional mean zero and variance $dt$ . Residual drift or correlations are warning signs that the measurement model, efficiency, phase, bandwidth, Hamiltonian, or noise calibration may be wrong.

Digital implementations add another layer. The continuous equation is then an approximation to a sampled loop with finite time step $\Delta t$ , filter memory, processing latency, and waveform-update constraints. The continuum model is reliable only when those scales are small compared with the controlled dynamics or are explicitly included.

Design Objectives

Feedback objectives should be stated in dynamical terms. Common choices include:

stabilizing a target state or subspace;
minimizing final occupation or energy;
tracking a desired trajectory;
suppressing detected jumps or correcting known syndromes;
increasing measurement sensitivity by adapting a probe;
maintaining a nonequilibrium steady state;
reducing variance of an estimated observable;
maximizing a time-averaged fidelity under realistic delay and efficiency.

For state stabilization, a natural score is the long-time average

\overline F = \lim_{T\to\infty} \frac1T \int_0^T \operatorname{Tr}[\rho_c(t)\rho_\star]\,dt,

or the ensemble steady-state fidelity

F_{\mathrm{ss}} = \operatorname{Tr}[\rho_{\mathrm{cl},\mathrm{ss}}\rho_\star],

where $\rho_{\mathrm{cl},\mathrm{ss}}$ is the closed-loop steady state after averaging over records. These scores answer different questions. The first describes performance along monitored trajectories; the second describes the average physical state if the record is ignored after the loop has acted.

Feedback Versus Other Control Strategies

Measurement-based feedback differs from open-loop control because the applied action depends on information gathered during the run. A pulse designed in advance may be robust, but it cannot react to a detector click that happens in one run and not another. For open-loop pulse design, see Pulse Sequences and Optimal Control.

It differs from Reservoir Engineering because the record is acquired, processed, and used as a classical signal. An engineered reservoir can stabilize a target autonomously after the couplings and losses are turned on. Feedback can be more flexible, but it pays for measurement, processing, delay, and actuator imperfections.

It differs from Coherent Feedback because the signal is classical. In coherent feedback, a quantum field or auxiliary system is routed back without first producing a measurement record. Measurement-based feedback discards some quantum information during readout, but it enables amplification, digital processing, nonlinear decision rules, and human-readable diagnostics.

Thermodynamic and Information Costs

Feedback loops use information as a physical resource. A detector records data, a controller stores or processes it, an actuator performs work, and the controller must eventually reset memory or dissipate heat. For many laboratory calculations those costs are negligible compared with the system energy scales; for thermodynamic claims they are central.

A feedback refrigerator, demon-like engine, or adaptive work-extraction protocol is incomplete unless it specifies:

what record is acquired;
where the record is stored;
what work is done by the actuator;
how the controller is reset;
which entropy flows are counted.

For the thermodynamic framing, see Maxwell Demon and Information.

Practical Validation Checklist

Before trusting a feedback claim, check the following.

Record: Which channel is monitored, and with what efficiency, bandwidth, and added noise?
Filter: Is the controller using a calibrated causal estimate, a raw signal, or an offline smoothed estimate?
Delay: Is latency negligible, modeled explicitly, or merely assumed away?
Actuator: What Hamiltonian, force, pulse, or dissipator does the controller actually implement?
Backaction: Which part of the measurement disturbance is observed, and which part is unobserved?
Baseline: What open-loop, passive-cooling, or reservoir-engineering protocol is used for comparison?
Robustness: Does the loop remain stable under detuning, gain error, model mismatch, and noise-rate uncertainty?
Ensemble check: Does the record-averaged closed-loop state match simulations or measured unconditional data?
Resource accounting: Are information, work, and reset costs included when thermodynamic conclusions are drawn?

The Approximation Checklist gives a broader audit trail for Markov, rotating-wave, detector, and coarse-graining assumptions.

Common Mistakes

Treating the noisy detector current as if it were directly equal to an expectation value.
Using an unconditional master-equation state in a controller that should use the filtered state.
Designing with zero delay and infinite bandwidth, then applying the result to a finite-latency experiment.
Claiming real-time feedback while using future records or postselected trajectories.
Ignoring no-click information in an event-based controller.
Assuming feedback can correct noise from unmonitored channels.
Comparing feedback cooling with passive cooling without reporting imprecision, backaction, efficiency, and actuator noise.
Moving a Markovian feedback formula between papers without checking record normalization and stochastic convention.
Treating feedback stabilization as error correction without identifying a protected subspace and recovery condition.

Exercises

Causal or acausal?

A controller sets $u(t)=g(\rho(t\mid\mathcal Y_{t+5\Delta t}))$ , where $\mathcal Y_{t+5\Delta t}$ includes five future digitizer samples. Is this a real-time feedback law?

Solution

No. The controller uses information not available at time $t$ . It is a smoothed or retrospective estimate. It may be useful for offline analysis or for a delayed controller acting after those samples arrive, but it is not a causal feedback law at time $t$ .

Inefficiency and feedback

Why does detector inefficiency limit feedback even if the controller and actuator are otherwise ideal?

Solution

Inefficiency means that only part of the environmental information reaches the record. The unobserved part still produces backaction, damping, or dephasing, but the controller cannot condition on its specific outcome. The filtered state is therefore less pure and less informative, and any feedback action can only respond to the observed component of the disturbance.

Feedback cooling tradeoff

A position-measurement feedback cooler reduces imprecision by increasing measurement strength. Why might the final occupation stop improving?

Solution

Stronger measurement usually increases measurement backaction, and with inefficiency some of that backaction is unobserved. Finite controller bandwidth and delay also prevent the actuator from using arbitrarily fast information. At some point additional measurement strength adds more disturbance or loop noise than useful damping, so the final occupation saturates or worsens.

Raw current versus innovation

For a diffusive record $dY_t=\mu_tdt+dW_t$ , what is the difference between feeding back $dY_t$ and feeding back $dY_t-\mu_tdt$ ?

Solution

Feedback from $dY_t$ reacts to the full measured increment, including the predicted signal $\mu_tdt$ and the noise. Feedback from $dY_t-\mu_tdt$ reacts to the innovation, the part of the record not predicted by the current filtered state. The two controllers can have different stability, noise injection, and steady-state behavior.

Cross-Links

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).
L. Bouten, R. van Handel, and M. R. James, “An introduction to quantum filtering,” SIAM Journal on Control and Optimization 46, 2199-2241 (2007).
R. Vijay, C. Macklin, D. H. Slichter, S. J. Weber, K. W. Murch, R. Naik, A. N. Korotkov, and I. Siddiqi, “Stabilizing Rabi oscillations in a superconducting qubit using quantum feedback,” Nature 490, 77-80 (2012).
R. Ruskov and A. N. Korotkov, “Quantum feedback control of a solid-state qubit,” Physical Review B 66, 041401(R) (2002).