Continuous Monitoring

Continuous monitoring is the limit in which a measuring apparatus obtains information from a quantum system through many weak time steps rather than one isolated projective readout. The output is not a single outcome. It is a time record, and the quantum state used for future predictions is conditioned on that record.

The central objects are:

• a measurement record $Y_{[0,t]}$;
• a conditional state $\rho_c(t)$ given that record;
• an unread or ensemble state $\rho(t)$ obtained by averaging over records;
• a measurement strength that sets the rate at which information and backaction accumulate;
• a stochastic update law that keeps track of the unpredictable part of the record.

Continuous monitoring is the operational bridge between weak measurement, open-system master equations, quantum trajectories, filtering, and feedback control. It is also the setting where the phrase “observing the system” becomes mathematically precise: one must specify which output channel is monitored, with what efficiency, and how the resulting record is used.

From One Outcome to a Record

For a single discrete measurement, an outcome $m$ is represented by an operation $\mathcal I_m$. If outcome $m$ is known, the state is updated to

$$\rho_m = \frac{\mathcal I_m(\rho)} {\operatorname{Tr}\mathcal I_m(\rho)}.$$

For monitoring, apply this idea over many short intervals of length $\Delta t$. A record over $[0,t]$ is a sequence

$$Y_{[0,t]} = (r_1,r_2,\ldots,r_n), \qquad t=n\Delta t.$$

At step $j$, the apparatus reports a small outcome $r_j$: a click or no-click result, a voltage increment, a homodyne current sample, or another short-time signal. The conditional state after that step has the same normalized-instrument structure as an ordinary selective measurement:

$$\rho_c(t+\Delta t) = \frac{\mathcal I_{r}^{\Delta t}(\rho_c(t))} {\operatorname{Tr}\mathcal I_{r}^{\Delta t}(\rho_c(t))}.$$

The superscript $\Delta t$ matters. A continuous measurement is not obtained by repeating a strong projective measurement infinitely often unless that is the physical model intended. In most laboratory continuous measurements, each short-time interaction is weak: one time step extracts little information, but the accumulated record can become decisive.

Weak Steps and the Continuous Limit

For a one-Kraus-outcome description in a short time interval, write

$$\mathcal I_r^{\Delta t}(\rho) = M_r^{\Delta t}\rho M_r^{\Delta t\dagger}.$$

The likelihood of the observed increment is

$$p(r|\rho_c(t)) = \operatorname{Tr} \!\left[ M_r^{\Delta t}\rho_c(t)M_r^{\Delta t\dagger} \right].$$

The continuous limit is arranged so that the nonselective map per time step is close to the identity:

$$\sum_r M_r^{\Delta t}\rho M_r^{\Delta t\dagger} = \rho + \mathcal L\rho\,\Delta t + o(\Delta t),$$

or, for a continuous-valued short-time record, the sum is replaced by an integral over $r$. The generator $\mathcal L$ is often a Lindblad–GKSL generator after Markov and rotating-wave approximations have been made. The measurement record then refines the unconditional master equation into many possible conditioned histories.

The scaling is the subtle point. In a diffusive record, the noise increment is of order $\sqrt{dt}$ even though the drift is of order $dt$. In a counting record, most intervals contain no event, but the probability of one event is of order $dt$. Both limits lead to stochastic equations rather than ordinary deterministic differential equations.

Conditional State

The conditional state is the state assigned after using the record up to time $t$:

$$\rho_c(t) = \rho\bigl(t\,\big|\,Y_{[0,t]}\bigr).$$

It is “conditional” in the same operational sense as an ordinary selective measurement state. It is the state used to predict future measurements for an observer who has access to the record. It is not the same object as the ensemble density matrix when different records are possible.

For a future outcome $m$ at time $t$, the observer with record $Y_{[0,t]}$ predicts

$$p(m|Y_{[0,t]}) = \operatorname{Tr} \!\left[ E_m\rho_c(t) \right],$$

where $E_m$ is the relevant effect for the future measurement. If another observer ignores the monitoring record, that observer generally assigns a different state $\rho(t)$ and therefore different conditional predictions.

This does not mean the two observers disagree about the experimental statistics after they condition on the same data. They are using different information sets. Continuous monitoring is a quantum version of filtering: update the state estimate as data arrive.

Ensemble State

If the record is ignored, the conditional states are averaged with their physical probabilities:

$$\rho(t) = \mathbb E_{\mathrm{records}}\!\left[\rho_c(t)\right].$$

The unread state $\rho(t)$ obeys the unconditional master equation,

$$\frac{d\rho}{dt} = \mathcal L\rho,$$

while the conditioned states obey stochastic equations. The average-over-records identity is the most important consistency check on any trajectory model:

$$\boxed{\rho(t)=\mathbb E[\rho_c(t)]}$$

should hold with the same generator $\mathcal L$ that describes the unmonitored open system.

This statement is not a license to average stochastic equations by dropping all noise by inspection. Normalized stochastic master equations are often nonlinear in $\rho_c$, and the Itô rules or counting rules are part of the calculation. The detailed forms are treated in Stochastic Master Equations.

Measurement Strength

Measurement strength is the rate at which the apparatus extracts information about the monitored observable or output channel. Different communities normalize this rate differently, but the physical role is stable:

• stronger monitoring separates possible records faster;
• stronger monitoring produces faster measurement-induced dephasing in incompatible coherences;
• inefficient detection keeps some backaction while discarding part of the information.

For a common diffusive convention for monitoring a Hermitian observable $A$,

$$dY_t = 2\sqrt{\eta\Gamma_{\mathrm m}}\, \langle A\rangle_c\,dt + dW_t,$$

where $0\le\eta\le1$ is the efficiency, $\Gamma_{\mathrm m}$ is a measurement-rate parameter, and $dW_t$ is a Wiener increment. The signal drift is proportional to the conditional expectation $\langle A\rangle_c$, while the record noise has variance $dt$:

$$\mathbb E[dW_t]=0, \qquad dW_t^2=dt.$$

If two candidate states have different values of $\langle A\rangle_c$, their integrated records separate only gradually. The signal-to-noise ratio grows like a square root in time:

$$\mathrm{SNR}(T) \sim 2\sqrt{\eta\Gamma_{\mathrm m}T}\, \bigl|\Delta\langle A\rangle\bigr|.$$

Numerical factors depend on record convention. The invariant lesson is that information is accumulated over a measurement time, not obtained as a noiseless instantaneous readout.

Counting and Diffusive Records

Two record types appear so often that they deserve separate names.

In a counting record, the increment $dN_t$ records whether an event occurred:

$$dN_t\in\{0,1\}, \qquad \mathbb E[dN_t|\rho_c(t)] = \lambda_t\,dt.$$

Here $\lambda_t$ is the conditional rate. Photon counting and some tunneling measurements are modeled this way. A click causes a jump update; a no-click interval also updates the state because absence of an expected event is information. The canonical event-conditioned treatment is Quantum Jump Trajectories.

In a diffusive record, the increment is a noisy continuous signal:

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

The drift $\mu_t$ depends on the conditional state, and the innovation

$$dW_t = dY_t-\mu_t\,dt$$

drives the state update. Homodyne detection, heterodyne detection, continuous weak qubit readout, and quantum-state diffusion are examples. The detailed stochastic forms are collected in Diffusive Trajectories.

Backaction Is Not Optional

Continuous monitoring is not passive observation. Even when each time step is weak, the measurement interaction changes the conditional state. Sometimes the change is dramatic, as in a photon-counting click. Sometimes it is gradual, as in a diffusive readout that slowly localizes the state in a measured basis.

Backaction has two related faces:

• information backaction: the state changes because the record changes the observer’s conditional predictions;
• physical disturbance: the system becomes entangled with detector or field degrees of freedom, producing dephasing or dissipation when those degrees of freedom are ignored.

The distinction matters because inefficient monitoring can leave physical disturbance without giving the observer all the corresponding information. For example, a decay channel with collapse operator $L$ may contribute $\mathcal D[L]\rho$ to the unconditional master equation even when only a fraction of the emitted photons are detected. The detected fraction conditions the state; the undetected fraction remains unmonitored decoherence.

This is why the right mathematical object is an instrument, not merely a POVM. The record probabilities and the post-record state updates must be specified together. See Measurement Backaction and Quantum Instruments.

Unraveling an Unconditional Equation

A Lindblad master equation does not by itself specify a unique monitored dynamics. The same unconditional generator can correspond to different measurement arrangements on the environment:

• counting emitted photons gives jump trajectories;
• measuring an output quadrature gives diffusive trajectories;
• ignoring the output gives no conditional trajectory at all;
• mixing channels, adding loss, or changing detector phase changes the record and the conditioned state.

This choice is called an unraveling of the master equation. It should be tied to a measurement model, not treated as a matter of taste. The unconditional density matrix describes what remains after records are ignored; it does not tell you which record was available.

The physical relation between system operators and monitored output fields is supplied, in quantum optics, by Input–Output Theory.

Simple Example: No-Click Information

Consider a two-level atom with spontaneous emission collapse operator

$$L=\sqrt{\gamma}\,|g\rangle\langle e|.$$

During a short interval $dt$, a detected photon corresponds to a jump toward $|g\rangle$. But suppose no photon is detected. For ideal monitoring and ignoring the Hamiltonian, the unnormalized no-click state is generated by the effective non-Hermitian Hamiltonian

$$H_{\mathrm{eff}} = - \frac{i\hbar}{2}L^\dagger L = - \frac{i\hbar\gamma}{2} |e\rangle\langle e|.$$

For a pure state

$$|\psi\rangle = \alpha|g\rangle+\beta|e\rangle,$$

the no-click update gives, to first order in $dt$,

$$|\tilde\psi_{\mathrm{no}}\rangle = \alpha|g\rangle + \left(1-\frac{\gamma dt}{2}\right)\beta|e\rangle.$$

After normalization, the excited-state weight is reduced. The detector did not click, but the absence of a click is evidence against the atom being excited. Continuous monitoring turns silence into data.

What Changes in the Laboratory

Real records are calibrated electrical or optical signals, not abstract increments by themselves. The mathematical record may be obtained after subtracting offsets, normalizing shot-noise units, coarse graining detector bandwidth, or choosing a rotating frame. A model should state:

1. Which output channel is monitored?
2. What is the record variable?
3. What is the detection efficiency?
4. Which channels are unmonitored?
5. Which stochastic convention and normalization are being used?
6. Does feedback use the record, or is the record only stored?

Without these choices, phrases such as “the system is continuously measured” are under-specified. They do not determine a unique stochastic equation.

Common Mistakes

• Treating continuous monitoring as nondisturbing because each step is weak.
• Confusing the noisy record $dY_t$ with its conditional mean $\mu_t\,dt$.
• Forgetting that no-click intervals are measurement outcomes.
• Averaging trajectories with the wrong record probabilities.
• Assuming a Lindblad equation determines a unique trajectory picture.
• Calling $\rho_c(t)$ the ensemble state after it has been conditioned on a particular record.
• Ignoring detector inefficiency and thereby overestimating how much information the observer has.
• Mixing Itô and Stratonovich conventions without transforming the equation.
• Using a strong projective update at each time step when the intended model is a weak continuous measurement.

Cross-Links

• State Update Rules for the hierarchy of projective, Kraus, instrument, and continuous-record updates.
• Weak Measurements for low-information measurement steps and their scaling.
• Measurement Records for photon counts, homodyne currents, heterodyne currents, qubit-readout voltages, calibration, and record imperfections.
• Quantum Filtering for recursive state estimation from the observed record.
• Photon Counting for direct-detection count records.
• Unravelings for why the same master equation can correspond to different monitored records.
• Stochastic Master Equations for normalized stochastic differential equations.
• Quantum Jump Trajectories for photon-counting and event-conditioned dynamics.
• Diffusive Trajectories for homodyne, heterodyne, and weak-readout records.
• Lindblad–GKSL Equation for the unconditional Markovian generator.
• Approximation Checklist for checks on Markov, secular, efficiency, and positivity assumptions.

References

• 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.
• H. J. Carmichael, An Open Systems Approach to Quantum Optics, Springer, 1993.
• C. W. Gardiner and P. Zoller, Quantum Noise, 3rd ed., Springer, 2004.
• 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.
• J. Dalibard, Y. Castin, and K. Mølmer, “Wave-function approach to dissipative processes in quantum optics,” Physical Review Letters 68, 580–583, 1992.

Exercises

1. Let $\{\mathcal I_r^{\Delta t}\}$ be the instrument for one short monitoring step. Show that if the record is ignored for two steps, the resulting state is obtained by summing over both short-time outcomes, not by choosing a representative record.

Solution

After one unread step,

$$\rho_1 = \sum_{r_1} \mathcal I_{r_1}^{\Delta t}(\rho_0).$$

After two unread steps,

$$\rho_2 = \sum_{r_2} \mathcal I_{r_2}^{\Delta t}(\rho_1) = \sum_{r_2,r_1} \mathcal I_{r_2}^{\Delta t} \!\left( \mathcal I_{r_1}^{\Delta t}(\rho_0) \right).$$

The unread evolution is therefore an average over all records $(r_1,r_2)$ with their instrument weights. A single typical record gives a conditional state, not the ensemble state.

2. For a counting measurement with collapse operator $c$, ignore the Hamiltonian and suppose no event is detected during $dt$. The unnormalized no-click update is

$$\tilde\rho_{\mathrm{no}} = \rho - \frac12 \{c^\dagger c,\rho\}\,dt.$$

Show that the probability of no click is $1-\operatorname{Tr}(c^\dagger c\rho)dt$ to first order.

Solution

The probability of the no-click outcome is the trace of the unnormalized state:

$$\begin{aligned} p_{\mathrm{no}} &= \operatorname{Tr}\tilde\rho_{\mathrm{no}} \\ &= \operatorname{Tr}\rho - \frac{dt}{2} \operatorname{Tr} \!\left( c^\dagger c\rho+\rho c^\dagger c \right). \end{aligned}$$

Using $\operatorname{Tr}\rho=1$ and cyclicity of the trace,

$$\operatorname{Tr}(\rho c^\dagger c) = \operatorname{Tr}(c^\dagger c\rho),$$

so

$$p_{\mathrm{no}} = 1-\operatorname{Tr}(c^\dagger c\rho)dt.$$

The complementary probability is the click probability to first order.

3. In the diffusive convention

$$dY_t = 2\sqrt{\eta\Gamma_{\mathrm m}}\, a\,dt + dW_t,$$

suppose the monitored system is known to be in one of two states with constant values $a_0$ and $a_1$ of the monitored observable. Compute the separation of the means of the integrated record $Y_T$ and compare it with the noise standard deviation.

Solution

Integrating from $0$ to $T$ gives

$$Y_T = 2\sqrt{\eta\Gamma_{\mathrm m}}\, aT + W_T.$$

For the two hypotheses, the mean separation is

$$\Delta\overline Y_T = 2\sqrt{\eta\Gamma_{\mathrm m}}\, |a_1-a_0|\,T.$$

Since $W_T$ has variance $T$, the noise standard deviation is $\sqrt T$. Thus

$$\mathrm{SNR} = \frac{\Delta\overline Y_T}{\sqrt T} = 2\sqrt{\eta\Gamma_{\mathrm m}T}\, |a_1-a_0|.$$

The distinguishability grows as $\sqrt T$, which is characteristic of noisy continuous monitoring.

4. A monitored pure state may remain pure along each ideal trajectory. Does that imply the unread ensemble state is pure? Explain.

Solution

No. The unread state is

$$\rho(t)=\mathbb E[\rho_c(t)].$$

Even if every $\rho_c(t)$ is a rank-one projector, the average of different projectors is generally mixed. Purity of each ideal conditioned trajectory is therefore compatible with decoherence in the unread ensemble.