Control Limits from Noise

Quantum control is limited by the information, energy, bandwidth, time, and noise resources actually available to the experiment. A control protocol can average slow noise, exploit measurement records, route quantum fields, engineer useful reservoirs, or optimize waveforms. It cannot make unobserved entropy disappear, deliver infinite-bandwidth pulses, infer a lost record, or validate a model it never tests.

The honest control question is therefore not:

How do we make the error vanish?

It is:

Which errors can this control resource move, average, monitor, or convert,
and which errors remain in the budget?

This page is the practical limit map for the chapter. It is not a replacement for Optimal Control, Dynamical Decoupling, Measurement-Based Feedback, or Reservoir Engineering. It explains what prevents those tools from being arbitrarily powerful.

What Counts as a Limit

A control limit is always relative to a model and a resource set. Relevant resources include:

allowed Hamiltonian terms and dissipative couplings;
maximum amplitude, slew rate, and bandwidth;
measurement efficiency, dynamic range, and latency;
bath spectra and correlation times;
Hilbert-space cutoffs and leakage channels;
calibration accuracy and drift time;
available time before relaxation or dephasing dominates.

A useful error budget often has the schematic form

\epsilon_{\mathrm{tot}} \approx \epsilon_{\mathrm{decoh}} + \epsilon_{\mathrm{leak}} + \epsilon_{\mathrm{bw}} + \epsilon_{\mathrm{meas}} + \epsilon_{\mathrm{cal}} + \epsilon_{\mathrm{model}}.

This is bookkeeping, not a theorem. The terms can be correlated, and the same physical imperfection may appear under several names. The point is to avoid hiding a dominant error behind a high-fidelity simulation that omitted it.

For channels and fidelity conventions, use Fidelity. For model-assumption checks, use the Approximation Checklist.

Time-Scale Hierarchy

Most control claims stand or fall on time scales. A protocol should state at least:

Symbol	Meaning	Typical failure if ignored
$T$	operation or stabilization time	decoherence accumulates during the protocol
$T_1$	energy-relaxation time	jumps or loss events are treated as reversible
$T_\phi$	pure-dephasing time	phase noise is mistaken for calibration error
$\tau_B$	bath correlation time	Markovian and non-Markovian regimes are confused
$\tau_{\mathrm{bw}}$	actuator or filter response time	programmed pulses are not delivered
$\tau_{\mathrm{lat}}$	feedback latency	controller reacts to stale information
$\tau_{\mathrm{drift}}$	calibration-drift time	optimized pulses overfit yesterday’s Hamiltonian

Fast control helps only when the relevant resource is actually fast compared with the error process. A pulse sequence can refocus slow detuning drift. It cannot refocus an unobserved Markovian jump after the emitted quantum has left the apparatus.

Decoherence During Control

Even a perfectly calibrated control Hamiltonian acts for a finite time. If a qubit suffers Markovian pure dephasing at rate $\Gamma_\phi$ , then in the dephasing basis

\rho_{01}(T) = e^{-\Gamma_\phi T} e^{-i\varphi(T)} \rho_{01}(0),

where $\varphi(T)$ includes any coherent phase from the control frame. A protocol that takes time $T$ cannot preserve arbitrary coherence better than this model permits unless it changes the noise model, monitors the environment, encodes the information, or uses a different physical degree of freedom.

For energy relaxation with instantaneous excited-state population $p_e(t)$ , the no-jump approximation gives a jump probability

p_{\mathrm{jump}} \simeq \int_0^T dt\, \Gamma_1 p_e(t),

when the probability is small. If the emitted photon, phonon, quasiparticle, or bath excitation is unobserved, an open-loop pulse generally cannot know which trajectory occurred. This is why reducing protocol duration and reducing excited-state occupation are real control objectives, not cosmetic optimizations.

The relevant canonical pages are Steady States and Relaxation, Dephasing versus Dissipation, and the Lindblad–GKSL Equation.

Spectral Limits

Control is most effective when the unwanted noise has structure. For dephasing noise filtered by a toggling function $y(t)$ , the Gaussian estimate is

\chi(T) = \frac12 \int_{-\infty}^{\infty} \frac{d\omega}{2\pi}\, S_{\xi\xi}(\omega) |Y(\omega,T)|^2,

with

Y(\omega,T) = \int_0^T dt\, y(t)e^{i\omega t}.

This formula says that pulses reshape exposure in frequency space. It does not say that pulses eliminate the spectrum. If $S_{\xi\xi}(\omega)$ is effectively flat over all frequencies sampled by the sequence, Parseval’s identity gives

\int_{-\infty}^{\infty} \frac{d\omega}{2\pi} |Y(\omega,T)|^2 = \int_0^T dt\,|y(t)|^2.

For ideal sign flips, $|y(t)|=1$ , so the total exposure remains $T$ . The sequence can move sensitivity around, but it cannot reduce truly memoryless white dephasing in this simple model.

For a slow spectrum, such as quasi-static drift or low-frequency $1/f$ -like noise, the same filter formalism explains why echo and dynamical decoupling can help. The distinction is spectral, not rhetorical. See Noise Spectra, One-Over-F Noise, and Dynamical Decoupling.

Amplitude and Speed

Finite amplitudes imply finite operation times. For a resonantly driven two-level system with rotating-frame control

H_c(t) = \frac{\hbar}{2} \left[ \Omega_x(t)X+\Omega_y(t)Y \right],

and amplitude bound

\sqrt{\Omega_x(t)^2+\Omega_y(t)^2} \le \Omega_{\max},

an ideal resonant $\pi$ rotation cannot be faster than

T_\pi \ge \frac{\pi}{\Omega_{\max}}.

More generally, closed-system quantum speed-limit bounds relate state separation to available energy scale. One Mandelstam–Tamm form is

T \ge \frac{\hbar\,\mathcal L}{\overline{\Delta E}}, \qquad \mathcal L = \arccos |\langle\psi_0|\psi_T\rangle|,

where $\overline{\Delta E}$ is a time-averaged energy uncertainty. Such bounds are useful sanity checks, but they do not by themselves certify an experimental pulse. A larger drive can shorten $T$ while increasing leakage, heating, counter-rotating terms, Stark shifts, or calibration sensitivity.

The practical limit is a tradeoff:

too slow -> decoherence dominates
too strong -> leakage and hardware errors dominate

Good control design searches the region between those failures.

Bandwidth and Delivered Waveforms

The waveform produced by a generator is not necessarily the waveform delivered to the system. A linear hardware model writes

u_{\mathrm{del}}(t) = \int_{-\infty}^{\infty} ds\, h(t-s)u_{\mathrm{prog}}(s),

or in frequency space,

\widetilde u_{\mathrm{del}}(\omega) = H_{\mathrm{hw}}(\omega) \widetilde u_{\mathrm{prog}}(\omega).

Sharp edges, narrow spikes, and fast phase jumps are filtered if the hardware transfer function is small in the corresponding band. A pulse optimized without the transfer function may be an excellent mathematical waveform and a poor physical waveform.

Bandwidth limits appear in feedback as well. If the useful record is $\mathcal Y_t$ but the loop has latency $\tau_{\mathrm{lat}}$ , the implementable law is

u(t) = f_t(\mathcal Y_{t-\tau_{\mathrm{lat}}}).

The controller acts on old information. For slow stabilization this may be acceptable. For fast quantum trajectories, the delay can turn feedback into added noise or instability.

Leakage and Off-Resonant Excitation

Few experimental Hilbert spaces are exactly two-dimensional. Suppose a drive intended for a transition also couples weakly to a nearby unwanted transition detuned by $\Delta_{\mathrm{leak}}$ . In a simple perturbative square-pulse estimate,

\epsilon_{\mathrm{leak}} \sim \left( \frac{\Omega}{\Delta_{\mathrm{leak}}} \right)^2,

up to pulse-shape and phase factors. This scaling is not universal, but the lesson is robust: increasing the Rabi rate can reduce decoherence time while increasing off-resonant population.

Leakage can be reduced by pulse shaping, derivative corrections, larger anharmonicity, optimized constraints, or dissipative removal. Each cure has a cost. Shaping usually lengthens the pulse or increases bandwidth demands. Dissipative removal may add noise. Strong confinement can create Zeno-like dynamics but leaves finite leakage corrections, often scaling inversely with the constraint rate.

The relevant pages are Driven Open Systems, Pulse Sequences, Optimal Control, and Quantum Zeno Dynamics.

Measurement Backaction

Measurement-based control is limited by the measurement’s information-disturbance tradeoff. A detector that learns more about one variable generally disturbs conjugate or incompatible variables, and inefficiency means some disturbance is not recorded.

For an idealized continuous linear position detector with no useful imprecision-backaction correlations, a quantum-limited form is

S_{xx}^{\mathrm{imp}}(\omega) S_{FF}^{\mathrm{ba}}(\omega) \ge \frac{\hbar^2}{4}.

Here $S_{xx}^{\mathrm{imp}}$ is imprecision noise and $S_{FF}^{\mathrm{ba}}$ is backaction force noise. More general detectors can have correlations and different conventions, but the structural message remains: information acquisition has a dynamical cost.

In continuous measurement notation, a monitored channel with efficiency $\eta$ contributes schematically

d\rho_c = \mathcal L\rho_c\,dt + \sqrt{\eta}\, \mathcal H[c]\rho_c\,dW_t.

The observed record improves the conditioned state only through the fraction captured by $\eta$ . The unobserved part still causes decoherence or diffusion. Feedback can act on the record it has, not on the record that was lost.

This is why statements such as “measure harder and feedback faster” need qualification. Stronger measurement can reduce imprecision, but it can also increase backaction, detector heating, bandwidth pressure, and loop instability. See Measurement Backaction and Measurement-Based Feedback.

Model Error and Calibration Drift

An optimized pulse is only as reliable as the model and validation set used to design it. Let $u_\star(t)$ be optimized at parameter value $\theta_0$ . The reported fidelity

F[u_\star;\theta_0]

does not answer how the pulse behaves under detuning, drift, amplitude miscalibration, crosstalk, bath-rate uncertainty, or transfer-function error. A more useful robustness score samples a test set:

F_{\mathrm{worst}} = \min_{\theta\in\Theta_{\mathrm{test}}} F[u_\star;\theta].

The test set should not be identical to the training grid used during optimization. Otherwise a pulse can overfit a discretized uncertainty model. Robust control therefore needs cross-validation in parameter space, baseline pulses, sensitivity plots, and experimental calibration checks.

Model error is especially dangerous in open systems because omitted channels are often invisible in a closed-system simulation. A pulse can look perfect in a Hamiltonian model while failing through relaxation, heating, leakage, non-Markovian memory, or measurement-induced diffusion.

Strategy-by-Strategy Limits

Strategy	What it can suppress	Typical limiting factor
Rabi and Ramsey control	calibrated rotations and spectroscopy errors	detuning, decay during pulses, amplitude noise
Dynamical decoupling	slow or structured dephasing	pulse errors, white noise, relaxation, bandwidth
Pulse shaping	leakage and spectral spillover	longer duration, transfer functions, calibration drift
Optimal control	constrained waveform-design errors	model mismatch, local optima, omitted noise channels
Measurement-based feedback	disturbances correlated with a record	inefficiency, latency, imprecision, backaction
Coherent feedback	routed quantum-field dynamics	loss, phase stability, network modeling errors
Reservoir engineering	entropy removal and autonomous stabilization	unwanted steady states, added noise, finite gap
Dissipative state preparation	attractive target states	nonunique dark spaces, slow convergence, natural decay
Zeno control	leakage between sectors	finite constraint rate, dephasing, anti-Zeno regimes

The table should be read as a diagnostic aid, not a taxonomy wall. Real experiments combine strategies, and the same physical component can be a control drive, reservoir, detector, and noise source depending on how it is used.

Practical Validation Workflow

A trustworthy control claim should report:

the target state, unitary, channel, observable, or steady state;
the physical model, frame convention, and Hilbert-space truncation;
all included decoherence channels and their calibration source;
amplitude, bandwidth, slew-rate, phase, and duration constraints;
measurement efficiency, imprecision, backaction, and latency for feedback;
delivered waveform or transfer-function model, not only programmed waveform;
leakage outside the intended subspace;
robustness sweeps over detuning, amplitude, drift, and noise rates;
comparison with a simple baseline protocol;
whether the quoted fidelity is simulated, measured, postselected, or conditioned.

If a paper or calculation omits one of these items, the result may still be valuable. But the omission should be treated as an open assumption rather than silently converted into evidence of unlimited controllability.

Common Mistakes

Treating ideal controllability as evidence of high-fidelity control in a noisy device.
Reporting a pulse fidelity without operation duration, decoherence rates, or leakage.
Assuming stronger drive always helps.
Optimizing a waveform before applying the hardware transfer function.
Claiming dynamical decoupling suppresses decoherence without specifying the noise spectrum.
Treating feedback as if it could correct unmonitored environmental noise.
Using measurement strength as a free resource while ignoring backaction and inefficiency.
Validating robustness on the same parameter grid used to train the pulse.
Comparing protocols with different postselection or conditioning conventions.
Calling a dissipatively stabilized state “protected” without checking the Liouvillian gap and unwanted steady states.

Exercises

Dephasing during a gate

A qubit starts in $\lvert +x\rangle$ . During an otherwise perfect identity operation of duration $T$ , it experiences Markovian pure dephasing in the $Z$ basis at rate $\Gamma_\phi$ . Compute the fidelity with $\lvert +x\rangle$ at the end.

Solution

The initial state is

\rho(0) = \frac12 \begin{pmatrix} 1 & 1\\ 1 & 1 \end{pmatrix}.

Pure dephasing leaves populations unchanged and multiplies the off-diagonal terms by $e^{-\Gamma_\phi T}$ :

\rho(T) = \frac12 \begin{pmatrix} 1 & e^{-\Gamma_\phi T}\\ e^{-\Gamma_\phi T} & 1 \end{pmatrix}.

The fidelity with $\lvert +x\rangle$ is

F = \langle +x\rvert\rho(T)\lvert +x\rangle = \frac12 \left( 1+e^{-\Gamma_\phi T} \right).

For $\Gamma_\phi T\ll1$ ,

1-F \simeq \frac{\Gamma_\phi T}{2}.

The error is proportional to operation time in this Markovian model.

Low-pass bandwidth

A first-order hardware response has

H_{\mathrm{hw}}(\omega) = \frac{1}{1+i\omega/\omega_c}.

For a Fourier component of the programmed waveform at angular frequency $\omega$ , what condition keeps the delivered amplitude at least $90$ percent of the programmed amplitude?

Solution

The magnitude is

|H_{\mathrm{hw}}(\omega)| = \frac{1} {\sqrt{1+(\omega/\omega_c)^2}}.

The condition $|H_{\mathrm{hw}}|\ge0.9$ gives

\frac{1} {\sqrt{1+(\omega/\omega_c)^2}} \ge 0.9.

Therefore

1+\left(\frac{\omega}{\omega_c}\right)^2 \le \frac{1}{0.9^2},

\frac{\omega}{\omega_c} \le \sqrt{\frac{1}{0.9^2}-1} \approx 0.48.

Fourier components near or above the cutoff are already substantially distorted.

Unobserved relaxation

During a control pulse, a qubit has average excited-state population $\overline p_e=1/2$ for time $T$ . It relaxes at rate $\Gamma_1$ , and emitted quanta are not monitored. Estimate the small-probability jump rate contribution to the error budget.

Solution

Using

p_{\mathrm{jump}} \simeq \int_0^T dt\, \Gamma_1p_e(t),

and replacing $p_e(t)$ by its time average gives

p_{\mathrm{jump}} \simeq \Gamma_1T\overline p_e = \frac{\Gamma_1T}{2}.

Since the emitted quantum is unobserved, an open-loop controller generally cannot know which run jumped. The exact fidelity loss depends on the target, but this probability is a necessary part of the error budget.

Feedback cannot use lost information

A continuous measurement has efficiency $\eta=0.25$ . Explain why a perfect classical controller cannot make the conditioned state as pure as it would be for $\eta=1$ .

Solution

Efficiency $\eta=0.25$ means only one quarter of the measurement information reaches the record used by the controller. The remaining backaction is real but unobserved. A perfect controller can process the available record without delay or computational error, but it cannot condition on outcomes that were lost to inefficiency. Thus the filtered state remains more mixed than in the ideal $\eta=1$ case, and feedback based on that state has a lower achievable stabilization fidelity.

Cross-Links

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).
A. A. Clerk, M. H. Devoret, S. M. Girvin, F. Marquardt, and R. J. Schoelkopf, “Introduction to quantum noise, measurement, and amplification,” Reviews of Modern Physics 82, 1155–1208 (2010).
C. Brif, R. Chakrabarti, and H. Rabitz, “Control of quantum phenomena: past, present and future,” New Journal of Physics 12, 075008 (2010).
S. J. Glaser, U. Boscain, T. Calarco, C. P. Koch, W. Kockenberger, R. Kosloff, I. Kuprov, B. Luy, S. Schirmer, T. Schulte-Herbrüggen, D. Sugny, and F. K. Wilhelm, “Training Schrödinger’s cat: quantum optimal control,” European Physical Journal D 69, 279 (2015).
N. Khaneja, T. Reiss, C. Kehlet, T. Schulte-Herbrüggen, and S. J. Glaser, “Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent algorithms,” Journal of Magnetic Resonance 172, 296–305 (2005).
L. Viola and S. Lloyd, “Dynamical suppression of decoherence in two-state quantum systems,” Physical Review A 58, 2733–2744 (1998).
L. Mandelstam and I. Tamm, “The uncertainty relation between energy and time in non-relativistic quantum mechanics,” Journal of Physics (USSR) 9, 249–254 (1945).
N. Margolus and L. B. Levitin, “The maximum speed of dynamical evolution,” Physica D 120, 188–195 (1998).
S. Deffner and S. Campbell, “Quantum speed limits: from Heisenberg’s uncertainty principle to optimal quantum control,” Journal of Physics A 50, 453001 (2017).