Quantum control asks a precise engineering question: given a system Hamiltonian and a limited set of controls, which quantum transformations are reachable, and how should the controls vary in time to reach the target accurately, quickly and robustly?
The problem has two layers. Controllability is structural: does the available Hamiltonian algebra generate the desired unitary group at all? Optimal control is constructive: among the reachable solutions, which pulse achieves the best fidelity subject to duration, amplitude, bandwidth, leakage and robustness constraints?
This guide develops bilinear control Hamiltonians, dynamical Lie algebras, piecewise-constant propagation, fidelity objectives, GRAPE gradients, line-search updates, robust ensemble optimisation, leakage penalties, bandwidth constraints and quantum-speed-limit caveats. It complements Guide 4 on unitary evolution and Guide 24 on benchmarking; here the owner is deliberate control synthesis.
Available generators → Lie closure → reachable set → discretise controls → propagate → differentiate fidelity → update pulse → validate against constraints and uncertainty.
1. Bilinear control model
A standard finite-dimensional model is
H(t)=H_0+Σ_{j=1}^m u_j(t)H_j.
H0 is the drift Hamiltonian. The Hj are control Hamiltonians. Real control amplitudes uj(t) are functions we may shape within hardware limits.
In units ℏ=1, the propagator obeys
dU/dt=−iH(t)U(t), U(0)=I.
The target may be one state, one unitary gate, an ensemble average, or a robust worst-case performance criterion.
2. Reachability before optimisation
An optimiser cannot create a transformation forbidden by the available generators.
Construct the skew-Hermitian generators
iH_0, iH_1, …, iH_m.
Take all real linear combinations and repeated commutators. The resulting smallest closed Lie algebra is the dynamical Lie algebra 𝔤.
If 𝔤=su(d) for a d-dimensional system, arbitrary special-unitary evolution is reachable under standard compact-system assumptions. If 𝔤=u(d), arbitrary unitary evolution including global phase is generated. Smaller algebras imply restricted reachable sets.
3. Why commutators create new directions
For short time ε, the group commutator
e^{Aε}e^{Bε}e^{-Aε}e^{-Bε}=e^{[A,B]ε²+O(ε³)}.
Alternating available controls therefore synthesises motion along the commutator [A,B], even if that direction does not appear directly in the laboratory Hamiltonian.
Nested commutators generate higher-order directions. Lie closure converts this local motion picture into a global controllability test.
4. Worked one-qubit controllability
Suppose the drift and control are proportional to Z and X:
H_0=(ω/2)Z, H_1=(1/2)X.
The skew-Hermitian generators contain iZ and iX. Their commutator is
[iZ,iX]=−[Z,X]=−2iY.
Thus {iX,iY,iZ} are generated, spanning su(2). The qubit is fully controllable up to global phase.
If only Z drift and Z control were available, all generators would commute and the reachable unitaries would be phase rotations about one Bloch-sphere axis only.
5. Two-qubit entangling control
Local single-qubit controls alone cannot create entanglement from a product state.
Add an interaction such as
H_int=J Z⊗Z.
Combined with sufficiently rich local controls, commutators generate a much larger algebra and can reach entangling gates.
A practical controllability calculation should exploit exact symmetries first. Conserved excitation number, parity or permutation symmetry can block the Hilbert space and reduce the relevant algebra.
6. Controllability is not the same as easy controllability
A target can be reachable in principle yet require:
- more time than the coherence window;
- control amplitudes beyond hardware limits;
- bandwidth the electronics cannot supply;
- intermediate population of leakage levels;
- extreme sensitivity to calibration errors.
Lie-algebraic reachability answers a yes/no structural question. Engineering feasibility is a constrained optimisation question.
7. Piecewise-constant controls
Divide total duration T into N intervals of width Δt=T/N.
Approximate each control as constant within interval k:
u_j(t)=u_{j,k}.
The interval Hamiltonian is
H_k=H_0+Σ_j u_{j,k}H_j
and propagator
U_k=e^{-iH_kΔt}.
The full evolution is ordered:
U(T)=U_NU_{N-1}…U_1.
8. Gate fidelity objective
For a target unitary W on dimension d, a convenient phase-insensitive overlap is
Φ=|Tr(W†U(T))|²/d².
0≤Φ≤1 and Φ=1 when U=eiφW.
This Hilbert–Schmidt overlap is useful in control optimisation but is not identical to every experimental average gate-fidelity definition. Conversion to average gate fidelity depends on dimension and whether the implemented map is truly unitary.
9. State-transfer objective
If only one input state |ψ₀⟩ must reach target |ψt⟩, use
Φ_state=|⟨ψ_t|U(T)|ψ_0⟩|².
State controllability can therefore be easier than full unitary controllability. A pulse that transfers one state perfectly may implement the wrong phases on every orthogonal state.
10. GRAPE
GRAPE—gradient ascent pulse engineering—optimises the discretised amplitudes uj,k by evaluating the gradient of a fidelity objective efficiently using forward and backward propagated products. Khaneja and collaborators introduced the influential NMR formulation in 2005. [1]
Instead of recomputing the entire propagator from scratch for every one of mN control variables, GRAPE reuses partial products.
11. Forward and backward products
Define forward products
F_k=U_kU_{k-1}…U_1
and backward products
B_k=U_NU_{N-1}…U_{k+1}.
Then
U(T)=B_k U_k F_{k-1}.
Changing one amplitude uj,k changes only Uk directly. The stored F and B factors make every gradient component cheap relative to a full independent propagation.
12. First-order GRAPE derivative
For sufficiently short Δt,
∂U_k/∂u_{j,k}≈−iΔt H_j U_k
when commutator corrections across the interval are neglected.
More accurate implementations use exact Fréchet derivatives of the matrix exponential or diagonalise Hk to include finite-step effects.
The approximation order must therefore match pulse bandwidth and discretisation size.
13. Gradient of the phase-sensitive overlap
Define
z=Tr(W†U(T)).
For Φ=|z|²/d²,
∂Φ/∂u=(2/d²)Re[z* ∂z/∂u].
With the forward/backward split,
∂z/∂u_{j,k}=Tr[W†B_k(∂U_k/∂u_{j,k})F_{k-1}].
This is the core computational economy of GRAPE.
14. Gradient ascent update
A basic update is
u_{j,k}^{new}=u_{j,k}^{old}+η ∂Φ/∂u_{j,k}
with step size η.
Modern optimisers may use conjugate gradients, L-BFGS, trust regions or adaptive learning rates. A larger step is not automatically better: overshoot can reduce fidelity or violate amplitude constraints.
15. Worked finite-difference gradient check
Before trusting an analytic GRAPE gradient, verify one component numerically:
g_FD=[Φ(u+εδ_{jk})−Φ(u−εδ_{jk})]/(2ε).
If the analytic gradient gives 0.031842 and ε=10−6 gives finite-difference 0.0318417, the relative mismatch is about 9×10−6.
Repeating this spot check for several random control variables is a strong implementation test. A gradient optimiser with a sign error can otherwise converge confidently in the wrong direction.
16. Amplitude constraints
Hardware typically enforces
|u_{j,k}|≤u_max.
One can project each update back into the allowed interval, reparameterise with a bounded function such as tanh, or use a constrained optimiser.
Post-hoc clipping after optimisation can destroy a high-fidelity pulse because the optimiser relied on forbidden peaks. Constraints should be present during synthesis.
17. Bandwidth and smoothness
Piecewise-constant pixels can alternate violently from one interval to the next, even when laboratory electronics cannot reproduce those edges.
Add a smoothness penalty such as
P_smooth=λΣ_k(u_{k+1}−u_k)²
or parameterise the pulse directly in a band-limited basis.
The optimisation target then becomes fidelity minus hardware-cost penalties rather than fidelity alone.
18. Leakage
Real qubits are often embedded inside larger Hilbert spaces. A transmon, for example, has |0⟩,|1⟩ logical levels plus |2⟩,|3⟩,… leakage levels.
If Pcomp projects onto the computational subspace, one can penalise population outside it:
P_leak=Σ_k [1−Tr(P_compρ_k)].
A final-state-only fidelity can hide large transient leakage that increases sensitivity to relaxation or calibration error.
19. Robust control as ensemble optimisation
Suppose the true system has uncertain detuning Δ or amplitude scale a.
Sample an ensemble of models θr and optimise
Φ_robust=(1/R)Σ_r Φ(θ_r).
The same pulse is propagated through every model. Gradients are averaged.
This sacrifices some nominal peak fidelity to improve performance over a calibration range.
20. Average robustness can hide a worst case
An average objective can tolerate one poor corner if other samples perform extremely well.
For safety-critical or thresholded performance, one may optimise a soft minimum, conditional value-at-risk, or explicitly maximise the worst sampled fidelity:
max_u min_r Φ_r(u).
Worst-case optimisation is harder and can be nonsmooth, but it matches a different engineering contract.
21. Control landscapes
For ideal unconstrained controllable finite systems, important quantum-control landscapes have favourable critical-point structure: suboptimal local traps can be absent under strong assumptions.
Real constraints change the landscape. Limited bandwidth, short duration, amplitude bounds, leakage, model error and discretisation can introduce difficult local optima or flat directions.
A failed optimiser does not by itself prove the target is uncontrollable; it may indicate a poor parameterisation or constrained landscape.
22. Quantum speed limits
Quantum speed-limit inequalities bound the minimum time for a state to move a specified distinguishability distance given energy or Hamiltonian-norm constraints.
The Mandelstam–Tamm form for time-independent H gives
T≥arccos|⟨ψ_0|ψ_T⟩| / ΔE
in ℏ=1 units under the corresponding assumptions.
Control-constrained minimum gate times depend on the allowed Hamiltonian set and are generally more specific than one universal speed-limit formula.
23. Pontryagin viewpoint
Optimal-control theory can also be formulated through Pontryagin’s maximum principle.
Introduce a costate and control Hamiltonian; an optimal trajectory must satisfy coupled state/costate equations plus a pointwise extremality condition on the control.
For bounded controls, this often explains bang-bang structure in time-optimal problems.
GRAPE is numerical gradient optimisation; Pontryagin supplies necessary structural conditions. They are complementary, not competing theories.
24. Dynamical decoupling
A fast sequence of control pulses can average unwanted Hamiltonian terms toward zero in a toggling frame.
For example, a π pulse X changes Z→−Z. Alternating free evolution before and after the pulse can cancel quasistatic Z detuning to first order.
Dynamical decoupling is therefore control used for noise filtering rather than target-gate synthesis, though robust gates often combine both ideas.
25. Filter-function language
For weak classical dephasing noise with power spectral density S(ω), a control sequence defines a filter function F(ωT).
Decoherence exponents often take a form schematically like
χ(T)∝∫_0^∞ dω S(ω)F(ωT)/ω².
Pulse design can therefore be interpreted spectrally: suppress filter weight where environmental noise is strongest.
26. Closed-loop calibration
Model-based optimal control is limited by model error.
Closed-loop calibration evaluates the pulse on hardware, estimates a cost, and updates parameters using experimental feedback.
Hybrid workflows use a trusted model to reach the correct region quickly, then tune a lower-dimensional pulse family in situ.
The optimisation target should match the observable that matters in hardware rather than only simulated state-vector overlap.
27. Common misconception: controllable means any gate can be implemented instantly
Controllability concerns reachability over admissible time evolution. Minimum time, amplitude, bandwidth and robustness are separate constraints.
28. Common misconception: higher simulated fidelity always means a better laboratory pulse
An optimiser can exploit neglected leakage, unmodelled bandwidth, calibration errors or unrealistic amplitude edges. Hardware-aware validation is part of the control problem.
29. Common misconception: GRAPE’s gradient is automatically exact
The common first-order derivative is accurate only when time pixels are sufficiently short. Exact matrix-exponential derivatives or smaller Δt may be needed for strong controls and coarse discretisation.
30. Worked synthesis problem
A qubit has drift H₀=0.4Z and one control H₁=X. Total gate time is split into N=100 equal pixels.
Step 1: Structural test. iZ and iX generate iY by commutator, so the dynamical Lie algebra is su(2); arbitrary one-qubit special unitaries are reachable in principle.
Step 2: Discretisation. Each pixel has H_k=0.4Z+u_kX and U_k=e^{-iH_kΔt}.
Step 3: Target objective. For target Hadamard up to global phase, optimise Φ=|Tr(Had†U)|²/4.
Step 4: Constraint. If hardware requires |uk|≤1, enforce that during every update rather than clipping only the final pulse.
Step 5: Robustness. Re-optimise or validate over drift values 0.38,0.40,0.42 and amplitude scales 0.98,1.00,1.02. The robust pulse may have slightly lower nominal Φ but better guaranteed hardware performance.
31. Practice set
- Write the bilinear control Hamiltonian.
- What is the dynamical Lie algebra?
- Why do commutators matter for reachability?
- What algebra implies full special-unitary controllability on dimension d?
- How is a piecewise-constant propagator built?
- State one phase-insensitive unitary-overlap objective.
- What computational trick makes GRAPE gradients efficient?
- When is
∂U_k/∂u≈−iΔtH_jU_konly approximate? - Why should amplitude limits appear inside optimisation?
- What does ensemble robust optimisation do?
- Why can average robustness hide a problem?
- What is the difference between controllability and a quantum speed limit?
Answers
H(t)=H₀+Σ_j u_j(t)H_j.- The real Lie algebra generated by iH₀ and the iHj under linear combinations and commutators.
- Alternating available generators synthesises new motion directions proportional to their commutators.
- su(d), up to global phase.
- Multiply interval exponentials in time order:
U_N…U_1. |Tr(W†U)|²/d².- Reuse forward and backward partial propagator products.
- When Δt is not small enough for commutator corrections to be negligible.
- A pulse depending on forbidden amplitudes may lose fidelity when clipped after the fact.
- Optimises one control waveform over a sampled family of uncertain Hamiltonians/calibrations.
- High-performing samples can compensate numerically for one poor corner.
- Controllability asks whether a target is reachable at all; a speed limit asks how quickly it can be reached under specified constraints.
Sources and further study
[1] Navin Khaneja, Timo Reiss, Cindie Kehlet, Thomas Schulte-Herbrüggen and Steffen 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). The foundational GRAPE formulation.
[2] Domenico D’Alessandro, Introduction to Quantum Control and Dynamics, CRC Press. A systematic treatment of dynamical Lie algebras, controllability and control synthesis.
[3] Herschel Rabitz and colleagues, Whither the future of controlling quantum phenomena?, Science 288, 824–828 (2000). A broad optimal-control perspective linking theory and experiment.
[4] Christiane P. Koch and colleagues, Quantum optimal control in quantum technologies, EPJ Quantum Technology 9, 19 (2022). A modern review of quantum-control methods, constraints and technology applications.
Continue through Quantum Mathematics
Guide 54: Continuous Quantum Measurement, Stochastic Master Equations, Quantum Filtering and Feedback closes the loop by updating state estimates from noisy observations. Guide 55: Quantum Error Mitigation, Zero-Noise Extrapolation, Probabilistic Error Cancellation and Symmetry Verification treats noisy outputs after control. Guide 56: Quantum Contextuality, Kochen–Specker Theorem, Compatibility Graphs and Contextual Fractions develops a foundational resource beyond Bell nonlocality.