Krotov Method for Quantum Control

• Krotov Method is an iterative optimization framework that guarantees monotonic improvement through sequential updates in quantum control problems.
• It efficiently handles both linear and nonlinear dynamics, addressing models like the Schrödinger, Liouville–von Neumann, and Gross–Pitaevskii equations.
• Its modular structure enables adaptations such as frequency filtering, noise-tailored controls, and numerical schemes for open-system and non-Markovian dynamics.

The Krotov method is a global iterative optimization strategy for optimal control problems in quantum dynamics, used to determine time-dependent external fields that steer a system toward a prescribed target while guaranteeing monotonic improvement of an optimization functional for approximately time-continuous controls (Goerz et al., 2019). In quantum control it is formulated as a sequential update algorithm, generally attributed to Krotov, and has been developed for a large class of problems that includes linear and nonlinear Schrödinger equations, non-unitary time evolution, nonlinear dependencies of the Hamiltonian on the control, time-dependent targets, and optimization functionals that depend to higher than second order on the time-evolving states (Schirmer et al., 2011, Reich et al., 2010).

1. Origins and problem domain

Within the review literature, the method is described as a global iterative optimization strategy that systematically improves the control function in optimal control problems, particularly for quantum systems described by the Schrödinger, Liouville–von Neumann, or Gross–Pitaevskii equations (Morzhin et al., 2018). It was initially proposed beyond quantum control in work by V.F. Krotov and I.N. Feldman, and was later used to develop a quantum-control formulation by D.J. Tannor, V. Kazakov, and V. Orlov, and by J. Somlói, V.A. Kazakov, and D.J. Tannor (Morzhin et al., 2018). In the quantum-control literature, it has become a standard monotonic algorithm for state-to-state transfer, quantum gate implementation, robust control, and open-system optimization (Goerz et al., 2019).

The scope of the method is unusually broad. The Konnov–Krotov generalization yields monotonically convergent algorithms not only for the standard linear, closed-system setting, but also for non-linear equations of motion, non-unitary time evolution, non-linear control dependence in the Hamiltonian, time-dependent targets, and higher-order functionals (Reich et al., 2010). Review treatments further emphasize modifications for defining the improvement function, constraints on control spectrum, constraints on the states of a quantum system, and regularizers, together with applications to molecular dynamics, Bose–Einstein condensates, and quantum gate generation (Morzhin et al., 2018).

A central conceptual point is that Krotov’s method is not merely a gradient formula. It is an iterative variational framework in which the control update is constructed together with forward and backward propagation so that the target functional improves monotonically under the assumptions of the chosen formulation (Reich et al., 2010). This distinguishes it from generic gradient-ascent procedures that rely on external line searches or quasi-Newton heuristics to enforce progress (Goerz et al., 2019).

2. Variational formulation and objective functionals

A standard formulation minimizes a functional of the form

$$J[\{\ket{\phi_k^{(i)}(t)}\}, \{\epsilon_l^{(i)}(t)\}] = J_T(\{\ket{\phi_k^{(i)}(T)}\}) + \sum_l \int_0^T g_a(\epsilon_l^{(i)}(t))\,dt + \int_0^T g_b(\{\phi_k^{(i)}(t)\})\,dt,$$

where $J_T$ is the final-time objective, $g_a$ is a running cost on the controls, and $g_b$ is an optional state-dependent running cost (Goerz et al., 2019). A common choice is a quadratic running cost on the change of the control from a reference field,

$$g_a(\epsilon_l^{(i)}(t)) = \lambda_{a,l} S_l(t) \left(\epsilon_l^{(i)}(t)-\epsilon_{l,\mathrm{ref}}^{(i)}(t)\right)^2,$$

with $S_l(t)$ an update shape function and $\lambda_{a,l}$ an inverse step size (Goerz et al., 2019). In application papers the same structure appears with different notational conventions, for example as a penalty on rapid changes or excessive values in the control field, or as a penalty on deviation from the previous iterate (Song, 16 Feb 2026, Jäger et al., 2014).

For closed systems, the dynamics are typically written in Schrödinger form with a Hamiltonian linear in the controls,

$$\frac{\partial}{\partial t}\ket{\Psi(t)} = -\frac{i}{\hbar}H(t)\ket{\Psi(t)}, \qquad H(t)=H_0+\epsilon_1(t)H_1+\epsilon_2(t)H_2+\dots,$$

where the control fields may represent laser amplitudes, phases, voltages, trap parameters, or exchange couplings (Goerz et al., 2019). In review treatments, equivalent formulations are given for density-matrix dynamics and for the Gross–Pitaevskii equation, the latter introducing a nonlinear mean-field term (Morzhin et al., 2018).

For open systems, the method is commonly embedded in a master-equation setting. A representative Markovian form is the Lindblad equation,

$$\frac{d\rho}{dt} = -\frac{i}{\hbar}[H,\rho] + \frac{1}{2}\sum_j \gamma_j \left( 2L_j\rho L_j^\dagger - L_j^\dagger L_j\rho - \rho L_j^\dagger L_j \right),$$

with collapse operators $L_j$ encoding dephasing, amplitude damping, or other dissipative channels (Fonseca et al., 2022). For non-Markovian systems, the method has been reformulated for a time-nonlocal master equation by transforming it to a set of coupled linear time-local equations in an extended auxiliary Liouville space (Hwang et al., 2012).

The final-time objective depends on the task. In state preparation it is often an infidelity such as $J_T$0 (Song, 16 Feb 2026). In gate optimization it may be expressed through overlaps between propagated basis states and target states, or through average-gate-fidelity functionals (Le et al., 23 Sep 2025). In continuous-variable and Gaussian settings, it can be a norm on the difference between final and target covariance matrices, or a mixed functional involving fidelity and logarithmic negativity (Chen et al., 2024, Luo et al., 31 Oct 2025).

3. Sequential update structure and monotonic convergence

The hallmark features of Krotov’s method are sequential field updates at each time step, monotonic convergence for approximately time-continuous controls, and the absence of a global step-size line search (Goerz et al., 2019). A typical first-order update has the form

$J_T$1

with

$J_T$2

Here $J_T$3 denotes the forward-propagated state under the updated controls, and $J_T$4 the backward-propagated co-state under the previous controls (Goerz et al., 2019). The terminal condition for the co-state is

$J_T$5

This forward/backward coupling is the operational core of the method (Goerz et al., 2019).

For the standard quantum-control problem of a convex final-time functional, linear equations of motion, and linear dependency of the Hamiltonian on the field, the second-order contribution is not required for monotonic convergence, although it can be used to speed up convergence (Reich et al., 2010). When the problem involves nonlinear equations of motion, non-unitary dynamics, higher-than-second-order functionals, or indefinite state-dependent costs, the Konnov–Krotov construction introduces a second-order term in the auxiliary functional through a parameter $J_T$6, yielding a provably monotonic algorithm for a substantially larger class of problems (Reich et al., 2010).

The method is often presented through an auxiliary functional or “improvement function” $J_T$7, linear or linear-quadratic in the states, from which the backward equation, control update, and monotonicity conditions are derived (Morzhin et al., 2018, Reich et al., 2010). In review treatments, this structure is explicitly connected to Pontryagin-type maximization of a control Hamiltonian or Pontryagin function (Morzhin et al., 2018, Morzhin et al., 2023).

A recurrent practical issue is the role of regularization and discretization. One line of work argues that, in the practically relevant case of finite time resolution, ad-hoc penalty terms are undesirable and unnecessary, and that theory should be developed directly in discrete time with exact discrete gradients and inexpensive dynamic search-length control (Schirmer et al., 2011). In that formulation, monotonicity is associated with the actual fidelity rather than merely a regularized surrogate, and the update can be understood as a sequential ascent in the discrete control parameters (Schirmer et al., 2011). This suggests that “Krotov’s method” names a family of sequential monotonic update schemes rather than a single immutable formula.

4. Constraints, open-system extensions, and specialized variants

The method has been repeatedly extended to handle experimentally motivated constraints and dynamical structures. A particularly important modification imposes strict limitations on the spectrum of the optimized laser fields without losing monotonic convergence, by incorporating a frequency-filter constraint directly into the Krotov functional and filtering the updated field after each iteration (0801.3935). In the review literature, additional modifications incorporate state constraints, regularizers penalizing deviation from the previous iterate, and spectral kernels that turn the control update into an integral equation (Morzhin et al., 2018).

Setting	Characteristic formulation	Representative work
Closed quantum systems	Schrödinger, Liouville–von Neumann, or Gross–Pitaevskii dynamics; linear or linear-quadratic improvement function	(Morzhin et al., 2018)
Bandwidth-limited fields	Frequency-filter constraint with monotonic convergence retained under spectral restrictions	(0801.3935)
Open Markovian systems	Lindblad master equation; unitary and non-unitary optimization compared directly	(Fonseca et al., 2022)
Non-Markovian systems	Extended auxiliary Liouville space or iteration-dependent recomputation of dissipative terms	(Hwang et al., 2012, Luo et al., 31 Oct 2025)
Gaussian-state covariance dynamics	Vectorized covariance-matrix propagation, avoiding Fock basis cutoffs	(Chen et al., 2024)

In open Markovian control, the difference between unitary and non-unitary optimization is task dependent. For target-state preparation, controls optimized with open-system dynamics included outperform controls obtained from unitary optimization and then deployed in noise (Fonseca et al., 2022). For quantum-gate implementation, however, this is not always true: if all states are equally affected by dissipation, or if there are no leakage states, unitary and non-unitary optimization can perform almost identically; if computational and leakage subspaces are differently affected by dissipation, non-unitary optimization becomes effective because it can exploit less dissipative pathways (Fonseca et al., 2022).

Non-Markovian settings require more substantial reformulation. One approach expands the bath correlation function into exponentials, promoting the time-nonlocal master equation to a time-local system in an extended auxiliary Liouville space and then applying Krotov updates to that enlarged dynamical system (Hwang et al., 2012). Another, developed for oscillator chains, recomputes the dissipative operator after every iteration because the dissipative terms depend on the precise time profile of the control fields; in that setting the update equation does not require the functional derivative of the dissipative operator with respect to the control, and monotonic convergence is retained (Luo et al., 31 Oct 2025). A plausible implication is that the method remains viable even when the environment is dynamically entangled with the control representation, provided the control-dependent memory kernel is handled consistently at each iteration.

5. Numerical realization, software, and relation to neighboring methods

The numerical behavior of Krotov’s method depends strongly on how time discretization, propagation, and update length are handled. In the discrete-time “unencumbered” formulation, control fields are parameterized directly on a finite grid, exact discrete gradients replace naive continuum-gradient scaling by $J_T$8, and local quadratic models are used for search-length control (Schirmer et al., 2011). That analysis shows that the penalty term traditionally introduced in the continuum theory is superfluous or even harmful in the finite-resolution regime, and that forward-only or split sequential updates can outperform the classical forward-backward pattern, especially near high fidelity (Schirmer et al., 2011).

A separate line of work compares Krotov with GRAPE for Bose–Einstein condensates. There, both methods are derived from the variational principle, but they differ in the way the control is updated, additional costs are accounted for, and second-order derivative information can be included (Jäger et al., 2014). The reported comparison is sharp: GRAPE produces smoother control fields and works in a black-box manner, whereas Krotov with a suitably chosen step-size parameter converges faster but can produce sharp features in the control fields (Jäger et al., 2014). This contrast recurs across applications: Krotov privileges monotonic sequential improvement, while GRAPE more readily absorbs derivative penalties and concurrent updates.

Software support has made the method operationally accessible. The open-source Python package krotov implements state-to-state transfer, quantum gate implementation, optimization toward an arbitrary perfect entangler, and interfaces with QuTiP objects and custom propagators (Goerz et al., 2019). The package foregrounds the standard first-order update, shape functions, inverse step sizes, and optional second-order constructions for non-convex objectives (Goerz et al., 2019).

Recent numerical analysis has also targeted the propagation step itself. Commutator-free Cayley integrators reformulate forward and backward propagation so that unitarity and symmetry are preserved at the discrete level while eliminating matrix exponentials and commutators (Wembe et al., 12 Mar 2026). In the reported experiments, the CF-Cayley formulation achieves the same accuracy as high-order exponential or Cayley–Magnus schemes at substantially lower cost, and in nonlinear Schrödinger and Gross–Pitaevskii settings it preserves norm and stability through a Cayley-polynomial interpolation strategy (Wembe et al., 12 Mar 2026). This locates Krotov’s method within a broader program that couples optimal-control updates to geometric integration.

6. Applications, empirical behavior, and recurrent misconceptions

The contemporary application record is diverse. In cold-atom interferometry, the method has been used to design robust Raman pulses by optimizing the temporal shape of both amplitude and phase. In numerical simulations, the optimized pulses maintain high atomic manipulation fidelity over an extended range of laser frequency detunings and intensity fluctuations, and in full interferometer sequences this robustness yields a significant enhancement in final fringe contrast under a systematic detuning (Song, 16 Feb 2026). The same study interprets the resulting gain as improved signal-to-noise ratio and precision for atomic sensors (Song, 16 Feb 2026).

In open-system qubit platforms, Krotov-based optimization has been used to adapt controls to specific noise sources. For exchange-coupled surface qubits, optimized operations with $J_T$9 are reported to be feasible, including $g_a$0 for NOT gates in two-qubit systems and $g_a$1 for CNOT gates in three-qubit systems under relevant experimental parameters (Le et al., 23 Sep 2025). In simpler open qubit and qutrit models, non-unitary optimization is found to outperform unitary optimization for state preparation, while the advantage for gate implementation depends on whether dissipation is uniform or whether leakage/non-computational levels are differently affected by noise (Fonseca et al., 2022). A common misconception is therefore that “including noise in the optimization is always better”; the more precise statement is conditional on the available dynamical pathways (Fonseca et al., 2022).

For non-Markovian transport and continuous-variable control, the method has been used to transfer entanglement in oscillator chains by tuning only on-site frequencies while keeping coupling strengths constant, and to generate optomechanical entanglement directly at the covariance-matrix level without Fock-basis cutoffs (Luo et al., 31 Oct 2025, Chen et al., 2024). In both cases, memory effects are reported to be beneficial rather than purely detrimental: the controlled non-Markovian system can reach smaller residuals than the Markov case in oscillator chains, and non-Markovian memory reduces entanglement decay in optomechanics (Luo et al., 31 Oct 2025, Chen et al., 2024). This suggests that Krotov’s method is particularly effective when the control objective can exploit dynamical structure in the environment instead of merely suppressing it.

The method has also been adapted to compilation problems. The Jenga–Krotov algorithm for exchange-only qubits uses a long, dense Krotov-optimized sequence followed by systematic pruning and re-optimization. For the Toffoli gate, it reduces the number of required exchange unitaries from 216 to 92 and compresses the time steps from 162 to 50 while maintaining target fidelity; under realistic noise, the accumulated gate error is an order of magnitude lower than in conventional decompositions (Wu et al., 16 Jul 2025). Here Krotov supplies the monotonic local optimization layer, while the “Jenga” stage performs global sequence reduction.

Comparative studies delimit the regime in which Krotov is most effective. In discretized state-preparation problems, deep Q-learning and policy-gradient methods outperform Krotov when the control is restricted to discrete values and when the problem scales up, whereas Krotov excels for continuous control with many time steps (Zhang et al., 2019). In the “Quantum Moves” controversy, a basic Krotov implementation seeded with random controls performs on par with GRAPE and clearly outperforms both human players and the KASS heuristic; the poor behavior of KASS is attributed to its seeding and sweep strategy rather than to an intrinsic limitation of Krotov’s method (Grønlund, 2019). Another recurrent misconception is therefore that adverse performance of a specific heuristic variant can be read as a verdict on the method itself.

Across these developments, a stable picture emerges. Krotov’s method is best understood as a family of monotonic, sequential optimal-control schemes whose practical character depends on the choice of functional, update law, regularization, and propagator. In smooth, continuously parameterized control landscapes it is often highly effective; in strongly discretized, heavily bounded, or basis-restricted settings it may lose ground to alternatives better matched to that control geometry (Jäger et al., 2014, Zhang et al., 2019). Its importance lies less in any single update equation than in the variational architecture that makes monotonic optimization compatible with closed, open, nonlinear, and even non-Markovian quantum dynamics.