Robust Control in Many-Body Quantum Systems

• Many-body robust control is a framework for engineering control protocols that maintain performance amid uncertainties such as decoherence, static perturbations, and parasitic couplings.
• It employs diverse techniques including gradient-based optimization, tensor-network methods, and reinforcement learning to tackle state transfer, entanglement control, and crosstalk suppression.
• Applications range from fixed-time state transfer in oscillator chains and GHZ state preparation in qubit arrays to engineered photonic lattices, all while addressing scalability and structural challenges.

Many-body robust control denotes the design of control fields, protocols, or engineered environments for interacting quantum systems such that a target task remains effective in the presence of uncertainty, decoherence, static perturbations, parasitic couplings, or model mismatch. In the recent literature, the topic encompasses fixed-time state transfer in coupled oscillator chains, GHZ-state preparation in interacting multi-qubit devices, suppression of many-body quantum crosstalk in qubit chains, adiabatic state preparation under static imperfections, self-correcting reinforcement-learning controllers, reduced-order non-Markovian control, density-based control derived from time-dependent density-functional theory, and nanophotonic or Kerr-soliton platforms where the many-body Hamiltonian itself is engineered for robustness (Yu et al., 2021, Le et al., 2021, Le et al., 4 Mar 2026, Zeng et al., 13 Jun 2025, Metz et al., 2022, Nielsen et al., 2014, Jin et al., 23 Apr 2026).

1. Definitions and problem classes

A central feature of the subject is that “robustness” is not defined by a single universal criterion. In a coupled harmonic-oscillator chain, robustness is defined operationally as the persistence of high fidelity for cat-state transfer and phase-space rotation when controls optimized for the closed system are applied unchanged to an open system coupled to a non-Markovian bosonic bath (Yu et al., 2021). In interacting multi-qubit systems for sensing, robustness is posed as a max–min problem: the control is chosen to maximize the worst-case fidelity over a hypercube of uncertain couplings, rather than the average over a nominal Hamiltonian (Le et al., 2021). In tensor-network control of quantum crosstalk, robustness is instead enforced by minimizing ensemble-averaged gate or state infidelity over random realizations of parasitic Heisenberg couplings (Le et al., 4 Mar 2026).

The task class is equally broad. Recent work includes transfer and rotation of Schrödinger cat states in a continuous-variable chain (Yu et al., 2021), GHZ preparation on a star graph of capacitively coupled transmons for Heisenberg-limited sensing (Le et al., 2021), parallel $X$ gates, parallel CNOT gates, GHZ preparation, and Heisenberg ground-state preparation on chains of up to 50 qubits under unknown crosstalk (Le et al., 4 Mar 2026), and adiabatic preparation of symmetry-broken many-body states in Ising chains and Rydberg arrays under static perturbations (Zeng et al., 13 Jun 2025). Earlier studies treated optimal ramps through the superfluid–Mott transition in the Bose–Hubbard chain (Doria et al., 2010), fixed-time preparation of Luttinger-liquid ground states with tunable interactions (Rahmani et al., 2010), and control of many-body entanglement itself, rather than a specific target state, through a local-unitary-invariant entanglement functional (Lucas et al., 2012).

A common misconception is that many-body robust control refers only to large qubit arrays. The literature uses the term for finite but genuinely interacting or high-dimensional many-body settings more generally. A chain of $N=3$ coupled harmonic oscillators is treated as a many-body continuous-variable system because the control problem lives in the multi-mode Hilbert space of coupled bosonic degrees of freedom (Yu et al., 2021). Conversely, work on 10-qubit and 8-transmon star graphs emphasizes that many-body difficulty arises not only from Hilbert-space size, but from interacting dynamics, multiple uncertain couplings, leakage, and nontrivial graph structure (Le et al., 2021).

2. System models and target transformations

The model classes span bosons, spins, qubits, localized systems, and driven-dissipative photonics. A representative continuous-variable example is the chain of coupled harmonic oscillators with Hamiltonian

$$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$

where the local frequencies $\tilde\omega_j(t)$ and couplings $\tilde k_j(t)$ are the controls. The demonstrated task is fixed-time transfer of a Schrödinger cat state from the first to the last oscillator while simultaneously rotating its phase-space angle by $45^\circ$ or $90^\circ$ (Yu et al., 2021). In open-system simulations, the same controls are tested under a bosonic bath coupled through $L=\lambda\sum_i q_i$, with non-Markovian Ornstein–Uhlenbeck correlations (Yu et al., 2021).

Interacting qubit models appear in several distinct forms. For robust sensing, a general many-body Hamiltonian

$$H(t)=\sum_j Q_j+\sum_{j,k}J_{jk}V_{jk}+\sum_{j\in\mathcal L}(\Omega_j^x(t)S_j^x+\Omega_j^y(t)S_j^y)$$

is specialized to a star graph of capacitively coupled transmons, with one driven central node and uncertain edge couplings $J_j$ (Le et al., 2021). The target is the canonical GHZ state $N=3$0, motivated by Heisenberg-limited phase estimation (Le et al., 2021). In the crosstalk setting, the controlled Hamiltonian combines local $N=3$1 drives, tunable nearest-neighbor $N=3$2 couplings, and an unknown parasitic Heisenberg term

$N=3$3

with targets including $N=3$4, $N=3$5, $N=3$6-qubit GHZ states, and the ground state of the antiferromagnetic Heisenberg chain (Le et al., 4 Mar 2026).

Other problem formulations shift the control variable rather than the target class. In the Luttinger-liquid setting, the scalar interaction $N=3$7 is the sole control, and the aim is to drive the ground state at $N=3$8 toward the ground state at a target $N=3$9 in a fixed time $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$0 (Rahmani et al., 2010). In density-based control, the control variable is the full scalar potential $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$1, and the target is a prescribed many-body density trajectory $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$2 for a given initial state and interaction; the method is used to implement translations and splittings of interacting and non-interacting densities in one and two dimensions (Nielsen et al., 2014). In many-body localization, the control objective is local preparation of an effective eigenstate and coherent manipulation of localized qubits using quantum phase estimation, without full microscopic Hamiltonian knowledge (Choi et al., 2015).

The same theme extends to photonic and hybrid platforms. Solid-state emitters in nanophotonic environments are described by effective exchange matrices $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$3 and collective decay matrices $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$4, enabling superradiant, subradiant, graph-state, and steady-state entanglement protocols in cavity and waveguide QED architectures (Daggett et al., 25 Nov 2025). Driven Kerr microresonators with photonic-crystal bandgaps realize a driven-dissipative many-mode bosonic system whose collective steady states interpolate between a Mott-insulator-like flattop comb and a superfluid-like spectrally modulated comb, controlled mainly by the bandgap strength $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$5, the detuning $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$6, and the pump amplitude $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$7 (Jin et al., 23 Apr 2026).

3. Control formalisms and numerical architectures

Gradient-based optimal control remains a principal methodology. In the oscillator-chain example, Krotov’s method is used with a terminal infidelity functional $$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$8 and a running cost penalizing control updates; the controls are the on-site frequencies and nearest-neighbor couplings (Yu et al., 2021). In the transmon star graph, the control fields are piecewise-constant quadrature drives on the central node, and robustness is handled by first optimizing the nominal point to near-unit fidelity and then performing either sequential convex programming on the minimum fidelity over extreme points or quasi-Newton optimization of the average fidelity over the uncertainty-set vertices (Le et al., 2021). The same work combines forward–backward propagation with Krylov subspace exponentiation to avoid explicit matrix exponentials in exponentially large Hilbert spaces (Le et al., 2021).

Tensor-network methods enlarge the controllable many-body regime. In the crosstalk study, matrix product states and matrix product operators are propagated by TEBD, while exact gradients of ensemble-averaged infidelity are computed within the tensor-network representation and optimized with L-BFGS (Le et al., 4 Mar 2026). In reinforcement-learning control, the many-body state is represented as an MPS and the action-value function is itself parameterized by a “QMPS” architecture, so that deep Q-learning can act directly on 1D many-body states while preserving linear-in-$$H(t)=\sum_{j=1}^N \tilde\omega_j(t)a_j^\dagger a_j+\sum_{j=1}^{N-1}\tilde k_j(t)(a_j^\dagger a_{j+1}+h.c.),$$9 scaling for area-law regimes (Metz et al., 2022). In reduced-order modelling, a tensor-network environment is truncated into a low-dimensional non-Markovian “digital twin” of a controlled subsystem, and Riemannian ADAM is used to optimize local unitaries on the unitary manifold of the subsystem controls (Luchnikov et al., 2022).

A distinct methodological line replaces direct wavefunction targeting by surrogate observables or local structure. The density-based approach derived from time-dependent density-functional theory constructs the external potential $\tilde\omega_j(t)$0 from the target density through an iterative fixed-point solution of a Sturm–Liouville equation for $\tilde\omega_j(t)$1 (Nielsen et al., 2014). The entanglement-control approach of “Tailoring many-body entanglement through local control” (Lucas et al., 2012) uses the algebraic lower bound

$\tilde\omega_j(t)$2

as the target functional and derives time-local controls by maximizing $\tilde\omega_j(t)$3 or, under locality constraints, $\tilde\omega_j(t)$4; the crucial point is that local control cannot change entanglement in first order, so the curvature contains the relevant interplay between local rotations, intrinsic interactions, and decoherence (Lucas et al., 2012). Earlier work on Luttinger liquids used simulated annealing over piecewise-constant interaction protocols $\tilde\omega_j(t)$5, and found nonmonotonic optimal ramps that outperform linear or power-law schedules (Rahmani et al., 2010). CRAB plus t-DMRG played the same role for the Bose–Hubbard superfluid–Mott ramp, using a truncated randomized Fourier basis for the lattice-depth trajectory (Doria et al., 2010).

4. Robustness mechanisms and performance criteria

The literature uses several inequivalent robustness metrics. Final-state fidelity and infidelity are standard in state-transfer and state-preparation tasks, with Uhlmann fidelity used for mixed states in the open-system cat-transfer problem (Yu et al., 2021). Worst-case fidelity over a parameter hypercube is the central figure of merit in robust GHZ preparation for sensing (Le et al., 2021). Ensemble-averaged gate infidelity and ensemble-averaged state infidelity are the key objectives in crosstalk suppression for large qubit chains (Le et al., 4 Mar 2026). The entanglement-based formulation uses the lower bound $\tilde\omega_j(t)$6 itself as a target functional and interprets slower loss of $\tilde\omega_j(t)$7 under Lindblad dephasing as enhanced robustness of the generated many-body state (Lucas et al., 2012). In Kerr-soliton control, the coefficient of variation of the comb intensities serves as an order parameter separating Mott-insulator-like and superfluid-like regimes (Jin et al., 23 Apr 2026).

Several distinct robustness mechanisms recur. One is explicit multi-scenario optimization: optimize against a set of Hamiltonians or noise realizations rather than a nominal model. This appears in worst-case transmon control under uncertain couplings (Le et al., 2021) and in ensemble-GRAPE over parasitic Heisenberg couplings for 50-qubit gate synthesis (Le et al., 4 Mar 2026). Another is interference-based cancellation of perturbative leakage. The adiabatic echo protocol makes the system traverse the ordered region twice so that leading-order amplitudes induced by a static perturbation cancel when the accumulated phase satisfies $\tilde\omega_j(t)$8 and the integrated couplings are balanced (Zeng et al., 13 Jun 2025). A third mechanism is state-feedback adaptation. In QMPS-based reinforcement learning, the controller selects a new action from the current many-body state after each perturbation, which yields on-the-fly correction under random wrong actions and noisy pulse durations (Metz et al., 2022).

Other works use many-body structure itself as a robustness resource. In the oscillator chain, non-Markovian memory slows the fidelity decay with increasing system–bath coupling $\tilde\omega_j(t)$9 relative to the more Markovian regime, so the control performs better at smaller $\tilde k_j(t)$0 in the Ornstein–Uhlenbeck bath (Yu et al., 2021). In many-body localization, quantum phase estimation projects the uncontrolled environment of other l-bits into an effective eigenstate, turning fluctuating interaction phases into static parameters and thereby stabilizing local two-qubit operations (Choi et al., 2015). In non-equilibrium Kerr solitons, photonic-crystal bandgap engineering suppresses cross-mode coupling while preserving self-mode Kerr interactions, so a broad low-CV region emerges without feedback control (Jin et al., 23 Apr 2026). In nanophotonic emitter arrays, decoherence-free dark states, collective subradiance, engineered dissipation, phononic bandgaps, Purcell enhancement, and spectral tuning all serve as robustness tools at the Hamiltonian or Lindbladian level rather than at the pulse-design level (Daggett et al., 25 Nov 2025).

An objective point of disagreement across the literature concerns what should count as robustness. Some studies address environmental decoherence but not parameter uncertainty, as in the cat-state transfer paper (Yu et al., 2021). Others study Hamiltonian uncertainty or static perturbations while neglecting explicit bath coupling, as in robust transmon control, crosstalk mitigation, and adiabatic echo protocols (Le et al., 2021, Le et al., 4 Mar 2026, Zeng et al., 13 Jun 2025). A broader reading of the field therefore treats robustness operationally: the protocol is robust if the task-specific figure of merit remains high under the error model actually included in the control design or in the a posteriori validation.

5. Scaling, complexity, and the role of structure

Scalability depends strongly on the structure of the many-body dynamics. “On the complexity of controlling quantum many-body dynamics” (Caneva et al., 2013) formulates a lower bound in terms of the dimension $\tilde k_j(t)$1 of the manifold supporting the dynamics: $\tilde k_j(t)$2 where $\tilde k_j(t)$3 is the number of control frequencies and $\tilde k_j(t)$4 the bit depth. The same work argues that control complexity grows roughly with $\tilde k_j(t)$5: approximately linearly in integrable models such as the LMG model or the transverse-field Ising chain with $\tilde k_j(t)$6, and exponentially in non-integrable settings where the effective manifold dimension scales exponentially with system size (Caneva et al., 2013). This provides a structural explanation for why robust control is comparatively tractable in some many-body systems and prohibitive in others.

Many practical schemes explicitly exploit this structural compression. The transmon star graph reduces the number of distinct worst-case scenarios from exponential in the number of uncertain couplings to linear in the number of boundary qubits because symmetry makes all vertices with the same number of upper-value couplings equivalent (Le et al., 2021). Tensor-network robust control assumes limited entanglement growth and achieves polynomial scaling in $\tilde k_j(t)$7, $\tilde k_j(t)$8, and $\tilde k_j(t)$9, enabling gate synthesis on 50 qubits and state preparation on 20–30 qubits (Le et al., 4 Mar 2026). QMPS reinforcement learning reaches 32-spin Ising chains by combining MPS environment states with an MPS-based value-function approximator (Metz et al., 2022). Reduced-order modelling compresses a 27-spin environment into an effective subsystem-plus-environment model of dimension $45^\circ$0, compared with a light-cone estimate of $45^\circ$1 for the full space (Luchnikov et al., 2022).

The role of intrinsic timescales is equally important. In the Luttinger-liquid control problem, a marked transition in controllability occurs when the ratio $45^\circ$2 exceeds a critical value; numerically, the critical ratio is reported as $45^\circ$3 for the representative choice $45^\circ$4, $45^\circ$5 (Rahmani et al., 2010). The interpretation is that the slowest mode with $45^\circ$6 imposes a minimum time scale $45^\circ$7, so near-perfect preparation becomes possible only once the control duration exceeds a linear-in-$45^\circ$8 threshold (Rahmani et al., 2010). A related message appears in the superfluid–Mott ramp, where CRAB-optimized pulses reduce the total time from about $45^\circ$9 ms to $90^\circ$0 ms while suppressing defects relative to simple ramps (Doria et al., 2010). These studies suggest that robust many-body control is governed not only by optimization algorithms, but by controllability windows set by collective many-body timescales.

6. Platforms, limitations, and research directions

The experimental and near-experimental platforms are correspondingly diverse. The oscillator-chain protocol is motivated by coupled cavities, trapped ions, and circuit-QED resonators (Yu et al., 2021). Robust GHZ preparation is formulated for capacitively coupled transmons with realistic three-level physics and leakage (Le et al., 2021). Tensor-network crosstalk control is framed around qubit chains with local drives and nearest-neighbor couplings, with clear relevance to near-term processors (Le et al., 4 Mar 2026). Adiabatic echo protocols are demonstrated in Ising spin chains, two-dimensional Rydberg atom arrays, and frustrated Rydberg lattices (Zeng et al., 13 Jun 2025). Density-based control is presented as compatible with large-scale TDDFT through Kohn–Sham systems (Nielsen et al., 2014). Solid-state emitter arrays and nanophotonic environments offer robust collective-state engineering at light–matter interfaces (Daggett et al., 25 Nov 2025), while Kerr microresonators show that the many-body Hamiltonian of a driven-dissipative photonic lattice can itself be programmed by nanophotonic bandgap design (Jin et al., 23 Apr 2026).

The limitations are equally specific. The cat-transfer study optimizes only for closed-system dynamics and then tests the controls a posteriori in a weakly non-Markovian bath; it does not address Hamiltonian uncertainty or distorted controls (Yu et al., 2021). Robust transmon GHZ preparation treats static coupling uncertainty but not decoherence, and the achievable fidelity is limited by leakage to the third level (Le et al., 2021). Tensor-network crosstalk mitigation assumes 1D geometry, static coherent parasitic couplings, and moderate entanglement growth (Le et al., 4 Mar 2026). Adiabatic echo theory is aimed at static perturbations and near-degenerate manifolds rather than fast time-dependent noise (Zeng et al., 13 Jun 2025). Reduced-order control is explicitly not applicable in the thermalized phase, where the effective environment dimension grows too rapidly (Luchnikov et al., 2022). Density-based control inherits the usual representability and approximation issues of TDDFT, especially when approximate exchange–correlation functionals are used for large interacting systems (Nielsen et al., 2014). The complexity analysis of (Caneva et al., 2013) indicates that non-integrability can make robust control exponentially hard in system size.

Current research directions therefore combine structural modelling, scalable numerics, and error-aware design rather than seeking a single universal algorithm. Several works point to fully consistent non-Markovian optimal control, integration of dynamical decoupling with optimal shaping, comparison of Krotov- and machine-learning-based schemes, and learning reduced-order models directly from experimental data (Yu et al., 2021, Le et al., 2021, Luchnikov et al., 2022). A plausible implication is that the field is converging on a layered view of robustness: Hamiltonian engineering to shape the many-body manifold, scalable optimization to exploit that structure, and explicit inclusion of dominant error channels only where they materially alter the target task. In that sense, many-body robust control is less a single technique than a research program for aligning controllability, many-body structure, and noise resilience within one design loop.