Analytical Series Expansion for Efficient Gradient Evaluation in Multi-Qubit Optimal Control

Technical Summary: Analytical Series Expansion for Efficient Gradient Evaluation in Multi-Qubit Optimal Control

Problem Statement
Quantum optimal control (QOC) is essential for calibrating large-scale quantum processors, particularly for mitigating errors arising from spurious interactions and crosstalk in multi-qubit systems. While gradient-based QOC methods (e.g., GRAPE, Krotov) are more efficient than gradient-free approaches, they face significant computational bottlenecks when scaling to large Hilbert spaces or open quantum systems. The core challenge lies in calculating the gradient of the time-ordered propagator, $U(t)$, with respect to arbitrary pulse parameters. Existing methods like GOAT (Gradient Optimization of Analytic conTrols) require co-propagating a set of derivatives alongside the propagator, which restricts parameterization choices or incurs high computational costs due to the repeated evaluation of matrix exponentials. Piecewise-constant ansätze (used in GRAPE) bypass some limitations but lack the flexibility of continuous parameterizations.

Methodology
The authors present a unifying framework derived from first principles to compute the gradient of the propagator for arbitrary pulse parameterizations. The methodology proceeds in three main stages:

1. Formal Solution via Heisenberg Picture:
   Starting from the Schrödinger equation, the authors derive an exact integral expression for the gradient of the propagator, $\nabla_\alpha U(t)$. This expression reveals that the gradient depends on the evolution of the derivative of the Hamiltonian in the Heisenberg picture:
   $$\frac{\partial U(t)}{\partial \alpha_k} = -i U(t) \int_0^t d\tau \frac{\partial u_m(\tau)}{\partial \alpha_k} H_m^{(H)}(\tau)$$
   where $H_m^{(H)}(\tau) = U^\dagger(\tau) H_m U(\tau)$. This formulation connects the gradient calculation to the study of operator evolution, a central topic in fields like operator spreading and quantum chaos.

2. Series Expansion:
   To avoid the expensive numerical integration of the Heisenberg operator at every time step, the authors integrate the expression by parts repeatedly. This yields a truncated series expansion:
   $$\frac{\partial U(t)}{\partial \alpha_k} \simeq \left[ \sum_{n=0}^N \Theta_n(t) \right] U(t)$$
   Here, $\Theta_n(t)$ consists of time-independent commutators of the Hamiltonian terms (e.g., $[H_{m'}, [H_{m''}, H_m]]$) multiplied by time-dependent coefficients $\beta^{(n)}_{m,m}(t)$. Crucially, the commutators are computed once at the start of the simulation, while only the coefficients (which are simple classical integrals) need to be updated during optimization.

3. Exploiting Locality:
   For systems with local interactions (e.g., 1D chains or 2D lattices), the authors leverage the graph structure of the Hamiltonian. By applying Lieb-Robinson bounds, they demonstrate that the influence of local operators spreads at a finite speed. This allows for the truncation of the series expansion, as commutators involving distant qubits become negligible. The method utilizes sparse representations (Pauli strings) and heuristic pruning strategies to further reduce the number of terms required, making the approach scalable for large multi-qubit platforms.

Key Contributions

• Unified Framework: The paper provides a formal solution for the gradient of the propagator that applies to any pulse parameterization (continuous or piecewise-constant) and can be extended to open systems and higher-order derivatives.
• Generalization of GRAPE: The derived series expansion is shown to be a generalization of the exact gradient used in GRAPE, extending its applicability to continuous pulse shapes.
• Computational Efficiency: By decoupling the time-independent commutators from the time-dependent coefficients, the method significantly reduces the number of matrix exponentials required.
• Scalability: The integration of locality constraints and sparse Pauli dynamics allows the method to handle large Hilbert spaces where traditional co-propagation methods fail.

Results
The authors validate their method through numerical simulations on multi-qubit systems:

• Accuracy: Comparisons with the GOAT method on a two-qubit system with a CRAB pulse ansatz show that the series expansion (both dense and sparse implementations) reproduces the exact gradient with high accuracy (relative error $< 10^{-4}$ for dense implementation).
• Performance: In simulations of qubit chains and ladder geometries (up to 10 qubits), the series expansion method demonstrates an order-of-magnitude speedup in wall time and a significant reduction in memory usage compared to GOAT.
• Application: The method was successfully applied to prepare a 4-qubit GHZ state on the center qubits of a 2D ladder geometry with 8 qubits (including spectator qubits) and stray ZZ interactions. The optimization, involving 264 parameters, achieved a fidelity of 99.95%, demonstrating the method's ability to handle non-trivial control landscapes with hundreds of parameters.

Significance
The paper claims that this work bridges the gap between quantum optimal control and the study of operator evolution. By mapping the gradient calculation to the evolution of Hamiltonian derivatives in the Heisenberg picture, the authors enable the use of established tools from operator spreading theory (such as tensor networks and sparse Pauli propagation) to optimize quantum systems. The authors emphasize that this approach offers a practical pathway for scaling QOC to larger quantum processors, addressing the computational overhead that currently limits the design of high-fidelity pulses for complex, interacting multi-qubit systems. The work also motivates future developments in sparse representations for operator and state evolution to further scale these optimization techniques.