Adaptive Tensor Network Sampling for Quantum Optimal Control
A tensor-network approach to quantum control finds good pulse sequences without gradients — potentially useful when standard optimization methods stall.
The Thesis
Quantum optimal control (QOC) is the engineering discipline of finding the right sequence of electromagnetic pulses to make a quantum processor execute a desired operation with high accuracy. Most modern QOC tools rely on gradient-based optimization — essentially calculus on a very complicated landscape — but that landscape is riddled with local minima where the solver gets stuck. This paper proposes a gradient-free alternative: use a matrix product state (MPS), a compact mathematical object borrowed from quantum physics simulations, to define a probability distribution over candidate pulse sequences, then iteratively improve that distribution by keeping the sequences that work best. The method is tested on a handful of standard benchmark tasks — including preparing entangled qubit pairs and controlling three-level quantum systems called qutrits — and holds its own against existing gradient-free baselines. The catch is that benchmarks are modest in scale, and the paper makes no claim of superiority over gradient-based methods, only that this approach is a 'viable heuristic.'
Catalyst
Tensor network methods — originally developed to simulate quantum many-body physics — have matured enough in software tooling and theoretical understanding to be repurposed as general optimization primitives. Simultaneously, quantum hardware providers are moving from a handful of qubits to systems with dozens or hundreds, making the pulse-design search space grow exponentially and rendering brute-force gradient methods increasingly fragile. The combination of better tensor-network libraries and harder hardware control problems creates a natural opening for this kind of approach.
What's New
Prior gradient-free QOC methods include evolutionary algorithms, reinforcement learning agents, and simulated annealing — all of which treat the search space as a black box. Earlier tensor-network methods were used primarily to simulate quantum systems, not to optimize control over them. This paper borrows the MPS data structure (a chain of low-rank numerical arrays that compresses high-dimensional probability distributions) and repurposes it as an adaptive sampler: the MPS encodes a score function over discrete pulse choices, which is updated as better sequences are found, biasing future sampling toward promising regions of the search space.
The Counter
The benchmarks here are small: single-qubit transfers, two-qubit Bell states, and a qutrit gate. Real quantum processors need to control tens to hundreds of qubits simultaneously, and the MPS compression that makes this method efficient on small problems may not survive that scaling — the tensor network's bond dimension (a measure of how much information it can represent) must grow exponentially in the worst case. The paper's own language is cautious: 'viable heuristic' and 'competitive' against baselines, not 'better.' Gradient-based methods like GRAPE (Gradient Ascent Pulse Engineering) remain the industry workhorse precisely because they converge faster when gradients are available. The gradient-free framing is also not new — evolutionary strategies and Bayesian optimization have been applied to QOC for years and are already well-tuned. Finally, there is no open code repository linked in the paper, which makes independent reproduction harder and limits community uptake.
Longs
- IONQ — direct beneficiary as a trapped-ion hardware company where pulse-level control quality determines gate fidelity
- RGTI (Rigetti Computing) — superconducting qubit company with active pulse-design toolchains
- QTUM (ETF) — broad quantum computing equity exposure
- IBM (IBM) — runs one of the largest public quantum fleets where QOC tooling improvements have direct operational value
- MCHP (Microchip Technology) — makes control electronics used in quantum hardware stacks
Shorts
- Q-CTRL — their core product is proprietary gradient-based and machine-learning-driven quantum control software; a compelling open-source gradient-free alternative could undercut their moat if it scales
- Vendors of classical optimal control software adapted for quantum (e.g., GRAPE-based toolkits) — this approach directly competes in the gradient-free niche they serve
Enablers (Picks & Shovels)
- ITensor and TensorNetwork open-source libraries — provide the MPS/tensor-train primitives this method builds on
- Qiskit Pulse (IBM) and Pulser (Pasqal) — hardware-level pulse definition frameworks where improved QOC could be deployed
- JAX and PyTorch — automatic differentiation backends that competing gradient-based methods rely on, giving context for what gradient-free approaches must match
- QuTiP (Quantum Toolbox in Python) — open-source quantum simulation library used in benchmarking open-system control problems
Private Watchlist
- Q-NEXT (DOE quantum center, not investable directly but signals government QOC funding direction)
- Qu & Co (quantum algorithms, private)
- Quantinuum (private, Honeywell spinout with active pulse-optimization research)
- Q-CTRL (private, Australian firm specializing specifically in quantum control software)
Resources
The Paper
Quantum optimal control (QOC) provides a systematic framework for achieving high-fidelity operations in quantum systems and plays a central role in tasks such as gate synthesis, state transfer, and pulse design. Existing QOC methods broadly fall into two categories: gradient-based and gradient-free algorithms. The associated optimization landscape is often high-dimensional, non-convex, and populated by numerous local minima, making efficient gradient-free search strategies essential. To address this, we introduce a gradient-free matrix product state/tensor train (MPS/TT) sampling heuristic for discrete quantum optimal control. In our approach, the MPS defines a score function over the space of discrete control parameters, which in turn induces a sampling distribution over candidate control sequences. This distribution is iteratively refined through selection of better performing sequences and local tensor updates to bias the search toward high-performing sequences. We evaluate the method on a range of benchmark problems, including single-qubit state transfer, Bell-pair preparation, qutrit gate implementation, and open-system population transfer. Across these tasks, the method exhibits stable convergence behavior and competitive empirical performance relative to established gradient-free baselines. These results suggest that tensor network sampling offers a viable heuristic framework for discrete quantum control.