Pulse-level variational quantum algorithms skip the digital gate layer. A classical optimizer shapes the raw microwave pulses that drive superconducting qubits and scores each candidate pulse by measuring an energy or cost on the device. This produces shorter, more accurate programs, but it also lets the optimizer propose fast, high-amplitude pulses. Those pulses kick the qubit out of its two usable levels into a third energy level, an effect called leakage, which silently corrupts the computation. Current pulse-level methods either ignore leakage or add a penalty for it to the loss. We propose instead that every pulse the optimizer proposes is first passed through a safety filter. The filter bends the pulse just enough that a model of the qubit provably keeps third-level population below a threshold at every instant of the pulse, not just at the end. The measured energy on the device then drives the outer optimization as usual.
The tool for the filter comes from safety-critical robotics. A control barrier function is a scalar "distance to danger" whose rate of change is constrained at each instant, so that a system never leaves a safe region. When the control input does not appear in the first time derivative of that distance, one needs high-order control barrier functions, which build a chain of nested buffer conditions. The situation is the same as a car whose braking acts on acceleration while the constraint is on position. Leakage is such a constraint once the pulse slope, which is what hardware and leakage physics actually care about, is treated as the input. The opportunity is to turn leakage from a tuned penalty into a hard, time-local guarantee built into how pulses are parameterized. This does for dynamic constraints what symmetry-preserving circuits did for static ones.
Core Claim
Constraining variational pulses by reparameterizing them through a high-order barrier filter yields lower peak leakage, at equal or better energy accuracy and convergence speed, than penalty-based or amplitude-capped pulse optimization. The key ingredient is the high-order structure: a first-order barrier loses its grip exactly where leakage starts.
The claim has two roots. Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms motivates it. That paper shows that soft penalties warp variational landscapes and fail to guarantee feasibility, while ansätze that confine evolution to the valid subspace reach errors below $10^{-10}$ on $H_2$, $LiH$, $H_2O$ and $NH_3$. That result covers only static symmetries such as fixed Hamming weight, and leakage is not a symmetry of the driven transmon. Control Barrier Functions for Systems with High Relative Degree supplies the method: its recursive chain $\psi_k = \dot\psi_{k-1} + \alpha_k(\psi_{k-1})$ enforces invariance when the input appears only after $m$ differentiations. It was built and tested for classical vehicles, not quantum systems.
Concretely, model each qubit as a three-level system with density matrix $\rho$, drift Hamiltonian $H_0$, and drive $u(t)H_1$. Define the safety function $h(\rho) = \epsilon - \mathrm{tr}(P_2\rho)$, where $P_2$ projects onto the third level and $\epsilon$ is the leakage budget. The drive amplitude $u$ is augmented into the state, and its slope $v = \dot u$ is the filtered input. Then $\dot h$ depends on $u$ but not on $v$, and the input first appears in
$$\ddot h = a(\rho,u) + b(\rho)\, v, \qquad b(\rho) \propto -\mathrm{tr}\big(i[H_1,P_2]\rho\big),$$
so the relative degree is 2. The coefficient $b$ is proportional to the coherence between the first and second excited levels and vanishes when that coherence is zero. This is the typical situation just before leakage begins. The problem therefore goes beyond "apply a second-order barrier". The relative degree is state-dependent and degenerate, and handling that degeneracy is part of the research. The mixed-relative-degree constructions of Wei Xiao et al. are the natural starting point.
Research Program
The pipeline has three pieces. The first is a nominal pulse, produced by the variational optimizer in the Legendre-polynomial basis of Differentiable Analog Quantum Computing for Optimization and Control. That paper recasts variational algorithms as optimal control of the Schrödinger equation and estimates gradients on hardware with a continuous parameter-shift rule.
The second piece is a classical filter. It integrates each qubit's local three-level model along the pulse and, at each time step, solves a one-constraint quadratic program that minimally changes $v$ so that $\psi_2 \ge 0$. With a single barrier plus slew and amplitude bounds, this program often has a closed-form solution. The explicit-solution approach of Explicit Solution of Tunable Input-to-State Safe-Based Controller Under High-Relative-Degree Constraints suggests this is tractable, which keeps the filter cheap and differentiable.
The third piece is execution of the filtered pulse on the device, here a noisy multi-level simulator standing in for hardware. The measured energy returns to the optimizer. Gradients combine two sources: parameter-shift derivatives of the energy with respect to the executed pulse, obtained on the device, and the filter's Jacobian, computed classically.
This division of labor is what makes the method hybrid rather than classical pulse design. The filter only needs a cheap per-qubit model, because leakage is local to each transmon. The globally entangled cost, which is exponentially expensive to simulate, is only ever measured. The file notes that variational loops absorb systematic model errors (Quantum algorithms for electronic structure calculations: Particle-hole Hamiltonian and optimized wave-function expansions), and we exploit that. The filter's guarantee holds only for its model. How much actual leakage exceeds modeled leakage under crosstalk and miscalibration is therefore a measured quantity, not an assumed one.
Three analysis questions accompany the build. First, how to treat the degenerate coefficient $b(\rho)$ without chattering or infeasibility near zero coherence. Candidates are a third-order chain, which takes over when $b$ is small, or a blended mixed-order barrier. Second, how the choice of class-$\mathcal{K}$ gains trades conservatism against energy accuracy. The high-relative-degree paper already found that square-root gains over-constrain a cruise-control vehicle even when it is far from danger, and we expect the analogous failure here. Third, whether amplitude and slew limits make the filter pointwise infeasible for short pulse durations. This infeasibility is a known limitation of high-order barriers under saturation.
We deliberately do not pursue the other bridges suggested by the literature. Quantum-Assisted Barrier Sequential Quadratic Programming for Nonlinear Optimal Control is the closest existing quantum-plus-barrier work. It runs the opposite direction: a fault-tolerant quantum linear solver speeds up a log-barrier SQP for classical control, with no invariance guarantees or high-order constraints. Using quantum solvers for barrier quadratic programs of this size would lose to classical solvers that already run in milliseconds. Matrix Control Barrier Functions could eventually express density-matrix constraints, but its semidefinite programs scale poorly with matrix size. It is a later extension, not the core.
Baselines
The do-nothing reference is the unconstrained pulse-level optimizer of Differentiable Analog Quantum Computing for Optimization and Control. It is run on the same three-level simulator, with the same Legendre basis, pulse duration, shot budget and iteration count. For this method, the file reports that on $H_2$ it is "10x more accurate than digital VQE", and that on a 4-vertex MaxCut it converges to loss $2.6\times10^{-6}$ where digital QAOA stalls at $0.08$. Those results were obtained on transmon models without any leakage accounting. No reported result was found on Lacuna for leakage population under that method or under NAPA. Establishing that number is itself the first deliverable.
The strongest competitor is the same optimizer with a leakage penalty added to the loss. The penalty uses measured final third-level population, with its weight tuned by grid search over the same total shot budget our method uses. This is exactly the soft-constraint strategy that Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms shows degrades landscapes in the static case, but it is also what practitioners would try first.
A second strong baseline is NAPA: Intermediate-level Variational Native-pulse Ansatz for Variational Quantum Algorithms. It tunes native pulse amplitudes and frequencies with derivative-free optimizers and progressive pulse growth, and reports $H_2$ accuracy of 96.482% and $HeH^+$ accuracy of 99.336%. We run it with a hard amplitude cap set so that its peak leakage matches our budget.
The ablation replaces the second-order barrier chain with a first-order barrier on $h$ using amplitude as input. This isolates the high-order ingredient. A class-$\mathcal{K}$ sweep with linear, quadratic and square-root gains isolates conservatism.
Our method beats a baseline if it satisfies three conditions at matched energy error, where error is measured against the chemical-accuracy threshold of $1.6\times10^{-3}$ Ha. It must keep peak in-pulse leakage below $\epsilon$ on the model in every iteration. It must show lower true leakage on the mismatched simulator. And it must need no more shots to converge.
Why not just…?
Why not just add a leakage penalty to the variational loss? A penalty on final population cannot see leakage that rises and returns mid-pulse, and that transient leakage still couples to decoherence. The penalty weight also has to be tuned, which the symmetry-preserving-circuit paper documents as fragile and landscape-warping. This alternative is a real threat on small problems, where a well-tuned penalty may match us. The penalty-weight grid at an equal shot budget, reported as a Pareto curve of energy error versus peak leakage, settles whether the filter's advantage survives fair tuning.
Why not just smooth or amplitude-cap the pulse basis? This is the most serious alternative. Bandwidth limits and amplitude caps are simple and do reduce leakage, and the file has no evidence on them either way. Their weakness is that they are state-blind: they must be conservative everywhere to be safe in the worst case. The filter restricts the pulse only when the model predicts approaching leakage. If a cap tuned to the same peak leakage matches our energy accuracy and pulse duration, the barrier machinery adds nothing, and we will report that. The capped-NAPA baseline and a capped Legendre baseline are the experiments that decide it.
Why not just design leakage-free pulses offline with classical optimal control like GRAPE? Offline design requires simulating the full multi-qubit, multi-level system, which is exponentially expensive. The resulting pulses are also fixed against a model that drifts. Differentiable Analog Quantum Computing for Optimization and Control reports $10^4$ times better X-gate accuracy than GRAPE and SLSQP. Our filter uses only local three-level models and leaves the entangled cost to measurement. The scaling experiment below tests whether this split holds as qubit count grows.
Proposed Experiments
All experiments use a multi-level transmon simulator with three levels per qubit, anharmonicity, and optional crosstalk and amplitude miscalibration. The filter's model omits crosstalk and miscalibration, so the model/device gap is controlled.
The tasks follow the file's benchmarks: $H_2$ and $HeH^+$ ground-state energies, a 4-vertex MaxCut, and X and CNOT gate synthesis. We then scale to the largest system the Leng et al. models reached, about 11 qubits, and to $LiH$ where feasible.
The metrics are six: energy error in Hartree, peak and final third-level population over the pulse, pulse duration, shots to convergence, and fraction of optimizer iterations that proposed any pulse exceeding $\epsilon$. The last is identically zero on the model for our method by construction and nonzero for the baselines.
Three outcomes are decisive. First, on the matched-leakage Pareto plot against the tuned penalty and the capped baselines, the filter must dominate at short pulse durations, where leakage pressure is highest. Second, the first-order ablation should show degraded protection, or filter infeasibility, concentrated at moments of near-zero coherence between the first and second excited levels. That would confirm that the degenerate high-order structure is the operative mechanism. Third, as crosstalk and miscalibration grow, true leakage under the filter should degrade gracefully, and energy accuracy should hold because the hardware-measured loop absorbs the mismatch. If true leakage instead tracks the unfiltered baseline once model error is realistic, the model-based guarantee is too brittle to matter. The proposal would then narrow to calibrated single-qubit gate synthesis.
References
- Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms
- Control Barrier Functions for Systems with High Relative Degree
- Differentiable Analog Quantum Computing for Optimization and Control
- Explicit Solution of Tunable Input-to-State Safe-Based Controller Under High-Relative-Degree Constraints
- Quantum algorithms for electronic structure calculations: Particle-hole Hamiltonian and optimized wave-function expansions
- Quantum-Assisted Barrier Sequential Quadratic Programming for Nonlinear Optimal Control
- Matrix Control Barrier Functions
- NAPA: Intermediate-Level Variational Native-Pulse Ansatz for Variational Quantum Algorithms