Skip to content
Quantum.world

Research

Multirobust Control Pulses Designed by Iterative Quantum Kolmogorov-Arnold-Moser Algorithm

Extends two-level DRAG pulses to multilevel hardware via KAM iteration, targeting leakage and parameter inhomogeneity without abandoning existing pulse designs.

Multirobust Control Pulses Designed by Iterative Quantum Kolmogorov-Arnold-Moser Algorithm
arXiv
2611.07802
Published
2026-11-05
Authors
Meri Harutyunyan, Stéphane Guérin

AI summary

Overview

Research area: Quantum control theory for multilevel quantum systems, specifically leakage suppression and error-robust pulse design for superconducting transmon qubits.

Technical level: Advanced. The paper builds on perturbation theory (quantum KAM transformations), the DRAG technique, and robust inverse optimization (RIO), and assumes familiarity with Hamiltonian control of qubits.

Scope (one sentence): The paper reformulates the DRAG leakage-suppression technique as an iterative quantum KAM algorithm so that control pulses designed for ideal two-level systems (here RIO pulses) can be transferred essentially intact into a multilevel system, yielding "RDK" pulses that are simultaneously robust against amplitude and detuning inhomogeneities and against leakage to higher states.

What This Paper Is About

Qubits are rarely isolated two-level systems: weakly anharmonic superconducting transmons have nearby higher levels into which population can leak during a gate. DRAG removes that leakage, and robust inverse optimization (RIO) produces pulses that tolerate unknown errors in the control amplitude (α) and detuning (δ), but combining both protections at once has previously required restrictive pulse ansätze or numerical optimization. The authors show that the iterative quantum Kolmogorov–Arnold–Moser (KAM) algorithm can implement DRAG in a way that fully incorporates any two-level control pulse, and they demonstrate the resulting multirobust RDK pulses on a resonant π-pulse, an α-robust RIO pulse, and an αδ-robust RIO pulse.

Key Contributions

  1. A generalized DRAG formulation via iterative quantum KAM. Instead of only compensating the derivative of the pulse envelope, the authors enforce that the qubit block of the effective Hamiltonian exactly reproduces the original two-level control, so DRAG becomes a method for replicating arbitrary two-level controls inside a multilevel system.

  2. The RDK pulse family. Combining the DRAG-KAM algorithm with RIO pulses yields pulses (RDK) that are simultaneously robust to amplitude inhomogeneities, detuning inhomogeneities, and leakage outside the qubit subspace, without a pulse ansatz or a gradient-based numerical search.

  3. A time-reparametrization procedure. Because RIO pulses do not vanish at the pulse boundaries, the authors introduce a smooth Gaussian reparametrization of time that preserves the pulse area and enforces Ω⁽¹⁾(t_i) = Ω⁽¹⁾(t_f) = 0, a requirement for the KAM transformation to reduce to the identity at the initial and final times.

  4. Characterization of convergence and cost. The authors show numerically that the error behaves as a superasymptotic series with an exponential form e^(−C/ε_σ), and that the added pulses are near-optimal in energy compared with the two-level RIO optimum.

Main Findings

  • Resonant π-pulse: An infidelity of order −4 (i.e. log₁₀ of the infidelity) is reached for pulses as short as σ = 0.75σ₀, with the normalization σ₀ = 2π/|d| and a negative anharmonicity d = −|d|; this is described as compliant with quantum technology requirements.

  • Leakage saturates fast: The leakage population P₂ = |⟨2|φ(t_f)⟩|² into the upper state reaches its minimum after a single iteration (labeled N = 2, with N = 1 the baseline) thanks to the optimal κ = 0.5; further iterations correct the transfer while preserving that minimum leakage.

  • Plateau behavior: Fidelity improves over roughly four successive iterations before plateauing. The number of iterations needed to reach the plateau grows with σ, as does the achievable fidelity. For sufficiently long pulses, machine-level accuracy is obtained; for pulses that are too short, the algorithm does not improve fidelity for N > 2.

  • Superasymptotic scaling: The minimum infidelity follows e^(−C/ε_σ) with ε_σ = 1/(σ|d|), with C = 2.5 for the π-pulse and C = 0.7 for α-RIO. The exponential form is observed for small enough ε_σ < 1/2.

  • α-RIO and αδ-RIO: An infidelity of −4 is achieved for σ = 2σ₀ (α-RIO) and σ = 2.5σ₀ (αδ-RIO); larger σ is needed than for the resonant pulse because a nonzero detuning increases the effect of higher levels. The algorithm was found to work for values of ε as large as approximately 0.5.

  • Robustness is recovered, not lost: DRAG-KAM recovers the two-level RIO robustness after three iterations for the α-profile (shown for α-RIO at σ = 2.5σ₀) and also for the δ-profile of αδ-RIO at σ = 2.5σ₀.

  • Near-optimal energy: Applying the algorithm modifies the optimal pulse area and energy of RIO by typically 10% with a hyper-Gaussian envelope and less than 5% for a longer Gaussian pulse.

  • Anharmonicity and higher levels: Removing leakage to the fourth and fifth levels in a multilevel anharmonic ladder was found to be negligible; the sign of d has no significant effect for the resonant π-pulse, while fidelities for the other pulses depend slightly on the sign of d.

  • Bandwidth requirements: For the shortest α-RIO pulses reaching infidelity −4, preserving that threshold requires a cutoff of at least 250 MHz for the hyper-Gaussian envelope and 146 MHz for the Gaussian envelope (the higher cutoff reflecting the more abrupt edges of the hyper-Gaussian shape).

  • Decoherence estimate: Using Lindblad dynamics with standard transmon coherence times T₁ = 200 μs, T₂ = 135 μs and d/2π = −300 MHz gives infidelities around −3.5 for both α-RIO and αδ-RIO pulses, for durations large enough to reach −4 closed-system infidelity.

  • Use as a GRAPE seed: RDK solutions serve as initial guesses for GRAPE, converging up to eightfold faster and reaching a nominal infidelity (α = δ = 0) two orders of magnitude lower than a RIO seed.

Methodology in Plain English

The system is modeled with the three lowest states |0⟩, |1⟩, |2⟩, a drive with Rabi frequency Ω and detuning Δ, an anharmonicity d, and a coupling λ = √2 to the upper state for a weakly anharmonic oscillator. The unwanted coupling both leaks population into |2⟩ and distorts the intended gate.

The KAM approach attacks this iteratively. The Hamiltonian is split into an unperturbed part H₀ and a perturbation εV₁, and a unitary transformation exp(εW₁) is constructed so that the off-diagonal (leakage) terms of order ε cancel, leaving only terms of order ε². The equations determining the transformation come from the standard cohomology relations D₁ = [H₀, W₁] + V₁ with [H₀, D₁] = 0, solved using the eigenvalues and eigenvectors of H₀. Repeating the procedure pushes the perturbation to higher and higher order; the authors use a simplified, non-superexponential version that keeps the same H₀ at every iteration and still converges exponentially in ε for an optimal iteration number.

The key design choice is that, at each iteration, the qubit-block elements of the correction are fixed so that the dynamics inside the qubit subspace are left untouched: the pulse seen by the qubit is exactly the two-level pulse that was originally designed. Remaining freedom in W_N is captured by a single free function, chosen proportional to Ω⁽¹⁾, giving one tuning parameter κ; the authors use κ = 0.5 as the best fidelity–robustness compromise (κ = 0 gives the best robustness, κ ≈ 1 gives lower infidelities at the expense of robustness). Because the transformation must reduce to the identity at the start and end of the pulse, RIO pulses—which do not vanish at the boundaries—are first reparametrized in time onto a smooth Gaussian while preserving the pulse area. Performance is quantified by the infidelity log₁₀(1 − |⟨φ_T|φ(t_f)⟩|²) for complete population transfer from |0⟩ to |1⟩, and the code is publicly available.

Why This Matters

Research impact: The work shows that leakage suppression and robustness to control errors need not be traded against each other, and it makes the large body of existing two-level control protocols (adiabatic passage, shortcuts to adiabaticity, composite pulses, geometric and RIO methods) usable in realistic multilevel hardware without redesign. It also establishes a link between KAM/superadiabatic perturbation theory and practical pulse shaping, and provides a semi-analytic alternative to gradient-based optimal control.

Real-world applications:

  • High-fidelity gate design for superconducting transmon quantum processors, where weak anharmonicity makes leakage to |2⟩ a dominant error channel.
  • Pulse-level calibration and error mitigation in cloud-accessible quantum computers and control electronics with limited bandwidth.
  • Other qubit platforms embedded in multilevel structures, and analogous multilevel leakage problems such as Bose–Einstein condensates.
  • Fast initial-guess generation for numerical optimal-control pipelines (GRAPE), reducing computation time in device calibration.

Industry relevance: The results are directly relevant to quantum-hardware vendors and control-software developers: the pulses are near-optimal in energy, compatible with existing two-level calibration protocols, and come with concrete hardware constraints (bandwidth cutoffs of 250 MHz and 146 MHz, and expected infidelities around −3.5 under realistic T₁ = 200 μs and T₂ = 135 μs).

Future Directions

  • Multiple simultaneous leakage channels. The authors outline how the W_j can be chosen to treat leakage both below and above the qubit space (e.g. the {|1⟩, |2⟩} qubit in a qutrit, or a d-level qudit) globally in a single step, rather than by recursive DRAG with multi-derivative shaping. This remains to be demonstrated in detail.

  • Limitations from time reparametrization. The reparametrization that enforces vanishing boundaries slightly degrades δ-robustness and is described as a potential limitation; finding boundary-satisfying alternatives or better reparametrizations is an open issue.

  • Incorporating other two-level protocols. The method is claimed to be versatile for any sufficiently smooth two-level control, including universal robust control techniques for arbitrary systematic unknown errors and geometric frameworks; testing these integrations is a natural next step.

  • Experimental validation and extension beyond complete population transfer. The analysis is performed for complete population transfer up to a phase, with generalization to partial transfers and general qubit gates stated but not shown; hardware demonstration with the reported bandwidth and decoherence constraints is the logical follow-up.

Target Audience

Specialists in quantum control and quantum information processing—particularly researchers and engineers working on superconducting qubit gate design, optimal control theory (GRAPE, RIO, shortcuts to adiabaticity), and adiabatic/perturbative methods such as KAM and superadiabatic theory. It is also relevant to quantum-hardware and control-software developers who need pulse shapes that are simultaneously leakage-free, error-robust, and bandwidth-feasible, and to graduate students with a background in quantum dynamics and perturbation theory.

Authors’ abstract

We introduce a multi-robust quantum control strategy designed to suppress multilevel leakage outside the computational qubit space while simultaneously ensuring robustness against unknown inhomogeneities in control parameters. It is based on a reformulation of the Derivative Removal by Adiabatic Gate (DRAG) technique which allows for the full incorporation of control pulses originally designed for two-level models by using the iterative quantum Kolmogorov-Arnold-Moser (KAM)…

The full abstract on arXiv
Read the original paper