Enhanced Maximum Independent Set Preparation with Rydberg Atoms Guided by the Spectral Gap

1. Problem Statement

The paper addresses the fundamental performance limitation of Adiabatic Quantum Computation (AQC) on Rydberg atom platforms when solving combinatorial optimization problems, specifically the Maximum Independent Set (MIS) problem.

• The Core Issue: The success of AQC relies on the system remaining in the ground state throughout the evolution. However, as system size and connectivity increase, the minimum spectral gap ($g_{min}$) between the ground state and the first excited state decreases (often superexponentially).
• Consequence: During finite-time evolution, this narrowing gap causes population leakage from the ground state to excited states, significantly reducing the probability of finding the correct MIS solution.
• Limitations of Existing Solutions:
  • Counterdiabatic driving: Requires adding complex, often non-local Hamiltonian terms that are difficult to implement on current hardware.
  • Variational/Bayesian optimization: Requires extensive sampling and repeated circuit executions, leading to high hardware overhead and long calibration times.
  • Standard schedules: Linear or simple polynomial detuning sweeps fail to adequately slow down evolution near the minimum gap, leading to errors.

2. Methodology: ADGLB

The authors propose a novel, hardware-efficient method called Adjusted Detuning for Ground-Energy Leakage Blockade (ADGLB). This method modifies the laser detuning schedule without introducing new Hamiltonian terms or iterative optimization loops.

• Theoretical Basis: The method is grounded in the adiabatic theorem, which states that population leakage is bounded by the inverse power of the spectral gap, $(\Delta E_{01})^{-j}$. To suppress leakage, the rate of change of the Hamiltonian ($dH/dt$) must be reduced when the spectral gap is small.
• Implementation Strategy:
  1. Focus on Detuning: The Rabi frequency $\Omega(t)$ follows the standard schedule. Only the laser detuning $\delta(t)$ is modified during the central evolution phase (where $\Omega$ is at its maximum).
  2. Gap-Guided Interpolation: The detuning path is constructed using a normalized interpolation function $\zeta_j(t)$ that depends on the spectral gap $\Delta E_{01}(t)$ raised to a power $j$:
     $$\frac{d\delta}{dt} \propto (\Delta E_{01}(t))^j$$
  3. Schedule Construction: The time interval is split at $t_{min}$ (where the gap is minimal). The detuning is interpolated such that the evolution slows down significantly near $t_{min}$ and speeds up elsewhere.
  4. Parameter $j$: A tunable parameter ($1 \le j \le 2$) controls the "steepness" of the slowdown. The authors found $j \approx 1.5$ to be optimal for balancing slowdown and total evolution time.

3. Key Contributions

• Hardware-Efficient Schedule Engineering: ADGLB achieves performance gains purely by reshaping the control pulse (detuning), avoiding the need for complex counterdiabatic terms or extensive classical optimization loops.
• Scalability and Transferability: The paper demonstrates that a schedule optimized for a small instance ($N=10$) can be directly applied to larger, more complex 2D lattices ($N=25, 37$) without re-optimization, provided the "hardness" of the problem is similar.
• Heuristic Extension for Harder Instances: For instances with significantly higher hardness parameters (where direct diagonalization is impossible), the authors introduce a heuristic offset ($\nu_d$) to the detuning value at the minimum gap. This allows the small-system optimized schedule to be adapted for larger, harder systems.

4. Experimental Results

The method was validated using QuEra's Aquila neutral-atom processor (256-qubit) via Amazon Braket.

• System 1: Quasi-1D Chain ($N=10$, $Q1D,10$)

  • Hardness: High ($H_P = 6.5$).
  • Result: The MIS preparation probability ($P_{MIS}$) increased from 28% (Standard Schedule) to 38% (ADGLB with $j=1.8$).
  • Observation: Numerical simulations predicted even higher gains ($>94\%$), but experimental imperfections (SPAM errors, decoherence) limited the absolute fidelity, though the relative improvement was clear.

• System 2: 2D Triangular Lattices ($N=25$ and $N=37$)

  • Hardness: Comparable to the $N=10$ chain ($H_P \approx 5.4 - 6.2$).
  • Result: The $N=10$ optimized schedule was applied directly.
    • $N=25$: $P_{MIS}$ increased from 17.7% to 24.2%.
    • $N=37$: $P_{MIS}$ increased from 3.0% to 5.6%.
  • Significance: This proves the schedule captures structural features of the problem hardness rather than just system size.

• System 3: Higher Hardness Instance ($N=23$, $Q'_{1D,23}$)

  • Hardness: Very high ($H_P = 34.2$).
  • Method: The $N=10$ schedule was used as a base, with a heuristic offset $\nu_d$ applied to the detuning at the minimum gap ($\delta'_{min} = \delta_{min} + \nu_d$).
  • Result: With an optimal offset of $\nu_d = 2\pi \times 0.2$ MHz, $P_{MIS}$ reached 2.54%, surpassing the standard schedule (1.45%) and the unmodified ADGLB schedule (1.1%).

5. Significance and Conclusion

• Scalability: ADGLB offers a scalable strategy for enhancing AQC on neutral-atom platforms. It does not require increasing the evolution time (which is limited by coherence) or adding complex control terms.
• Generalizability: The finding that small-system schedules transfer to larger systems is crucial for near-term quantum advantage, as it reduces the calibration overhead for large-scale problems.
• Hardware Efficiency: By improving the raw probability distribution at the hardware level, this method also aids classical post-processing and error mitigation strategies, making the overall quantum optimization pipeline more effective.

In summary, the paper demonstrates that spectral-gap-guided schedule engineering is a powerful, low-overhead technique to mitigate adiabatic leakage, significantly boosting the performance of Rydberg atom processors on hard combinatorial optimization problems.