Direct entanglement ansatz learning (DEAL) with ZNE on error-prone superconducting qubits

1. Problem Statement

Quantum combinatorial optimization algorithms, particularly the Quantum Approximate Optimization Algorithm (QAOA), face significant hurdles when deployed on current Noisy Intermediate-Scale Quantum (NISQ) hardware. The primary challenges include:

• Complex Optimization Landscapes: QAOA often suffers from "barren plateaus" (vanishing gradients) and numerous local minima, making convergence to the global optimum difficult.
• Hardware Noise: Superconducting qubits are susceptible to crosstalk, coherence errors ($T_1$, $T_2$), and readout errors, which degrade solution quality.
• Ansatz Sensitivity: The performance of QAOA is highly dependent on the initialization of parameters and the structure of the quantum circuit (ansatz). Random or fixed initialization often fails to capture problem-specific structures efficiently.
• Mapping Inefficiencies: Standard approaches often ignore the specific connectivity and error rates of the target Quantum Processing Unit (QPU), leading to suboptimal qubit mapping and increased error accumulation.

2. Methodology: Direct Entanglement Ansatz Learning (DEAL)

The authors propose DEAL, a framework designed to enhance QAOA on noisy superconducting hardware (specifically IBM QPUs). The methodology integrates three core components:

A. Qubit-Prioritized Normalization (QPN) & Direct Mapping

Instead of random or uniform parameter initialization, DEAL derives initial parameters directly from the problem structure (the QUBO matrix).

• Importance Scoring: It calculates an importance score ($s_i$) for each qubit based on the sum of absolute values of its corresponding row in the QUBO matrix ($Q$).
• Normalization: These scores are normalized to create a probability distribution ($w_i$).
• Angle Encoding: The mixer Hamiltonian angles ($\beta$) and cost Hamiltonian angles ($\gamma$) are initialized using non-linear trigonometric functions of these weights:
  • $\Phi^\gamma_{k,i} = \lambda_\gamma \cdot \frac{k}{p} \cdot \arccos(1 - 2w_i)$
  • $\Phi^\beta_{k,i} = \lambda_\beta \cdot (1 - \frac{k}{p}) \cdot \arcsin(\sqrt{w_i})$
• Dynamic Layering: Early layers emphasize cost unitaries (larger $\gamma$), while later layers prioritize mixing (larger $\beta$), creating a bias-variance trade-off that avoids saturation at boundaries (0 or $\pi$).

B. Dynamic Physical Qubit Mapping

DEAL employs a dynamic mapping strategy that sorts qubits by their importance weights ($w_i$) and maps them to the physical QPU topology to minimize the product of edge weights and physical distance/error rates. This leverages the SABRE layout algorithm to select qubits with minimal error rates for high-importance variables.

C. Adaptive Zero-Noise Extrapolation (ZNE)

To mitigate noise without requiring ancilla qubits or full error correction:

• Pauli Twirling & Gate-Flip: The method uses Pauli twirling and gate-flip corrections.
• Bayesian Refinement: Unlike standard polynomial fitting for ZNE, DEAL uses a Bayesian posterior update to refine noise scaling parameters ($\lambda$).
• Weighted Noise Scaling: The noise scaling factor is defined as $\lambda_{ij} = w_i \cdot w_j \cdot d_{ij}$, where $d_{ij}$ is the qubit distance. This ensures that high-importance qubits receive preferential treatment in error mitigation and circuit routing.

3. Key Contributions

1. Direct Parameter Initialization: The paper introduces a deterministic initialization scheme (QPN) that maps QUBO coefficients directly to circuit parameters. This reduces the search space for classical optimizers and mitigates the barren plateau problem by incorporating problem-specific bias.
2. End-to-End Noise Mitigation: The integration of adaptive ZNE with problem-aware qubit mapping allows for scalable operations on noisy hardware with minimal overhead, specifically targeting crosstalk and coherence errors.
3. Dynamic Hamiltonian Learning: The framework treats the ansatz construction as a learning process where the Hamiltonian parameters are dynamically adjusted based on the problem graph and hardware constraints.
4. New Metric (QNRE): The authors define the Quantum Noise-Limited Relative Error (QNRE) to quantify how close the observed energy is to the optimal ground state relative to the noise floor.

4. Experimental Results

The authors evaluated DEAL on IBM's Torino (133 qubits) and Marrakesh (156 qubits) superconducting processors, as well as high-performance simulators.

• Success Rate: DEAL consistently outperformed "vanilla" QAOA. On the Torino QPU, DEAL achieved a success rate improvement of over 14% in finding ground-state energies for 5-vertex problems across 10 layers.
• Convergence: DEAL demonstrated faster convergence rates and lower variance in optimization compared to random initialization.
• Problem Versatility: The method was successfully applied to three NP-hard problems:
  • MaxCut: Showed improved probability concentration on optimal solutions.
  • Traveling Salesman Problem (TSP): Successfully navigated the complex energy landscape.
  • Knapsack Problem (KP): Demonstrated effective exploration of latent spaces.
• Noise Resilience: While performance degraded significantly beyond 43–45 CZ gates due to hardware limits, DEAL maintained better stability than standard QAOA up to 7 layers. The use of Echoed Cross-Resonance (ECR) gates helped mitigate phase errors in deeper circuits.
• Eigenvalue Analysis: Numerical simulations showed that DEAL could extract near-optimal ground-state energies (lowest eigenvalues) for TSP, MaxCut, and KP, effectively avoiding inefficient local minima.

5. Significance

This work represents a significant step toward the practical applicability of quantum optimization on current noisy hardware.

• Bridging the Gap: By directly mapping problem structures to ansatz parameters, DEAL reduces the reliance on classical optimization loops, which are often the bottleneck in NISQ algorithms.
• Hardware Awareness: The methodology acknowledges that not all qubits are equal; by prioritizing low-error qubits for high-importance variables, it maximizes the utility of imperfect hardware.
• Scalability: The combination of direct mapping and adaptive ZNE offers a pathway to solve larger NP-hard problems on existing quantum processors without waiting for full fault tolerance.
• Paradigm Shift: It moves beyond treating QAOA as a generic black-box optimizer, instead proposing a "learning" framework where the ansatz is intrinsically linked to the problem's mathematical structure and the hardware's physical constraints.

In conclusion, DEAL provides a robust, noise-resilient framework for solving combinatorial optimization problems, demonstrating that tailored ansatz learning and advanced error mitigation can significantly enhance the performance of quantum algorithms on current superconducting devices.