Embedding-Aware Noise Modeling of Quantum Annealing

Seon-Geun Jeong, Ming Cong, Dae-Il Noh, Quoc‐Viet Pham, Won‐Joo HwangarXiv 2025 · DOI · arXiv

Quantum annealing (QA) is one of the most accessible paths to solving complex optimization problems today, but it faces a persistent "connectivity gap." While we want to solve problems where every variable might interact with every other variable, real-world hardware is sparse. To bridge this gap, researchers use minor embedding, which bundles multiple physical qubits into "chains" to act as a single logical variable.

The catch? The longer the chain, the more likely it is to "break" due to hardware noise. A new paper, Embedding-Aware Noise Modeling of Quantum Annealing, provides a rigorous mathematical bridge between the overhead of these embeddings and the reliability of the hardware.

TL;DR

This paper develops a mathematical framework to predict how hardware noise scales with problem size in D-Wave’s latest Zephyr topology. By linking chain length to Gaussian control errors, the authors derive closed-form laws for chain breaks and prove that to maintain reliability, chain strength must scale sub-linearly with problem size—a finding they validated on real quantum hardware.

What Problem Does This Paper Address?

In an ideal quantum annealer, every qubit would be connected to every other qubit. In reality, D-Wave’s Advantage2 system uses the Zephyr topology, where each qubit has a limited number of neighbors. To map a "dense" problem onto this "sparse" hardware, we create chains of qubits.

However, these chains are the Achilles' heel of QA. If the noise in the system (Integrated Control Error, or ICE) overpowers the "chain strength" (the magnetic glue holding the qubits together), the chain breaks. A broken chain means the hardware is no longer solving the problem you intended. Until now, tuning these chains was largely a matter of trial and error. This paper seeks to replace that "heuristic guesswork" with predictive physics.

Methods

The authors built their framework on three main pillars:

  1. Structural Analysis: They utilized the fact that in "clique embeddings" (the worst-case scenario where every variable connects to every other), the average chain length grows linearly with the problem size.
  2. Statistical Noise Modeling: They treated hardware errors as independent Gaussian perturbations (zero-mean noise) affecting local fields and couplers.
  3. The Failure Criterion: They defined a "chain break" as occurring when the accumulated noise along a chain exceeds a specific stabilizing margin.

By combining these, they derived closed-form expressions for Chain Break Probability (CBP) and Chain Break Fraction (CBF) using the complementary error function ($erfc$). Crucially, they derived a theoretical "scaling rule": to keep a chain stable as it grows, the chain strength ($k$) should scale as the square root of the chain length ($k \sim O(\sqrt{\ell})$).

Datasets

The researchers didn't use a static dataset but rather a synthetic benchmark suite:

  • Random QUBO Instances: They generated Quadratic Unconstrained Binary Optimization (QUBO) problems of varying sizes ($L = 5$ to $100$ logical variables).
  • Hardware: All experiments were run on the D-Wave Advantage2_system1.6 QPU, which features the Zephyr topology.
  • Baselines: They compared the QPU's performance against classical solvers, including Simulated Annealing (SA) and industrial-grade Mixed-Integer Linear Programming (MILP) solvers like Gurobi and PuLP.

Key Insights and Ideas

  • Linear Growth, Linear Noise: The paper proves that the variance of accumulated error grows linearly with chain length. This formalizes the intuition that longer chains are "noise magnets."
  • The "erfc" Relationship: The probability of a chain breaking follows a Gaussian tail distribution. This means that once noise reaches a certain threshold, reliability doesn't just dip—it falls off a cliff.
  • The Theory-Practice Gap: While the math suggested a $\sqrt{\ell}$ scaling for chain strength, experiments on real hardware showed a steeper requirement ($\ell^{0.82-0.85}$). This is a major "aha" moment: it suggests that real-world hardware noise isn't just independent white noise; it’s likely correlated, meaning errors in one part of the chip are "talking" to others.
  • Predictable Performance: Unlike classical MILP solvers (which can take exponentially longer as problems grow), the Quantum Annealer provided solutions in a nearly constant, predictable time window while maintaining competitive energy quality.

Key Results

  • Model Accuracy: The theoretical model for Chain Break Fraction (CBF) matched the hardware data with a very low Sum of Squared Errors (SSE $\approx 0.0021$), proving the Gaussian noise assumption is a powerful tool for predicting hardware behavior.
  • Stability Boundaries: The researchers successfully mapped the "sweet spot" for chain strength. They showed that increasing chain strength ($k$) exponentially suppresses breaks but warned that if $k$ is too high, it "washes out" the actual problem logic, leading to poor solutions.
  • Competitive Scaling: In head-to-head tests, the D-Wave system matched the best energy results of Gurobi and SA for all problem sizes, but it significantly outperformed the open-source PuLP solver, which began to fail as the problem size exceeded 55 variables.

Limitations and Future Work

While the Gaussian model is highly effective, the authors acknowledge it is an approximation.

  • Correlated Noise: The discrepancy in scaling exponents suggests that future models need to account for global or correlated noise sources (like calibration drift or flux offsets).
  • Beyond ICE: The current model focuses on control errors (ICE) but doesn't explicitly model thermal noise or decoherence (quantum "fuzziness" caused by the environment).
  • Next Steps: The authors suggest integrating these analytic laws into "hardware-aware" simulators. This would allow researchers to "pre-tune" their quantum experiments on a laptop before ever spending expensive time on a real QPU.

Summary updated

Abstract

Quantum annealing provides a practical realization of adiabatic quantum computation and has emerged as a promising approach for solving large-scale combinatorial optimization problems. However, current devices remain constrained by sparse hardware connectivity, which requires embedding logical variables into chains of physical qubits. This embedding overhead limits scalability and reduces reliability as longer chains are more prone to noise-induced errors. In this work, building on the known structural result that the average chain length in clique embeddings grows linearly with the problem size, we develop a mathematical framework that connects embedding-induced overhead with hardware noise in D-Wave's Zephyr topology. Our analysis derives closed-form expressions for chain break probability and chain break fraction under a Gaussian control error model, establishing how noise scales with embedding size and how chain strength should be adjusted with chain length to maintain reliability. Experimental results from the Zephyr topology-based quantum processing unit confirm the accuracy of these predictions, demonstrating both the validity of the theoretical noise model and the practical relevance of the derived scaling rule. Beyond validating a theoretical model against hardware data, our findings establish a general embedding-aware noise framework that explains the trade-off between chain stability and logical coupler fidelity. Our framework advances the understanding of noise amplification in current devices and provides quantitative guidance for embedding-aware parameter tuning strategies.
BibTeX
@misc{jeong2025embedding,
  author = {Seon-Geun Jeong and Ming Cong and Dae-Il Noh and Quoc‐Viet Pham and Won‐Joo Hwang},
  title = {{Embedding-Aware Noise Modeling of Quantum Annealing}},
  year = {2025},
  eprint = {2510.04594},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.04594},
  howpublished = {arXiv},
}

Loading related content...

Figure