⚛️ quantum physics

Distance Prediction for Quantum Codes

This paper introduces BB-Net, an attention-based neural network that significantly outperforms state-of-the-art heuristics in predicting the distance of Bivariate Bicycle quantum codes with high accuracy and minimal computational cost, thereby accelerating the discovery of optimal quantum error-correcting codes.

Original authors: Nadav Schneider, Eliya Nachmani

Published 2026-10-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nadav Schneider, Eliya Nachmani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Quantum computers promise to solve problems that would take classical machines thousands of years, from cracking unbreakable codes to simulating complex molecules. Yet, the physical building blocks of these machines, known as qubits, are incredibly fragile. They are so sensitive to heat, vibration, and electromagnetic noise that they lose their information almost instantly, a phenomenon called decoherence. To build a useful machine, scientists must wrap these delicate qubits in layers of protection, a system known as quantum error correction. This system works by spreading a single piece of information across many physical qubits, creating a safety net that can detect and fix mistakes without destroying the data. The strength of this safety net is measured by a single number called the "distance." Think of this distance as the number of independent errors a code can withstand before the information becomes unrecoverable. A higher distance means a more robust code, capable of running longer and more complex calculations.

Finding the best codes is a massive search problem. Scientists have developed families of codes, such as those based on bivariate bicycle structures, which offer a promising balance between efficiency and protection. However, determining the exact distance of a specific code is notoriously difficult. It is a computational puzzle so complex that even the most powerful supercomputers struggle to solve it for large codes. Researchers currently rely on two main approaches: exact solvers that guarantee a correct answer but take days or weeks to run, or faster heuristic searches that provide a best-guess estimate but can sometimes miss the true limit. This bottleneck slows down the discovery of better codes, as researchers cannot quickly screen the millions of potential designs to find the few that are truly exceptional.

In a new study, researchers Nadav Schneider and Eliya Nachmani from Ben-Gurion University of the Negev have introduced a different way forward. Instead of trying to solve the distance puzzle from scratch every time, they trained an artificial intelligence to predict the answer. They built a dataset of nearly 59,000 quantum codes, calculating the distance for each one using expensive, high-budget computer searches. These calculations served as the "answers" for a machine learning model they named BB-Net. The model was designed to look at the mathematical blueprint of a code—specifically the arrangement of its protective rules—and learn the patterns that correlate with high or low distances. By analyzing the structure of the code directly, rather than running a slow simulation, the AI learned to make rapid predictions.

The results show that this approach is remarkably effective. When tested on a set of codes it had never seen before, BB-Net predicted the distance with an accuracy of 96.39 percent. This performance was achieved in a fraction of a second per code. In contrast, the best existing computer methods, which are allowed to run for up to ten thousand times longer, reached an accuracy of only 94.86 percent. On a smaller subset of codes where the exact distance was known with absolute certainty, the AI's accuracy climbed to 98.94 percent. The study demonstrates that the AI can outperform traditional, computationally heavy methods while using a tiny fraction of the time. This speed allows researchers to screen vast numbers of potential codes almost instantly, filtering out the weak ones and saving the expensive, slow calculations for only the most promising candidates.

The researchers did not just rely on raw speed; they also engineered the way the computer "reads" the code to make the task easier for the AI. They developed a method to translate the complex algebraic description of a code into a compact list of tokens, similar to how a sentence is broken down into words. They created two different views of the code: one that looks at the local connections between the rules, and another that looks at the global, repeating patterns of the entire structure. By combining these two perspectives, the model learned a deep understanding of what makes a code strong. They also taught the model to recognize that shifting the starting point of a code does not change its strength, ensuring it focused on the essential properties rather than arbitrary details.

This work represents a significant shift in how scientists approach the design of quantum error-correcting codes. For years, the field has been limited by the time it takes to verify a code's strength. By introducing a tool that can predict this strength with high confidence in milliseconds, the researchers have removed a major barrier to discovery. The study confirms that machine learning can learn the hidden rules of quantum error correction from data, offering a practical alternative to repeated, expensive computations. While the AI provides estimates rather than absolute mathematical proofs for every single case, its ability to quickly identify high-quality codes accelerates the search for the robust, scalable systems needed to build the quantum computers of the future. The code for this new tool is now available to the public, inviting other scientists to use it to explore the vast landscape of quantum possibilities.

Technical Summary: Distance Prediction for Quantum Codes (BB-Net)

Problem Statement

The minimum distance ( $d$ ) of a quantum error-correcting code (QECC) is the primary metric determining its error-correction capability. For Quantum Low-Density Parity-Check (qLDPC) codes, specifically the Bivariate Bicycle (BB) family and their Perturbed (PBB) extensions, computing this distance is a significant computational bottleneck. The decision problem of determining if a code has a distance below a certain threshold is NP-complete, and exact computation is NP-hard.

Current approaches rely on:

Exact Solvers: (e.g., Mixed Integer Linear Programming) which are computationally intractable for moderate-to-large code sizes.
Heuristic Searches: (e.g., QDistRnd, QDistEvol, BP-OSD) which provide distance upper bounds but require substantial computational budgets (often thousands of trials) to approach tight bounds.

This computational cost severely restricts the ability to screen large design spaces of candidate codes, hindering the discovery of optimal qLDPC codes necessary for scalable fault-tolerant quantum computing.

Methodology: BB-Net

The authors introduce BB-Net, an attention-based neural network architecture designed to predict the distance of BB and PBB codes directly from their algebraic specifications, bypassing the need for iterative search.

1. Data Construction

Dataset: A corpus of 58,704 code examples (CSS BB and non-CSS PBB types) with block lengths $n \le 360$ .
Labels: Distance upper bounds obtained via high-budget heuristic searches (10,000 trials per code using BP-OSD, QDistRnd, and QDistEvol).
Ground Truth: A subset of 2,366 test codes where the distance was certified exactly using an MILP solver within a 2-hour time limit.

2. Input Representation (Tokenization)

To avoid constructing large parity-check matrices, the authors propose a compact tokenization scheme exploiting the cyclic algebraic structure of BB codes:

Polynomial Tokens: Represent the local algebraic relations. Each monomial in the sparse defining polynomials ( $A, B, C, D$ ) is tokenized with a one-hot polynomial identity and positional embeddings using sine/cosine functions of the cyclic coordinates (Fourier harmonics).
Tanner Spectral Tokens: Represent global structural properties. These encode the responses of the full polynomials at each character of the dual group $\mathbb{Z}_\ell \times \mathbb{Z}_m$ . Features include normalized polynomial magnitudes, spectral energy, and cross-polynomial interactions.

3. Architecture

BB-Net employs a dual-branch encoder followed by a fusion mechanism:

Polynomial Encoder: Processes polynomial tokens using multi-head self-attention. It is conditioned on global code parameters ( $\ell, m, n, k$ ) via FiLM (Feature-wise Linear Modulation) layers.
Tanner Spectral Encoder: Processes spectral tokens using a parallel attention stack.
Fusion: The latent representations from both branches are combined via a learned gated residual connection: $h_{fuse} = h_{poly} + \tanh(g) \odot h_{tanner}$ .
Head: A classification head predicts the distance as a categorical distribution over possible distance values.

4. Training Augmentation

To ensure invariance to cyclic shifts and symmetries:

Canonicalization: Inputs are mapped to a deterministic representative of their cyclic-shift equivalence class.
Augmentation: Random automorphisms of the underlying group and cyclic translations are applied to the polynomial supports during training.

Key Results

Accuracy and Efficiency

On a test benchmark of 5,870 codes, BB-Net demonstrates superior efficiency and competitive accuracy compared to state-of-the-art heuristics:

Accuracy: BB-Net achieves 96.39% accuracy (within $\pm 2$ of the reference distance) on the full test set. This outperforms the best heuristic baseline (QDistEvol) at 94.86%, despite QDistEvol using up to $10^4$ times more computational runtime.
Certified Subset: On the subset of 2,366 codes with certified exact distances, BB-Net achieves 98.94% accuracy.
Speed: BB-Net evaluates a single code in 0.005 seconds, compared to 1.0s for QDistEvol, 3.3s for QDistRnd, and 19.1s for BP-OSD.

Distance Screening Performance

In the context of screening pipelines (filtering candidates by a distance threshold $t$ ), BB-Net achieves Area Under the Curve (AUC) scores comparable to or slightly below the best heuristics at intermediate thresholds but with negligible inference time. For high-distance thresholds ( $d \ge 20$ ), BB-Net maintains an AUC of 0.990, comparable to QDistEvol (0.991).

Ablation Studies

Experiments confirm the necessity of specific architectural components:

Removing the Tanner spectral branch increases the Negative Log-Likelihood (NLL) from 0.683 to 1.039.
Removing augmentation, shift canonicalization, or FiLM conditioning significantly degrades performance, highlighting the importance of symmetry handling and parameter conditioning.

Significance and Claims

The paper positions BB-Net as a practical alternative to repeated, computationally expensive distance estimation. Its primary contributions are:

First Learned Model: It is the first work to predict quantum code distance directly using a learned model, framing the problem as a machine learning task trained on expensive heuristic labels.
Bottleneck Relief: By reducing the distance estimation cost by orders of magnitude, BB-Net enables the rapid screening of large code spaces, particularly within ML-driven search pipelines.
Accelerated Discovery: The authors claim that by alleviating the distance-calculation bottleneck, this approach can accelerate the discovery of novel, high-quality quantum error-correcting codes, a capability essential for advancing scalable fault-tolerant quantum computing.

The authors note that while heuristics provide upper bounds, BB-Net learns to predict these bounds effectively, offering a "fast filter" that allows researchers to focus expensive exact solvers only on the most promising candidates.