arXiv:2610.25213
LiftDiff: Guided Discrete Diffusion for Quantum Code Construction
Yotam Peled, Eliya Nachmani
Abstract
The realization of fault-tolerant quantum computation relies on Quantum Error-Correcting Codes (QECC). Of particular interest are codes that combine a high encoding rate with sufficient distance. Existing algebraic constructions, such as lifted product (LP) codes, build a large quantum low-density parity-check (qLDPC) code from a classical base matrix. In this work we introduce LiftDiff, a generative approach to the discovery of QECC. A discrete denoising diffusion model is trained to generate base matrices over the group algebra $\mathbb{F}_2[\mathbb{Z}_\ell]$, which are then expanded into LP codes. The reverse diffusion process is guided by an energy function that evaluates quantum code properties, including girth, degree, sparsity and decoding performance, steering the sampler toward matrices that yield LP codes with favorable parameters. Our method discovers codes with parameters such as $[[169,13,11]]$, $[[143,15,9]]$, and $[[260,24,10]]$, including a $[[247,19,14]]$ code that achieves the highest ${kd^2}/{n}$ among the baselines. Moreover, several discovered codes outperform the Gross code across a broad range of the tested physical error rates under both code-capacity and circuit-level noise.
AI Summary
LiftDiff addresses the search for finite-size lifted-product (LP) quantum LDPC codes with favorable rate, distance, check weight, and decoding performance. It generates base matrices over $\mathbb{F}_2[\mathbb{Z}_\ell]$ and uses the self-product construction $LP(M,M)$ to enforce CSS commutation. A discrete denoising diffusion model is trained on progressive-edge-growth supports with random polynomial entries. During sampling, gradients through a continuous relaxation of the LP construction guide generation using penalties for short cycles, degree, sparsity, spectral properties, rank loss, check weight, and error-reconstruction loss from an unrolled belief-propagation decoder. Candidates are screened using exact binary ranks and distance upper bounds; the paper reports MILP certification of final candidates’ distances. The reported set contains 32 codes with $65\le n\le730$, $5\le d\le14$, rates up to $0.310$, and maximum check weights from 5 to 12. Examples include $[[143,15,9]]$, $[[169,13,11]]$, and $[[260,24,10]]$. The $[[247,19,14]]$ code has $kd^2/n\approx15.1$, the largest among the compared codes with check weight at most 8. Under independent-$X$ code-capacity noise and BP-OSD decoding, the $[[260,24,10]]$ code has about twentyfold lower block logical error rate than the $[[144,12,12]]$ baseline at $p=0.01$. Circuit-level tests find $[[247,19,14]]$ below that baseline for $p=0.001$–$0.003$, while $[[260,24,10]]$ remains within its confidence intervals. The spectral distance proxy is explicitly heuristic for continuous, potentially irregular matrices. Guidance weights are tuned per run, analysis plots use distance upper bounds, and exact distance computation is acknowledged to be NP-hard in general.
Error found: The abstract claims several discovered codes outperform the Gross code under both noise models, but the circuit-level results establish an advantage for only one generated code over part of the tested range; the other tested generated code is statistically indistinguishable from the baseline.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.