Analytical Theory of Greedy Peeling for Bivariate Bicycle Codes and Two-Shot Streaming Decoding

TL;DR

This paper develops an analytical theory for greedy peeling decoding of bivariate bicycle codes, demonstrating significant latency reduction and establishing a closed-form collision resolution factor that predicts decoding success across various codes and noise levels.

Contribution

It introduces a parameter-free analytical model for collision resolution in peeling decoding, linking it to code structure and enabling accurate performance predictions.

Findings

Decoding latency is reduced by 330x compared to belief propagation at p=10^{-3}.

The collision resolution factor A_0 closely matches empirical values for different codes.

Two-shot streaming decoding achieves 89% success with low latency (~50 ns).

Abstract

We present an analytical theory of greedy peeling decoding for bivariate bicycle (BB) codes under circuit-level noise. The deferred greedy decoder achieves 330x latency reduction over belief propagation (BP) at p = 10^{-3} while maintaining identical logical error rate. Our main theoretical contribution is a closed-form collision resolution factor A_0 = |true collisions| / |birthday collisions|, derived from XOR syndrome analysis with no free parameters, that quantifies the fraction of detector-sharing fault pairs genuinely blocking iterative peeling. For the [[144,12,12]] Gross code, A_0 = 0.8685 (within 0.5% of the empirical value), with shared-2 pairs (4-cycles) always resolving under peeling. We show A_0 depends on the mean fault-graph degree d-bar rather than code size: A_0 = 0.87 for d-bar = 52 (Gross family) versus A_0 = 0.76 for d-bar = 17 ([[32,8,6]]). We establish a syndrome code stopping distance d_S = n/4.5 for the Gross family and demonstrate that [[32,8,6]] (d_S = 4) enables two-shot streaming decoding: T = 2 rounds achieve 89% peeling success with 1.29 +/- 0.03 LER ratio versus T = 12, at estimated latency ~50 ns. The full formula P_peel = exp(-A_0 * gamma_analytic * exp(-BTp) * n * p^2) is validated across five BB codes, four noise levels, and four values of T with R^2 = 0.86. Cross-platform reproduction of the Kunlun [[18,4,4]] experiment matches their hardware LER within 0.73 percentage points.