Bivariate-Bicycle Quantum Codes

• Bivariate-Bicycle Codes are quantum low-density parity-check codes that use two sparse circulant matrices to achieve high encoding rates and constant-weight stabilizers.
• They enable efficient decoding through methods such as BP+OSD and MBBP-LD, significantly reducing logical error rates in practical implementations.
• Their algebraic structure supports fault-tolerant operations and scalable deployment on modern quantum architectures like superconducting processors and neutral atom arrays.

Bivariate-Bicycle Codes

Bivariate-bicycle (BB) codes are a class of quantum low-density parity-check (QLDPC) codes that generalize classical bicycle codes to quantum settings using a pair of sparse circulant (or group-algebraic) matrices, enabling efficient quantum error correction with relatively high encoding rates and constant-weight stabilizers. Originating from constructions over cyclic and group-algebraic rings, BB codes admit systematic algebraic analysis of their code parameters, support efficient decoding algorithms, and, through extensions, offer practical trade-offs for implementation architectures and specialized noise environments. They also serve as a testing ground for advanced techniques in fault-tolerant quantum computation, including morphing circuits, erasure decoding, and architectural modularity.

1. Algebraic Construction and Code Parameters

A standard bivariate-bicycle code is defined over a finite group algebra, typically $R = \mathbb{F}_2[x, y]/(x^\ell-1, y^m-1)$, corresponding to the bicyclic group $\mathbb{Z}_\ell \times \mathbb{Z}_m$ (Rabeti et al., 4 Nov 2025, Postema et al., 24 Feb 2025). Two polynomials $A(x,y)$ and $B(x,y)$, each sparse and of bounded Hamming weight, specify circulant (or block-circulant) matrices by interpreting powers of $x$ and $y$ as translation operators along the two torus axes.

The Calderbank-Shor-Steane (CSS) parity-check matrices for a BB code take the form: $$H_X = [A \;|\; B], \quad H_Z = [B^\top \;|\; A^\top],$$ with the commutation condition $A B^\top + B A^\top = 0$ imposed either explicitly or by construction. The code encodes

$$n = 2 \ell m, \qquad k = 2 \deg \gcd(A(x,y), B(x,y), x^\ell-1, y^m-1),$$

logical qubits, with code rate $R = k / n$. The minimum distance $\mathbb{Z}_\ell \times \mathbb{Z}_m$0 is typically determined numerically for given $\mathbb{Z}_\ell \times \mathbb{Z}_m$1 and may, in favorable cases, scale linearly with system size for moderate $\mathbb{Z}_\ell \times \mathbb{Z}_m$2 (Rabeti et al., 4 Nov 2025, Postema et al., 24 Feb 2025).

In the coprime case ($\mathbb{Z}_\ell \times \mathbb{Z}_m$3), a ring isomorphism reduces the construction to a univariate case $\mathbb{Z}_\ell \times \mathbb{Z}_m$4, where all code parameters can be characterized in terms of a greatest common divisor polynomial $\mathbb{Z}_\ell \times \mathbb{Z}_m$5 (Rowshan, 3 Jan 2026, Postema et al., 24 Feb 2025).

2. Decoding Algorithms and Single-Shot Performance

BB codes admit efficient decoding strategies leveraging their LDPC structure. The belief-propagation (BP) decoder, often combined with ordered-statistics decoding (BP+OSD), is the standard approach. Recent work introduces the Multiple-Bases Belief-Propagation List Decoder (MBBP-LD), which outperforms BP+OSD, achieving up to 40% lower logical error rate for moderate-size BB codes while maintaining linear-time decoding (Rabeti et al., 4 Nov 2025).

For BB codes with additional structure, specifically the coprime subclass, the stabilizer redundancy is tightly linked to the quantum rate. Single- and few-shot decoding schemes are governed by the redundancy density $\mathbb{Z}_\ell \times \mathbb{Z}_m$6, with

$\mathbb{Z}_\ell \times \mathbb{Z}_m$7

and measurement-robustness constrained by classical code bounds (BCH and Singleton). The redundancy in X- and Z-type stabilizers enables syndrome cleaning and single-shot correction, but the maximum achievable syndrome distance is fundamentally bounded by the code's rate, creating a rate-distance trade-off for measurement-limited architectures (Rowshan, 3 Jan 2026).

Machine learning approaches, such as transformer-based decoders, have demonstrated improved logical error rates and lower latency over conventional BP+OSD for short BB codes, making them promising for practical medium-scale QLDPC decoding (Blue et al., 17 Apr 2025).

3. Variants: Univariate, Coprime, Self-Dual, and Bias-Tailored Constructions

Several subclasses and extensions of BB codes provide specialized trade-offs:

• Univariate Bicycle (UB) Codes: These restrict the defining polynomials by setting $\mathbb{Z}_\ell \times \mathbb{Z}_m$8 and $\mathbb{Z}_\ell \times \mathbb{Z}_m$9 (mod $A(x,y)$0), reducing the search space to $A(x,y)$1 and automatically satisfying the CSS condition. Such codes enable efficient search for low-overhead codes with comparable or improved BP decoder performance (Rabeti et al., 4 Nov 2025).
• Coprime BB Codes: By constructing codes where the factors of the group-algebra modulus are coprime, the code rate can be fixed in advance by the choice of factors, facilitating highly efficient algebraic and numerical search for good parameter sets (Wang et al., 2024, Rowshan, 3 Jan 2026).
• Self-Dual BB Codes: Formed when $A(x,y)$2, these codes support full sets of transversal Clifford gates, including CNOT, Hadamard, and $A(x,y)$3 gates, with weight-8 stabilizers on twisted tori yielding parameters such as $A(x,y)$4 (Liang et al., 6 Oct 2025).
• Bias-Tailored and Clifford-Deformed Codes: Codes constructed to exploit asymmetries in physical error channels (noise-bias) combine two classical LDPC codes (e.g., cellular automaton rules) on separate sublattices, then apply a Clifford deformation (Hadamard on one sublattice), resulting in mixed (half-X, half-Z) stabilizers. In high-bias regimes, the effective code distance approaches that of the underlying classical codes, leading to improved logical error suppression relative to both standard BB codes and bias-tailored surface/XZZX codes (Leroux et al., 30 May 2025).

4. Scaling Laws, Limits, and Code Coverings

While BB codes outperform planar surface codes in encoding efficiency and offer high rates for modest block lengths, they are asymptotically "bad" in the sense of the Bravyi-Poulin-Terhal bound—neither rate nor relative minimum distance remain positive in the limit $A(x,y)$5 (Postema et al., 24 Feb 2025). Specifically,

$A(x,y)$6

for fixed check weights. However, infinite sequences of BB codes with systematically scaled parameters can be generated via covering graphs, with $A(x,y)$7-fold covers yielding new codes with $A(x,y)$8, $A(x,y)$9, and $B(x,y)$0 (for odd $B(x,y)$1), reducing search complexity and inducing explicit homological relations between coverings and logical operators (Symons et al., 17 Nov 2025).

Recent results leverage ring-theoretic approaches (Gröbner bases for the two-variable ideal generated by stabilizer polynomials and boundary conditions) to compute code parameters, analyze logical operator structure, and optimize locality/stabilizer weight under twisted torus boundary conditions (Liang et al., 5 Mar 2025, Halla, 4 Feb 2026). These methods have produced superior rate–distance trade-offs for finite-length codes, including qudit generalizations with improved $B(x,y)$2 (Halla, 4 Feb 2026).

5. Implementation, Circuit Design, and Architectural Integration

BB codes are compatible with modern circuit architectures, including superconducting processors with long-range coupling and neutral atom tweezer arrays. Experimental demonstrations include a distance-4 BB code on a 32-qubit transmon processor, achieving a per-logical-qubit error rate of 8.9% per cycle—comparable to, but at substantially lower overhead than, distance-4 surface codes (Wang et al., 14 May 2025). The practical syndrome extraction protocols for BB codes exploit the periodicity and circulant structure for efficient, constant-depth implementation, with morphing circuits enabling reduced connectivity (degree 5 rather than 6) and biplanar layouts (Shaw et al., 2024).

Modular architectures, such as the "bicycle architecture," use BB code modules with resource-optimized circuit synthesis, enabling an order-of-magnitude improvement in logical capacity relative to surface code architectures for fixed physical qubit budgets (Yoder et al., 3 Jun 2025). Importantly, systematic pruning procedures allow the construction of planar BB codes with open boundary conditions, retaining check locality and fold-transversal Clifford gates (Eberhardt et al., 2024).

Under erasure noise, the BiBiEQ framework produces erasure-aware memory circuits and two decoding engines (exact and independence-approximate), facilitating quantitative assessment of BB performance under various erasure check schedules and confirming sub-threshold exponential scaling in logical error with code distance (Bhave et al., 7 Feb 2026).

6. Topological Structure and Logical Operator Analysis

The stabilizer structure of BB codes admits a rigorous topological order framework. The classification of code parameters, logical degeneracies, and anyon-type excitation content is carried out using homological algebra (chain complexes) and algebraic geometry (e.g., BKK theorem via Newton polytopes). Phenomena such as topological frustration (the ground-state degeneracy of finite-size BB codes is typically lower than the asymptotic anyon number) and quasi-fractonic mobility (restricted anyon transport) are identified, with logical operators and anyon-type strings related via Koszul complex duality (Chen et al., 6 Mar 2025).

Fold-transversal Clifford gates, essential for efficient fault-tolerant logical operations, arise naturally in symmetric BB codes (e.g., those with $B(x,y)$3), and their action survives planarization and pruning, ensuring that medium-scale patches can maintain full logical universality without increasing circuit depth (Eberhardt et al., 2024, Eberhardt et al., 2024, Liang et al., 6 Oct 2025).

7. Practical Performance and Applications

Simulations and experiments consistently demonstrate that BB codes can achieve logical error rates and overheads favorable to surface codes at short and moderate block sizes ($B(x,y)$4), particularly when compiled to architectures that support long-range interactions or efficient routing (Berthusen et al., 2024, Wang et al., 14 May 2025). Thresholds under phenomenological and circuit-level noise are competitive ($B(x,y)$5), and specialized variants dramatically improve logical performance under biased noise or erasure-dominated noise models (Leroux et al., 30 May 2025, Bhave et al., 7 Feb 2026).

The combination of algebraic transparency, transversal gate support, and decoder performance makes BB codes attractive for moderate-scale fault-tolerant computation, experimental exploration of advanced QEC concepts (single-shot, morphing circuits), and as a laboratory for comparing decoding algorithms from classical BP to contemporary machine learning methods (Rabeti et al., 4 Nov 2025, Blue et al., 17 Apr 2025).