Bivariate Bicycle & Quantum Tanner Codes

• The paper presents a formal framework comparing BB and QT qLDPC codes, highlighting their algebraic constructions, decoding strategies, and performance benchmarks.
• Bivariate Bicycle codes utilize polynomial rings to generate explicit logical operator bases and enable transversal Clifford gate implementations.
• Quantum Tanner codes employ lifted base codes on Cayley complexes to achieve constant encoding rates and linear minimum distance, validated through empirical studies.

Bivariate bicycle (BB) codes and quantum Tanner (QT) codes are two pivotal quantum low-density parity-check (qLDPC) code families, each providing a distinct algebraic-combinatorial framework for constructing CSS (Calderbank–Shor–Steane) codes with constant-weight sparse parity constraints and high asymptotic performance. Both have played central roles in advancing practical and theoretical quantum error correction, offering flexible construction, enhanced logical rates, explicit bases for logical operators, and in many cases Clifford gates with transversal or fold-transversal realization. This article details the formal structure, algebraic conditions, decoding considerations, logical operator bases, and empirical behavior of these families, and surveys recent results in erasure and depolarizing-channel decoding, code performance benchmarks, and gate implementations.

1. Algebraic Structure and Construction

1.1. Bivariate Bicycle Codes

BB codes generalize the univariate “bicycle code” to two spatial directions via the group algebra $R = \mathbb{F}_2[x,y]/(x^\ell - 1, y^m - 1)$, isomorphic to $\mathbb{F}_2[\mathbb{Z}/\ell \times \mathbb{Z}/m]$. Two “check polynomials” $c(x,y), d(x,y) \in R$ define the code through a chain complex whose middle cohomology gives the logical space:

$$C(c,d) : \quad R \xleftarrow{(c,-d)} R_h \oplus R_v \xleftarrow{(d,\,c)^T} R.$$

Explicitly, the CSS parity-check matrices are

$$H_X = \begin{pmatrix} A & B \end{pmatrix},\qquad H_Z = \begin{pmatrix} B^T \ A^T \end{pmatrix},$$

where $A, B$ are $nm \times nm$ circulants induced by $c(x,y), d(x,y)$. The CSS condition $H_X H_Z^T = 0$ is equivalent to the commutativity $c d = d c$ in $R$ (Eberhardt et al., 2024).

The parameters are $n = 2\ell m$, $k = \dim_\mathbb{F} H_1(C(c,d))$ with an explicit homological formula, and $d$ is the minimum weight of a nontrivial logical operator. Principal BB codes are those for which the annihilator ideals $\operatorname{ann}(c), \operatorname{ann}(d)$ are each principal, yielding a highly regular logical operator basis.

1.2. Quantum Tanner Codes

QT codes are constructed by a combinatorial group action (lifting) of a small CSS “base code” on an $n_A \times n_B$ grid over the vertex-, edge-, and face-complex of a left–right Cayley square complex associated to a finite group $G$ and generating multisets $A, B \subset G$:

• Each qubit is indexed by $(i,j,g)$ with $1 \leq i \leq n_A$, $1 \leq j \leq n_B$, $g \in G$.
• $H_X$ and $H_Z$ result as direct sums and “twists” of Kronecker products of base-code generators and permutation matrices for left-right regular actions of $A$ and $B$ (Leverrier et al., 23 Dec 2025).

The base code uses parity-check matrices of the form

$$H_X^\text{base} = \begin{pmatrix} H_0 \otimes G_0' \ H_1 \otimes G_1' \end{pmatrix}, \qquad H_Z^\text{base} = \begin{pmatrix} G_0 \otimes H_1' \ G_1 \otimes H_0' \end{pmatrix},$$

and the full, lifted code has generators structured by the permuted fibers of $G$. Given good expander properties and robust intersection in local codes, QT codes achieve constant encoding rate and linear minimum distance as $|G|\to\infty$ (Leverrier et al., 23 Dec 2025).

2. Parity-Check Forms and Homological Properties

2.1. Summary Table

Code	Parity Checks ($H_X,H_Z$)	Structure Basis	Typical Parameters
BB	$[A|B]$, $[B^T|A^T]$ (bivariate circulants)	$R = \mathbb{F}_2[x,y]$	$[[2\ell m,k,d]]$
QT	Kronecker & lifted base code plus group act.	$n_A\times n_B$ base code	$[[n,k,d]]$, $d=\Omega(n)$

• For BB, all commutators vanish due to the group-algebra commutativity.
• For QT, commutation is enforced via orthogonality of base code projections and the structure of the Cayley complex (Freire et al., 15 Jan 2026, Leverrier et al., 23 Dec 2025).

2.2. Logical Operator Structure

BB codes with the purity condition $(c)\cap(d) = (cd)$ admit a decomposition of logical operators into “horizontal” and “vertical” components, producing a “nice” basis analogous to the toric code’s logical loops (Eberhardt et al., 2024). For principal BB codes, all logicals are generated from two seed polynomials (annihilators of $c$ and $d$), giving $k=2\dim_\mathbb{F} R/(c,d)$. Similar logical operator locality and row/column structure are seen in QT base codes, where each logical $X$ ($Z$) can be chosen to lie on a single row (column) of the $n_A\times n_B$ array (Leverrier et al., 23 Dec 2025).

3. Decoding Methods and Erasure Performance

3.1. Quantum Maxwell Decoder

The quantum Maxwell erasure decoder is a symbolic, affine-peeling-based algorithm applicable to any CSS code, including BB and QT (Freire et al., 15 Jan 2026). It proceeds in rounds of:

• Peeling: resolve checks of residual degree 1 as in standard BP.
• Guessing: if peeling stalls and $|\mathrm{Pivots}|<G_\text{max}$, guess an erased variable and introduce a symbolic pivot.
• Pivot Demotion: whenever a check closes with a nonzero symbolic syndrome, solve for the pivot, eliminate and reimburse the guess.

The algorithm output is any assignment respecting remaining pivots. For any fixed $G_\text{max}$, runtime is $O(|E_0|)$ (Freire et al., 15 Jan 2026).

3.2. Performance Guarantees

The Maxwell decoder can match the ML exponent as $\varepsilon\to0$ if $G_\text{max}\geq d-s+1$, where $d$ is the minimum distance and $s$ the stopping distance (Freire et al., 15 Jan 2026). Empirical results show that for BB and QT codes at lengths $n\in\{360,576\}$, $G_\text{max}=6$ brings logical error rate $p_L^\text{QM}(G_\text{max};\varepsilon)$ indistinguishable from ML in the waterfall region; a modest performance gap remains at high erasure but closes steadily with higher guess budget. Decoding remains linear time for fixed $G_\text{max}$, and practical CPU times per block are comparable to cluster decoding (Freire et al., 15 Jan 2026).

4. Finite-Length Performance and Benchmarks

4.1. Comparative Analysis

Head-to-head comparisons for codes around $n=392$ reveal that generalized bicycle codes with optimized parameters (including BB as a special case) can achieve strictly higher minimum distance and better logical error curves under MBP$_4$+ADOSD$_4$ decoders than quantum Tanner codes at the same $(n,k,\rho)$ (Mostad et al., 9 May 2025). For example, a $[[392,32,14]]$ GB code outpaces a $[[392,32,12]]$ QT code; the minimum distance dictates performance in the low-noise regime via $P_L\lesssim C\epsilon^d$.

4.2. Parameter Flexibility and Rate

GB codes allow tuning of check weight $\rho$ to optimize $d$, at the cost of higher-weight stabilizers. QT codes retain fixed LDPC structure ($\rho=O(\Delta)$), but $d$ increases only logarithmically in $n$ unless large degree is used (Mostad et al., 9 May 2025). BB codes with uniform check-weight (e.g., 6 for $[n,k,d]$) remain competitive at moderate error rates but can present higher error floors; increasing $\rho$ via the GB construction eliminates this, maintaining or improving rate.

4.3. Small-Group Instances and Asymptotic Scaling

Explicit small-group QT instances, e.g., $\llbracket144,12,11\rrbracket$, $\llbracket576,28,\leq24\rrbracket$ with weight 9, already outperform toric codes of similar length and rate. Asymptotically, QT codes with robust base codes and expanders exhibit $d/n\geq\Delta>0$ and rate $R>0$ (Leverrier et al., 23 Dec 2025).

5. Gate Implementations and Logical Operator Bases

Principal and pure BB codes support explicit bases of logical operators corresponding to torus cycles, with logical $Z$ and $X$ operators arising from seed polynomials $P$ and $Q$. Fold-transversal Clifford gates—including swap-type, Hadamard-type, and CZ-type—are attainable with zero overhead; group-theoretic analysis reveals these generate substantial subgroups of the symplectic group on $H\oplus H^*$, implementing generalized Clifford actions at the logical level (Eberhardt et al., 2024). Explicitly, this includes symplectic group actions corresponding to the logical blocks, e.g., $\text{Sp}_2(\mathbb{F}_{2^3})$ for $n=98, k=6$.

For small quantum Tanner codes, a plausible implication is that logical operator support can also be localized due to the underlying base code structure, though explicit transversal or fold-transversal gate sets have not been established in the referenced QT literature.

6. Bias-Tailored Variants and Novel Constructions

“Romanesco codes” are bias-deformed, Clifford-transformed bivariate bicycle qLDPC codes leveraging two reflected classical cellular automaton codes on a bipartite hexagonal lattice; their stabilizers are non-CSS (each with half $X$, half $Z$), and locality derives directly from the input rules (Leroux et al., 30 May 2025). Under strong noise bias ($p_X\ll p_Z$), the code effectively decouples: logical error is dictated by the classical code's distance, and the quantum code outperforms both standard BB and 2D-local surface/XZZX variants over relevant physical error rates. The effective minimum distance $d'\rightarrow d_c$ in the limit $\eta\to\infty$, with logical error $p_L\sim p_Z^{\lfloor d_c/2\rfloor}$, potentially violating the Bravyi–Poulin–Terhal $k d^2 = O(n)$ bound for balanced CSS codes (Leroux et al., 30 May 2025).

7. Research Directions and Implementation Considerations

Recent results establish Maxwell-peel decoding as a universal and efficiently tunable erasure decoder for both BB and QT codes, with empirical performance indistinguishable from ML for practical block sizes and moderate guessing budgets (Freire et al., 15 Jan 2026). In the depolarizing channel, GB and QT codes are both highly competitive, with the GB/BB approach offering greater parameter flexibility for finite-length optimization (Mostad et al., 9 May 2025). BB codes additionally support explicit Clifford gate sets, enhancing their applicability to fault-tolerant quantum computing. For bias-tailored error models, Clifford-deformed bicycle codes demonstrate strong suppression of logical error leveraging fractal code input and offer robust alternatives to balanced local CSS codes (Leroux et al., 30 May 2025). Decoding for large-scale quantum Tanner codes and fold-transversal gates for generalized group lifts remain areas of active study.

---

References:

• (Eberhardt et al., 2024) for explicit algebraic construction of BB codes, logical operators, and gates
• (Mostad et al., 9 May 2025) for finite-length performance and benchmarks between BB/GB and QT codes
• (Leverrier et al., 23 Dec 2025) for QT code lifting from Cayley complexes, exact formulas, and instance enumeration
• (Leroux et al., 30 May 2025) for bias-tailored Clifford-deformed bivariate bicycle codes
• (Freire et al., 15 Jan 2026) for Maxwell decoder, runtime analysis, and simulations on BB/QT codes