Bicycle Bivariate Codes Overview

• Bicycle Bivariate codes are CSS quantum LDPC codes constructed from two commuting cyclic structures that yield sparse parity checks and nontrivial encoding rates.
• They utilize a two-variable polynomial quotient ring framework that supports univariate reductions, symmetry analysis, and diverse decoding strategies.
• Key developments include open-boundary, subsystem variants and tailored decoders that address noisy-syndrome and circuit-level error models.

Bivariate bicycle (BB) codes are a family of CSS quantum low-density parity-check codes built from two commuting cyclic structures, typically presented through parity-check matrices

$H_X=[A\mid B],\qquad H_Z=[B^\top\mid A^\top],$

with $A$ and $B$ obtained from a two-variable group algebra or polynomial quotient ring on a torus (Postema et al., 24 Feb 2025). In the standard bivariate description, one works over

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$

with $x=S_\ell\otimes \mathbb{I}_m$ and $y=\mathbb{I}_\ell\otimes S_m$, so the blocklength is $n=2\ell m$ and the stabilizer weight is bounded when $A(x,y)$ and $B(x,y)$ are sparse (Postema et al., 24 Feb 2025). BB codes have become a central QLDPC family because they combine sparse checks, nontrivial encoding rate, and multiple algebraic representations that support code construction, symmetry analysis, noisy-syndrome decoding, open-boundary variants, and several distinct decoder architectures (Rowshan, 3 Jan 2026).

1. Algebraic construction and foundational formalism

A standard BB code is defined from commuting shifts in two directions. Let $S_\ell$ be the $A$0 cyclic shift matrix and set

$A$1

Using the trinomial ansatz,

$A$2

the CSS parity checks are

$A$3

and CSS commutation follows automatically from the generalized bicycle form (Postema et al., 24 Feb 2025). Because $A$4 and $A$5 are trinomials, the stabilizer weight is at most $A$6, so this subclass is LDPC (Postema et al., 24 Feb 2025).

The same structure can be expressed in group-algebra language. For

$A$7

the code is specified by two sparse algebra elements, and the qubits naturally split into two blocks or, geometrically, into horizontal and vertical edges of a periodic $A$8 lattice (Eberhardt et al., 2024). This two-block viewpoint is the basis for later interpretations of BB codes as group-algebra codes, balanced-product-like objects, and translationally invariant 2D CSS codes (Eberhardt et al., 2024).

A recurring operational distinction is between the abstract algebra and the finite torus geometry. In the polynomial formalism used for logical-operator and matching-decoder work, periodic boundaries may be twisted through relations such as

$A$9

so the same local stabilizer pattern can realize distinct finite codes through different torus parameters (Sahay et al., 26 Feb 2026). This suggests that BB codes are best understood not as a single construction, but as a construction framework in which sparse polynomial data and boundary conditions are equally important.

2. Coprime BB codes, univariate reduction, and structural limits

A particularly important subclass is the coprime odd case, where $B$0 and $B$1 are odd and coprime. Then the bivariate ring becomes equivalent to a univariate cyclic ring: $B$2 with $B$3 and explicit images for $B$4 and $B$5 determined by modular inverses (Postema et al., 24 Feb 2025). This reduction is the key algebraic simplification behind the coprime theory.

In this coprime setting, the central invariant is the gcd of the two defining polynomials with the torus relation. One paper denotes it

$B$6

and proves the dimension formula

$B$7

for blocklength $B$8 (Postema et al., 24 Feb 2025). A later noisy-syndrome analysis uses the notation

$B$9

and obtains the equivalent rank statement

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$0

for coprime BB codes in the univariate quotient ring $$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$1 (Rowshan, 3 Jan 2026). Notation differs across these works, but both treat the gcd polynomial as the master invariant of the coprime subclass.

The coprime theory also gives explicit existence constraints. The nontrivial dimension condition reduces to asking which irreducible factors of $$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$2 divide trinomials. For prime cyclotomic factors, the relevant criterion states that such minimal polynomials divide trinomials only when the prime is a Mersenne prime or one of the listed outlier primes

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$3

and the composite case is reduced to the prime case by cyclotomic identities (Postema et al., 24 Feb 2025). Connectedness is characterized by

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$4

for trinomials $$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$5, $$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$6 (Postema et al., 24 Feb 2025).

The same work proves a negative asymptotic result. Any sequence of constant-weight coprime BB codes of trinomial form is asymptotically bad; specifically,

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$7

equivalently

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$8

Thus the coprime trinomial subclass is excluded as a route to asymptotically good QLDPC codes, even though moderately long instances remain practically attractive (Postema et al., 24 Feb 2025). A plausible implication is that BB codes are primarily a finite-length rather than asymptotic design family.

3. Noisy measurements, syndrome codes, and single-shot decoding

The algebra of BB codes becomes more intricate when stabilizer measurements are noisy. For a CSS code with

$$\mathcal{S}=\frac{\mathbb{F}_2[x,y]}{\langle x^\ell-1,\; y^m-1\rangle},$$9

the syndrome layer is controlled by the left nullspaces of $x=S_\ell\otimes \mathbb{I}_m$0 and $x=S_\ell\otimes \mathbb{I}_m$1. If $x=S_\ell\otimes \mathbb{I}_m$2, then the noiseless $x=S_\ell\otimes \mathbb{I}_m$3-syndromes form the classical linear code

$x=S_\ell\otimes \mathbb{I}_m$4

and similarly for $x=S_\ell\otimes \mathbb{I}_m$5 (Rowshan, 3 Jan 2026). In the coprime BB family, these syndrome codes are completely determined by the gcd polynomial: $x=S_\ell\otimes \mathbb{I}_m$6 so both are cyclic length-$x=S_\ell\otimes \mathbb{I}_m$7 codes generated by $x=S_\ell\otimes \mathbb{I}_m$8 (Rowshan, 3 Jan 2026).

This leads to an exact identity between logical rate and stabilizer redundancy density. Since

$x=S_\ell\otimes \mathbb{I}_m$9

one has

$y=\mathbb{I}_\ell\otimes S_m$0

The same polynomial that fixes the logical dimension also fixes the number of redundant same-type stabilizer relations (Rowshan, 3 Jan 2026). The syndrome distance $y=\mathbb{I}_\ell\otimes S_m$1 then quantifies one-round correction of measurement-bit flips, with bounded-distance radius

$y=\mathbb{I}_\ell\otimes S_m$2

Because the syndrome codes are ordinary cyclic codes generated by $y=\mathbb{I}_\ell\otimes S_m$3, one can import BCH-style lower bounds: if $y=\mathbb{I}_\ell\otimes S_m$4 has $y=\mathbb{I}_\ell\otimes S_m$5 consecutive roots, then $y=\mathbb{I}_\ell\otimes S_m$6 (Rowshan, 3 Jan 2026). At the same time, Singleton gives the upper bound

$y=\mathbb{I}_\ell\otimes S_m$7

This is the structural bottleneck identified for single-shot decoding in the coprime ansatz: high quantum rate and syndrome robustness are algebraically locked together (Rowshan, 3 Jan 2026).

The paper gives explicit small examples. For

$y=\mathbb{I}_\ell\otimes S_m$8

the code has parameters $y=\mathbb{I}_\ell\otimes S_m$9 and syndrome distance $n=2\ell m$0. For

$n=2\ell m$1

the code has parameters $n=2\ell m$2 and syndrome distance $n=2\ell m$3 (Rowshan, 3 Jan 2026). Under a phenomenological model with data error probability $n=2\ell m$4 and syndrome-bit flip probability $n=2\ell m$5, BP on the syndrome code followed by BP+OSD-2 on the CSS code shows that the $n=2\ell m$6 example can track a repeated $n=2\ell m$7 baseline closely in a useful low-noise regime, though the paper emphasizes that the result is a finite-radius guarantee rather than a full threshold theorem (Rowshan, 3 Jan 2026).

4. Boundaries, open-boundary descendants, and subsystem realizations

Standard BB codes are naturally defined with periodic boundary conditions because $n=2\ell m$8 and $n=2\ell m$9 are cyclic shifts on a torus (Eberhardt et al., 2024). Several later constructions address how to obtain open-boundary or hardware-friendlier descendants without abandoning the underlying BB algebra.

One approach is pruning. For the univariate hypergraph-product subclass

$A(x,y)$0

if

$A(x,y)$1

then the periodic code

$A(x,y)$2

admits a pruned open-boundary version

$A(x,y)$3

that is as local as the original code on a lattice with open boundary conditions (Eberhardt et al., 2024). In this theorem, distance is preserved exactly, the number of logical qubits is halved, and the physical-qubit count stays of the same order (Eberhardt et al., 2024).

A second line of work studies bivariate-bicycle-surface (BBS) codes, i.e. planar/open-boundary descendants defined from Laurent polynomials $A(x,y)$4. Their dimension is controlled by the quotient

$A(x,y)$5

or equivalently by the number of common roots $A(x,y)$6 with finite nonzero coordinates, counted with algebraic multiplicity (Wang et al., 7 Jun 2026). Roots at $A(x,y)$7 or $A(x,y)$8 signal that specialized generators are needed near the corresponding boundaries, and monomial automorphisms reveal tilted-boundary structure. The paper formulates a prescription for rectangular, diagonal, and arbitrarily tilted boundaries and states that no corner corrections are needed when the orientation-specific edge conditions are satisfied (Wang et al., 7 Jun 2026).

A third development is the subsystem realization of BB logical structure. Subsystem bivariate bicycle (SBB) codes are translation-invariant CSS subsystem codes with local weight-$A(x,y)$9 gauge measurements whose stabilizer syndromes are inferred by multiplying gauge outcomes (Liang et al., 5 May 2026). When a determinantal-ideal criterion excludes nonlocal stabilizers, a finite-depth Clifford circuit decouples one gauge qubit per unit cell and identifies the protected subsystem with a corresponding BB stabilizer code (Liang et al., 5 May 2026). The paper proves that an SBB code is topological if and only if the corresponding BB code is topological, and reports examples including

$B(x,y)$0

(Liang et al., 5 May 2026).

5. Logical operators, self-duality, and transversal Clifford structure

The algebraic study of logical operators reveals a toric-code-like sector structure in many BB codes. In the two-block group-algebra formalism, the logical $B(x,y)$1-operators form the homology group

$B(x,y)$2

and a code is called pure if $B(x,y)$3 is generated by horizontally and vertically pure classes (Eberhardt et al., 2024). For odd $B(x,y)$4, the ring $B(x,y)$5 is semisimple, and every BB code is pure and principal (Eberhardt et al., 2024). In semiperiodic cases, annihilator ideals are generated explicitly; if $B(x,y)$6, then

$B(x,y)$7

generates $B(x,y)$8 (Eberhardt et al., 2024). This yields explicit “nice bases” of logical operators analogous to horizontal and vertical loop operators in the toric code.

These bases support fold-transversal Clifford gates when additional symmetries are present. For symmetric BB codes with

$B(x,y)$9

the paper constructs swap-type, Hadamard-type, and phase/CZ-type fold-transversal gates, and gives explicit logical bases and logical gate groups for the $S_\ell$0 and $S_\ell$1 BB codes (Eberhardt et al., 2024). The significance is that BB codes can admit nontrivial no-overhead Clifford operations through lattice symmetries rather than through code switching or added ancillas.

A related but distinct development is the self-dual BB subclass on the honeycomb lattice. Here the self-duality condition is

$S_\ell$2

with $S_\ell$3 (Liang et al., 6 Oct 2025). The paper proves that the weight-$S_\ell$4 self-dual case reduces to the color code or disjoint copies of it, and therefore focuses on weight-$S_\ell$5 self-dual BB codes of the form

$S_\ell$6

These codes support transversal CNOT, Hadamard, and $S_\ell$7 gates, and the search up to $S_\ell$8 yields examples including

$S_\ell$9

(Liang et al., 6 Oct 2025). This suggests that self-duality restores color-code-like transversal Clifford structure while retaining the rate advantages associated with the BB framework.

6. Decoding, circuit-level behavior, and later extensions

A large part of the BB literature concerns decoding under increasingly realistic noise models. Under circuit-level noise, one notable result is an analytical theory of greedy peeling for BB codes. For the $A$00 Gross code, the paper derives a collision-resolution factor

$A$01

shows that shared-2 pairs always resolve under peeling, and validates

$A$02

across five BB codes with $A$03 (Pakhunov, 13 Apr 2026). At $A$04, deferred greedy achieves identical logical error rate to BP+OSD for $A$05 while giving a reported $A$06 latency reduction, and the $A$07 code with $A$08 supports two-shot streaming decoding with $A$09 peeling success and a $A$10 LER ratio versus $A$11 (Pakhunov, 13 Apr 2026).

Several decoder families have been developed specifically for BB codes. A recurrent transformer-based decoder for circuit-level noise achieves a logical error rate almost $A$12 times lower than BP-OSD on $A$13 at $A$14, and remains an order of magnitude faster than the worst-case times of a benchmark BP-OSD implementation, though on $A$15 it has worse logical error rates while retaining the speed advantage (Blue et al., 17 Apr 2025). A different line extends multiple-bases BP to the Multiple-Bases Belief-Propagation List Decoder (MBBP-LD); on $A$16 it achieves up to $A$17 lower logical error rate than BP-OSD while preserving the linear-time complexity profile of plain BP under parallel scheduling (Rabeti et al., 4 Nov 2025). Matching-based decoding has also entered the field: a recent MWPM decoder uses the equivalence of BB codes to copies of the toric code and introduces the “cylinder trick,” with additional gains from BP and “over-matching” on gross codes, cyclic hypergraph-product codes, generalized toric codes, and directional codes (Sahay et al., 26 Feb 2026).

The BB framework has also expanded in several orthogonal directions. Covering-graph methods generate infinite BB-code sequences from a base Tanner graph; for odd cover degree $A$18, the paper proves

$A$19

and if $A$20, then

$A$21

(Symons et al., 17 Nov 2025). The Gross code $A$22 is reinterpreted there as a 2-cover of $A$23, and the same framework yields new weight-$A$24 BB codes such as $A$25 and $A$26 (Symons et al., 17 Nov 2025). Beyond the bivariate setting, multivariate generalizations recover BB codes as the $A$27 member of a broader family (Voss et al., 2024), while independent trivariate bicycle codes preserve the same bicycle/CSS block form and produce examples including $A$28 with code-capacity pseudothreshold $A$29 and circuit-level pseudothreshold $A$30 (Galimova, 18 Mar 2026). Other hardware-oriented extensions include Stairway codes, which Floquetify Abelian two-block group algebra codes including the BB family (Jacoby et al., 27 Feb 2026), and BiBiEQ, which compiles BB memory circuits into erasure-aware circuits and shows that under the 4EC schedule the approximate and exact conversion engines remain close while most subthreshold gains are already realized by the $A$31 BB instance (Bhave et al., 7 Feb 2026).

Taken together, these results present BB codes as a broad algebraic platform rather than a single static code family. The same sparse two-block structure supports univariate reductions, explicit logical-operator theory, noisy-syndrome cyclic codes, open-boundary and subsystem descendants, fold-transversal or transversal Clifford gates in special subclasses, and a wide range of decoders under phenomenological, circuit-level, Floquet, and erasure-aware noise models (Postema et al., 24 Feb 2025). A plausible synthesis is that BB codes are most significant as a finite-length, algebraically programmable QLDPC design space whose strengths and weaknesses are controlled as much by symmetries, gcd structure, and boundary conditions as by raw check sparsity alone.