XZZX Code: Bias-Tailored Quantum Error Correction

Updated 21 January 2026

The XZZX code is a quantum error-correcting scheme that modifies standard surface code stabilizers using an alternating XZZX pattern to optimally exploit biased noise.
It achieves exceptional error thresholds—up to 50% in the infinite-bias regime—by mapping dominant errors to independent repetition codes along diagonal strips.
The design supports versatile implementations across superconducting, spin, and neutral atom platforms with efficient MWPM decoding and scalable resource overhead.

The XZZX code is a high-performance family of quantum error-correcting codes that optimally exploits biased noise by modifying the stabilizer structure of the standard surface code. It is defined on a 2D (or higher-dimensional) lattice and incorporates stabilizer checks composed of alternating Pauli- $X$ and $Z$ operators in the canonical XZZX pattern. The code achieves threshold performance at or above the random-coding (hashing) bound for all single-qubit Pauli channels, and in the infinite-bias regime achieves a threshold of 50%, surpassing all previously known topological codes. Its design admits efficient minimum-weight perfect matching (MWPM) decoding under bias, near-optimal resource scaling, and compatibility with a wide range of quantum hardware platforms.

1. Code Definition, Lattice Structure, and Stabilizer Generators

The XZZX code is most naturally formulated on a square lattice, with qubits residing on either the vertices or edges, depending on the implementation. Each face (plaquette) of the lattice is assigned a weight-4 stabilizer operator with the canonical pattern $XZZX$ ordered cyclically around the four surrounding qubits. Two primary notational conventions exist:

On a face $f$ bounded by qubits $i,j,k,\ell$ in cyclic order:

$S_f = X_i\,Z_j\,Z_k\,X_\ell$

In a checkerboard coloring, one may designate "black" and "white" faces, interchanging $X$ and $Z$ placements:

$S_f^{(b)} = X_{\rm NW} Z_{\rm NE} Z_{\rm SE} X_{\rm SW}, \quad S_f^{(w)} = Z_{\rm NW} X_{\rm NE} X_{\rm SE} Z_{\rm SW}$

For planar, toric, or rotated boundary conditions, the pattern is preserved and weight-2/3 stabilizers appear near the edges. The full stabilizer group is generated by all such $S_f$ over the lattice faces.

Logical operators are noncontractible strings of Pauli- $X$ (for $\bar X$ ) or $Z$ (for $\bar Z$ ) operators traversing the lattice diagonally from one boundary to its opposite. In the square or rotated square geometry with distance $d$ , both $\bar X$ and $\bar Z$ have weight $d$ and the code corrects up to $\lfloor(d-1)/2\rfloor$ errors. For generalized toric or cyclic variants, stabilizer and logical structure generalizes naturally with the same XZZX pattern (Ataides et al., 2020, Xu et al., 2022, Forlivesi et al., 2023).

2. Exploiting Noise Bias: Thresholds, Bias Tailoring, and Effective Distance

The XZZX code is uniquely optimized for biased noise, such as when phase errors ( $Z$ ) far outnumber bit-flip ( $X$ ) or $Y$ errors—a generic situation in superconducting, semiconductor, and bosonic platforms. The code's structure maps strongly biased Pauli- $Z$ noise to independent one-dimensional repetition codes ("strip symmetry"), yielding robust decoupling of error correction along $d$ diagonal strips. This underpins its ultrahigh threshold in the infinite-bias regime (Ataides et al., 2020, Rowshan, 7 Jan 2026):

For depolarizing noise ( $r_X = r_Y = r_Z = 1/3$ ), the code-capacity threshold is $p_c \approx 18.7\%$ .
As $Z$ (or $Y$ or $X$ ) bias diverges, $p_c \to 50\%$ .
Under combined data and measurement errors ("phenomenological noise"), the threshold approaches 10–10.3%.
The finite-size logical error is analytically linked to phase transitions in the random-bond Ising model, with exact solutions available at special disordered points (Xiao et al., 2024).

The effective distance $d_{\mathrm{eff}}$ of the code under bias increases: logical operators predominantly consist of the least biased error (e.g., $Z$ strings under dephasing), allowing for optimized aspect ratios or code layouts that minimize overhead for a desired logical error rate. In strong bias, the code achieves resource scaling near the theoretical minimum (Xu et al., 2022, Forlivesi et al., 2023). This scaling persists under concatenation with bosonic codes or LDPC generalizations (Wu et al., 3 Jul 2025, Zhang et al., 2022).

3. Syndrome Extraction, Decoding Strategies, and Algorithmic Implications

Syndrome extraction in the XZZX code differs from CSS codes in that each check is of mixed $XZZX$ type, requiring ancilla-based circuits that measure these four-body operators. The measurement can be performed using bias-preserving gates (CX and CZ), and for platforms with restricted connectivity (spin qubits, heavy-hex lattices), circuits are explicitly adapted for low-degree or flag-based readout (Hetényi et al., 2023, Kim et al., 2022).

The XZZX syndrome graph admits highly efficient decoding:

Minimum-Weight Perfect Matching (MWPM): Standard under symmetric noise; under bias, the decoder weights edges according to the dominant error probability along strips, achieving ultrafast performance. MWPM naturally exploits the code's one-dimensional strip structure under strong bias, further reducing decoding complexity (Ataides et al., 2020, Sakashita, 22 Sep 2025, Rowshan, 7 Jan 2026).
Simulated Annealing (SA) Decoder: For strongly $Y$ -biased channels, SA with randomized greedy initialization matches the performance of optimal integer-programming (CPLEX) decoders, yielding the fastest known high-bias decoder in parallel implementations (Sakashita, 22 Sep 2025).
Tensor-Network Decoders: Accurate threshold estimates at large sizes and across bias regimes (Xiao et al., 2024).
Decoding at the logical level factorizes into independent repetition codes along strips in the infinite-bias limit, giving analytic tractability and substantial algorithmic speedup.

4. Generalizations, Concatenations, and Measurement-Based Variants

The XZZX code serves as the foundation for higher-dimensional and bias-tailored code families:

Hierarchical and 3D Generalizations: Syndrome-encoded hypergraph product LDPC codes incorporate XZZX as a 2D special case. By "rotating" (Hadamard acting on code blocks), bias tailoring is extended to 3D cluster codes and single-shot LDPC codes, with quadratic scaling in distance and 50% threshold under extreme bias. Variants such as BSH, SSH, and their bias-rotated versions make explicit trade-offs of overhead versus single-shot robustness (Wu et al., 3 Jul 2025).
Concatenation with Bosonic Codes: XZZX underlies concatenated GKP or cat code architectures. E.g., four-legged cat–XZZX concatenation achieves quadratic suppression of residual errors, effective code distance doubling, and full cQED implementability with realistic hardware limitations (Babla et al., 5 Aug 2025, Zhang et al., 2022). Thresholds in these architectures are dominated by bosonic hardware parameters, and efficient decoding leverages real-valued syndrome information from the underlying bosonic layer.
Cluster-State (MBQC) Realizations: Bias-preserving foliations yield "XZZX cluster states" for measurement-based computation, achieving double or higher thresholds under bias than standard Raussendorf-Harrington-Goyal (RHG) clusters. Resource state fusion and hybrid-fusion techniques enable high-threshold, bias-tolerant MBQC even without fully bias-preserving CX gates (Claes et al., 2022, Sahay et al., 2023).

5. Hardware Implementations and Practical Performance

XZZX code performance has been studied in detail across multiple hardware platforms:

Superconducting Circuits: Highly Z-biased noise (dephasing) matches the XZZX orientation, yielding code-capacity thresholds $\sim 18.7\%$ (depolarizing), up to 50% (infinite bias), and circuit-level thresholds $\sim$ 1–2% with bias-preserving gates. In Kerr-cat/superconducting qubits, concatenated architectures achieve $p_{\mathrm{CX}} \sim 6.5\%$ at moderate bosonic cat size (Darmawan et al., 2021).
Spin Qubits: Under strong dephasing bias, XZZX achieves order-of-magnitude higher thresholds for idling error ( $p_T$ up to $\sim 15.1\%$ ) compared to the surface code, and withstands lower $p_G$ (gate errors) due to its mixed stabilizers (Hetényi et al., 2023).
Neutral Atom Arrays: XZZX performs optimally under biased erasure noise, achieving two-qubit gate thresholds above 8% (metastable $^{171}$ Yb) and up to 10.3% in the fully biased erasure model, without requiring bias-preserving entangling gates (Sahay et al., 2023).
Heavy-Hexagon Structures (IBM Quantum): XZZX is embedded into heavy-hex topology using flag-qubit circuits for stabilizer measurement. Under dephasing bias ( $\eta \to \infty$ ), the threshold climbs to $\approx0.33\%$ , $\sim25\%$ higher than the standard surface code on the same lattice (Kim et al., 2022).

Effective resource scaling for the XZZX code is favorable under bias. For target logical error rates, the required physical qubits or GKP qubit overhead is much lower than for unbiased surface codes—XZZX surface-GKP can achieve $P_L\sim2.5\times10^{-7}$ with 291 bosonic modes, compared to over 3000 physical qubits for the standard code at similar fidelity (Zhang et al., 2022).

6. Fault Tolerance, Gate Implementation, and Overhead Optimization

All standard Clifford logical gates, including single-qubit Paulis and two-qubit CNOTs, are available in the XZZX code via lattice surgery and patchwise operations, with twist defects enabling $Y$ -type measurements and Clifford completeness. Measurements are designed to be bias-preserving, suppressing correlated hook errors and enabling thresholds above 10% under strong bias even with noisy syndrome extraction (Ataides et al., 2020, Xu et al., 2022).

Under circuit-level noise, the code retains large effective distance owing to the suppression of the dominant error by the code geometry or lower-level error correction layers (as in cat or GKP concatenation). Flag-based syndrome circuits preserve effective distance with minimal overhead: a single additional flag qubit per stabilizer suffices (Xu et al., 2022). For hardware concatenations (e.g., four-legged cat–XZZX), the architecture suppresses residual error rates quadratically, leading to effective code distance doubling ( $d_{\mathrm{eff}} \approx 2d$ ) and realistic operation below threshold for plausible hardware imperfections (Babla et al., 5 Aug 2025).

MWPM-based decoders and parallel SA or MPS decoders deliver scalable error correction matching theoretical performance, and their computational load is further reduced by the factorization over diagonal strips in the high-bias regime (Ataides et al., 2020, Sakashita, 22 Sep 2025, Rowshan, 7 Jan 2026).

7. Limitations, Variants, and Design Guidelines

While XZZX outperforms standard surface codes in the presence of even moderate bias, certain code layout choices can reduce effectiveness:

Simultaneous application of rotation and XZZX modifications on rectangular lattices can symmetrize distances such that advantages from bias are lost; e.g., a $d_x \times d_z$ rectangular code with $d_z > d_x$ collapses to $d_x$ -distance for both logicals under XZZX, negating the native asymmetric protection (Forlivesi et al., 2023).
In the absence of noise bias, XZZX performs comparably to the standard surface code, imposing no penalty.
For highly asymmetric layouts, the standard code on a rectangle is preferable.

The code family is highly versatile, supporting cyclic, toric, planar, LDPC, 3D, and single-shot generalizations, with analytic performance characterization across all regimes. Threshold calculations, logical error exponents, and optimal resource scaling have all been rigorously derived, providing concrete design rules for code selection in practical quantum processors (Ataides et al., 2020, Xu et al., 2022, Forlivesi et al., 2023).

References:

Bonilla Ataides et al., "The XZZX Surface Code" (Ataides et al., 2020)
Bombin, Haah, et al., "Bias-tailored single-shot quantum LDPC codes" (Wu et al., 3 Jul 2025)
Puri et al., "Fault-tolerant Fusion-based Quantum Computing with the Four-legged Cat Code" (Babla et al., 5 Aug 2025)
Hirata et al., "Fast and Accurate Decoder for the XZZX code Using Simulated Annealing" (Sakashita, 22 Sep 2025)
Hamma et al., "Exact results on finite size corrections for surface codes tailored to biased noise" (Xiao et al., 2024)
Hetényi & Wootton, "Tailoring quantum error correction to spin qubits" (Hetényi et al., 2023)
Darmawan et al., "Practical quantum error correction with the XZZX code and Kerr-cat qubits" (Darmawan et al., 2021)
Guillaud & Mirrahimi, "Concatenation of the Gottesman-Kitaev-Preskill code with the XZZX surface code" (Zhang et al., 2022)
Xu et al., "Tailored XZZX codes for biased noise" (Xu et al., 2022)
Forlivesi et al., "Logical Error Rates of XZZX and Rotated Quantum Surface Codes" (Forlivesi et al., 2023)
Higgott et al., "Design of Quantum error correcting code for biased error on heavy-hexagon structure" (Kim et al., 2022)
Sahay et al., "Strip-Symmetric Quantum Codes for Biased Noise: Z-Decoupling in Stabilizer and Floquet Codes" (Rowshan, 7 Jan 2026)
Nickerson et al., "Tailored cluster states with high threshold under biased noise" (Claes et al., 2022)
Schindler et al., "High threshold codes for neutral atom qubits with biased erasure errors" (Sahay et al., 2023)