WorksPhysics and Astronomy
Logical error prediction and geometry selection for rectangular surface codes under biased circuit-level noise
Abstract
Realistic quantum hardware often exhibits biased noise, but at the circuit level this bias may affect only a subset of operations and need not be transferred uniformly to logical errors. The logical response to a given physical bias is therefore central to selecting resource-efficient surface code geometries. Here, we study unrotated rectangular surface codes under standard depolarizing and superconducting-inspired circuit-level noise models. We first construct geometry-dependent logical error rate models from unbiased circuit-level simulations. We then use logical sensitivities to compare the responses to localized idle-bias and global physical error perturbations, defining logical-error-type-dependent response coefficients from an unbiased reference circuit without fitting to finite-bias data. These coefficients are incorporated through effective error parameters, yielding a median fold error of 1.09 over the evaluated finite-bias conditions. Geometries selected by the model are validated through direct biased circuit-level simulations. At a physical error rate of 0.004 and a target logical error rate of 10^{-3}, the validated rectangular codes reduce physical qubit overhead by up to 42.4% and 26.2% under the two noise models, respectively, relative to unrotated square codes. These results provide a practical link between circuit-level bias response and resource-efficient surface code design.
Introduction
The surface code is a leading candidate for fault-tolerant quantum computing because it relies on local interactions in two dimensions and exhibits a high error threshold [1, 2, 3, 4]. In many physical qubit platforms, however, noise is asymmetric, with dephasing-related errors occurring more frequently than bit-flip errors [5, 6]. Rectangular surface codes can exploit such asymmetry by independently adjusting the logical and distances, thereby reducing unnecessary protection along the less error-prone direction [7, 8]. Realizing this resource advantage requires quantitative prediction of the logical and error rates under circuit-level noise, since the appropriate geometry is determined by their relative suppression.
Previous studies have addressed this problem under different noise assumptions. Code-capacity analyses assume perfect syndrome extraction and neglect faults introduced by stabilizer-measurement circuits [9]. Under biased code-capacity noise, Lee et al. showed that optimizing the two lattice dimensions can substantially reduce physical qubit overhead [7]. At the circuit level, Higgott et al. demonstrated that bias-tailored geometries can remain advantageous when syndrome-extraction faults are included [8]. However, the physical bias of the underlying qubits is not necessarily preserved uniformly throughout a realistic syndrome-extraction circuit. In particular, bias-preserving CNOT operations are not generally available for two-level qubits [10, 11, 12], motivating hybrid biased–depolarizing (HBD) descriptions that distinguish physical processes retaining the underlying bias from circuit operations with effectively depolarizing faults [12]. Bias degradation during faulty syndrome extraction can therefore reduce the resource advantages predicted under simpler noise assumptions [13].
These circuit-level effects impose two coupled requirements on bias-aware geometry design. First, the dependence of logical and errors on lattice geometry and physical error rate must be characterized. Second, the logical response to a bias acting only on selected circuit locations must be distinguished from the response produced by an overall variation of the circuit error rate. In the CNOT-based syndrome-extraction circuits considered here, we adopt an HBD-based noise prescription in which the prescribed dephasing bias is retained only at explicit idle locations, while gates, measurements, and resets remain governed by unbiased, operation-dependent fault rates [12, 14, 15]. Under this noise prescription, the nominal physical bias alone is insufficient to determine the resulting logical error asymmetry. Direct biased circuit-level simulations can reliably evaluate individual configurations, but independently simulating and decoding every geometry–bias condition becomes increasingly demanding as the design space expands. This motivates the use of a transferable prediction model for candidate screening, while direct circuit-level simulation is retained for final validation.
The central question of this work is whether the logical response to such a localized circuit-level bias can be inferred from an unbiased reference circuit and transferred to finite-bias conditions without geometry- or bias-specific response fitting. We consider unorated rectangular surface codes under standard depolarizing (SD) and superconducting-inspired (SI) circuit-level noise models. Unbiased baseline models and are first constructed to characterize the dependence of the logical error rates on lattice geometry and physical error rate. The effect of the idle-only bias is then incorporated through separate effective error parameters, and , for logical Z and X errors, respectively.
Recent logical-sensitivity approaches provide mechanism-resolved measures of the response of logical failure probabilities to variations in detector-error probabilities [16, 17]. These sensitivities, however, do not by themselves establish a mapping between a localized perturbation acting on a subset of circuit locations and an equivalent variation of the global physical error parameter used in a baseline logical error model. To establish such a mapping, we compare the logical responses along two physically distinct perturbation directions: a localized idle-bias perturbation and a global variation of the baseline physical error rate. Their relative response defines the logical-error-type-dependent coefficient . The coefficients and are evaluated entirely from an unbiased reference circuit, without fitting to finite-bias logical error rate data, and are subsequently used to construct the effective error parameters. The resulting predictions achieve a median fold error of 1.09, compared with 1.28 when the bias response is neglected and 3.18 when the full idle-channel change is transferred to the effective baseline parameter.
Finally, the bias-dependent predictions are used to identify minimum-overhead rectangular geometries subject to a target logical error rate, and the selected geometries are independently validated by direct biased circuit-level simulations. At and , the validated rectangular codes reduce physical qubit overhead by up to 42.4% under the SD model and 26.2% under the SI model relative to the corresponding unrotated square code baselines over the tested bias ratios . These results demonstrate, within the investigated regime, a practical connection between unbiased circuit characterization, localized bias-response estimation, and resource-efficient geometry selection under biased circuit-level noise.
Results
The overall workflow of the proposed rectangular surface code design framework is summarized in Figure 1. First, baseline logical error rate (LER) models are constructed from unbiased circuit-level simulations under the SD and SI noise models. An effective-error mapping is then constructed using response coefficients obtained from an unbiased reference circuit, without fitting to biased LER data, and applied to the

Figure 1. Overall workflow of the rectangular surface code design framework. Baseline logical error rate models are constructed from unbiased SD and SI circuit-level simulation data and extended to biased noise through an effective error mapping. The resulting bias-dependent model identifies the minimum-overhead lattice geometry among candidates predicted to satisfy the target logical error rate. The selected geometry is then independently evaluated using direct biased circuit-level simulations to verify target satisfaction.
baseline models for bias-aware LER prediction and geometry optimization. Finally, the selected lattice sizes are independently evaluated by direct biased circuit-level simulations to validate their logical performance and resource efficiency.
Baseline logical error rate models
Figure 2 shows the simulated logical error rates and the corresponding fitted curves under the SD and SI circuit-level noise models. To characterize the dependence of the logical error rate on lattice geometry and the physical error rate, we construct an empirical baseline model. Motivated by the approximately exponential suppression of logical errors with increasing code distance in the low-error regime, we model using a term linear in with a -dependent coefficient. Additional terms in and are included to capture residual geometry-dependent behavior. The resulting expression is

Figure 2. Logical error rates under the Standard Depolarizing (SD) and Superconducting-Inspired (SI) circuit-level noise models. Results are shown for fixed values of (a) 13, (b) 15, (c) 17 and (d) 19, with . Markers represent the simulation data, and lines represent the fitted logical error rates obtained from (1). A longer logical operator () leads to improved suppression of logical errors in the low physical error rate regime.
Here, are -independent coefficients determined by fitting the unbiased circuit-level simulation data. This expression serves as an empirical surrogate for logical error rate prediction rather than a theoretical scaling law, and the individual regression coefficients are not assigned independent physical interpretations.
The fitting accuracy of the baseline model is evaluated against the simulated logical error rates. For lattice geometries beyond the fitting domain, the model is used to screen candidate configurations, while the final resource comparisons are established through independent direct circuit-level simulations.
Similarly, the logical error rate is obtained by modeling the scaling behavior along the code’s horizontal direction, as the logical operator connects smooth boundaries. The resulting expression is
Here, and denote -independent regression coefficients for the logical and error rate models, respectively. Both expressions serve as empirical surrogates for logical error rate prediction rather than theoretical scaling laws, and their individual regression coefficients are not assigned independent physical interpretations.
The coefficients were estimated by ordinary least-squares regression using the unbiased circuit-level simulation data. Specifically, was used as the response variable, and the corresponding design matrix was constructed from , , , and according to Eq. (1). Because the model is linear in the coefficients after the logarithmic transformation, the coefficients were estimated by minimizing the residual sum of squares between the simulated and fitted values of .
Similarly, the coefficients were obtained by applying the same least-squares procedure to the simulated logical error rate data using Eq. (2). The same fitting procedure was applied separately to the SD and SI circuit-level noise models. Further details of the logical model and its fitting results are provided in the Supplementary Note on logical error rate estimation. The fitted coefficients for the SD and SI models are listed in Supplementary Table S2.
Across the unbiased fitting domain, the empirical baseline models reproduce the simulated logical error rates with -scale root-mean-square errors below 0.06 for all four combinations of noise model and logical error type. The corresponding median fold errors range from 1.048 to 1.087, indicating typical multiplicative deviations of approximately 5–9% between the fitted and simulated logical error rates. These results show that, within the fitted parameter range, the same empirical modeling framework provides a consistent description of both logical error types under the SD and SI circuit-level noise prescriptions. Detailed fitting statistics are provided in Supplementary Table S3.
Eqs. (1) and (2) therefore provide empirical baseline models that quantitatively describe the geometry-dependent logical error rates within the fitted domain under unbiased circuit-level noise. We further use the same models, without geometry-specific refitting, to extrapolate beyond the fitted lattice range and guide candidate selection. To avoid relying solely on these extrapolations, the final logical error rates and physical qubit comparisons reported in § 2.3 are established by independent direct circuit-level simulations. In the following section, these unbiased baseline models are extended to biased circuit-level noise through response coefficients obtained independently from an unbiased reference circuit.
Prediction of biased logical error rates from an unbiased reference circuit
The baseline models in Eqs. (1) and (2) are parameterized by the physical error rate parameter used in the unbiased SD and SI circuit-level simulations. In the hybrid biased–depolarizing (HBD)-based noise model (§ 4.1), however, the -dependent Pauli asymmetry is introduced only at explicit idle locations. Therefore, directly replacing in the baseline models by one of the Pauli components of the biased idle channel would incorrectly apply the idle-channel asymmetry to all circuit fault locations. Our goal is therefore to represent the effect of the idle-only physical bias as separate equivalent shifts of the baseline error rate parameter for logical and errors, while determining this mapping from an unbiased reference circuit without calibrating its response coefficients to biased LER data.
We therefore introduce an effective-error mapping that converts the prescribed idle-noise bias into separate equivalent baseline error parameters for the logical and error models. For the idle-channel Pauli components in Eq. (8), the relevant Pauli combinations for the logical and errors are and , respectively. At the unbiased point , both reduce to . Their relative changes are therefore
Here, describes the relative change of the Pauli error components relevant to logical error type at the physical idle channel level. To translate this local change into the baseline error rate parameter that characterizes the complete circuit, we define
The central quantities in Eq. (4) are the response coefficients and . Unlike , which is determined by the Pauli composition of the biased idle channel, accounts for how strongly this prescribed idle-bias perturbation is reflected in the logical response of the complete circuit. Crucially, these coefficients are not obtained by fitting Eq. (4) to biased logical error rate data. Instead, they are evaluated from an unbiased reference circuit at , , and . For each logical error type , we define
where indexes detector error mechanisms and denotes the probability of mechanism . The logarithmic sensitivity quantifies the response of the logical failure probability to a relative change in [16, 17]. The quantities and describe the responses of to the prescribed idle-bias perturbation and to , respectively. All derivatives are evaluated at the same unbiased reference circuit, with the decoder held fixed when defining . The numerator therefore represents the local logical response to the idle-bias perturbation, whereas the denominator represents the response to a fractional change in the baseline physical error rate. The precise definitions and numerical estimation procedure are given in Section 4.4 and the Supplementary Note on the estimation of the reference-circuit response coefficients.
The resulting coefficients are for the SD model and for the SI model. The corresponding statistical uncertainties are of order . These coefficients are all substantially smaller than unity, showing that applying the full relative change of the biased idle channel directly to the baseline error parameter would overestimate its influence on the complete circuit. However, is not interpreted as a literal fraction of physical bias transmitted to the logical level. It instead quantifies the local response of the logical error probability to the specified idle-channel perturbation relative to an overall variation of the baseline error parameter.
Once is determined, the coefficients are kept fixed. The biased logical error rates are predicted by evaluating the unbiased baseline models, Eqs. (1) and (2), at the corresponding effective error parameters:
We evaluated this fixed-coefficient model against direct HBD circuit-level simulations at finite bias. To assess whether the response obtained from an unbiased reference circuit remains predictive when transferred to different lattice geometries and finite-bias conditions, we used a matched 132-point evaluation set. The evaluation set comprises 66 conditions per noise model, matched in , and spans , nine lattice geometries, and .
The -based predictions give a median fold error of 1.09. In contrast, neglecting the bias response by setting gives a median fold error of 1.28, whereas assuming that the full idle-channel change is transferred to the effective baseline parameter, , increases the median fold error to 3.18. These results indicate that the response obtained from an unbiased reference circuit can be transferred across the evaluated finite-bias and lattice-geometry conditions without geometry- or bias-specific recalibration, with substantially smaller prediction error than either neglecting the bias response or assuming its full transfer. Furthermore, excluding all conditions associated with the reference geometry leaves 112 validation conditions, for which the fixed-coefficient model achieves an MFE of 1.077 (SD: 1.091; SI: 1.072). These results assess the predictive performance of the fixed-coefficient model on geometries other than the reference configuration, with accuracy varying across geometries as shown in Supplementary Table S6. The prediction errors reflect the combined performance of the empirical baseline models and the response mapping; the present comparison does not isolate their respective error contributions.
We therefore treat as a fixed reference parameter within the evaluated domain, not as a universal circuit constant; its transferability is assessed empirically through the logical error rate predictions over the matched finite-bias and lattice-geometry conditions above. Additional assessments of the fixed-coefficient predictions, including results resolved by bias, logical error type, and lattice geometry, are provided in Supplementary Tables S4–S6.
In the following section we use the fixed mapping to identify candidate rectangular geometries predicted to satisfy the target logical error rate, and then independently verify their target satisfaction through direct biased circuit-level simulations.
Geometry optimization and direct circuit-level validation
To determine the lattice size for a given target logical error rate , physical error rate , and bias , we minimized the physical qubit overhead subject to the logical error rate predicted by the fixed- bias-dependent LER estimator in Eq. (6). The optimization was performed by an exhaustive search over odd lattice dimensions :
Here, denotes the logical error rate estimated using Eq. (6).
Accordingly, denotes the minimum-overhead lattice geometry within the specified search space that is predicted to satisfy the target logical error rate. Because the empirical LER model is used to rank candidate geometries, particularly when the selected sizes extend beyond the baseline fitting range, the model prediction alone is not used as the final criterion for target satisfaction. The resulting candidate geometries were therefore subsequently evaluated using independent direct biased circuit-level simulations, as described in the Supplementary Note on direct simulation and geometry extrapolation validation.
We then compared the physical qubit counts of the model-selected rectangular codes and the square baselines satisfying a fixed target logical error rate, as summarized in Table 1. We first considered at and bias values of and . The rectangular geometries reported in Table 1 are the minimum-qubit solutions among the candidate geometries predicted by Eq. (7) to satisfy the target logical error rate. Their target satisfaction was subsequently confirmed by independent direct circuit-level simulations using the upper bounds of the 95% confidence intervals. The square surface codes were evaluated under the same SD or SI noise model as the corresponding rectangular codes, and each square baseline is the lowest-overhead directly validated square code satisfying the same condition. Additional lower-overhead geometries were directly simulated to assess prediction accuracy, with the results reported in the Supplementary Note on direct simulation and geometry extrapolation validation.
Table 1 Direct-simulation comparison of selected rectangular surface codes and square code baselines. Results are shown for the SD and SI circuit-level noise models at and . The rectangular geometries were selected using the -based logical error rate estimator and subsequently evaluated through direct biased circuit-level simulations. The square codes serve as reference baselines under the corresponding noise conditions. The reported logical error rates are obtained from direct memory experiments with syndrome-extraction rounds. The corresponding 95% confidence intervals are reported in Supplementary Table S7.
(a) SD circuit-level noise model
| # of qubits | # of qubits | ||||||||
| 1 | 0.004 | 1845 | 2025 | ||||||
| 2 | 0.004 | 1813 | 2401 | ||||||
| 3 | 0.004 | 1617 | 2809 | ||||||
| 3 | 0.003 | 525 | 625 | ||||||
| 3 | 0.003 | 1073 | 1369 |
(b) SI circuit-level noise model
| # of qubits | # of qubits | ||||||||
| 1 | 0.004 | 2793 | 2809 | ||||||
| 2 | 0.004 | 2565 | 3249 | ||||||
| 3 | 0.004 | 2745 | 3721 | ||||||
| 3 | 0.003 | 725 | 841 | ||||||
| 3 | 0.003 | 1353 | 1681 |
Table 1.
The implementation of rectangular surface codes leads to a substantial reduction in the total physical qubit overhead under the biased circuit-level conditions considered here (Figure 3). For the SD model, the physical qubit reduction relative to the corresponding square baseline increases from approximately 8.9% at to 24.5% at and 42.4% at . At , the rectangular code uses 1617 physical qubits and yields , whereas the corresponding square code requires 2809 qubits and yields .

Figure 3. Physical qubit reduction achieved by rectangular surface codes. Comparison of the physical qubit overhead of the rectangular geometries and corresponding square baselines listed in Table 1 at and . Rectangular codes require fewer qubits than the corresponding square baselines, with reductions of up to approximately 42.4% under the SD model and 26.2% under the SI model over the bias values considered.
Under the SI model, the reduction is smaller at weak bias but becomes more pronounced as the bias increases. The rectangular codes reduce the physical qubit count by approximately 0.6%, 21%, and 26.2% at , 2, and 3, respectively. At , the rectangular geometry uses 2745 physical qubits and yields . In contrast, the corresponding square baseline requires 3721 physical qubits and yields .
The SI model generally requires larger lattices and higher total qubit counts than the SD model at the same , , and target logical error rate. For example, at and , the selected SI rectangular code requires 2565 physical qubits, whereas the corresponding SD rectangular code requires 1813 qubits. This difference reflects the combined effect of the operation-dependent fault rates in the SI prescription, including its different reset, measurement, gate, and idle error probabilities, rather than a uniform rescaling of the SD model. These results emphasize that operation-dependent circuit-level fault assumptions can materially affect the resource-efficient lattice geometry.
The resource advantage is not limited to the primary operating point of and . At and , the directly validated rectangular geometries continue to require fewer physical qubits than the corresponding square baselines for both and . These additional operating points also show that the selected lattice geometry depends on both the physical error rate and the target logical error rate. The corresponding 95% confidence intervals are reported in Supplementary Table S7.
As a complementary analysis, Table 2 compares rectangular and square surface codes with comparable physical qubit overhead. In this comparison, the square lattices were chosen to use a similar or slightly larger number of physical qubits than the corresponding rectangular lattices. For the conditions examined in Table 2, the rectangular codes show lower directly simulated logical error rates than the corresponding square baselines under both the SD and SI noise models. For example, at and , the rectangular code uses 1353 qubits and yields under the SD model and under the SI model. In comparison, the square code uses 1369 qubits but yields higher logical error rates of and under the SD and SI models, respectively. These results provide a complementary indication that, at comparable qubit cost, rectangular geometries can improve logical performance under the investigated hybrid biased circuit-level noise conditions.
| Rectangular | # qubits | SD | SI | Square | # qubits | SD | SI | ||
| 3 | 0.005 | 825 | 841 | ||||||
| 3 | 0.004 | 609 | 625 | ||||||
| 2 | 0.005 | 1073 | 1089 | ||||||
| 2 | 0.004 | 1353 | 1369 |
Table 2. Logical performance at comparable physical qubit overhead. Directly simulated logical error rates are compared between rectangular and square surface codes under the SD and SI noise models. For each case, the square lattice uses a comparable or slightly larger number of physical qubits than the corresponding rectangular lattice.
Discussion
In this work, we investigated resource-efficient rectangular surface code design under biased circuit-level noise, focusing on whether the logical response to a localized physical bias can be inferred from an unbiased reference circuit. The unbiased baseline models described the geometry-dependent logical and error rates under the SD and SI noise models with median fold errors of approximately 1.05–1.09. We then introduced separate effective error parameters for logical and errors using the response coefficients and , evaluated from an unbiased reference circuit rather than fitted to finite-bias data. Despite being determined solely from the unbiased reference circuit, the fixed-coefficient model achieved a median fold error of 1.09 over the evaluated finite-bias conditions. For the same finite-bias evaluation set, the zero-response assumption, , and the full-response assumption, , yielded median fold errors of 1.28 and 3.18, respectively. These results indicate that neither neglecting the bias response nor assuming its full transfer adequately captures the logical response of the complete circuit.
This comparison also clarifies why the nominal physical bias alone is insufficient to determine logical performance under circuit-level noise. In the HBD-based prescription used here, the prescribed Pauli asymmetry is applied only to explicit idle locations, whereas gate, reset, and measurement faults retain their corresponding unbiased circuit-level channels [12, 18]. The resulting coefficients, for SD and for SI, are all substantially below unity. From Eq. (5), should not be interpreted as a literal fraction of the physical bias transmitted to the logical level. Instead, it quantifies the sensitivity-weighted logical response of the bias-responsive idle mechanisms relative to the response produced by an overall variation of the baseline physical error parameter. This construction uses mechanism-resolved logical sensitivities to establish a directional response mapping that connects an unbiased reference circuit to finite-bias logical error prediction. The systematically smaller coefficients obtained for SI are therefore consistent with a smaller relative contribution of the biased idle subset under its operation-dependent fault prescription.
Applying this fixed mapping to geometry selection resulted in substantial physical qubit savings relative to square codes. At and , the reduction increased with bias and reached 42.4% for SD and 26.2% for SI at . Smaller reductions of 13.8–21.6% were obtained at the additional operating points under the same acceptance criterion. The selected geometries were independently verified by direct circuit-level simulations, with target satisfaction assessed using the upper bounds of the corresponding 95% confidence intervals. Thus, the reported resource reductions do not rely solely on extrapolation of the empirical LER model.
Taken together, these results suggest a practical separation between two components of biased-noise surface code design: the geometry dependence of the logical error rate in an unbiased circuit, described by and , and the response of that circuit to a localized bias perturbation, summarized by and . Within the range studied here, combining these quantities enabled finite-bias LER prediction and geometry selection without separately fitting the bias response for every and lattice geometry. This separation is particularly relevant at the circuit level, where physical bias may be retained only at a subset of fault locations rather than uniformly across the syndrome-extraction circuit [12, 14, 15].
Several limitations define the scope of these conclusions. The response coefficients were obtained from a single reference circuit at , and their transferability was evaluated only over the noise parameters and lattice geometries considered here. They should therefore be regarded as fixed reference parameters within the tested regime rather than universal constants. A new reference characterization may be required when the syndrome-extraction circuit or operation-dependent noise prescription changes, or when the set of circuit locations carrying the physical bias differs materially from that used to determine . In addition, the present simulations employ stochastic Pauli fault channels and do not include leakage, coherent errors, or temporally correlated noise that may arise in realistic quantum devices [19, 20]. The study is also restricted to unrotated rectangular surface codes with MWPM decoding.
Future work should examine the transferability of the reference-circuit response across different code distances, circuit constructions, and distributions of biased locations. Extending the framework to rotated surface codes [21] and more realistic hardware noise models would further test its generality. More broadly, the present results indicate that circuit-level noise characterization can be connected to resource-efficient code design through a compact intermediate response model, while direct simulation remains the final validation step for the selected geometry.
Methods
Circuit-level noise models
We consider two circuit-level noise prescriptions, the standard depolarizing (SD) and superconducting-inspired (SI) models [14, 15]. Both models assign stochastic faults to the elementary operations of the syndrome-extraction circuit, including gates, reset, measurement, and idling, but differ in how the corresponding fault rates are distributed across operation types. In the SD model, all operations are assigned the same reference fault rate . In contrast, the SI model assigns operation-dependent fault rates to account for the different error tendencies associated with the durations of the corresponding circuit operations. The fault rates used in this work are summarized in Table 3.
| Qubit operation | Standard Depolarizing (SD) | Superconducting-Inspired (SI) |
| CNOT | ||
| Hadamard () | ||
| Reset to () | ||
| Measurement | ||
| Idle during gates on other qubits | ||
| Idle during reset or measurement of other qubits |
Table 3. Circuit-level fault rates for the SD and SI noise models.
In quantum systems, Pauli , , and errors do not necessarily occur with equal probabilities, and one type of error may dominate the others [5, 6]. In particular, in dephasing-dominated environments, phase-flip () errors occur more frequently than or errors. Such asymmetry is referred to as biased noise. We describe biased noise using an asymmetric Pauli channel with error probabilities , , and , and define the bias ratio as . A depolarizing channel () corresponds to , whereas larger values of correspond to increasingly dephasing-dominated noise.
For biased circuit-level simulations, we employ an HBD-based noise model [12]. In two-level qubit systems, gate operations such as Hadamard and CNOT do not generally preserve the underlying dephasing bias throughout syndrome extraction. Applying the same biased Pauli channel to all circuit locations can therefore overestimate the amount of bias retained at the circuit level. Following the HBD framework, we retain the dephasing bias only at explicit idle locations.
Idle faults were separately assigned to qubits idling during gate layers and to qubits idling during measurement/reset periods, with the corresponding total fault probabilities determined by the SD or SI rates in Table 3. Specifically, the bias parameter modifies only the relative Pauli composition of an idle fault, while its total fault probability remains determined by the corresponding SD or SI prescription. Denoting the total fault probability of an idle channel by , the Pauli components are assigned as
The unbiased case corresponds to , for which . This unbiased setting is used for both baseline LER model fitting and the estimation of the reference circuit response coefficients . Hadamard and CNOT faults are modeled by symmetric one- and two-qubit depolarizing channels, respectively, while reset and measurement faults retain their SD- or SI-specific rates without additional bias reweighting. Further details of the syndrome-extraction schedule and circuit-level fault assignments are provided in the Supplementary Note on circuit-level implementation of the biased noise models.
Rectangular surface code
We consider an unrotated planar surface code defined on an asymmetric lattice (Figure 4). The minimum weights of the logical and logical operators are and , respectively, allowing the two logical distances to be adjusted independently. A square surface code corresponds to , whereas a rectangular geometry enables different levels of protection against the two types of logical errors [7].

Figure 4. Square and rectangular surface code geometries and stabilizer circuits. Illustration of (a) logical operators and stabilizer measurement circuits in square and (b) rectangular surface codes. On the left, the logical operators and are shown as strings of Pauli operators that connect the boundaries of the same type. The blue (X-type) vertex and orange (Z-type) plaquettes correspond to stabilizers that detect phase-flip and bit-flip errors, respectively. On the right, the corresponding quantum circuits depict how X and Z stabilizers are measured.
Under dephasing-dominated noise, choosing increases the logical distance and provides stronger protection against logical failures while avoiding unnecessary protection along the less affected logical direction. This anisotropic degree of freedom motivates the use of rectangular surface codes under biased noise [7]. For an unrotated rectangular surface code, the total number of physical qubits, including both data and measurement qubits, is . For a square surface code with , this reduces to .
Numerical simulations
In our analysis, the logical error rate is characterized in terms of the lattice sizes ( and ) and the physical error rate . We simulated unrotated rectangular surface code memory experiments under the SD and SI circuit-level noise models. Simulations were performed using Stim [22], a tool for modeling and analysis of stabilizer circuits; PyMatching [23], a fast minimum-weight perfect-matching (MWPM) decoder; and Sinter [22] for multiprocessing-based batch sampling and decoding of Stim circuits.
To construct the circuits, we adapted the built-in generators of Stim such that whenever specific qubits were idle during a given time step, explicit idling faults were applied to those qubits. Each generated circuit was converted into a detector error model, and the resulting syndrome data were decoded using PyMatching on the corresponding detector graph.
To construct the baseline LER models, unbiased simulations () were performed over the physical error rate range , with and . The baseline model coefficients were determined by ordinary least-squares regression of the logarithmic logical error rates using the corresponding model basis functions. Independent fits were performed for the logical and errors under each noise model. Only unbiased simulation data were used for baseline LER fitting, without calibration using finite-bias LER data.
To reduce the statistical uncertainty of the estimated logical error rates, Monte Carlo sampling was performed using Sinter. For each simulation point, at least shots were collected. For all memory experiments, including the unbiased simulations used to construct the baseline LER models and the direct biased simulations, the number of syndrome extraction rounds was set to , which reduces to for square codes. Therefore, all logical error rates reported in this work correspond to the full memory-experiment failure probability over syndrome extraction rounds. The direct biased circuit-level simulations used the same circuit construction, sampling, and decoding procedure, with the HBD-based noise model described in § 4.1.
Estimation of reference-circuit response coefficients
The logical-error-type-dependent response coefficients , with , were evaluated from an unbiased reference circuit for each noise model. Specifically, we employed an unrotated surface code memory circuit at and . The reference circuit was converted into a detector error model (DEM), where each error mechanism has an associated probability .
We first considered the relative changes in the idle-channel Pauli components relevant to logical and errors, defined as , where is given in Eq. (3). This perturbation changes only the Pauli composition of explicit idle faults while keeping their total fault probabilities fixed. The corresponding logarithmic response of each DEM mechanism was defined as
The derivatives were estimated using central finite differences of the DEM mechanism probabilities at .
To quantify the effect of these mechanism-level changes on the logical error rate, we evaluated the logarithmic logical sensitivity . The sensitivities were estimated from Monte Carlo samples of the unbiased DEM using a fixed MWPM decoder constructed from the reference circuit. Let denote the number of logical failures and the number of those failures in which mechanism occurred. The corresponding score-function estimator is
This procedure estimates the sensitivities of individual DEM mechanisms without separately perturbing each mechanism. Following [16, 17], Eq. (10) provides mechanism-resolved logical sensitivities from a common set of Monte Carlo samples. In the present work, these sensitivities are subsequently combined with the responses of the detector-error mechanisms to two physically distinct perturbations, corresponding to a localized variation of the idle-channel bias and a fractional variation of the baseline physical error rate. This construction allows the first-order logical responses along the two perturbation directions to be compared within the same reference circuit.
The second perturbation direction is defined by a fractional variation of the baseline physical error rate, , which changes the operation-dependent fault probabilities according to the SD or SI prescription. The corresponding logarithmic response is defined as
and is estimated using central finite differences with . Using the same logical sensitivities, the first-order logical response along each perturbation direction is obtained by weighting the corresponding mechanism-level responses by . The response coefficient is then defined by the ratio of these two logical responses, as given in Eq. (5). Accordingly, quantifies the relative logical response to the localized idle-bias perturbation with respect to a fractional variation of the baseline physical error rate and maps the former onto an equivalent variation of the baseline error parameter.
For each noise model and logical error type, shots were sampled and divided into 40 independent chunks. The statistical standard error of was estimated using a leave-one-chunk-out jackknife procedure. Once determined, the coefficients were kept fixed for finite-bias LER prediction and geometry optimization, without fitting to biased logical error rate data. Further numerical details of the sensitivity estimation and finite-difference response calculations are provided in the Supplementary Information.
Statistical analysis and validation
Logical error rates were estimated from Monte Carlo simulations of logical and memory experiments. For each logical error type , the logical error rate was estimated as , where denotes the number of observed logical failures and is the total number of shots. The combined logical error rate was approximated as . Statistical uncertainties were quantified using the Clopper–Pearson exact confidence intervals. A Bonferroni correction was applied to obtain a simultaneous confidence level of at least 95%, corresponding to two-sided 97.5% confidence intervals for each logical error type. The confidence limits of the combined logical error rate were obtained by applying the same combination formula to the corresponding lower and upper confidence bounds for logical and errors.
Prediction accuracy was evaluated using the median fold error (MFE), defined as
where and denote the predicted and directly simulated logical error rates, respectively. Baseline fitting accuracy was assessed using unbiased simulation data, whereas finite-bias prediction accuracy was evaluated over 132 conditions, comprising 66 conditions for each noise model. The same evaluation set was used to compare the predictions obtained with , , and . The reference coefficients were determined exclusively from unbiased reference circuits and remained fixed throughout the finite-bias prediction evaluations.
For geometry optimization, candidate lattice sizes were evaluated using the predicted logical error rates and physical qubit overhead. Selected rectangular geometries were subsequently validated through independent biased circuit-level simulations. A geometry was considered to satisfy the target logical error rate only when the upper bound of its combined 95% confidence interval did not exceed . The square code baselines were determined using the same statistical criterion. Resource efficiency was compared between the model-selected rectangular geometries and the square baselines, both of which satisfied this criterion.
AI-assisted manuscript preparation
ChatGPT (OpenAI) was used to assist with English-language editing, manuscript formatting, and preliminary review of statistical descriptions. The authors retain full responsibility for all scientific analyses, interpretations, and conclusions.
Code availability
The custom code used for circuit generation, simulation, and response-coefficient estimation is available from the corresponding author upon reasonable request.
Data Availability
The datasets generated and/or analysed during the current study are available from the corresponding author upon reasonable request.
Acknowledgements
This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP), Information Technology Research Center (ITRC) grant funded by the Korean government (Ministry of Science and ICT) (IITP-2026-RS-2021-II211810).
Author contributions
Y. Chung and J. Heo developed the theoretical idea. Y. Chung performed the main analysis and wrote the manuscript. Y. Kang provided technical support. All authors contributed to the analysis of the results.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary Information accompanies this manuscript as a separate document.
References
- [1]Kitaev, A. Y. Fault-tolerant quantum computation by anyons. Ann. Phys. 303, 2–30 (2003).DOI
- [2]Bravyi, S. B. & Kitaev, A. Y. Quantum codes on a lattice with boundary. arXiv:quant-ph/9811052 (1998).DOI
- [3]Raussendorf, R. & Harrington, J. Fault-tolerant quantum computation with high threshold in two dimensions. Phys. Rev. Lett. 98, 190504 (2007).
- [4]Fowler, A. G., Stephens, A. M. & Groszkowski, P. High-threshold universal quantum computation on the surface code. Phys. Rev. A 80, 052312 (2009).DOI
- [5]de Marti i Olius, A., Martinez, J. E., Fuentes, P., Crespo, P. M. & Garcia-Frias, J. Performance of surface codes in realistic quantum hardware. Phys. Rev. A 106, 062428 (2022).
- [6]Seis, Y., Brown, B. J., Sørensen, A. S. & Goodwin, J. F. Improving trapped-ion-qubit memories via code-mediated error-channel balancing. Phys. Rev. A 107, 052417 (2023).DOI
- [7]Lee, J., Park, J. & Heo, J. Rectangular surface code under biased noise. Quantum Inf. Process. 20, 231 (2021).
- [8]Higgott, O., Bohdanowicz, T. C., Kubica, A., Flammia, S. T. & Campbell, E. T. Improved decoding of circuit noise and fragile boundaries of tailored surface codes. Phys. Rev. X 13, 031007 (2023).DOI
- [9]Watson, F. H. E. & Barrett, S. D. Logical error rate scaling of the toric code. New J. Phys. 16, 093045 (2014).DOI
- [10]Guillaud, J. & Mirrahimi, M. Repetition cat qubits for fault-tolerant quantum computation. Physical Review X 9, 041053 (2019).DOI
- [11]Puri, S. et al. Bias-preserving gates with stabilized cat qubits. Science advances 6, eaay5901 (2020).
- [12]Martinez, J. E. et al. Leveraging biased noise for more efficient quantum error correction at the circuit level with two-level qubits. Phys. Rev. Appl. 25, 014021 (2026).
- [13]Benito, C., Velázquez-Reséndiz, I. J. & Bermudez, A. Optimizing bias-tailored quantum error correction beyond code-capacity noise. arXiv preprint arXiv:2606.17709 (2026).arxiv.org/abs/2606.17709
- [14]Gidney, C., Newman, M., Fowler, A. & Broughton, M. A fault-tolerant honeycomb memory. Quantum 5, 605 (2021).DOI
- [15]Gidney, C., Newman, M. & McEwen, M. Benchmarking the planar honeycomb code. Quantum 6, 813 (2022).DOI
- [16]Fu, W., Staples, J. W. & Thompson, J. D. Efficient Estimation of Logical Sensitivities Through Fault-Counting. arXiv preprint arXiv:2610.10531 (2026).arxiv.org/abs/2610.10531
- [17]Shao, Y. & Liu, Z.-W. Optimizing Quantum Error Correction Through Error Attribution. arXiv preprint arXiv:2610.03399 (2026).arxiv.org/abs/2610.03399
- [18]Meinking, A., Campos, J. & Brown, K. R. Leveraging correlated decoding for bias-tailored compass codes. arXiv preprint arXiv:2605.27598 (2026).arxiv.org/abs/2605.27598
- [19]Harper, R. & Flammia, S. T. Learning correlated noise in a 39-qubit quantum processor. PRX Quantum 4, 040311 (2023).DOI
- [20]McEwen, M. et al. Removing leakage-induced correlated errors in superconducting quantum error correction. Nature Commun. 12, 1761 (2021).
- [21]Xiao, Y., Srivastava, B. & Granath, M. Exact results on finite size corrections for surface codes tailored to biased noise. Quantum 8, 1468 (2024).DOI
- [22]Gidney, C. Stim: A fast stabilizer circuit simulator. Quantum 5, 497 (2021).DOI
- [23]Higgott, O. PyMatching: A Python package for decoding quantum codes with minimum-weight perfect matching. ACM Trans. Quantum Comput. 3, 1–16 (2022).arxiv.org/abs/2105.13082