Circuit-Level Local Stochastic Error Model

• The circuit-level local stochastic error model attaches noise to specific circuit operations, defining errors through classically sampled fault paths with support bounds.
• It encompasses variants from iid stochastic Pauli insertions to adversarial and locally decaying models, each tailored for distinct simulation and threshold analyses.
• This framework informs the design of fault-tolerant circuits and decoders by bridging microscopic hardware assumptions with logical-level error behavior.

Searching arXiv for the cited papers on circuit-level local stochastic error models and closely related formulations. Searching arXiv for “circuit-level local stochastic noise quantum error correction”. A circuit-level local stochastic error model is a fault model in which noise is attached to elementary operations, wait locations, preparations, measurements, or circuit layers, rather than only to data qubits or syndrome bits, and in which locality is imposed at circuit locations while stochasticity means that fault paths are classically sampled rather than coherently superposed. In the literature, the phrase does not denote a single formal object. It can refer to a restricted iid stochastic Pauli insertion model embedded in a fault-tolerant circuit (Barnes et al., 2017), to an adversarial support-bounded model with $\Pr[F\subseteq \supp(E)]\le p^{|F|}$ (Choe et al., 2024), to an elementary-location Pauli fault model whose decoder-relevant statistics are organized by a circuit error-equivalence group (Pryadko, 2019), or to a locally decaying variant used in threshold-style analyses of Knill error correction (Murphy et al., 5 Mar 2026). Taken together, this indicates that the term is best understood as a family of circuit-attached stochastic noise models rather than a uniquely standardized definition.

1. Formal definitions and scope

In its most theorem-oriented form, local stochastic noise is defined through support bounds. For a random subset $I\subseteq [n]$, $I$ is a local stochastic set of parameter $p$ if

$$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$

If $E\in \mathcal P_n$ is a random $n$-qubit Pauli with support $\supp(E)\subseteq [n]$, then $E$ is local stochastic noise of strength $p$ if

$I\subseteq [n]$0

written $I\subseteq [n]$1. In the circuit-level version, an adaptive circuit is decomposed into layers, and Pauli faults are inserted between layers so that each marginal layer error satisfies $I\subseteq [n]$2. This model explicitly accounts for imperfect single-qubit preparation, single- and two-qubit Clifford gates, identity or wait locations, and measurement faults (Choe et al., 2024).

A closely related formulation is the “locally decaying” model. For a finite set $I\subseteq [n]$3, a probability distribution over subsets of $I\subseteq [n]$4 is locally decaying with rate $I\subseteq [n]$5 if, for every $I\subseteq [n]$6,

$I\subseteq [n]$7

The same support-decay condition is then imposed either on Pauli-error supports or on subsets of circuit fault locations. This is structurally a local-stochastic-style bound, but expressed through subset-containment probabilities rather than an iid assumption (Murphy et al., 5 Mar 2026).

By contrast, some circuit-level studies use “local stochastic” more operationally. In that usage, errors are “randomly inserted in a quantum circuit with classical probabilities,” and the defining physical distinction is that “there [are] no interfering pathways between them.” This broader operational notion includes sampled Pauli fault patterns, but does not automatically imply the adversarial support-bounded formalism used in threshold theory (Barnes et al., 2017).

2. Restricted iid stochastic Pauli insertion models

A particularly clear example appears in the comparison between stochastic and coherent circuit noise for fault-tolerant Steane and surface-code circuits. That work recalls the density-matrix stochastic Pauli channel

$I\subseteq [n]$8

and refers to it as the “stochastic Pauli error model.” In the actual simulations, however, the model is narrower: at selected data-qubit locations before syndrome measurements, it “randomly applies an $I\subseteq [n]$9 and $I$0 error each with probability $I$1. Since $I$2, all Pauli errors can occur.” For one such location, the realized event is no fault with probability $I$3, $I$4 only with probability $I$5, $I$6 only with the same probability, and both $I$7 and $I$8 with probability $I$9 (Barnes et al., 2017).

This is local because faults occur independently by circuit location and act on single qubits, and circuit-level because the insertions are tied to locations in the syndrome-extraction circuit rather than to a phenomenological syndrome-bit-flip model. At the same time, it is explicitly not a full depolarizing circuit model over all operations. The choice is described as “conservative”: errors are restricted to codeword qubits and only before each syndrome measurement; measurements are perfect and instantaneous; the Steane cat-state ancilla is “always error free”; and no stochastic faults are placed on ancilla preparation, ancilla verification, idle locations, measurements, or state preparation (Barnes et al., 2017).

The simulation itself is implemented by sampled unitary insertions. “Insertions of errors in the stochastic Pauli (SP) error model are treated as unitary qubit gates,” so each Monte Carlo trajectory samples a classical Pauli fault pattern and then evolves under a pure-state circuit. The nonunitary element resides only in the classical sampling over runs, plus projective measurement updates. This distinguishes the model from both explicit Kraus simulation and from coherent gate-imperfection models in which alternatives remain in superposition (Barnes et al., 2017).

3. Elementary circuit locations, induced correlations, and exact decoding

A different formalization begins from elementary circuit locations. For a Clifford syndrome-measurement circuit, the relevant noise variables are Pauli faults on the elementary wire segments between gates or between gates and endpoints. A circuit fault is a Pauli assignment

$p$0

represented as a binary vector $p$1. The physical model need not be iid: the probability distribution can be written in Ising/Wegner form,

$p$2

so independent-per-location faults are only a special case (Pryadko, 2019).

Because the circuit is Clifford, any circuit fault can be propagated to an output data Pauli and a pattern of ancilla-measurement flips. Distinct fault histories can nevertheless be operationally equivalent. The trivial differences generated by local propagation identities, ancilla-preparation relations, and measurement relations form the circuit error-equivalence group (EEG), equivalently the gauge group of an associated circuit subsystem code. Two fault patterns are equivalent when

$p$3

where $p$4 is the binary generator matrix of the EEG. The exact probability of an equivalence class is then a partition-function sum over all physically distinct circuit faults in that class, and maximum-likelihood decoding compares logical sectors

$p$5

by maximizing the corresponding partition functions (Pryadko, 2019).

This viewpoint makes precise a point that is often obscured in simplified phenomenological models: locality at the circuit-location level does not imply independence at the effective data-and-syndrome level. Even when faults at elementary locations are independent, fault propagation through a Clifford measurement circuit and summation over EEG-equivalent histories induce structured correlations among effective data errors and syndrome errors. A common misconception is therefore that a circuit-level local stochastic model reduces naturally to independent final data faults plus independent measurement flips; the exact EEG formulation shows that, in general, it does not (Pryadko, 2019).

4. Common instantiated circuit-level models in simulation studies

Many numerical studies adopt explicit per-location stochastic Pauli or depolarizing models rather than adversarial support bounds. For the rotated planar surface-code Hadamard gate, a “popular superconducting-inspired circuit-level Pauli noise model” is used: after each entangling gate, a two-qubit depolarizing channel with strength $p$6; measurement outcomes flipped with probability $p$7; after each Hadamard, reset, or measurement gate, a one-qubit depolarizing channel with strength $p$8; and on each idle qubit in each layer, a one-qubit depolarizing channel with strength $p$9. The model is local, independent, Markovian, and circuit-level, and the central metric is effective distance, defined as the minimum number of fault locations required to create an error (Gehér et al., 2023).

For repeated surface-code syndrome extraction in decoder-engineering studies, a standard circuit-level depolarizing model is often used in which each single-qubit gate is followed by $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$0, $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$1, or $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$2 with probability $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$3; each two-qubit gate fails with probability $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$4 and is followed by a uniformly random nontrivial two-qubit Pauli; $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$5 preparation is replaced by $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$6 with probability $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$7, $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$8 by $$\Pr[F \subseteq I] \leq p^{|F|}\qquad \text{for all}\qquad F\subseteq [n].$$9 with the same probability; measurement outcomes are flipped with probability $E\in \mathcal P_n$0; and idles fail with probability $E\in \mathcal P_n$1 followed by a uniformly random nontrivial single-qubit Pauli. The resulting fault patterns are local at the physical level but generate space-time correlated detection events, including the “vertical pairs” emphasized in local-decoder postprocessing (Chamberland et al., 2022).

Surface-code decoding with an Ising machine uses a related six-step CNOT-based circuit model: idling data qubits receive $E\in \mathcal P_n$2, $E\in \mathcal P_n$3, or $E\in \mathcal P_n$4 with total probability $E\in \mathcal P_n$5; each ideal CNOT is followed by a uniformly random non-identity two-qubit Pauli with total probability $E\in \mathcal P_n$6; ancilla reset or preparation is flipped with probability $E\in \mathcal P_n$7; and measurement outcomes are flipped with probability $E\in \mathcal P_n$8. The corresponding decoder solves a 3D parity-constrained Ising/QUBO optimization rather than a phenomenological 2D problem (Fujisaki et al., 2023).

Measurement-free local error-correction circuits provide another specialized example. There, data qubits receive ambient stochastic errors at rate $E\in \mathcal P_n$9 each round, and independent stochastic Pauli $n$0 faults are inserted after every parallelized gate layer at rate $n$1 on data and ancilla qubits, with perfect ancilla reset and no measurement noise by design. This is circuit-level and local in the simulation sense, but narrower than a full theorem-oriented local-stochastic model (Park et al., 2024). A hardware-instantiated extension appears in trapped-ion surface-code memory, where initialization, single-qubit gates, two-qubit gates, and measurements are modeled stochastically, while idling and transport carry coherent $n$2-axis rotations whose angle is proportional to compiler-resolved elapsed time (LeBlond et al., 19 Aug 2025).

5. Relation to coherent and non-Pauli noise

The principal limitation of circuit-level local stochastic Pauli models is that they do not generically reproduce the logical statistics of coherent circuit errors. In the Steane and tilted-17 surface-code comparison, the stochastic Pauli model yields narrow or effectively binomial logical-failure distributions, whereas coherent pulse-area noise produces “very broad” and “heavy-tailed” distributions. The paper’s central conclusion is explicit: “There is no simple map between the Pauli error model and the pulse area error model.” It further rejects the approximation that syndrome measurements cut off coherence between fault paths, emphasizing that one “cannot rely on the assumption that syndrome measurements cut off the interfering pathways” (Barnes et al., 2017).

A more perturbative generalization begins from a local Markovian CPTP circuit model and maps it onto a detector error model (DEM). In that framework, ordinary stochastic Pauli circuit noise is recovered as the $n$3-only special case, while coherent and non-Pauli components are represented in an error-generator basis and compressed into detector events. At leading order, a CPTP model yields nonnegative DEM event rates, but beyond leading order correlated detector events and even negative DEM hyperedge weights can appear. In surface-code threshold studies within this framework, two physical models with the same generator infidelity can have markedly different logical behavior: approximately $n$4 CNOT infidelity for a purely stochastic model versus approximately $n$5 when $n$6 of CNOT generator infidelity comes from coherent errors (Hines et al., 19 Mar 2026).

The trapped-ion study shows a complementary regime distinction. When coherent idling and transport dephasing rates are near current trapped-ion values, logical error rates in the mixed coherent-plus-stochastic model align closely with those of the fully stochastic Pauli-twirled approximation up to $n$7. At larger dephasing rates, however, the mixed model develops nonzero off-diagonal logical PTM terms, coherent logical rotations about all three logical axes, larger diagonal logical-error components, and a reduced threshold-like crossover (LeBlond et al., 19 Aug 2025). This suggests that the adequacy of a circuit-level local stochastic approximation is regime dependent rather than absolute.

6. Fault-tolerance, compilation, and decoder architecture

One major use of the formal adversarial local-stochastic model is as a transportable abstraction across architectures. A recent locality-compilation theorem shows that an arbitrary adaptive circuit $n$8 can be transformed into geometrically local 3D or quasi-2D circuits such that any noisy implementation under arbitrary local stochastic noise of strength $n$9 is equivalent to a noisy implementation of the original circuit with local stochastic noise of strength $\supp(E)\subseteq [n]$0. The resulting localized circuits have quantum depth $\supp(E)\subseteq [n]$1, with qubit overhead $\supp(E)\subseteq [n]$2 in 3D and $\supp(E)\subseteq [n]$3 in quasi-2D (Choe et al., 2024).

Knill error correction supplies a different structural use of circuit-level support-bounded noise. Under locally decaying circuit noise, the noisy transversal Bell measurement induces an effective locally decaying Pauli error on the measured blocks with rate

$\supp(E)\subseteq [n]$4

Because the Bell-measurement qubits are destructively measured, the online decoding problem reduces to the same code-capacity decoder used for ideal syndrome extraction, rather than a spacetime decoder over repeated noisy rounds. The residual logical-error accumulation obeys

$\supp(E)\subseteq [n]$5

so for $\supp(E)\subseteq [n]$6 rounds one has $\supp(E)\subseteq [n]$7 (Murphy et al., 5 Mar 2026).

Decoder design under explicit circuit-level models reflects these structural differences. Repeated syndrome extraction under circuit-level depolarizing noise generates space-time correlated highlighted vertices, diagonal fault signatures, and large numbers of vertical pairs after local corrections, motivating 3D convolutional local decoders, syndrome collapse, and vertical cleanup as preprocessing layers before a global decoder (Chamberland et al., 2022). The broader implication is that “circuit-level local stochastic error model” is not only a noise-specification phrase; it also determines whether decoding is naturally formulated as code-capacity inference, detector-graph matching, correlated maximum-likelihood summation over circuit-fault classes, or a hybrid architecture.

Taken together, these developments fix the modern meaning of the concept. A circuit-level local stochastic error model is local with respect to circuit structure, stochastic in the sense of classical sampling or support-bounded random fault sets, and useful precisely because it interpolates between microscopic hardware assumptions and logical-level analysis. Its mathematical content, however, depends strongly on which of its several established variants is intended.