Pauli Error Estimation

• Pauli error estimation is a method that infers error rates and logical noise from experimental data using population recovery techniques and Pauli-basis eigenvalue analysis.
• It employs full-distribution recovery, relative-precision estimation, and entanglement-free protocols, ensuring accurate channel characterization even in SPAM-noisy environments.
• These techniques enhance error mitigation and fidelity estimation, supporting advanced quantum error correction and metrology applications.

Pauli error estimation is the set of methods used to infer Pauli error rates, Pauli-transfer eigenvalues, or logical Pauli noise from experimental data. In its standard $n$-qubit form, a Pauli channel is written as

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$

while an equivalent diagonal representation uses Pauli-basis eigenvalues $\lambda_Q=\sum_P p_P\,\chi(P,Q)$, where $\chi(P,Q)\in\{\pm1\}$ records commutation or anticommutation. The subject includes full-distribution recovery in $\ell_\infty$, relative-precision estimation of rare Pauli rates, learning logical Pauli channels from syndrome statistics, and mitigation-oriented schemes that estimate Pauli-detectable error components through parity checks, extrapolation, or quasi-probability inversion (Flammia et al., 2021, Flammia et al., 2019, Wagner et al., 2022, Langfitt et al., 2024).

1. Formal models and estimation targets

The central objects of estimation vary with the application. One line of work targets the full vector $p=(p_P)_{P\in\mathcal{P}_n}$ of a Pauli channel, often with guarantees such as

$\lVert \hat p-p\rVert_\infty\le \epsilon,$

or, in the near-identity regime $p(I^{\otimes n})=1-\eta$, additive precision $\epsilon\eta$, equivalently multiplicative precision $1\pm\epsilon$ on nonzero rates. A second line targets Pauli-basis eigenvalues $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$0, which diagonalize the channel in the Pauli transfer matrix and are related to $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$1 by a Walsh–Hadamard transform. A third line targets logical Pauli noise, namely the decoder-independent channel induced on encoded degrees of freedom after averaging over stabilizer cosets. These formulations are all explicit in the modern literature and are not interchangeable without model assumptions (Flammia et al., 2021, Flammia et al., 2019, Wagner et al., 2022).

Generalized Pauli channels extend the same logic beyond qubits. On a $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$2-dimensional Hilbert space, the Heisenberg–Weyl operators $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$3 define a generalized Pauli channel

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$4

with $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$5 and $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$6. In that setting the estimation target is the $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$7-component probability vector $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$8, recovered from transition probabilities measured in eigenbases of selected Weyl operators (Rehman et al., 2021).

Single-parameter Pauli estimation remains important as a metrological subcase. For a known axis $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$9, the channel

$\lambda_Q=\sum_P p_P\,\chi(P,Q)$0

reduces Pauli error estimation to inference of the scalar $\lambda_Q=\sum_P p_P\,\chi(P,Q)$1. In this regime the relevant figures of merit are Fisher information, quantum Fisher information, and Cramér–Rao bounds. An absolute upper bound derived for the phase-flip channel is

$\lambda_Q=\sum_P p_P\,\chi(P,Q)$2

which excludes Heisenberg-like scaling for this noise-estimation task (Collins, 2012).

2. Entanglement-free estimation and Population Recovery

A major development is the reduction of Pauli-channel learning to classical Population Recovery. In the $\lambda_Q=\sum_P p_P\,\chi(P,Q)$3-qubit protocol of Flammia and O’Donnell, one chooses $\lambda_Q=\sum_P p_P\,\chi(P,Q)$4 uniformly, prepares the product state whose $\lambda_Q=\sum_P p_P\,\chi(P,Q)$5-th qubit is the $\lambda_Q=\sum_P p_P\,\chi(P,Q)$6 eigenstate of $\lambda_Q=\sum_P p_P\,\chi(P,Q)$7, applies the channel once, and measures each qubit in the $\lambda_Q=\sum_P p_P\,\chi(P,Q)$8-basis. If the realized Pauli error is $\lambda_Q=\sum_P p_P\,\chi(P,Q)$9, the readout is $\chi(P,Q)\in\{\pm1\}$0, where $\chi(P,Q)\in\{\pm1\}$1 when the corresponding Paulis anticommute and $\chi(P,Q)\in\{\pm1\}$2 when they commute. When $\chi(P,Q)\in\{\pm1\}$3 is uniform, each coordinate behaves as a $\chi(P,Q)\in\{\pm1\}$4-channel with crossover probability $\chi(P,Q)\in\{\pm1\}$5. This yields an unbiased estimator for any fixed $\chi(P,Q)\in\{\pm1\}$6,

$\chi(P,Q)\in\{\pm1\}$7

Combined with a branch-and-prune procedure over prefixes, the algorithm learns the full distribution to $\chi(P,Q)\in\{\pm1\}$8-accuracy with

$\chi(P,Q)\in\{\pm1\}$9

channel uses, and in the near-identity regime reaches

$\ell_\infty$0

channel uses for additive precision $\ell_\infty$1 (Flammia et al., 2021).

The same framework was extended to strong SPAM noise. In that model, state preparation and measurement are each depolarized with known retention parameters $\ell_\infty$2 and $\ell_\infty$3, so the combined retention is $\ell_\infty$4 and $\ell_\infty$5. The probe protocol then induces the concatenated classical channel $\ell_\infty$6. The resulting SPAM-tolerant estimator remains entanglement-free and, for $\ell_\infty$7, achieves

$\ell_\infty$8

in the mixed-noise regime, while reverting to

$\ell_\infty$9

in the low-SPAM regime. The same work gives evidence that $p=(p_P)_{P\in\mathcal{P}_n}$0-type dependence is unavoidable for SPAM-tolerant Pauli error estimation under this model (O'Donnell et al., 30 Sep 2025).

For generalized Pauli channels on $p=(p_P)_{P\in\mathcal{P}_n}$1-dimensional systems, an entanglement-free protocol prepares a single eigenstate of each selected $p=(p_P)_{P\in\mathcal{P}_n}$2, measures in the same eigenbasis, and reconstructs the parameter vector by a precomputed linear inverse $p=(p_P)_{P\in\mathcal{P}_n}$3. The number of measurement configurations $p=(p_P)_{P\in\mathcal{P}_n}$4 scales linearly with $p=(p_P)_{P\in\mathcal{P}_n}$5: $p=(p_P)_{P\in\mathcal{P}_n}$6 suffices for prime $p=(p_P)_{P\in\mathcal{P}_n}$7, $p=(p_P)_{P\in\mathcal{P}_n}$8 was sufficient for composite $p=(p_P)_{P\in\mathcal{P}_n}$9 tested up to $\lVert \hat p-p\rVert_\infty\le \epsilon,$0, and $\lVert \hat p-p\rVert_\infty\le \epsilon,$1 is a universal worst-case bound. The summed variance and MSE scale as $\lVert \hat p-p\rVert_\infty\le \epsilon,$2, and Pauli noise on probes can be modeled as measurement error, with the depolarizing correction

$\lVert \hat p-p\rVert_\infty\le \epsilon,$3

This connects Pauli error estimation directly to detector-calibration methods (Rehman et al., 2021).

3. Structured estimation, memory, and maximum likelihood

When the target is not the full $\lVert \hat p-p\rVert_\infty\le \epsilon,$4-dimensional distribution but either a structured subset or a channel with locality constraints, sharper resource bounds are available. A cycle-benchmarking-based framework estimates the full $\lVert \hat p-p\rVert_\infty\le \epsilon,$5-qubit Pauli channel to relative precision $\lVert \hat p-p\rVert_\infty\le \epsilon,$6 with $\lVert \hat p-p\rVert_\infty\le \epsilon,$7 single-qubit measurements, a specified set of $\lVert \hat p-p\rVert_\infty\le \epsilon,$8 Pauli error rates with $\lVert \hat p-p\rVert_\infty\le \epsilon,$9 measurements, and a Pauli channel given by a Markov random field with at most $p(I^{\otimes n})=1-\eta$0-local correlations using $p(I^{\otimes n})=1-\eta$1 measurements. The same framework is SPAM-robust because sequence-length ratios cancel the SPAM coefficients that appear in the decay amplitudes (Flammia et al., 2019).

Another direction asks how much coherent memory is needed to estimate all Pauli eigenvalues efficiently. A concatenating protocol with $p(I^{\otimes n})=1-\eta$2 ancillas estimates every $p(I^{\otimes n})=1-\eta$3 to additive error $p(I^{\otimes n})=1-\eta$4 using $p(I^{\otimes n})=1-\eta$5 measurements. By contrast, any zero-ancilla protocol, even if concatenating and adaptive, must use at least $p(I^{\otimes n})=1-\eta$6 measurements, and any protocol with $p(I^{\otimes n})=1-\eta$7 ancillas requires $p(I^{\otimes n})=1-\eta$8 queries. The protocol’s architecture combines stabilizer POVMs, control channels, a purification primitive, and online convex optimization over Choi states (Chen et al., 2023).

For sparse, 1D-local Pauli-Lindblad channels, maximum likelihood estimation has recently become computationally tractable. In that setting the channel factorizes as

$p(I^{\otimes n})=1-\eta$9

and the likelihood can be rewritten as a low-treewidth Bayesian network whose exact evaluation uses belief propagation. The paper reports that, for a $\epsilon\eta$0-qubit $\epsilon\eta$1D $\epsilon\eta$2-local example with $\epsilon\eta$3, MLE reached a target MSE per parameter with roughly one third the samples required by the empirical Pauli fidelities estimator. With $\epsilon\eta$4 shots distributed across nine bases, per-parameter accuracy was nearly independent of $\epsilon\eta$5, and in a probabilistic error-cancellation application the learned model supported accurate mitigated magnetization to much longer times with MLE than with EPF (Belkin et al., 2 Jun 2026).

4. Syndrome data, Pauli checks, and mitigation-oriented estimation

In stabilizer quantum error correction, Pauli error estimation can be performed at the logical level directly from syndrome data. For a physical Pauli distribution $\epsilon\eta$6, the decoder-independent logical channel is

$\epsilon\eta$7

constant on stabilizer cosets. The theory developed for arbitrary stabilizer codes, subsystem codes, and data syndrome codes shows that the logical channel is identifiable from syndrome measurements as long as the code can correct the noise. The analysis uses Fourier moments

$\epsilon\eta$8

and canonical moments obtained by Möbius inversion; with positivity assumptions such as $\epsilon\eta$9, the estimation problem becomes a log-linear inversion from measured stabilizer moments to logical Pauli probabilities (Wagner et al., 2022).

A distinct mitigation-oriented line uses Pauli parity checks to estimate and suppress residual bias. In Pauli Check Sandwiching, a payload circuit is bracketed by identical check pairs, and a shot is accepted only if the checks pass. A check detects an error precisely when the error anticommutes with the check operator. This leads to an acceptance probability $1\pm\epsilon$0 for $1\pm\epsilon$1 check pairs and to observable estimates $1\pm\epsilon$2 that typically increase with $1\pm\epsilon$3 because undetected error contributions decrease. Pauli Check Extrapolation replaces direct operation at very large $1\pm\epsilon$4 by extrapolation to the “maximum check” limit $1\pm\epsilon$5. The explicit linear fit used in the paper’s figure is

$1\pm\epsilon$6

for measured values $1\pm\epsilon$7, $1\pm\epsilon$8, and $1\pm\epsilon$9, which yields

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$00

at $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$01. Under a Markovian ansatz, the alternative model is

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$02

The method was applied to shadow estimation for VQE-prepared states and was reported to achieve higher fidelities than Robust Shadow while eliminating the need for a calibration procedure (Langfitt et al., 2024).

Pauli error estimation also enters probabilistic error cancellation on Clifford circuits. If each inverse noise channel is expanded in Pauli corrections and then propagated through a Clifford subcircuit, the fused inverse satisfies

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$03

so quasi-probability overhead decreases by destructive interference among correction paths. In a reported VQE commuting-group measurement example, $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$04, $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$05, and $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$06, illustrating the dependence of mitigation cost on the quality and representation of the learned Pauli model (Scheiber et al., 2024).

5. Pauli twirling, model construction, and threshold studies

Many estimation protocols target the Pauli projection of a general channel rather than the full non-Pauli dynamics. The Pauli twirling approximation averages a completely positive channel over the Pauli group and produces

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$07

with probabilities given by the diagonal $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$08-matrix entries,

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$09

This yields a direct route from process tomography or physical noise models to Pauli error rates. In the stabilizer-measurement circuit studied in the PTA analysis, predictions were “essentially perfect” when decoherence was present without gate errors and remained “excellent” when both decoherence and intrinsic gate errors were present (Geller et al., 2013).

Twirling need not use the full Pauli group. For a channel with Pauli support $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$10, a smaller twirling set $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$11 suffices provided

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$12

where $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$13 records commutation or anticommutation. The resulting size bounds are

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$14

The same work shows that one-gate twirling with $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$15 is equivalent to a stabilizer measurement of $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$16 with the outcome discarded, giving an operational link between twirling and syndrome-extraction circuitry (Cai et al., 2018).

Threshold studies show both the utility and the limitations of Pauli-based estimation. For arbitrary single-qubit Pauli noise, numerical lower bounds on quantum-capacity thresholds vary strongly across the Pauli simplex, and different graph-state code families dominate in different bias regions: repetition codes in Z- and Y-biased regions, cat codes near the unbiased center, and tree codes in broad X-biased regions (Bausch et al., 2019). At the same time, beyond-Pauli simulations using the Pauli Frame Sparse Representation show that coherent-noise thresholds computed at circuit level up to $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$17 are systematically overestimated by a Pauli-twirling approximation by a factor of about $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$18. For rotated surface-code memory, the reported values were $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$19 and $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$20, whereas amplitude damping at the phenomenological level remained well captured by the Pauli approximation, with $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$21 (Tuloup et al., 15 Mar 2026).

6. Precision limits, correlations, and related Pauli-measurement tasks

The role of quantum correlations in Pauli error estimation depends strongly on the probe model. For pure probes in the single-axis channel $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$22, independent unentangled strategies already attain the absolute bound

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$23

For mixed probes, however, a correlated-state protocol can outperform every independent-state protocol with the same number of channel uses. The reported gains persist even when the post-channel states are separable and, in some cases, when quantum discord is absent after channel invocation. This regime is explicitly connected to NMR, where the dephasing parameter obeys $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$24 (Collins, 2012).

Direct fidelity estimation is a related but distinct Pauli-measurement task. For a target pure state $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$25, with Pauli coefficients $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$26 and $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$27, the fidelity is

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$28

Sampling $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$29 from the importance distribution $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$30 gives the unbiased estimator

$$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$31

For stabilizer targets, the required number of settings and shots is $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$32, independent of the Hilbert-space dimension, and the same formalism extends to entanglement fidelity of channels. In Pauli-diagonal noise models, the sampled ratio $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$33 is exactly the Pauli-transfer eigenvalue $$\mathcal{E}(\rho)=\sum_{P\in\mathcal{P}_n} p_P\,P\rho P, \qquad p_P\ge 0,\quad \sum_P p_P=1,$$34, so the method estimates noise directly on the support emphasized by the target state (Flammia et al., 2011).

Across these formulations, a recurring limitation is model fidelity. Population-recovery methods assume Pauli-diagonal structure or an effectively twirled channel; logical-noise recovery assumes correctability of the relevant supports; check-based extrapolation assumes monotone improvement with additional detectability; and Pauli twirling can suppress coherent interference that matters at circuit level. Pauli error estimation is therefore best understood not as a single estimator, but as a hierarchy of inference problems whose validity is determined by the commutation structure, locality assumptions, and operational task that define the underlying Pauli model.