Probabilistic Error Cancellation in Quantum Computing

• Probabilistic Error Cancellation (PEC) is a quantum error-mitigation method that reconstructs ideal outcomes by representing inverse noise channels through quasi-probability distributions over noisy gates.
• PEC relies on precise noise characterization techniques, such as Gate Set Tomography and sparse Pauli-Lindblad learning, to decompose ideal operations into experimentally feasible noisy operations.
• The method incurs a sampling overhead and variance amplification that grow exponentially with circuit depth, necessitating advanced aggregation and adaptive mitigation strategies.

Searching arXiv for recent and foundational PEC papers to support the article. Probabilistic Error Cancellation (PEC) is a quantum error-mitigation protocol that estimates noiseless expectation values from noisy circuit executions by representing the inverse of the noise channel, or equivalently the ideal operation, as a quasi-probability distribution over implementable noisy operations. Its defining property is unbiasedness with respect to the assumed noise model: for an observable measured at circuit end, PEC constructs an estimator whose expectation equals the ideal value, while incurring a sampling overhead governed by the negativity of the quasi-probability decomposition (Mari et al., 2021). In practical implementations, PEC is closely tied to noise characterization—often through gate-set tomography or sparse Pauli-Lindblad learning—and to the structure of the target circuit, because the total overhead typically grows multiplicatively with depth and noise strength (Berg et al., 2022).

1. Formal definition and estimator

Let an ideal circuit be composed of superoperators $\{\mathcal{G}_i\}$, and let each hardware gate be implemented as a noisy operation $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$. For a terminal Hermitian observable $O$, the target quantity is $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$. PEC rewrites either each ideal gate or the inverse noise map as a signed linear combination of implementable operations. In the gate-based form,

$$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$

where the coefficients $\eta_{\alpha,i}$ can be positive or negative (McDonough et al., 2022). In the equivalent inverse-channel form, one writes

$\Lambda^{-1}=\sum_j \eta_j\,G_j,$

or, in a Pauli-channel specialization,

$$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$

with real quasi-probabilities $\eta_k$ (Berg et al., 2022).

Sampling is organized by the absolute values of these coefficients. For each gate, the per-gate overhead is

$\Gamma_g:=\sum_\alpha |\eta_\alpha|,$

with probabilities

$\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$0

A sampled circuit instance selects one term per gate, runs the corresponding noisy circuit, and records the measurement outcome $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$1. The standard PEC estimator takes the form

$\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$2

with

$\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$3

and is unbiased under the assumed decomposition (McDonough et al., 2022).

The same structure appears in more abstract Monte Carlo formulations. If $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$4 is the sampled sign and $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$5 the measured outcome from the sampled noisy circuit, then with $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$6 i.i.d. samples

$\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$7

and in expectation this recovers the ideal observable value after the global normalization is included (Mari et al., 2021). Unbiasedness, however, does not imply low mean-squared error: the central cost of PEC is variance amplification induced by negative quasi-probabilities (Annis et al., 3 Jun 2026).

2. Noise representations and characterization

Practical PEC depends on a noise model rich enough to support inversion but structured enough to remain learnable. Two representations recur in the literature covered here.

The first is a gate-based description obtained from Gate Set Tomography (GST). GST reconstructs gate-wise Pauli transfer matrices for implementable gates, up to gauge, after which ideal gates are decomposed over the reconstructed noisy operations. In the automated framework of “Automated quantum error mitigation based on probabilistic error reduction” (McDonough et al., 2022), PyGSTi is used for GST, gauge optimization is used to make $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$8 virtually noiseless on Rigetti Aspen-11, and the resulting superoperators are passed to Mitiq for quasi-probability decomposition.

The second is a sparse Pauli-Lindblad description learned layer by layer. After Pauli twirling, the layer noise channel is diagonal in the Pauli basis and can be written as

$\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$9

where $O$0 and $O$1 is the Lindblad rate associated with Pauli jump operator $O$2 (Berg et al., 2022). The corresponding Pauli fidelities satisfy

$O$3

and the rates are extracted by non-negative least squares from benchmarking data. This representation is designed to capture local and correlated noise, including crosstalk, while keeping the parameter count sparse (Berg et al., 2022).

For circuits with mid-circuit measurements and classically controlled feedforward, the representation must be extended to measurement-based operations. “Probabilistic error cancellation for dynamic quantum circuits” develops such an extension by enforcing projective measurements with a post-measurement random $O$4 and showing that, after twirling, the noisy measurement-plus-feedforward channel can be expressed as

$O$5

so that PEC can be applied to dynamic circuits using the same sparse Pauli-Lindblad logic, now augmented with measurement crosstalk and control-flow idle errors (Gupta et al., 2023).

3. Overhead, variance, and the core bottleneck

The defining limitation of PEC is the multiplicative growth of sampling overhead. If $O$6 is the per-gate or per-layer quasi-probability norm, then for a circuit of $O$7 corrected gates or layers,

$O$8

in the homogeneous case, and the variance grows with $O$9 (McDonough et al., 2022). Standard Monte Carlo analysis then gives sample complexity scaling as

$$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$0

or, equivalently, $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$1 at fixed precision for single-gate overhead $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$2 (Mari et al., 2021).

Within the sparse Pauli-Lindblad model, the per-layer norm is explicit:

$$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$3

so the total cost scales exponentially in accumulated Lindblad weight (Berg et al., 2022). This is the source of the familiar statement that PEC is exponentially expensive in noise-weighted circuit volume.

Several papers sharpen this point. In a VQE ansatz-selection study on $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$4 and $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$5, the practical overhead associated with $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$6 was linked directly to two-qubit gate count and realistic noise rates. With per-gate cost factor estimated as $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$7 under the study’s noise model, circuits with $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$8 two-qubit gates incurred $$\langle O\rangle_{\mathrm{ideal}}=\mathrm{Tr}[O\,U(\rho)]$$9, and circuits with $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$0 incurred $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$1; under the fixed sampling budget used there, PEC increased error in 11 of 12 $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$2 circuits and all 6 $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$3 circuits (Annis et al., 3 Jun 2026). This suggests that unbiasedness is not, by itself, a reliable predictor of practical accuracy under finite shot budgets.

PEC also relies on stable, approximately gate-local, time-stationary, Markovian noise. Drift, non-Markovianity, and unmodeled crosstalk degrade performance, while full GST scales exponentially in qubit count and is therefore practically limited to small subsystems (McDonough et al., 2022). Sparse Pauli-Lindblad learning improves scalability, but the model assumptions remain consequential (Berg et al., 2022).

4. Structural and algorithmic extensions

A large body of work attempts to preserve PEC’s bias properties while reducing variance, reducing calibration burden, or exploiting circuit structure.

Noise scaling provides one of the main bridges between PEC and zero-noise extrapolation (ZNE). “Extending quantum probabilistic error cancellation by noise scaling” defines a generalized framework in which ideal operations are expanded over noise-scaled implementations, encompassing both PEC and ZNE as special cases (Mari et al., 2021). For a canonical decomposition with positive and negative parts $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$4 and $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$5, the noise-scaled gate is

$$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$6

and the associated one-norm obeys

$$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$7

This underlies probabilistic error reduction (PER), which trades exact unbiasedness for lower overhead, and virtual ZNE, which extrapolates PER data back toward zero noise (Mari et al., 2021). The automated framework in (McDonough et al., 2022) operationalizes this idea with GST/Mitiq and sparse Pauli tomography.

Several works reduce PEC cost by changing the granularity of inversion. MoSAIC aggregates multiple noisy layers into “noise-aligned” blocks, learns an effective block channel variationally, and applies one quasi-probabilistic inverse per block rather than per layer. On IBM Heron processors, MoSAIC reported relative energy-per-site errors of $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$8, $$\mathcal{G}_i=\sum_\alpha \eta_{\alpha,i}\,\mathcal{O}_\alpha,$$9, and $\eta_{\alpha,i}$0 for 14, 30, and 50 qubits, versus $\eta_{\alpha,i}$1, $\eta_{\alpha,i}$2, and $\eta_{\alpha,i}$3 for standard PEC under identical sampling budgets; the corresponding experimentally extracted $\eta_{\alpha,i}$4 values were substantially smaller for MoSAIC than for standard PEC (Ma et al., 27 Mar 2026).

“Faster Probabilistic Error Cancellation” reorganizes the inverse channel into an identity part and a non-identity “inverse generator,” expands the circuit by powers of that generator, and allocates shots deterministically across grouped terms. The resulting bias from truncation is explicitly bounded by the tail weight

$\eta_{\alpha,i}$5

and 2D TFIM simulations reported up to approximately $\eta_{\alpha,i}$6 variance reduction relative to standard PEC (Chen, 4 Jun 2025).

Variance can also be reduced statistically without changing the quasi-probability decomposition. CV4Quantum imports control variates into QPD-based estimators, keeping unbiasedness through leave-one-out coefficient estimation. In more than 50% of PEC-based estimates in the study, the number of samples needed for a given precision was reduced by more than 50% (Shyamsundar et al., 12 Feb 2025).

Circuit structure provides another route. “Locality and Error Mitigation of Quantum Circuits” introduces a lightcone-restricted PEC estimator for local observables:

$\eta_{\alpha,i}$7

where $\eta_{\alpha,i}$8 is the observable’s light cone. The estimator remains unbiased, but its variance depends only on in-cone noise, not the full circuit (Tran et al., 2023). “Lightcone shading for classically accelerated quantum error mitigation” tightens this idea by computing per-channel bias bounds and reported that, on a 127-qubit TFIM Trotter circuit, shaded-lightcone PEC reduced the required shots to below $\eta_{\alpha,i}$9, compared with approximately $\Lambda^{-1}=\sum_j \eta_j\,G_j,$0 for conventional lightcone PEC and $\Lambda^{-1}=\sum_j \eta_j\,G_j,$1 for full PEC (Eddins et al., 2024).

Hardware-specific bias can help as well. For cat-qubit architectures with exponentially suppressed bit-flip errors, “Low bit-flip rate probabilistic error cancellation” shows that Block-PEC can exploit Pauli-$\Lambda^{-1}=\sum_j \eta_j\,G_j,$2 compatibility and bias-preserving gates to remove the exponential dependence of the quantum sampling overhead on depth for compatible blocks, shifting the cost to classical preprocessing that scales as $\Lambda^{-1}=\sum_j \eta_j\,G_j,$3 (Rennela et al., 2024).

5. Adaptive, feed-forward, and code-assisted PEC

Because PEC is model-based, several extensions address cases where the baseline assumptions fail.

Under non-stationary noise, static quasi-probabilities become stale. “Adaptive mitigation of time-varying quantum noise” models the Pauli-channel coefficients with a Dirichlet prior and updates the PEC weights online via Bayesian inference, reporting that Bayesian PEC outperformed a non-adaptive approach by a factor of $\Lambda^{-1}=\sum_j \eta_j\,G_j,$4 in Hellinger distance from the ideal distribution (Dasgupta et al., 2023). In a separate 5-qubit Bernstein–Vazirani experiment on ibm_kolkata, Bayesian adaptive PEC improved accuracy by 42% on average and stability by 60% on average relative to non-adaptive PEC (Dasgupta et al., 2024).

A different failure mode arises when the inserted recovery operations are themselves noisy. “Feed-Forward Probabilistic Error Cancellation with Noisy Recovery Gates” shows that conventional PEC can then become biased, because the recovery-gate noise reintroduces the channel being inverted. FFPEC instead constructs the inverse over the noisy recovery basis itself, restoring unbiasedness while only slightly increasing $\Lambda^{-1}=\sum_j \eta_j\,G_j,$5 relative to conventional PEC under the studied depolarizing models (Kurosawa et al., 2024).

Logical encodings and symmetry measurements can weaken the effective channel that PEC must invert. “Optimizing Symmetry Informed Probabilistic Error Cancellation” combines PEC with quantum error detection (QED), measuring selected stabilizers and applying PEC only to the undetectable residual channel. The total sampling overhead becomes

$\Lambda^{-1}=\sum_j \eta_j\,G_j,$6

or includes an additional symmetry-circuit mitigation term when those circuits are themselves mitigated (O'Leary et al., 1 Jul 2026). For GHZ-state preparation, the optimization over symmetry-measurement configurations was reported as essential; for generalized superfast-encoded Fermi–Hubbard dynamics, PEC+QED improved observable estimation on a $\Lambda^{-1}=\sum_j \eta_j\,G_j,$7 lattice and reduced overhead further for larger systems by measuring only subsets of stabilizers (O'Leary et al., 1 Jul 2026).

At the logical-qubit architecture level, “Co-Designing Error Mitigation and Error Detection for Logical Qubits” derives the break-even condition

$\Lambda^{-1}=\sum_j \eta_j\,G_j,$8

where $\Lambda^{-1}=\sum_j \eta_j\,G_j,$9 is physical error, $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$0 is post-selected logical error, and $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$1 is rejection probability (Kumar et al., 21 Apr 2026). The same work identifies a first-cycle transient error profile that makes naive PEC+QED integration worse than QED alone, and introduces steady-state extraction to isolate the steady-state channel, reducing estimation bias by up to $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$2 (Kumar et al., 21 Apr 2026).

A related QEDC-based construction, “Zeno-Enhanced Probabilistic Error Cancellation with Quantum Error Detection Codes,” uses post-selection to map physical noise to a weaker accepted logical channel and then constructs a perturbative inverse for that normalized post-selected map. For fixed truncation order $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$3, the number of retained fault branches is reduced from $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$4 to $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$5 per block, and the order-$$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$6 residual scales as $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$7 in the block noise weight $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$8 (Yuan et al., 12 May 2026). In first order, the protocol reached $$\mathcal{E}^{-1}\approx \sum_k \eta_k\,\mathcal{B}_k,$$9 physical qubits in logical GHZ preparation, lowering sampling overhead by three to four orders of magnitude relative to standard PEC while maintaining fidelity approximately $\eta_k$0 under the paper’s ideal-stabilizer assumptions (Yuan et al., 12 May 2026).

6. Applications, empirical performance, and practical controversies

PEC has been demonstrated on several platforms and workloads, but the empirical record is mixed in exactly the way its theory predicts.

On superconducting hardware and simulators, automated PER/virtual-ZNE pipelines have been used to approximate PEC-level accuracy at lower cost. For a noisy $\eta_k$1 gate on Rigetti Aspen-11, the GST-derived per-gate overhead was $\eta_k$2 (or $\eta_k$3 with gauge-enforced noiseless $\eta_k$4); for an 8-gate circuit, this corresponds to $\eta_k$5 under PEC versus approximately $\eta_k$6 under PER at $\eta_k$7 (McDonough et al., 2022). In a 15-step TFIM Trotter example on FakeVigoV2, the overhead dropped from $\eta_k$8 for PEC to $\eta_k$9 for PER, and the estimated sample complexity fell from approximately $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$0 to approximately $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$1 when PER was combined with virtual ZNE (McDonough et al., 2022).

Sparse Pauli-Lindblad PEC has also been implemented directly on hardware with correlated noise. “Probabilistic error cancellation with sparse Pauli-Lindblad models on noisy quantum processors” demonstrated layerwise PEC on IBM ibm_hanoi, recovering high-weight observables in Ising dynamics up to 10 qubits and 15 Trotter steps, while explicitly learning crosstalk terms in the layer noise model (Berg et al., 2022).

On trapped-ion hardware, PEC has been benchmarked in digital simulation of interacting fermions. In a two-qubit spinless Fermi–Hubbard benchmark, the per-gate cost for the noisy $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$2 gate was reported as $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$3, and the fitted per-gate population fidelity improved from $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$4 to $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$5 after mitigation (Chen et al., 2023). Three- and four-qubit trapped-ion experiments required stronger auxiliary constraints—specifically maximum-likelihood projection onto valid probabilities and symmetry-based postselection—to counteract variance and model mismatch, but they still recovered the expected fermionic dynamics, including different charge and spin behavior in the spinful Fermi–Hubbard case (Chen et al., 2023).

Variational algorithms illustrate both the promise and the fragility of PEC. “Variational quantum algorithms with invariant probabilistic error cancellation on noisy quantum processors” introduces invariant-PEC (IPEC), which fixes the sampled PEC circuits across VQA iterations, and adaptive partial PEC (APPEC), which modulates the cancellation strength during optimization. On a superconducting processor, APPEC reduced sampling cost by $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$6 and improved the fidelity of the final state distribution from $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$7 to $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$8 in the reported 3-qubit QAOA experiment (Chi et al., 8 Jun 2025). By contrast, the VQE ansatz-selection study cited earlier found PEC predominantly harmful in its tested regimes: mean error rose from $\Gamma_g:=\sum_\alpha |\eta_\alpha|,$9 to $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$00 Ha for $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$01 and from $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$02 to $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$03 Ha for $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$04, while circuit rankings were actively reordered under PEC (Annis et al., 3 Jun 2026). This suggests that PEC’s practical value is highly sensitive to circuit topology, noise strength, and shot budget, rather than being uniformly beneficial across NISQ workloads.

Taken together, these results position PEC as a rigorously defined but resource-sensitive mitigation primitive. It is strongest when the noise model is accurate, the circuit structure can be exploited, and the residual variance can be controlled through aggregation, locality, bias-aware design, adaptive inference, or code-assisted filtering. It is weakest when the effective $\tilde{\mathcal{G}} = \mathcal{N}\circ\mathcal{G}$05 is pushed far from 1 by depth, entangling-gate count, noisy recovery operations, drift, or post-selection-induced channel correlations that are not accounted for in the inversion model.