Qiskit's Client-Side Estimator Makes PEC Noise Learning Explicit

October 15, 20265 min readBy Mert Efe Sensoy

quantum-computing error-mitigation Qiskit IBM Quantum

Switching to Qiskit's client-side Estimator makes you supply the noise model for probabilistic error cancellation (PEC) and probabilistic error amplification (PEA). IBM researchers identify drift between learning and mitigation as a leading source of PEC fluctuations (Kim et al.). Here are three ways to manage that gap.

What qiskit-ibm-runtime 0.50.0 moved to the client

Version 0.50.0, released on 24 September 2026, added client-side Sampler and Estimator classes that run pre- and post-processing locally and hand execution to the Executor primitive (release notes). The server-side EstimatorV2 and SamplerV2 are deprecated. Their top-level imports will switch to the client-side classes when the deprecation period ends; upgrading to 0.50.0 alone does not switch existing code.

IBM's announcement emphasizes inspectable, customizable mitigation and says most code needs only an import change (IBM Quantum blog):

# legacy, server-side: from qiskit_ibm_runtime import Estimator
from qiskit_ibm_runtime.executor_estimator import Estimator

PEA and PEC require explicit noise learning

For PEA and PEC, changing the import is not enough. IBM's migration guide requires separately learned layer-noise models. Find the circuit's unique layers, learn their noise with NoiseLearnerV3, convert the result to Pauli-Lindblad maps and assign them to resilience.layer_noise_model.

This PEC example assumes mode is a real backend or an active Batch or Session, and pubs contains ISA-transpiled circuits and their mapped observables:

from qiskit_ibm_runtime import NoiseLearnerV3
from qiskit_ibm_runtime.executor_estimator import Estimator

estimator = Estimator(mode=mode)
estimator.options.resilience.pec_mitigation = True

layers = estimator.find_unique_layers(pubs)
learning_job = NoiseLearnerV3(mode=mode).run(layers)
maps = learning_job.result().to_pauli_lindblad_maps()
estimator.options.resilience.layer_noise_model = zip(layers, maps)
job = estimator.run(pubs)

learning_job_id = learning_job.job_id()

Twirled readout error extinction (TREX) still learns measurement noise implicitly. The server-side Estimator could learn layer noise automatically and expose the learned model in result metadata (PEA tutorial); it also accepted supplied models (release notes, 0.28.0). The client-side Estimator makes explicit model supply mandatory for PEA and PEC, so your code must manage reuse and freshness.

Why noise-model drift can bias PEC results

PEC inverts a learned sparse Pauli-Lindblad model. Each Pauli term k carries a rate λk; the channels can be inverted independently, with γ = exp(Σ 2λk) and PEC sampling cost proportional to γ² for fixed precision (Kim et al., Nature Communications 16, 8439). The estimate is unbiased in the absence of model inaccuracy, so the model must represent the noise during mitigation.

Kim and colleagues, all at IBM Quantum, used a six-qubit device with electrodes that tune interactions with defect two-level systems (TLS). Its T1 values fluctuated on average by over 300% over 60 hours. During roughly 50 hours of tracking, model parameters in the fixed-bias control fluctuated strongly around the 13-hour mark, correlating with a strong qubit-TLS interaction.

The authors attribute a leading source of spread in mitigated results to noise drift between learning and mitigation. Predicted and observed deviations from the ideal observable value were correlated. They also identify other possible biases, including quasi-static noise, so shortening the gap cannot guarantee accuracy.

Noise strength affects cost too. Their worst and best γ values were about 1.13 and 1.06. For the benchmark's repeated unit-layer depth N = 20, the paper estimates about 13 times more PEC circuits in the worst case; at N = 40, about 167 times more.

The drift is not specific to one IBM device

Klimov and colleagues at Google and UCSB found direct evidence linking TLS defects to large fluctuations in superconducting-qubit relaxation times (Physical Review Letters 121, 090502).

Researchers at Tecnun (University of Navarra) and the University of Delaware summarize studies reporting T1 and T2 variations of up to 50% of their mean value (npj Quantum Information 7, 115). Their modeled T1 process is nearly constant over intervals much shorter than its minute-scale correlation time. This does not establish a universal expiry time for PEC models.

That paper concerns channel models and error correction. My inference for PEC is that a fixed fit to noise measured during a learning window can become unrepresentative as hardware drifts.

Local mode cannot rehearse the noise-learning step

The client-side Estimator applies TREX locally when resilience_level = 1; the server-side implementation ignored resilience options on simulators (IBM Quantum blog). This lets you test that mitigation path locally.

However, NoiseLearnerV3 has no local testing mode and accepts only a real Backend, Session or Batch (migration guide). Local noisy simulation does not reproduce hardware results. The guide also confirms that seed_estimator has no client-side equivalent, so that seed no longer provides reproducibility.

Keep each learned noise model representative of its workload

IBM recommends learning once over the union of layers and reusing the maps across split Estimator jobs in a Batch. A Batch can reduce delays, but other workloads and calibration jobs can intervene. An active Session provides exclusive access, while noise can still drift (execution modes).

My practical recommendations are:

Learn near the Estimator jobs. Use the same Batch or Session where appropriate; recheck after delays or recalibration, even within one batch.
Record the learning job's ID and execution window with the maps, along with the backend identity and workload execution times. Submission time alone does not measure model age.
Compare model predictions with a control without PEC. Kim et al. interleaved non-PEC Clifford benchmark runs with mitigation runs; readout-error mitigation was applied to both.

Using ConvertISAToClifford for a surrogate check is my suggestion. The pass rounds RX, RZ and RZZ angles (API reference), so confirm model applicability to the surrogate's layers and account for sampling uncertainty. Agreement on the surrogate does not establish accuracy for the original workload.

Kim et al. also reduce spread by averaging independent learning-mitigation cycles, at increasing cost for larger circuits. The error detection plus PEC post explains the sampling cost. Read SuperconducTED's calibration-drift approach before reusing a noise model across runs or calibration changes.