Error Detection Plus Probabilistic Error Cancellation: Up to a 63-Fold Cut in Sampling Overhead

Probabilistic error cancellation pays for accuracy in shots, and the bill grows exponentially with circuit depth. An IBM Research preprint posted on 11 September 2026 reports a way to make error detection absorb part of that cost, cutting the inferred sampling overhead from 85,545 to 1,359 on a six-step Trotter circuit.

Why the PEC shot bill grows exponentially with depth

PEC learns a model of the processor's noise, then inverts it by running modified circuits and combining their results with positive and negative weights. The sparse Pauli-Lindblad protocol that made this practical on hardware (Nature Physics 19, 1116) models correlated noise with a tractable number of parameters; it learned the noise of a 20-qubit gate layer and mitigated Ising dynamics on chains of four and ten qubits.

The price is a sampling overhead that compounds layer by layer. Takagi, Tajima and Gu showed this is a property of mitigation rather than of PEC: their universal lower bounds (Physical Review Letters 131, 210602) apply to general mitigation protocols, including ones yet to be discovered, and imply that for a wide class of them the sampling cost must grow exponentially with circuit depth under various noise models.

So there is no algorithmic escape inside mitigation alone. What remains available is lowering the base of the exponent.

Error detection buys accuracy by discarding shots

The other lever adds check qubits, measures them, and throws away every run whose syndrome fires. Microsoft and Quantinuum reported logical error rate improvements of 11x to 800x over physical circuit baselines on Quantinuum's trapped-ion hardware (Nature 654, 349) by combining two error-correcting codes, a 12-qubit carbon code and a 16-qubit tesseract colour code, with error detection and post-selection.

Post-selection has its own currency: acceptance probability. A discarded shot still consumed device time, so detection trades one exponential for another.

The nonlinearity that blocked composing the two

The PEC estimator is linear in its samples: each randomized circuit contributes a signed weight, and the average is the mitigated expectation value. Post-selection normalizes each post-selected expectation by its own acceptance probability, which can depend on the input state and the randomization; that makes it generally nonlinear. Normalize every randomization separately and the ensemble being averaged stops matching the one the PEC inversion assumed; the preprint demonstrates the residual bias with an amplitude damping example.

The fix is to keep the post-selected maps unnormalized. Each randomization's expectation is multiplied by its own acceptance probability alpha_r, the results are averaged, and only then is the total divided by the common alpha. Normalization happens once, after averaging, where it stays compatible with a linear estimator.

What a spacetime Pauli-Lindblad model tracks that a per-layer model cannot

Post-selection also changes the shape of the noise left behind, and that is the second obstacle. The preprint introduces an auxiliary Pauli space with one tensor factor per circuit layer, so every elementary fault carries a label of the form (Pauli, layer).

That labelling splits the post-selected noise cleanly. Faults no check catches contribute at first order, exactly as in ordinary PEC. Faults that are individually detectable but whose syndromes cancel survive together, generating higher-order terms that couple distinct locations in space and time.

A per-layer noise model has no vocabulary for a correlation between a fault in layer two and a fault in layer five. That is the gap the spacetime representation fills: the residual channel is constructed as a spacetime object from the learned physical noise model and the circuit's syndrome information, then inverted.

The numbers from the IBM Aachen processor

Validation ran on a 24-qubit ring of 20 data and 4 check qubits under Clifford circuits, where those four checks detected 65% of the modelled elementary faults. The physics run used 22 data and 27 check qubits on transverse-field Ising dynamics, reaching 648 CZ gates at six Trotter steps and recovering the expected mean magnetization.

The inferred sampling overhead, PEC alone against PEC with detection:

• Two Trotter steps: 44.1 falls to 12.1, a 3.7x improvement
• Four steps: 1,941 falls to 122, a 15.9x improvement
• Six steps: 85,545 falls to 1,359, a 63x improvement

Read those three improvement factors in order. The advantage roughly quadruples for every two additional Trotter steps, which is the signature of a smaller base in the exponent rather than a fixed discount. Deeper circuits should widen the gap.

One caveat on the metric: the overhead is an inferred quantity, gamma_L squared divided by alpha, computed from the learned noise models and the measured syndrome acceptance rate, which falls from 0.534 at two steps to 0.197 at six. Both columns also leave out the cost of a filter for leakage and other non-Markovian errors, applied to every method, whose typical survival rate was 16% to 24% of shots. It is a sampling-cost measure, and no wall-clock runtime is claimed.

The limits on the 63x figure

Three limits are worth carrying out of the paper (arXiv:2609.13108):

• Checks are measured only at the end of the circuit, and the authors list midcircuit post-selection as future work
• The inversion is first order, with the second-order bias bounded below the statistical error for the observables tested
• The result is a preprint, run on IBM hardware and measured by IBM, so the 63x figure is reported rather than independently confirmed

Every number above rests on a learned noise model. The learning step is the same sparse Pauli-Lindblad characterization standard PEC uses, and the post-selected logical model is computed from it and the circuit's syndrome information. In the Nature Physics demonstration, the authors relearned the noise model every few hours to counter time-dependent fluctuations in the noise, so how long a learned model stays accurate is the quiet constraint under all of this.

That shelf-life question is what SuperconducTED was built to study: it uses fuzzy inference on IBM calibration data to build Qiskit Aer noise models, aiming for models that stay accurate across calibration cycles. Read the SuperconducTED calibration-drift approach if model staleness is the part of PEC that worries you.