Skip to content
Tutorials Center
My learning

Topics

Lesson 27 of 39 ยท Noise and Error Correction

NISQ Devices and Error Mitigation

Understand the NISQ era and learn practical error mitigation: readout calibration, zero-noise extrapolation, probabilistic error cancellation and more.

John Preskill coined the term NISQ in 2018: Noisy Intermediate-Scale Quantum. It describes machines with roughly 50 to a few thousand physical qubits and no full error correction. They are too big to simulate exactly on a laptop in some regimes, but too noisy to run the long, clean circuits that algorithms like Shor's need.

Between "no error handling" and "full error correction" sits a toolbox called error suppression and mitigation. These methods do not make individual runs correct. Instead, they combine many noisy runs (or reduce noise at the pulse level) so that the averaged results are closer to the ideal ones. This lesson covers the main techniques with small simulations.

Suppression, mitigation, correction

Level Idea Cost Examples
Suppression Make each run less noisy Little or none Better calibration, dynamical decoupling, optimized compilation, pulse shaping
Mitigation Post-process many noisy runs to estimate the ideal expectation value Exponential in noise (more shots) Readout correction, zero-noise extrapolation, probabilistic error cancellation, symmetry checks
Correction Encode logical qubits and fix errors during the run Many physical qubits Surface code

Mitigation estimates expectation values (average outcomes of observables). It cannot fix a single run, so it suits tasks like variational algorithms, where only averages matter.

Readout error mitigation

Measurement is often the noisiest operation, with a 1% to 5% chance of misreading a qubit. Fix it by calibration: prepare each basis state, measure, and record how often each outcome appears. This builds a confusion matrix M with M[observed, true]. Real counts are observed = M @ true, so you invert M to estimate the true distribution.

Python Correcting readout errors with a confusion matrix

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

For n qubits the full matrix has 4^n entries, so large devices use per-qubit models or tensored calibrations (as above), or measure only the qubits and subsets that matter.

Zero-noise extrapolation (ZNE)

Idea: you cannot reduce the noise, but you can increase it deliberately. Run the circuit at noise scale factors 1, 2, 3 (by stretching pulses or by inserting extra gate pairs G G^dagger that do nothing ideally), record the expectation value at each, fit a curve, and extrapolate back to a noise scale of 0.

Python Zero-noise extrapolation with Richardson

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The raw value is biased low (around 0.53), while extrapolation brings the estimate much closer to 0.8. Notice the trade-off: extrapolating amplifies the shot noise, so you need more shots. If the true noise model is not what the fit assumes, extrapolation can overshoot.

Probabilistic error cancellation (PEC)

PEC represents the inverse of a noise channel as a combination of implementable noisy operations with positive and negative weights: Noise^-1 = sum of c_i * O_i. You sample operations O_i with probability proportional to |c_i|, run, and multiply each result by sign(c_i) * gamma where gamma = sum |c_i|. The estimate is unbiased, but the variance grows like gamma^2 per layer, so the number of shots grows exponentially with circuit size (gamma is a little above 1 per gate). It needs an accurate noise model, which is usually learned with tomography of the noise (for example Pauli twirling plus sparse Pauli-Lindblad fits).

Other tools

  • Dynamical decoupling. Insert pairs of X pulses into idle periods; they cancel slow noise and extend effective T2.
  • Pauli twirling / randomized compiling. Randomly dress gates with Paulis so that coherent errors become simple stochastic Pauli noise, which is easier to model and mitigate.
  • Symmetry verification and post-selection. If your problem conserves a quantity (particle number, parity), discard results that violate it.
  • Clifford data regression, learning-based methods. Train a model on classically simulable circuits to correct noisy expectation values of similar circuits.
  • Noise-aware compilation. Choose the qubits and gates with the best calibration data, and reduce the two-qubit gate count.
flowchart TB
    A["Noisy hardware"] --> B["Suppression: decoupling, compilation"]
    B --> C["Run many circuits (scaled noise, twirled, calibration)"]
    C --> D["Mitigation: extrapolate, cancel, post-select"]
    D --> E["Better estimate of the expectation value"]

Mitigation pipeline: run more circuits, process the results classically.

Limits

Mitigation has a sampling overhead that grows exponentially with circuit depth and error rate. That is why it stretches today's chips (maybe hundreds to a few thousand two-qubit gates) but cannot replace error correction for algorithms that need billions of gates. It is also the reason claims of "quantum advantage" on noisy devices are fiercely debated: classical simulation methods such as tensor networks keep improving, and mitigated results have to be checked against them.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. What does NISQ stand for?

    Preskill's term for today's 50-to-thousands qubit machines without full error correction.
  2. 2. How does zero-noise extrapolation work?

    Results at scale factors 1, 2, 3 are fit with a curve evaluated at scale 0.
  3. 3. What is the main downside of probabilistic error cancellation?

    The cost factor gamma^2 per noisy layer multiplies across the circuit.
  4. 4. Mitigation techniques estimate which quantity?

    They process many noisy runs to estimate averages, not to repair individual runs.

Summary

  • NISQ machines are noisy and have no full error correction; mitigation squeezes more out of them.
  • Suppression reduces noise on each run; mitigation post-processes many runs to estimate expectation values.
  • Readout calibration inverts a confusion matrix; ZNE extrapolates from amplified noise; PEC cancels noise with signed sampling.
  • All mitigation costs extra shots that grow exponentially, so it cannot replace error correction for long algorithms.

Next: Qiskit Basics