Zero-Noise Extrapolation: Estimate the Ideal Signal from Noisy Circuits

Zero-Noise Extrapolation: Estimate the Ideal Signal from Noisy Circuits

Zero-noise extrapolation estimates an ideal observable from several deliberately noise-amplified circuit runs. Learn the gate-folding idea and its tradeoffs.

Quantum hardware cannot run a circuit without noise, but sometimes we can estimate how an observable might look if that noise were smaller. Zero-noise extrapolation, or ZNE, does this by collecting expectation values at several noise levels and extrapolating them toward a zero-noise limit.

It is a way to mitigate bias in an estimate, not a way to make the device physically noiseless. Here is the basic idea and a small numerical example.

Three steps from noisy runs to an estimate

First, choose a circuit and an observable, such as the average value of Pauli Z on one qubit. Run the circuit several times to estimate that expectation value.

Next, create logically equivalent versions of the circuit that experience more noise. One digital technique is gate folding. A unitary gate U can be replaced by U U-dagger U. Since U-dagger U is the identity, the ideal operation is still U, but the folded version uses three physical gate operations and gives noise more opportunities to act. Larger noise factors use more folding.

Finally, plot or fit the measured expectation values against the noise factor and extend the trend back to factor zero. Zero noise is not another circuit run. It is the intercept estimated from the noisy data.

A small linear-fit example

The following values are synthetic and chosen to make the arithmetic easy. Suppose the expectation estimates are 0.70, 0.50, and 0.30 at noise factors 1, 3, and 5. A straight-line fit has an intercept of 0.80. The code below demonstrates only the fitting step. It does not generate noise-amplified circuits or run on quantum hardware.

import numpy as np

noise_factors = np.array([1.0, 3.0, 5.0])
expectations = np.array([0.70, 0.50, 0.30])

slope, intercept = np.polyfit(noise_factors, expectations, 1)
zero_noise_estimate = intercept
print(round(zero_noise_estimate, 3))

Expected output:

0.8

In a real workflow, the expectation values come from executing the original and folded circuits. The noise factors describe an amplification strategy, not a promise that every physical error scales perfectly by that number. The choice of fit also matters. A linear model is easy to inspect, while other workflows may use exponential or higher-order fits.

What ZNE can and cannot promise

ZNE often improves expectation estimates, but it is not guaranteed to be unbiased. The extrapolation can be wrong if the assumed curve does not describe the data or if noise amplification is inaccurate. The extra circuits also cost time and shots. For example, sampling three noise factors can mean roughly three times the circuit executions before accounting for other overheads.

The practical lesson is to compare the mitigated result with the raw estimate, keep track of uncertainty, and spend the additional sampling budget only when the observable matters enough. ZNE is most useful when you need an expectation value and can tolerate the cost of several related circuit runs.

Further reading