⚛️ quantum physics

Benchmarking Error Mitigation: Artefactual Improvements in Zero-Noise Extrapolation

This paper reveals that Richardson Zero-Noise Extrapolation can produce misleading, artefactual improvements on current quantum hardware by collapsing into a fixed rescaling of noisy data rather than reflecting true physics, and proposes specific negative controls and a reporting checklist to prevent such invalid benchmarking.

Original authors: Dominik Köster, Wolfgang Mauerer

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dominik Köster, Wolfgang Mauerer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to hear a faint whisper in a very noisy room. You have a special trick called Zero-Noise Extrapolation (ZNE). The idea is simple: you shout the same whisper louder and louder (amplifying the noise) to see how the sound changes, then you use a math formula to guess what the whisper sounded like when the room was perfectly quiet. It's like taking a blurry photo, zooming in until it's super blurry, and then using a computer program to guess what the original sharp image looked like.

For a long time, scientists thought this trick was a reliable way to fix errors in quantum computers. But this paper, written by Dominik Köster and Wolfgang Mauerer, pulls back the curtain to show a sneaky trap: sometimes, the math trick doesn't fix the noise; it just invents a fake improvement.

The "Horoscope" Effect

The authors call this a "horoscope effect." Think of it like a horoscope that says, "You will have a great day!" It sounds nice, but it's just a generic guess that happens to feel right, not a real prediction based on your actual life.

In the quantum world, when the noise gets too loud, the signal (the real information) gets destroyed. It hits a "floor" where the computer can't hear anything anymore. When this happens, the math formula doesn't stop working; instead, it starts doing something silly. It takes the one noisy measurement it has left and simply rescales it (multiplies it by a fixed number) to make it look like it got better.

The paper proves that this "improvement" is an artefact. It's not real. It's a mathematical ghost. The formula says, "Look, we improved by 21%!" but in reality, the computer just guessed a number that happened to look good.

The "Garbage" Test

To prove this isn't just a fluke, the researchers played a clever trick. They created a fake version of their noise-amplifying method called "garbage-folding."

Imagine you are trying to clean a dirty window.

Genuine Folding: You actually wipe the window with a cloth.
Garbage Folding: You take a piece of trash, crumple it up, and press it against the window. It adds the same amount of "stuff" (cost) as the cloth, but it doesn't clean anything; it just makes the window dirtier.

The researchers ran both methods on real quantum hardware (a machine called IQM Euro-Q-Exa).

When they used the genuine method on deep circuits, the signal got destroyed, and the math produced a fake improvement.
When they used the garbage method, the signal was also destroyed, and the math produced an even bigger fake improvement.

This is the smoking gun. If the "garbage" method (which does nothing useful) claims to fix the problem better than the real method, then the "fix" isn't real. It's just the math playing tricks. In their tests, the garbage method produced an apparent improvement ratio of 2.77, while the real method was only 0.99. The garbage actually looked "better" because it destroyed the signal more effectively, triggering the math glitch.

The "Negative Probability" Alarm

How do you know if you've fallen into this trap? The authors found a free, instant check that uses data you already have.

When the math goes wrong, it starts predicting negative probabilities. In the real world, a probability can't be negative (you can't have a -50% chance of rain). But when the signal is destroyed, the math formula gets confused and spits out numbers like -0.039.

The researchers checked this on a 6-qubit Grover circuit.

With the real method, 0 out of 64 states had negative probabilities.
With the garbage method, 29 out of 64 states had negative probabilities.

This negative number is a giant red flag. It's a "zero-cost" alarm that tells you, "Hey, the signal is gone, and this improvement is fake."

What They Found on Real Hardware

The paper isn't just a simulation; they tested this on a real quantum computer. They ran a 4-qubit Quantum Trotter Circuit and increased its depth (making it more complex).

At low depth, the signal was clear, and the math worked perfectly.
At higher depths (specifically depth 3 and depth 5), the signal collapsed.
At depth 3, the math claimed an improvement that overshoot the ideal value by 21% (predicting a value of 1.16 when the true ideal was 0.96).

This proves that even with standard, trusted settings (using scale factors of 1, 3, and 5), the method can fail if the circuit is too deep or the noise is too high.

The Takeaway

The paper doesn't say ZNE is useless. It says we need to be careful. Just because a number looks like an improvement doesn't mean it's real.

The authors suggest a simple checklist for anyone reporting these results:

Check the signal: Did the signal hit the "floor" (become too noisy to hear) at the higher noise levels?
Check for negatives: Did any of the probability estimates turn negative? If yes, the result is likely an artefact.
Check the overshoot: Did the result go higher than the maximum possible value? If yes, it's a fake improvement.

In short, if your quantum computer starts telling you it's solved a problem with a "magic" improvement that looks too good to be true, check the negative probabilities. You might just be reading a horoscope instead of a scientific result.

Technical Summary: Benchmarking Error Mitigation: Artefactual Improvements in Zero-Noise Extrapolation

Problem Statement
Reliable benchmarking of Quantum Error Mitigation (QEM) requires distinguishing genuine physical improvements from mathematical artefacts introduced during post-processing. The paper focuses on Richardson Zero-Noise Extrapolation (ZNE), a widely used technique where expectation values at different noise amplification levels ( $\lambda_k$ ) are fitted to a polynomial to estimate the zero-noise limit ( $\hat{E}(0)$ ).

The core problem identified is a specific failure mode: when noise amplification drives the measured signal below the usable regime (approaching the "floor" of the observable, e.g., $f=0$ for parity or $f=1/2^N$ for probabilities), the extrapolation no longer reflects the underlying physics. Instead, it collapses into a deterministic rescaling of a single noisy measurement ( $E(\lambda_1)$ ). This creates a "horoscope effect"—an apparent improvement that is independent of the actual noise amplification mechanism and can result in unphysical overshoots (values exceeding the ideal) or bogus improvements that are indistinguishable from genuine mitigation without further diagnostics.

Methodology
The authors employ a combination of theoretical derivation, simulation, and empirical hardware experiments to characterize and validate this failure mode.

Theoretical Derivation: The paper derives a closed-form characterization (Equation 2) showing that if expectation values for amplified noise levels ( $k > 1$ ) collapse to the floor ( $f$ ), the Richardson estimate $\hat{E}(0)$ becomes a linear function of only the first measurement $E(\lambda_1)$ and the floor $f$ . Specifically, $\hat{E}(0) = c_1 E(\lambda_1) + (1-c_1)f$ . Since $c_1 > 1$ for standard scale factors (e.g., $\{1, 3, 5\}$ ), any signal near the floor is rescaled upward, creating an artificial improvement.
Signal-Retention Taxonomy: The authors define three regimes based on the relationship between the signal and the observable floor:
- Signal Retained: The signal remains well above the floor; extrapolation is valid.
- Signal Decayed: The signal approaches the floor; recovery degrades.
- Signal Destroyed: The signal collapses to or below the floor; extrapolation degenerates into the artefact described above.
Garbage-Folding Negative Control: To isolate the effect of signal destruction from the amplification mechanism, the authors introduce a "garbage-folding" technique. This involves inserting gate sequences that do not reduce to the identity (unlike standard folding) but match the gate count and residual error rate of genuine folding. This maximizes signal destruction at $\lambda > 1$ while maintaining identical computational costs.
Hardware Validation: Experiments were conducted on the 54-qubit IQM Euro-Q-Exa machine using a 4-qubit Quantum Trotter Circuit (QTC) and a 6-qubit Grover circuit. The study utilized scale factors $\{1, 3, 5\}$ and compared genuine folding against garbage folding and identity folding.
Zero-Cost Diagnostic: The authors propose a diagnostic check based on the per-basis-state probability distribution. Since Richardson extrapolation is a linear combination of probabilities, a valid distribution must have non-negative probabilities for all states. The authors define a "negative-probability weight" ( $W_{neg}$ ) to quantify violations of this physical constraint.

Key Results

Empirical Collapse: On the IQM Euro-Q-Exa hardware, as circuit depth increased, the amplified parity values collapsed to the floor ( $f=0$ ). At this point, the Richardson estimate $\hat{E}(0)$ decoupled from the true ideal value and tracked the theoretical rescaling prediction ( $\frac{15}{8}E(\lambda_1)$ ).
Artefactual Overshoot: In the signal-destroyed regime, the method produced a 21% overshoot of the ideal value ( $\hat{E}=1.16$ vs $E_{ideal}=0.96$ ) despite the signal being effectively zero.
Garbage-Folding Superiority: In simulations and hardware runs, "garbage folding" (which carries no usable signal) produced larger apparent improvements than genuine folding. For instance, on the Grover circuit, garbage folding yielded a recovery ratio $\rho \approx 0.56$ compared to $0.44$ for genuine folding. In the signal-retained regime, garbage folding produced extreme unphysical values ( $\rho = 16$ and $125$).
Diagnostic Efficacy: The negative-probability weight ( $W_{neg}$ ) successfully distinguished between valid and artefactual regimes. Valid folding resulted in $W_{neg} \approx 0.02\text{--}0.04$ , while garbage folding in the destroyed regime resulted in $W_{neg} \approx 0.74$ . The two distributions did not overlap (AUC = 1.0), making $W_{neg}$ a robust, zero-cost indicator of the artefact.

Key Contributions

Closed-Form Characterization: A mathematical proof demonstrating that Richardson ZNE degenerates to deterministic rescaling when $E(\lambda_{>1})$ reaches the observable floor, rendering the reported improvement meaningless.
Signal-Retention Taxonomy: A framework allowing benchmarks to determine their operational regime (retained, decayed, or destroyed) using their own per- $\lambda$ data.
Matched-Cost Negative Control: The introduction of "garbage-folding" to demonstrate that larger reported improvements can be manufactured solely through signal destruction, proving that the magnitude of improvement is not evidence of correctness.
Zero-Cost Diagnostic and Checklist: A practical method to flag artefacts using existing data (per-state counts) via the $W_{neg}$ metric, alongside a reporting checklist for ZNE benchmarks.

Significance and Claims
The paper argues that the validity of many empirical evaluations in quantum computing is threatened by this rarely considered failure mode. The authors claim that:

The "horoscope effect" (artefactual improvement) is distinct from the well-known ill-conditioning caused by closely spaced scale factors; it affects even variance-optimal scale factor sets like $\{1, 3, 5\}$ once the signal is destroyed.
Reported improvements in the signal-destroyed regime are not evidence of successful mitigation but are mathematical artefacts of the post-processing arithmetic.
A credible benchmark must ensure that improvements are attributable to the noise amplification mechanism rather than extrapolation arithmetic.
The proposed checklist (reporting signal retention status, $W_{neg}$ , and physical bounds) allows the community to distinguish between what a ZNE pipeline claims to measure and what it actually measures, without requiring additional computational resources.

The authors maintain a modest scope, noting that while hardware evidence is currently limited to one device and specific circuit families, the theoretical derivation depends only on general properties of linear extrapolators and observable floors, suggesting the failure mode is generalizable across devices and circuit depths.