⚛️ quantum physics

Comment on "Beyond-classical computation in quantum simulation"

This comment challenges a recent finding that Neural Quantum States (NQS) are inferior to quantum processors in simulating quantum annealing, arguing that NQS can actually achieve competitive accuracy when properly accounting for Monte-Carlo noise and large autocorrelation times.

Original authors: Wladislaw Krinitsin, Nikita Alert, Matteo Rizzi, Markus Schmitt

Published 2026-07-13

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Wladislaw Krinitsin, Nikita Alert, Matteo Rizzi, Markus Schmitt

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 predict the final pose of a giant, tangled ball of yarn after you've shaken it for a specific amount of time. In the world of quantum physics, this "yarn" is a collection of tiny magnets (spins) on a grid, and the "shaking" is a process called annealing. Recently, a team of researchers used a super-advanced quantum computer (a QPU) to shake this yarn and claimed that a popular computer simulation method called Neural Quantum States (NQS) couldn't keep up. They said the simulation was too messy to match the quantum computer's accuracy.

But in this new paper, the authors take a second look and suggest the simulation wasn't actually losing the race—it just got confused by the noise in its own measurements.

Here is what they found:

The "Static" in the Signal
The researchers focused on a specific grid size of $L = 8$ and an annealing time of $t_a = 7$ ns. They noticed that when the NQS method tried to measure how far apart the magnets were "talking" to each other (correlations), the results would flatten out and stop getting better once the distance got too large.

Think of it like trying to hear a whisper across a crowded stadium. If you stand close to the speaker, you hear them clearly. But if you stand far away, the background noise of the crowd drowns out the whisper. The authors realized that the NQS method wasn't failing to understand the physics; it was just hitting a wall of "statistical noise." Because the method uses a sampling technique called Monte Carlo (which is like taking a million snapshots to guess the average), there is a natural limit to how quiet the background noise can get. With a budget of $N_{MC} = 10^6$ samples, that noise floor sits right around $10^{-3}$ . Once the actual signal (the correlation between distant magnets) dropped below this level, the noise drowned it out, making the simulation look like it had stopped working.

The Real Race
To prove this, the authors ran a new test. They treated the "noise" and the "real error" as two separate things. They found that if you look at the data carefully, the NQS method actually converges perfectly to the true answer for short distances, just like a high-quality reference simulation (MPS) does. The messiness only happened at long distances where the signal was too weak to be heard over the static.

When they mathematically stripped away the noise to see the "pure" performance, the results were surprising. The NQS method didn't just match the quantum processor; for this specific setup, it actually produced a lower error rate than the quantum processor itself.

What This Means (and What It Doesn't)
The authors are careful to say this isn't a total victory for classical computers over quantum ones. They point out that their analysis was limited to a simple square grid and short times ( $t_a \le 7$ ns). The original study looked at more complex shapes and longer times (up to $t_a = 20$ ns), where things might get much harder. They also noted that if the "yarn" is made of a rougher, more chaotic material (using bimodal couplings), classical methods might struggle more.

So, the main takeaway is a suggestion: The claim that Neural Quantum States fail to compete with quantum processors might be an illusion caused by not accounting for the "static" in the measurement. When you clean up that static, the simulation holds its own, and in this specific case, it even outperformed the quantum hardware. The authors hope this encourages more experiments to see where these classical methods truly shine and where they might still need a quantum boost.

Technical Summary: "Beyond-classical computation in quantum simulation"

Problem Statement
This comment addresses a recent benchmark study [1] (King et al., Science 388, 199-204, 2025) which compared various classical numerical methods against a superconducting quantum processor unit (QPU) in simulating quantum annealing dynamics for a spin glass Hamiltonian. The original study concluded that Neural Quantum States (NQS), a variational ansatz based on artificial neural networks, failed to achieve the same level of accuracy as the QPU for specific instances (specifically on an $L=8$ two-dimensional square lattice with an annealing time of $t_a = 7$ ns). The authors of this comment argue that the reported failure of NQS may be an artifact of statistical sampling limitations rather than a fundamental limitation of the method itself.

Methodology
The authors revisit the specific disorder realization and parameters used in the original study, focusing on the time-dependent variational principle (TDVP) evolution of NQS. Their analysis involves:

Re-evaluation of Correlations: Instead of relying solely on the aggregate correlation error metric ( $\epsilon_c$ ) from the original paper, they analyze the distance-resolved magnitude of two-point correlations, $C(d) = \langle |c_{i,j}| \rangle$ , where $c_{i,j} = \langle \sigma^z_i \sigma^z_j \rangle$ .
Statistical Noise Analysis: They investigate the impact of Monte Carlo (MC) sampling noise on the estimation of correlations. They derive that for long-range correlations where the true value is small, the MC noise scales as $\sigma_{MC} \approx 1/\sqrt{N_{MC}}$ . With the original sampling budget of $N_{MC} = 10^6$ , the noise floor is approximately $10^{-3}$ .
Error Decomposition: The authors propose an analytical model for the correlation error $\epsilon_c(N_{MC})$ as a function of the sampling budget:
$\epsilon_c(N_{MC}) = \sqrt{\epsilon_{sys} + \frac{\epsilon_{stat}}{N_{MC}}}$
Here, $\epsilon_{sys}$ represents the systematic, method-specific error, and $\epsilon_{stat}$ represents the statistical error contribution.
Fitting and Extrapolation: By varying the number of samples and fitting the data to the model above, they extrapolate the correlation error to the infinite sampling limit ( $N_{MC} \to \infty$ ) to isolate the systematic error inherent to the NQS method.

Key Results

Plateauing Behavior: The authors observe that NQS results for distance-resolved correlations plateau at $C(d) \approx 10^{-3}$ for large distances. They demonstrate that this plateau coincides exactly with the expected magnitude of MC noise for $N_{MC} = 10^6$ , indicating that the "failure" to capture long-range correlations is due to the signal being drowned out by statistical noise, not a failure of the variational ansatz.
Convergence Analysis: When analyzing correlations grouped by distance classes, short-range correlations converge rapidly to the Matrix Product State (MPS) ground-truth reference. In contrast, long-range correlations follow the $1/\sqrt{N_{MC}}$ scaling expected of statistical noise, confirming they have not converged with the original sampling budget.
Revised Accuracy: Upon fitting the error model and extrapolating to the infinite sampling limit, the systematic error ( $\epsilon_{sys}$ ) of the NQS method is found to be lower than the error reported for the QPU in the original study. Furthermore, the extrapolated NQS error is lower than the NQS error reported in the original paper (which was based on finite sampling).
Autocorrelation Considerations: The authors note that for longer annealing times (e.g., $t_a = 7$ ns), the final state develops a highly peaked Born distribution, leading to long autocorrelation times in Markov chain MC sampling. While this complicates sampling, the magnitude of correlations increases, reducing the relative error at fixed $N_{MC}$ .

Significance and Claims
The paper claims to contextualize the findings of King et al. by demonstrating that the apparent underperformance of NQS in the original benchmark was significantly influenced by the sampling budget and the resulting statistical noise, particularly for vanishing long-range correlations.

Competitive Performance: The authors conclude that NQS can provide competitive results for quantum annealing protocols. In the specific instance analyzed, the NQS method, when accounting for sampling noise, yields a correlation error that surpasses that of the QPU.
Limitations and Scope: The authors maintain a modest scope, acknowledging that their analysis is restricted to the simplest lattice geometry ( $L=8$ ) and annealing times up to $t_a = 7$ ns. They note that the original study covered more complex geometries and longer times ( $t_a = 20$ ns) and different coupling distributions (bimodal vs. uniform).
Future Direction: Rather than claiming universal superiority, the paper motivates further numerical experiments to better understand the strengths and limitations of NQS across a broader scope, citing previous work on larger diamond-structured lattices [4] as a precedent for successful application.

In summary, the paper argues that NQS is not inherently inferior to QPUs for the studied problem but requires careful accounting for Monte Carlo noise and autocorrelation times to reveal its true systematic accuracy.

Technical Summary: "Beyond-classical computation in quantum simulation"

More like this