⚛️ quantum physics

Sufficient quantum provenance: retained fields and certified recording precision

This paper introduces a framework for defining "sufficient quantum provenance" by deriving circuit-based certificates that determine the minimum execution fields and recording precision required to guarantee a bounded Hellinger distance between outcome distributions, thereby ensuring reliable comparison of quantum computations without full distribution simulation.

Original authors: George Kourousias, Sergio Carrato, Roberto Pugliese, Aljoša Hafner, Francesco Guzzi, Matteo Billè, Fulvio Billè, Panagiotis Dimitrakis

Published 2026-10-08

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: George Kourousias, Sergio Carrato, Roberto Pugliese, Aljoša Hafner, Francesco Guzzi, Matteo Billè, Fulvio Billè, Panagiotis Dimitrakis

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

When scientists build a quantum computer, they are not just running a program; they are conducting a delicate experiment where the environment, the machine's internal quirks, and the way the program is translated into machine code all shape the final result. Two researchers might run the exact same logical task on two different machines, or even the same machine at different times, and get slightly different patterns of answers. In the world of classical computing, we often assume that if the code is the same, the result is the same. But in quantum computing, the path taken to get the answer matters just as much as the answer itself. To trust a quantum result, we need to know exactly what conditions were present during the run. The question becomes: how much detail do we need to write down to prove that two runs are comparable? If we leave out a small detail, like a tiny shift in the machine's noise or a slight change in how the computer routes information, could we be comparing two completely different realities?

A team of researchers from Italy and Greece has tackled this problem by creating a new way to decide which details are essential and which can be safely ignored. They asked a simple but profound question: if two quantum experiments look the same on paper, how close must their actual outcomes be to count as a match? To answer this, they developed a method to measure the "distance" between the probability patterns of two experiments. If this distance is small enough, the experiments are considered equivalent for the purpose of the study. Their work provides a certificate—a set of rules—that tells scientists exactly how many digits of precision they need to record for things like noise levels and rotation angles to guarantee that their comparisons are valid.

The researchers tested their ideas on small quantum circuits, specifically looking at a type of algorithm known as QAOA, which is used to solve optimization problems. They simulated thousands of different scenarios, changing the layout of the computer chips, the strength of the noise, and the angles at which the quantum bits were rotated. They found that simply recording the broad strokes of an experiment is often not enough. For instance, in one specific test case, two runs had identical records for the layout and the noise strength, but they differed slightly in how the computer routed data and in a tiny bias in the rotation angles. Even though these differences seemed minor, they caused the final outcomes to drift apart by more than the allowed limit. This meant that to guarantee a fair comparison, the researchers had to record those specific routing details and angle biases, not just the general setup.

To solve this systematically, the team created a mathematical framework that acts like a sieve. It filters out the details that don't matter and keeps only the ones that change the outcome. They discovered that for a specific four-qubit circuit, a record consisting of just thirteen bits of information was enough to guarantee that the difference between any two matching runs would stay below a strict threshold. This is a surprisingly small amount of data, but it is precise. The team showed that if you try to save space by recording fewer details, you risk hiding a difference that makes the comparison invalid. They also found that the way you group these details matters; recording the noise and the angles together in a specific way is more efficient than recording them separately, because the errors in quantum systems often combine in complex ways that simple addition cannot predict.

The study went beyond simulations to test these ideas on real hardware. The researchers ran a new experiment on a six-qubit processor, using a task that involved refining the boundaries of a synthetic image. They compared the raw data coming directly from the machine against the data after it had been processed by a decoder. They found that for the raw data, they needed to record two specific labels to ensure the runs were comparable. However, once the data was processed by the decoder, the complexity dropped, and only one label was needed. This confirmed that the level of detail required depends entirely on what you are looking at: the raw machine behavior or the final, interpreted result. The experiment proved that their method works in the real world, identifying exactly which pieces of information are necessary to trust a comparison.

The researchers also addressed what happens when we try to be too clever by omitting details. They showed that a common shortcut—checking if removing one detail at a time causes a problem—is not enough. Sometimes, removing two details that seem harmless on their own creates a large error when they are missing together. Their method catches these hidden traps by looking at the worst-case scenario for every possible combination of missing information. This ensures that the record is robust. They also clarified that this method is about ensuring two runs are comparable, not about proving that the result is the "correct" answer to a problem. Two runs can be perfectly comparable and still both be wrong if the machine is biased in a certain way. The goal is simply to know when two experiments are talking about the same thing.

In the end, this work provides a clear map for the future of quantum research. It tells scientists that they do not need to record every single scrap of data from an experiment to make it reproducible, but they must be very careful about which scraps they keep. By using a precise, mathematically derived certificate, they can define a "sufficient record" that guarantees agreement within a declared tolerance. This means that as quantum computers grow larger and more complex, researchers will have a reliable way to know if their results are consistent, without being buried under an impossible amount of data. The study establishes that with the right level of precision, we can trust that a quantum experiment today can be fairly compared to one tomorrow, even as the machines and the noise around them change.

Technical Summary: Sufficient Quantum Provenance

Problem Statement
Physical implementations of identical quantum tasks often yield different outcome laws due to variations in compilation, hardware noise, and measurement conditions. While existing provenance frameworks (e.g., PROV, QProv) effectively preserve execution context, they do not specify the resolution at which context must be retained to guarantee that two executions are comparable within a declared tolerance. The core challenge is determining which execution details (fields) must be recorded and to what precision (for continuous fields) to ensure that the Hellinger distance between the underlying outcome laws of any two executions sharing the same record does not exceed a specified threshold $\epsilon$ .

Methodology
The authors propose a framework to define and certify a "sufficient record" based on the largest Hellinger distance ( $R_\Omega(S)$ ) between outcome laws that share a specific retained field set $S$ . The methodology is divided into two complementary approaches:

Finite-Class Analysis (Separation Certificates):
- For a finite set of declared contexts $\Omega$ , the problem is formulated as a weighted hitting-set problem. A field set $S$ is $\epsilon$ -sufficient if and only if it intersects the set of distinguishing fields for every pair of contexts whose outcome laws violate the tolerance $\epsilon$ .
- The authors evaluate various omission strategies (e.g., reference-mixture rules, singleton-threshold rules) against exhaustive enumeration of field sets to identify minimum-cost sufficient records.
- This approach uses exact density-matrix propagation to compute raw outcome laws without shot sampling, ensuring the analysis targets the underlying laws rather than statistical estimates.
Continuous Precision Certification (Channel-Derived Bounds):
- For continuous fields (e.g., noise strength, angle bias), exact retention is impossible in finite digital records. The authors derive a certificate based on the quantum channel description rather than exhaustive simulation.
- Proposition 1 establishes a joint stochastic-coherent bound: The squared Hellinger distance between two executions is bounded by $1 - a_{st} \cos(\min\{D, \pi/2\})$ , where $a_{st}$ is the affinity of independent latent noise flags (stochastic component) and $D$ is the accumulated difference in coherent rotations.
- This bound allows the calculation of a "whole-cell" risk for a binning of continuous parameters without simulating the outcome distributions. It connects recording precision to the length of a future measurement transcript via the exact equivalence between Hellinger distance and affinity.

Key Contributions

Definition of Sufficient Retention: The paper defines a sufficient record not as a minimal set of fields for a single execution, but as a set that guarantees agreement within tolerance for an entire class of executions.
Analytic Channel Certificate: A novel bound that separates independent stochastic noise (via affinity) from coherent rotation differences (via conditional quantum fidelity). This allows for precision guarantees over continuous parameter spaces without exhaustive simulation.
Separation Witnesses: A method to certify the necessity of specific fields by identifying "violating pairs" that become indistinguishable if a specific field is omitted.
Operational Distinction: The framework distinguishes between comparison (agreement between implementations) and accuracy (agreement with an ideal target), noting that a sufficient record guarantees the former but not necessarily the latter.

Results

Finite-Class Study (QAOA and HEA):
- Using four-qubit QAOA (depth 1 and 2) and hardware-efficient ansatz (HEA) circuits, the authors evaluated 64 field sets across 324 contexts per design.
- QAOA-1: At $\epsilon=0.05$ , there are two tied minimum-cost sufficient sets: $\{L, R, G, M\}$ and $\{L, G, M, A\}$ . However, when extended to the full family of 28 archived QAOA instances, the set $\{L, G, M, A\}$ is the unique minimum-cost sufficient set.
- Failure of Heuristics: Simple omission rules failed to guarantee the global tolerance. The reference-mixture rule retained only readout (for ansatz) or gate noise (for QAOA), resulting in record-cell diameters of 0.1215–0.2924, far exceeding the 0.05 tolerance. The singleton-threshold rule was closer but still failed for QAOA-1: omitting routing alone (risk 0.0417) and omitting angle bias alone (risk 0.0497) both passed individual checks, but their joint omission resulted in a risk of 0.0514, violating the tolerance.
- Cost: The optimal record required 4 fields, whereas the reference-mixture heuristic required only 1 field but failed the tolerance guarantee.
Continuous Precision (QAOA-1):
- For a continuous noise box (Gate noise $G$ , Readout $M$ , Angle bias $A$ ), a 13-bit fixed-length record (using specific bin allocations for layout, routing, and noise parameters) guarantees a Hellinger distance below 0.05 uniformly over programmed angles.
- The bound scales as $O(\log N)$ for an $N$ -shot transcript, providing a channel-envelope rate.
Hardware Confirmation (Six-Qubit CT Task):
- A fresh experiment on an IBM Kingston processor (Heron R2) tested a synthetic image-segmentation task.
- Raw Laws: Two labels (e.g., layout) were necessary and sufficient to meet $\epsilon=0.15$ .
- Decoded Laws: After applying a fixed decoder, only one label (empty record) was sufficient.
- This demonstrates that application-specific decoding can reduce the minimum comparison-label budget, a finding conditional on the specific finite context class and sampling model.

Significance and Claims
The paper claims to provide a rigorous, operational framework for determining "what matters" in quantum execution records.

Guarantee over Simulation: The continuous certificate offers a way to bound outcome-law differences across a continuum of parameters without simulating every possible outcome distribution, replacing exhaustive checks with a finite precision budget derived from channel properties.
Contextual Necessity: The study demonstrates that the necessity of specific fields (e.g., routing vs. angle bias) is not universal but depends on the specific circuit family and the declared tolerance.
Modest Scope: The authors explicitly state that their results are conditional on stationary independent shots and specific physical models. The analytic certificate does not validate the model itself; if unmodeled correlated environments exist, the certificate may fail. The framework specifies the evidence required to support a comparison guarantee but does not certify the accuracy of the computation relative to an ideal target.
Separation of Concerns: The work clarifies that task definitions and evidential archives are separate from the optimization of the comparison record. A sufficient record ensures comparability, but the full archive remains necessary for investigating how a result was produced.

Technical Summary: Sufficient Quantum Provenance

More like this