Shot- and Noise-Aware Pauli Correlation Encoding for Max-Cut Experiments on IBM Heron Quantum Processor

In the world of modern computing, some problems are so vast that checking every possible answer would take longer than the age of the universe. These are known as combinatorial optimization problems, where a computer must find the single best arrangement among a dizzying number of options. A classic example is the "maximum cut" problem: imagine a network of cities connected by roads, where each road has a weight representing its importance. The goal is to split the cities into two groups so that the total weight of the roads connecting the two groups is as large as possible. This task is notoriously difficult for classical computers, which often have to rely on clever shortcuts rather than perfect solutions.

Quantum computers have been proposed as a new way to tackle these challenges. Unlike classical machines that process information as bits of zero or one, quantum machines use qubits that can exist in complex states of both zero and one simultaneously. However, current quantum computers are imperfect. They are noisy, meaning their calculations are easily disturbed by the environment, and they can only run for a short time before errors accumulate. Because of these limitations, scientists use a method called variational quantum algorithms. In this approach, a quantum computer runs a specific circuit, measures the result, and a classical computer adjusts the circuit's settings to try and find a better answer. The process repeats, hoping to climb toward the best possible solution.

A recent study by Achraf Boussahi addresses a critical flaw in how these quantum experiments are currently designed. The research focuses on a specific technique called Pauli correlation encoding, which allows a quantum computer to solve problems with many variables using far fewer qubits than usual. The standard way to train these circuits involves a mathematical "sharpness" setting that controls how closely the computer mimics the real problem. Previous research suggested that making this setting as sharp as possible always led to better results, but those experiments were run in perfect, noise-free simulations. Boussahi's work reveals that this assumption breaks down when the experiment is performed on a real, imperfect machine. By running tests on an actual quantum processor, the study demonstrates that the best settings for a perfect simulation are often the worst settings for a real device, and it provides a new rule for tuning these machines to work effectively in the real world.

The core of the problem lies in how quantum computers measure their answers. To get a result, the machine must run its calculation many times, a process known as taking "shots." Each shot produces a slightly different outcome due to statistical noise, much like flipping a coin many times to see if it is fair. The more shots taken, the clearer the true answer becomes. The researchers discovered that the "sharpness" setting used to train the quantum circuit is secretly tied to this number of shots. If the sharpness is set too high, the circuit tries to find a solution that requires more shots than the machine can realistically provide. In a perfect simulation, where there is no limit on shots, a very sharp setting works well because the computer can resolve tiny details. But on a real device with a limited number of shots, a setting that is too sharp forces the computer to guess the signs of its answers based on noise, leading to random, useless results.

To prove this, the team trained quantum circuits on five different complex graphs, each with 800 variables, using a quantum processor with only 13 qubits. They compared the old, aggressive training schedules against a new, "shot-aware" schedule that adjusted the sharpness based on the actual number of shots available. The results were striking. On the real quantum processor, the old schedules performed poorly, often failing to do better than a random guess. In contrast, the new schedule, which was tuned to the specific limits of the machine, consistently outperformed every published method. In 150 direct comparisons across different graphs and shot counts, the new method won every time. It managed to find solutions that were significantly closer to the best possible answer, retaining up to 91.9 percent of the optimal cut value, whereas the previous best methods collapsed into randomness.

The study also uncovered how the physical noise of the machine affects the calculation. The researchers found that the noise in the processor acts like a dimmer switch, reducing the strength of the signals the computer reads. This effect is mathematically equivalent to having fewer shots than actually taken. By measuring this dimming effect on the specific machine they used, they could predict exactly how to adjust the sharpness setting to compensate. They developed a simple rule: set the sharpness based on the number of shots and the measured noise level. When they tested this rule on a new graph before even running the experiment, their predictions of the final result were accurate to within a tiny fraction of a percent. This ability to predict performance before running the code is a major step forward for using quantum computers practically.

Another crucial finding concerned the design of the quantum circuit itself. The best results in previous literature came from circuits that connected every qubit to every other qubit, a design that works well on machines with perfect connectivity. However, the real quantum processor used in this study has a specific layout where qubits are only connected to their immediate neighbors. When the researchers tried to run the "all-to-all" design on this machine, the computer had to shuffle data around using extra gates, which introduced so much noise that the final answer was no better than a random guess. By switching to a "brickwork" design that matched the machine's natural layout, they avoided these extra steps. This hardware-native circuit, combined with the new shot-aware training, allowed the machine to retain its signal strength and deliver high-quality results.

The research concludes that for quantum optimization to work on current hardware, scientists must stop treating the machine as a perfect simulator and start treating it as a noisy, shot-limited device. The old strategy of pushing for maximum sharpness is counterproductive in the real world. Instead, the key is to match the training schedule to the specific constraints of the hardware: the number of shots available and the level of noise present. By doing so, researchers can extract meaningful solutions from today's imperfect machines. The study used only a few minutes of actual quantum computing time to reach these conclusions, suggesting that with the right tuning, these devices are already capable of solving complex problems better than previously thought, provided we stop asking them to do the impossible and start asking them what they can actually do.