Quantum error correction can constantly recalibrate a processor

量子纠错技术可实现处理器的持续校准

There are some obvious big picture issues that stand between us and useful quantum computing. Issues like whether we can make enough high-quality hardware qubits to connect into the error-corrected logical qubits we need, and how we generate the states needed to perform universal computation on those logical qubits. But there are also many less prominent challenges that will need to be solved before we can perform calculations. 在通往实用量子计算的道路上，存在一些显而易见的宏观障碍。例如，我们能否制造出足够多高质量的硬件量子比特，并将它们连接成所需的纠错逻辑量子比特；以及我们如何生成在这些逻辑量子比特上执行通用计算所需的状态。但除此之外，在能够真正进行计算之前，还有许多不太引人注目的挑战需要解决。

One of those challenges, which only affects some types of hardware, is calibration. For devices we manufacture, like superconducting qubits, there are always subtle variations among individual qubits. (This is not true when we use something like an atom to hold the qubit, but the lasers that control them can drift.) As a result, this hardware is put through a process called calibration, where we test different frequencies and amplitudes of the microwave pulses that control them to find the combination that produces the lowest error rates, and then save those settings for use in calculations. 其中一个仅影响特定硬件类型的挑战是“校准”。对于我们制造的设备（如超导量子比特），单个量子比特之间总是存在细微差异。（虽然使用原子等作为量子比特载体时情况并非如此，但控制它们的激光器可能会发生漂移。）因此，这些硬件必须经过一个称为“校准”的过程：我们测试控制它们的微波脉冲的不同频率和振幅，以找到能产生最低错误率的组合，然后保存这些设置以供计算使用。

However, you can’t perform the typical calibration process while you’re doing calculations, which means drift becomes an issue for long and complicated algorithms. Google, though, has figured out that it’s possible to do calibration using the same data that’s used for error correction. 然而，在进行计算时无法执行常规的校准过程，这意味着对于冗长而复杂的算法来说，漂移会成为一个问题。不过，谷歌已经发现，利用用于纠错的相同数据进行校准是可行的。

Reinforcement learning

强化学习

The hardware that Google and a number of other companies rely on are transmons. They consist of a loop of superconducting wire connected to a resonator, and they’re controlled by pulses of microwave photons. Those pulses are controlled by hardware that is kept outside of the refrigeration, including classical computers and the microwave sources they control. This hardware is used to test different combinations of wavelengths and amplitudes during calibration. 谷歌和其他多家公司所依赖的硬件是超导量子比特（transmons）。它们由连接到谐振器的超导线环组成，并由微波光子脉冲控制。这些脉冲由制冷机外部的硬件控制，包括经典计算机及其控制的微波源。在校准过程中，这些硬件被用来测试波长和振幅的不同组合。

This equipment can also drift from its initial settings due to random factors, such as the hardware heating up as it’s used. And that could be an issue for the sorts of complicated algorithms we ultimately intend to run on quantum computers, like those that could crack current encryption. Currently, if the system shows signs of drifting away from calibration, Google says that it simply stops the computations and recalibrates. However, that is not going to be an option partway through a complicated calculation. 由于硬件在使用过程中发热等随机因素，这些设备也可能偏离其初始设置。对于我们最终打算在量子计算机上运行的复杂算法（例如可能破解当前加密技术的算法）来说，这可能是一个问题。目前，如果系统显示出偏离校准的迹象，谷歌表示只能停止计算并重新校准。然而，在复杂的计算过程中，这种方法是行不通的。

These computations will be taking place using error-corrected qubits, in which measurements on a subset of the hardware qubits are used to detect and characterize any errors that occur on the ones that hold the data. As the Google researchers point out in their paper, some of the errors they’ll detect will be the product of calibration failures: “errors from imperfect calibrations produce detectable syndromes just like all other errors.” 这些计算将使用纠错量子比特进行，其中通过对部分硬件量子比特进行测量，来检测和表征数据存储量子比特上发生的任何错误。正如谷歌研究人员在论文中指出的那样，他们检测到的一些错误正是校准失败的产物：“来自不完美校准的错误会像所有其他错误一样产生可检测的征兆（syndromes）。”

In theory, we could use the same error detection to identify both random errors and those produced by calibration issues. The challenge is telling the two apart. The team’s solution? Reinforcement learning, in which the computer tries different configurations of the 1,000 or so control parameters it has access to, and scores their effectiveness at limiting errors. 理论上，我们可以使用相同的错误检测机制来识别随机错误和由校准问题引起的错误。挑战在于如何区分两者。团队的解决方案是什么？是强化学习。计算机尝试其可访问的约 1000 个控制参数的不同配置，并对其限制错误的效果进行评分。

“We deliberately apply small, simultaneous perturbations to all control parameters during the computation to explore the control space,” the team wrote. “These perturbations translate into subtle changes in the statistics of error-detection events.” Using that information, the system can infer how adjusting these parameters can minimize certain errors. If those errors start to show up, it can make the appropriate adjustments. And that can be done in parallel with the error detection and correction system that manages the logical qubit. “我们在计算过程中故意对所有控制参数施加微小的同步扰动，以探索控制空间，”团队写道。“这些扰动会转化为错误检测事件统计数据中的细微变化。”利用这些信息，系统可以推断出如何调整这些参数以最小化特定错误。如果这些错误开始出现，系统就能做出相应的调整。这可以与管理逻辑量子比特的纠错系统并行完成。

The system was put in charge of two logical qubits hosted on a calibrated system. The two were using different error correction schemes (a surface code and a color code). These were set in a specific state, and the error-correction system was then used with and without reinforcement-learning-driven corrections. Having the system active led to a 20 percent increase in the ability to detect and correct errors in the logical qubits. 该系统被用于管理托管在校准系统上的两个逻辑量子比特。这两个量子比特使用了不同的纠错方案（表面码和色码）。它们被设置为特定状态，然后分别在有和没有强化学习驱动校准的情况下使用纠错系统。结果显示，启用该系统后，逻辑量子比特检测和纠正错误的能力提高了 20%。

Going real time

实现实时化

The limitation of this approach is that it works only if the drift keeps the system reasonably close to the state the system was trained in. The corrections that might bring things back into alignment from one state might not be effective when the system’s in a significantly different state. The solution to this is to constantly re-evaluate the effectiveness of different changes. But this has an obvious problem: You can’t simply randomize all the potential control configurations in the middle of a calculation. 这种方法的局限性在于，它仅在漂移使系统保持在训练状态附近时才有效。在一种状态下能使系统恢复对齐的校准，在系统处于完全不同的状态时可能无效。解决办法是不断重新评估不同调整的有效性。但这有一个明显的问题：你不能在计算过程中简单地随机化所有潜在的控制配置。

Even with limited variation, the system will necessarily operate outside its optimal error correction. So, the question was whether the frequent sub-optimal error correction paid off by keeping drift from causing even larger problems. “The favourable resolution of the exploration–exploitation trade-off would mean that the aggregate performance of all sampled policy candidates, most of which are worse than [the optimal one], is still better than the performance without reinforcement learning steering,” the researchers write. 即使变化有限，系统也必然会在非最优纠错状态下运行。因此，问题在于频繁的次优纠错是否值得，因为它防止了漂移导致更大的问题。“探索与利用权衡的有利解决意味着，所有采样策略候选者的总体表现（其中大多数比最优策略差）仍然优于没有强化学习引导的表现，”研究人员写道。

Performing many simulations with a very small error-corrected qubit showed that the trade-off worked out, provided that drift was slow enough. The team showed that it could work in real time with a large error-corrected qubit, in which the reinforcement learning system had control over roughly 40,000 parameters. This is clearly not a solution for the present; we can only keep systems operating for long enough to perform relatively short, simple algorithms, so drift isn’t even a concern. Ultimately, our intention is to build hardware that can perform the sorts of calculations where issues like this will matter. 通过对极小的纠错量子比特进行多次模拟表明，只要漂移足够缓慢，这种权衡是可行的。该团队展示了它可以在大型纠错量子比特上实时工作，其中强化学习系统控制着大约 40,000 个参数。这显然不是当前的解决方案；我们目前只能让系统运行足够长的时间来执行相对简短、简单的算法，因此漂移甚至还不是一个问题。最终，我们的目标是构建能够执行此类问题至关重要的计算的硬件。