量子计算机还没到来，量子算法已投入使用

微软公司的研究人员最近已经用上了一种专门为目前还不存在的量子计算机设计的算法，以提高医学成像的速度和质量。

微软表示，这种技术或将能够改善对乳腺癌和其他一些疾病的治疗。比如它可以让医生几天内就确定肿瘤是否因为化疗而缩小了，而不用像以往一样等上几周甚至几个月。

量子计算机如果真被生产出来了，当代的所谓超级计算机跟它一比，立即就成了算盘一样的老古董。最近，有不少研究人员都在尝试用针对量子计算机设计的算法提高目前现有硬件的计算效率。比如利用量子算法管理电网负载，改善交通拥挤的城市的快递配送路线，控制投资组合的风险和回报等等。

提高医学成像的质量和效率

最近，微软与位于克里夫兰的凯斯西保留地大学的科学家一起展示了双方的合作成果。凯斯西保留地大学十分擅长一种叫做磁共振指纹（MRF）的技术。像更常见的磁共振成像（MRI）技术一样，MRF也是靠强大的磁场和无线电波来生成人体内部器官和软组织的影像的。不过传统的MRI技术只能够识别明暗区域，放射科医生必须加以一定的主观推断，而MRF则可以精确区分人体组织的类型，从而获得更详细和更容易解读的影像。

凯斯西保留地大学的MRF项目负责人马克·格里斯沃尔德将这项技术比作唱诗班的合唱，人体内的组织好比歌手。在传统的MRI技术中，整个唱诗班就好像在唱同一首歌，听众只能分辨出是否某人的声音是否比别人略大或略小，音调是否比别人略高或略低，是否有人跑了调。而有了MRF技术，整个唱诗班好像每个歌手都在各自唱一首不同的歌曲，听众就能够将某一首歌曲从其它声音中识别出来，并且用它定位到某一个歌手。

不过，配置一台扫描仪来找到某种特定的组织类型（也就是分离出某首特定的“歌曲”）是非常耗时的。但凯斯西保留地大学的科研人员发现，在微软量子算法的帮助下，他们只需要以前的三分之一到六分之一的时间就能够完成扫描，同时将扫描的精度提高了25%以上。格里斯沃尔德表示：“精确度的提高是非常重要的，因为它可以让我们看到组织中越来越小的变化。”

微软近年来一直在强调量子算法的潜力，这在一定程度上也是要给未来的量子计算机市场播下种子。同时微软也高度重视用量子算法编写的软件。毕竟微软的某些竞争对手已经搞出了量子计算机的原型机，而微软虽然在量子计算机领域也搞了好几年的研发，但迄今还没有搞出什么花哨的硬件可以当作噱头来展示。

量子计算的崛起

量子计算机利用了量子力学特性来表达和处理信息。在传统计算机中，信息是以二进制处理的，又称为比特，其值为0或1。每个比特的值独立于计算中使用的其他所有比特。而在量子计算机中，信息是用量子比特表示的。量子比特可以用任何数量的具有量子特性的现象来创建（比如电子的自旋或光子的偏振）。

与比特不同，量子比特可以同时表示0和1。在某些情况下，甚至可以是0和1之间的任何值。更重要的是，每个量子比特的值都会影响到系统中其他量子比特的值。因此，量子计算机不必像传统计算机一样按顺序运算，而是具备了瞬间完成运算的可能。这两种特点让量子计算机从理论上对传统计算机形成了巨大的优势——量子计算机每增加一个量子比特，其性能就会呈指数型增长，而非线性增长。一台足够大的量子计算机所能做的事情，就能够远超当今最大的超级计算机——比如找到更节能的化肥生产流程，或者破解全世界大多数数据的加密保护程序。

量子计算机一度只在科幻小说里存在。然而在2011年，加拿大公司D-Wave Systems推出了全球第一台商用量子计算机。（不过，该量子计算机只能用于计算某些数学问题的子集。）自此，IBM、谷歌和Rigetti Computing（加州伯克利的一家创业公司）等公司都研发了用于通用用途的量子计算机，顾客已经可以通过互联网访问它们。英特尔还推出了几款量子处理器，不过它们目前尚未被提供给商用客户。

到目前为止，这些量子计算机都还没有强大到可以做到传统计算机做不到的事情，不过很多人认为，谷歌可能已经接近所谓“量子时代”的门槛了。不过即便果真如此，现阶段对于大多数商业应用来说，它的量子计算机依然太小，计算也很容易出错。

量子竞赛

在过去一年里，中国的阿里巴巴宣布将制造量子处理器，亚马逊也悄悄聘请了一支量子计算专家团队，表明它也在从事量子计算机的研发。另外，从事量子计算机硬件研发的初创公司少说也得有六七家。

微软的首席执行官萨蒂亚·纳德拉已经将量子计算称为三大突破性技术之一——其他两个分别是增强现实技术和人工智能技术，这三大技术也对微软未来的发展至关重要。在他的领导下，微软对量子技术也算下了血本。它从全球各地招募了一支由物理学家、数学家、计算机科学家和工程师组成的团队，并任命了一名在工程领域最有经验的老将——Xbox游戏机和HoloLens混合现实头盔项目的负责人托德·霍姆达尔来负责量子计算项目。

不过，微软为量子计算机的量子比特选择了一种以前从未测试过的架构，该架构基于一种非常难以捉摸的亚原子粒子，物理学家直到2017年前，都不能100%的确定该粒子的存在。与IBM、谷歌和Rigetti公司的架构相比，微软使用的这些亚原子粒子组成了发辫状，这种形状应该会让它们更加稳定，从而更不容易受到周围电磁力的冲击干扰。这种冲击干扰会导致量子计算机产生计算错误，必须加以纠正。由于这种架构理论上的犯错率更低，微软的设计对商业应用来说应该是更安全的选择。但是，微软首先要证明的是，它可以可靠地创造这种辫状结构，然后用它们形成量子比特——不过这一点它至今还没有做到。

与此同时，微软也有一群数学家和计算机科学家正在研究针对量子计算机的编程方法。事实证明，一些利用量子计算机的奇异特性开发的算法，也可以在普通计算机上体现出很大的优势。

定制算法

格里斯沃尔德表示，MRF技术的难点，在于如何控制扫描仪传输的无线电脉冲的强度、频率和角度。找到正确的脉冲模式，是扫描仪识别不同组织类型的关键——用格里斯沃尔德的比喻来说，就是分离出唱诗班里每个歌手所唱的歌曲。他表明，有一种数学上的优化模式，可以让扫描仪仅提取某一固定的组织类型，精确度甚至可以达到提取单个细胞的级别。但是要找到这种模式，则涉及太多的变量，靠传统计算机的运算能力无法达到。因此，研究人员以往基本是完全靠猜测来计划每次扫描的脉冲模式。不过即便是用这种不完美的方法，MRF仍然能够得到比普通的MRI更详细的图像。

格里斯沃尔德表示，为了取得进一步的进展，他需要超越人的直觉。他的团队申请了一笔拨款，用于研究如何使用传统算法技术优化MRF扫描技术，但是申请却被拒绝了，理由是要解决这样一个数学难题根本就是不可能的。

后来格里斯沃尔德听说微软正在寻找合作伙伴，为量子算法创建类似的演示案例——微软曾经与凯斯西保留地大学的医学影像专家密切合作，对它开发的HoloLens增强现实眼镜进行测试。近20年来，格里斯沃尔德本人也一直在密切关注量子计算的发展，而且认识一些为微软工作的研究人员，他意识到，自己的机会可能来了。

参与了MRF项目的微软量子计算项目研究人员马蒂亚斯·特罗耶表示：“我们喜欢看似不可能的问题。”更重要的是，优化MRF算法的挑战虽然看似不可能，但是针对这种问题的量子算法却已经被设计出来了。

特罗耶表示，针对格里斯沃尔德提出的需求，现有的量子算法必须要做出一些调整。“我们想强调的是，要想真正充分利用量子优化器的功能，就必须定制一个专门的解决方案。”特罗耶表示，对于MRF来说，最难的部分就是从构建MRF图像所涉及的数千个变量中，找出算法应该优化哪些因素的子集。他表示，只要这样做了，“一开始不可能的事情就开始变为可能。”

他还表示，虽然在传统计算机上运行量子算法，可以显著提高MRF扫描的速度和精度，但如果是在一台足够大的量子计算机上运行，结果会更加令人振奋。“那样会快得多。”他说。

不过，特罗耶所说的那种量子计算机需要大约需要100万个逻辑量子比特。要想造出那么大的量子计算机，可能还得几十年甚至更久。（财富中文网）

译者：朴成奎

computer to enhance the speed and quality of medical imaging.

The advance may one day improve the treatment of breast cancer and other diseases, the company says. For instance, it might allow doctors to determine within days whether a tumor is shrinking in response to chemotherapy, rather than having to wait weeks or months.

The development is one of a number of recent cases in which researchers have used algorithms designed for future quantum computers, machines that would make today’s supercomputers look like abacuses, to improve calculations running on today’s existing hardware. Other examples include using quantum algorithms to find better ways to manage the load across an electrical grid, improve delivery routes in a crowded city, and control risks and returns in an investment portfolio.

Better medical scans, more quickly

In the most recent illustration, Microsoft worked with scientists at Case Western Reserve University in Cleveland, who specialize in a type of medical imaging called magnetic resonance fingerprinting (or MRF.) Like the more familiar magnetic resonance imaging (MRI), the technique uses powerful magnetic fields and radio waves to create images of internal organs and soft-tissue. But while traditional MRIs can only identify areas of light or dark, which a radiologist must then subjectively evaluate, MRF can differentiate precisely between tissue types, allowing for more detailed and interpretable images.

Mark Griswold, a pioneer in MRF at Case Western Reserve who led the project, likes to use the analogy of trying to listen to a choir, where the tissues in the body are the singers: With a conventional MRI, it is as though the entire choir is all singing the same song, and the listener can only determine if one singer is a bit louder or softer than others, a bit higher or lower pitched, and maybe if they are out of tune. With MRF, on the other hand, it is like listening to a choir in which each singer has his or her own unique song, and the listener is able to isolate that song from the other voices in the choir and use it to identify the singer.

Configuring a scanner to find a particular tissue type—to isolate those individual songs—is time-consuming. With help from Microsoft’s quantum algorithm, the Case researchers found they could produce the scans in one third to one sixth of the time it took previously, while simultaneously boosting the precision of the scans by more than 25%. “The increase in precision is really important because it allows us to see smaller and smaller changes in the tissue,” Griswold says.

Microsoft has been highlighting the potential of quantum algorithms in part to seed the market for its future quantum computer. But it has also been emphasizing its quantum-inspired software because, unlike some rivals, it doesn’t yet have any fancy quantum hardware to show off, despite years of development.

The rise of quantum computing

Quantum computers use quantum mechanical properties to represent and manipulate information. In a conventional computer, information is processed in a binary format called bits, which have a value of either 0 or 1. The value of each bit is independent from all the other bits being used in the calculation. In a quantum computer, information is represented using quantum bits, or qubits. These qubits can be created using any number of phenomena that have quantum properties (for instance, the spin of electrons or the polarization of photons).

Unlike bits, qubits can represent both a 0 and a 1 at the same time—or in some cases, any value between 0 and 1. What's more, the value of each qubit affects the value of other qubits in the system, opening the door to nearly instantaneous solutions instead of having to process information in a serial fashion. These two factors, in theory, give quantum computers an enormous advantage over conventional ones: Each additional qubit added to a quantum computer increase its power not linearly, but exponentially. A sufficiently large quantum computer ought to be able to do things that are beyond the ability of even today’s biggest supercomputers—like find much more energy efficient processes for manufacturing fertilizer or break the encryption that protects much of the world’s data.

Quantum computers were once the stuff of sci-fi novels. But in 2011, D-Wave Systems, a Canadian company, debuted the first commercially available quantum computer. (Its machine, however, can only be used for a certain sub-set of mathematical problems.) Since then, IBM, Google, and Rigetti Computing, a Berkeley, Calif.-based startup, have all built more general-purpose quantum computers that customers can access over the Internet. Meanwhile, Intel has unveiled quantum processors, although these are not yet available to commercial customers.

So far, none of these quantum computers are powerful enough to do something a conventional computer can’t, although it is believed Google may be close to crossing this threshold, which is known as “quantum supremacy.” Even when that happens, the quantum machines will still be too small and their calculations too prone to errors to be useful for most commercial applications.

The corporate quantum race

In the past year, Chinese company Alibaba has announced that it would build a quantum processor, Amazon has quietly hired a team of quantum computing experts, signaling it too may be building a machine, while at least a half dozen startups are also working on quantum hardware.

At Microsoft, CEO Satya Nadella has described quantum computing as one of three groundbreaking technologies —along with augmented reality and artificial intelligence—that will be essential to the company’s future. Under his leadership, the company has made a big bet on quantum: hiring a team of physicists, mathematicians, computer scientists and engineers from around the globe and placing one of its most experienced engineering executives, Todd Holmdahl, a veteran of both the Xbox game console and the HoloLens mixed reality headset, in charge of the effort.

The company has chosen an untested architecture for the qubits of its quantum computer, based on an elusive sub-atomic particle physicists weren’t even 100% sure existed until 2017. Those sub-atomic particles form a braid, and this shape should make them much more stable and less susceptible to buffeting interference from surrounding electromagnetic forces than those being used by IBM, Google, and Rigetti. That buffeting creates errors in a quantum computer’s calculations, which then have to be corrected. With a theoretically lower error rate, Microsoft’s design ought to be a safer bet for commercial applications. But, first, the company has to prove it can reliably create these braids and use them to form qubits—something it hasn’t yet done.

In the meantime, Microsoft has a whole group of mathematicians and computer scientists looking at ways to program quantum computers. And, as it turns out, some of the algorithms developed to take advantage of the weird properties of quantum computers can also be used to great advantage on normal ones.

A custom algorithm

With MRF, the trick is figuring out exactly how to tune the strength, frequency, and angle of the radio pulses the scanner transmits, Griswold says. Finding the right pulse pattern is what enables the scanner to identify tissue types—to isolate the song of each singer in the choir to use Griswold’s analogy. There is a mathematically optimal pattern that would allow the scanner to pick up only that tissue type with a precision down to the individual cell—but finding it involves so many variables that it is beyond the computational power of a conventional computer, he says. So researchers have relied almost entirely on educated guesswork to plan the pattern of pulses for each scan, he says. Even with this imperfect method, he says, MRF still results in much more detailed images than a typical MRI.

To get further improvements, Griswold says, he needed to get beyond human intuition. But when his team applied for a grant to research how to optimize the MRF scans using conventional algorithmic techniques, the application was rejected on the grounds that solving such a mathematically-challenging problem was simply impossible.

Then Griswold heard that Microsoft, which had worked closely with medical imaging experts at Case on a test case for its HoloLens augmented reality goggles, was looking for partners to create similar demonstration cases for quantum algorithms. Griswold, who had followed quantum computing developments closely for 20 years and knew some of the researchers now working on Microsoft's efforts, realized this might be his chance.

“We like problems that are seemingly impossible,” Matthias Troyer, Microsoft quantum computing researcher who worked on the MRF project, says. What's more, Troyer, says, MRF was the kind of seemingly impossible problem—an optimization challenge—for which quantum algorithms already existed.

Troyer says the existing quantum algorithm, however, had to be tweaked for Griswold’s exact problem. “What we like to stress it to really get the full power of the quantum optimizer, one really has to make a bespoke solution,” he says. In this case, Troyer says, the hard part was figuring out, from the several thousand variables involved in building an MRF image, which subset of factors the algorithm should try to optimize. Once you do this, he says, “the initially impossible begins to look possible.”

He also says that even though running the quantum algorithm on a conventional computer resulted in a significant increase in the speed and precision of the MRF scans, the results would have been even more impressive on a large-enough quantum computer. “It would have been much faster,” he says.

But the size quantum computer Troyer is talking about would require about one million logical qubits. And machines of that size are still many years, if not a few decades, away.