IBM Q

Making Quantum Computers More Accurate

Dr. Abhinav Kandala discusses his paper, published this month in Nature, on extending the reach and utility of quantum computers.

Q: Nature just published your paper about your progress in error mitigation for quantum computers. It gets pretty technical fast, so before diving in, let’s start with some context. What’s the problem you set out to solve?

A: Let me start by providing some motivation for quantum computing in general. In the scientific community, we think we know quite well the equations that describe how natural processes work at the microscopic level. These equations form the laws of quantum mechanics, right? But solving these equations is extremely hard for a regular or classical computer. So we take shortcuts called approximations. We make predictions about the properties of molecules, of proteins, and so on by simplifying the equations to the point where they become accessible to a classical computer.

Q: Approximations, by definition, can only be so accurate.

A: Right. And there are classes of problems we can’t even approximate well. Which is where quantum computers come in. Since Richard Feynman posed this challenge in the early 80s, scientists have thought, what if we could make a programmable quantum system to mimic or simulate the properties of natural quantum systems like molecules, for instance? Fast-forward to 2019, and we now have these. We call them quantum computers, and they’re available in the cloud. But they’re extremely sensitive, very fragile, and the encoded information gets lost in a few hundred microseconds. This loss of information happens in all of today’s so-called Noisy Intermediate Scale Quantum Computers, or NISQ machines. They’re all susceptible to decoherence, which basically just means that the computations have errors.

Q: So you want to prevent or correct those errors?

A: Exactly! This is something we know how to do, at least theoretically. As far back as the 90s, Peter Shor and others developed the theory for quantum error correction. What it came down to was this: if your error rates are sufficiently small, you could use a large number of physical qubits to detect and correct the errors. This is what we would call a fault tolerant universal quantum computer. But the resource requirements to build one are well beyond our grasp for the near-term. So we were looking to create something useful in the meantime. We asked ourselves, can we actually achieve accurate computations using NISQ machines?

Q: And the answer is …

A: We demonstrate, yes, it is possible, to some extent. We’ve developed a technique where repeating a desired computation at varying levels of noise lets us extrapolate to what the quantum computer would have calculated in the absence of noise. It’s called zero-noise extrapolation. Our team published the theoretical idea in 2017, but this paper is an experimental demonstration of the technique. We simulated hydrogen with two qubits and lithium hydride with four qubits and we showed that you can significantly enhance the quality of the computation. In effect, you’re able to extract more out of the machine – by using many wrongs, to reconstruct the right answer.

Q:The title of your paper is Extending the computational reach of a noisy superconducting quantum processor. How far of an extension are we talking about?

A: Well, for the simulations that we addressed in this work, we were able to achieve accuracies that were previously inaccessible to the hardware. The technique basically let us benefit from longer sequences of quantum operations, since the effect of the associated decoherence could be mitigated.

Q: Where do you think this technique is going to be most useful?

A: I believe anyone attempting to use a noisy device to do quantum computation based on what we call expectation values is going to find benefit in employing techniques like this, in particular since it doesn’t require any additional hardware resources. Quantum chemistry calculations and machine learning tasks are some examples of problems that rely on expectation values, and are areas of current interest. But it’s really a general-purpose technique that can be easily extended to computations in, say, high-energy physics and material science.

Learn more about how Noise Amplification Squeezes More Computational Accuracy From Today’s Quantum Processors