Quantum computing's hype problem

Prof. Sankar Das Sarma provides a lucid perspective on much-hyped near-term applications of quantum computing.

I think a lot of this hype can be blamed on commercial interests or non-specialist science journalists who hide the nuances of the field in search of clickbait.

From what I've seen in talks by researchers from hardware companies such as Google, Xanadu, etc., they do try to focus on the interesting physics of the current generation of quantum processors, rather than trying to overhype unlikely or far-off commercial applications.

I think the same is true for academic researchers working on quantum computing. Unfortunately funding for basic science at universities is scarce, and many governments are keen to promote partnerships with industry even when the science may be at a far too early stage.

Perspectives on quantum machine learning

Is quantum advantage the right goal for quantum machine learning?

This perspective article is a must-read for anyone working on or interested in quantum machine learning. In it, the authors argue that current research should not focus on attempting to "beat" existing classical machine learning approaches; rather, we should aim to better understand the fundamentals of how quantum learning may work. For example, what is the quantum analogue of a perceptron, and how can we quantum machine learning algorithms be related to better-understand classical learning theory?

In my opinion, unless someone can come up with a quantum version of the backpropagation algorithm, quantum neural networks will never be competitive with classical deep neural networks, because they will be too slow and expensive to train. 

Even avoiding this training issue (e.g. using kernel methods), the bottleneck of transferring data into and out of the quantum circuit will be huge in practical applications and may overwhelm any quantum speedups.

I am most excited about topological data analysis-based approaches because quantum TDA algorithms seem to avoid this data input bottleneck.

 

arXiv moderation

The core goal of arXiv is to make preprints quickly available for anyone to read, avoiding the long processing times of peer-reviewed journals.

 

Some moderation is necessary to prevent the listings from being swamped with spam or pseudoscience. But in recent years the volume of submissions and moderators' workloads have ballooned, leading to delays in processing submissions. For example, last year one of our preprints was flagged as being incorrectly classified and spent one and a half weeks in the moderation queue (as if we didn't know who our intended audience was).

Recently, Prof. Jorge Hirsch was banned from submitting articles for 6 months on the basis of comments he posted on a claimed observation of room temperature superconductivity: Preprint server removes ‘inflammatory’ papers in superconductor controversy

I read one of the comments when it first appeared (arXiv:2201.07686v1). My interest was piqued by the last sentence of the abstract, "We conclude that the published data have been manipulated, making it impossible to draw any conclusions about the susceptibility of the material from the reported numbers." This statement has been removed from the latest version of the manuscript.

I am not an expert on superconductivity, so I'm not sure how much credence to give this claim. Whether the claim is correct or not is beside the point. arXiv is not a journal and moderators are not journal editors responsible for peer review. Manuscripts by established researchers should not be censored in this manner. Why is arxiv stifling academic debate?

arxiv highlights

Quantum persistent homology

A generalization of the quantum algorithm by Lloyd et al., which provides an exponential speed up for computing Betti numbers, enabling the computation of persistent Betti numbers. Like the original algorithm, however, it assumes the existences of a QRAM allowing the input data to be queried in a quantum superposition. It is an open question whether the exponential speedup remains when the data-encoding overhead is taken into account. See also the related arXiv:2111.00433.

 

Anomalous single-mode lasing induced by nonlinearity and the non-Hermitian skin effect

Highlighting a nice collaboration I was involved in. One limitation of many topological or PT-symmetric models for single mode lasers is that they require a structured pump. Such structured pumping will necessarily lower the device efficiency (in terms of output power / device size). Here we show counterintuitively how nonlinear gain saturation can lead to the emergence of stable single mode lasing in uniformly-pumped systems exhibiting the non-Hermitian skin effect. Due to the non-Hermitian skin effect, most of the linear modes become localized to the boundary of the system. However, a few (non-extensive) delocalized bulk modes remain and can be used as large volume lasing modes.

CAFQA: Clifford Ansatz For Quantum Accuracy

This is a neat approach for solving the barren plateau problem that makes quantum neural networks (and other variational quantum algorithms) expensive to train. The idea is to the initialize the circuit as a set of Clifford gates, which are efficiently simulable using classical computers. Therefore a classical computer can be used to find the best Clifford circuit approximation to the solution of the problem. The gate parameters are then allowed to deviate from those corresponding to Clifford gate, producing classically intractable states, and are optimized by running the quantum circuit. The better Clifford starting ansatz allows the quantum circuit optimization to converge more quickly, minimizing the number of expensive quantum circuit evaluations.

Physics-Informed Neural Networks for Quantum Eigenvalue Problems

Physics-informed neural networks are a new approach for solving partial differential equations, based on minimizing a cost function measuring the deviation from the equation being fulfilled at a set of points in the bulk and at the edges of the domain of interest. Once trained one can obtain the field value at any desired point x within the domain. Mesh-free solutions of Maxwell's equations are one promising application. This preprint applies the physics-informed neural network approach to solve the time-independent single particle Schrodinger equation. I wonder whether this approach will also be useful for the many-body problem, since it does not require storing all the wave function components in memory, one can instead query the wavefunction at the desired set of points.

The quest for high refractive index materials

Expanding the Photonic Palette: Exploring High Index Materials

An entertaining and provocative article on why we should care about broadening the palette of high refractive index materials and where we might find them.

 

While graphene is singled out as the "king" of exotic materials with limited practical applications in optics and photonics, it is worth emphasizing that the huge interest in graphene has been a source of funding, techniques, and inspiration for studies of related quasi-2D materials. Among these, transition metal dichalcogenides seem quite promising for photonics applications including high index metasurfaces.

Thanks to nanoscale views for making me aware of this article.