Feng Pan and Pan Zhang uploaded a new paper on the arXive “Simulating the Sycamore supremacy circuits.” with an amazing announcement.
Abstract: We propose a general tensor network method for simulating quantum circuits. The method is massively more efficient in computing a large number of correlated bitstring amplitudes and probabilities than existing methods. As an application, we study the sampling problem of Google’s Sycamore circuits, which are believed to be beyond the reach of classical supercomputers and have been used to demonstrate quantum supremacy. Using our method, employing a small computational cluster containing 60 graphical processing units (GPUs), we have generated one million correlated bitstrings with some entries fixed, from the Sycamore circuit with 53 qubits and 20 cycles, with linear cross-entropy benchmark (XEB) fidelity equals 0.739, which is much higher than those in Google’s quantum supremacy experiments.
Congratulations to Feng Pan and Pan Zhang for this remarkable breakthrough!
Of course, we can expect that in the weeks and months to come, the community will learn, carefully check, and digest this surprising result and will ponder about its meaning and interpretation. Stay tuned!
Here is a technical matter I am puzzled about: the paper claims the ability to compute precisely the amplitudes for a large number of bitstrings. (Apparently computing the amplitudes is even more difficult computational task than sampling.) But then, it is not clear to me where the upper bound of 0.739 comes from? If you have the precise amplitudes it seems that you can sample with close to perfect fidelity. (And, if you wish, you can get a F_XEB score larger than 1.)
Update: This is explained just before the discussion part of the paper. The crucial thing is that the probabilities for the 2^21 strings are distributed close to Porter-Thomas (exponentials). If you take samples for them indeed you can get samples with F_XEB between -1 and 15. Picking the highest 10^6 strings from 2^21 get you 0.739 (so this value has no special meaning.) Probably by using Metropolis sampling you can get (smaller, unless you enlarge 2^21 to 2^25, say) samples with F_XEB close to 1 and size-biased distribution (the distribution of probabilities of sampled strings) that fits the theoretical size biased distribution. And you can also use metropolis sampling to get a sample of size 10^6 with the correct distribution of probabilities for somewhat smaller fidelity.
The paper mentions several earlier papers in this direction, including an earlier result by Johnnie Gray and Stefanos Kourtis in the paper Hyper-optimized tensor network contraction and another earlier result in the paper Classical Simulation of Quantum Supremacy Circuits by a group of researchers Cupjin Huang, Fang Zhang, Michael Newman, Junjie Cai, Xun Gao, Zhengxiong Tian, Junyin Wu, Haihong Xu, Huanjun Yu, Bo Yuan, Mario Szegedy, Yaoyun Shi, and Jianxin Chen, from Alibaba co. Congratulations to them as well.
I am thankful to colleagues who told me about this paper.
Scientists Say They Used Classical Approach to outperform Google’s Sycamore QC (“The Quantum” Written by Matt Swayne with interesting quotes from the paper. )
Another axe swung at the Sycamore (Shtetl-Optimized; with interesting preliminary thoughts by Scott; )