Two weeks ago, I blogged about the striking claim, by the group headed by Chaoyang Lu and Jianwei Pan at USTC in China, to have achieved quantum supremacy via BosonSampling with 50-70 detected photons. I also did a four-part interview on the subject with Jonathan Tennenbaum at Asia Times, and other interviews elsewhere. None of that stopped some people, who I guess didn’t google, from writing to tell me how disappointed they were by my silence!
The reality, though, is that a lot has happened since the original announcement, so it’s way past time for an update.
I. The Quest to Spoof
Most importantly, other groups almost immediately went to work trying to refute the quantum supremacy claim, by finding some efficient classical algorithm to spoof the reported results. It’s important to understand that this is exactly how the process is supposed to work: as I’ve often stressed, a quantum supremacy claim is credible only if it’s open to the community to refute and if no one can. It’s also important to understand that, for reasons we’ll go into, there’s a decent chance that people will succeed in simulating the new experiment classically, although they haven’t yet. All parties to the discussion agree that the new experiment is, far and away, the closest any BosonSampling experiment has ever gotten to the quantum supremacy regime; the hard part is to figure out if it’s already there.
Part of me feels guilty that, as one of reviewers on the Science paper—albeit, one stressed and harried by kids and covid—it’s now clear that I didn’t exercise the amount of diligence that I could have, in searching for ways to kill the new supremacy claim. But another part of me feels that, with quantum supremacy claims, much like with proposals for new cryptographic codes, vetting can’t be the responsibility of one or two reviewers. Instead, provided the claim is serious—as this one obviously is—the only thing to do is to get the paper out, so that the entire community can then work to knock it down. Communication between authors and skeptics is also a hell of a lot faster when it doesn’t need to go through a journal’s editorial system.
Not surprisingly, one skeptic of the new quantum supremacy claim is Gil Kalai, who (despite Google’s result last year, which Gil still believes must be in error) rejects the entire possibility of quantum supremacy on quasi-metaphysical grounds. But other skeptics are current and former members of the Google team, including Sergio Boixo and John Martinis! And—pause to enjoy the irony—Gil has effectively teamed up with the Google folks on questioning the new claim. Another central figure in the vetting effort—one from whom I’ve learned much of what I know about the relevant issues over the last week—is Dutch quantum optics professor and frequent Shtetl-Optimized commenter Jelmer Renema.
Without further ado, why might the new experiment, impressive though it was, be efficiently simulable classically? A central reason for concern is photon loss: as Chaoyang Lu has now explicitly confirmed (it was implicit in the paper), up to ~70% of the photons get lost on their way through the beamsplitter network, leaving only ~30% to be detected. At least with “Fock state” BosonSampling—i.e., the original kind, the kind with single-photon inputs that Alex Arkhipov and I proposed in 2011—it seems likely to me that such a loss rate would be fatal for quantum supremacy; see for example this 2019 paper by Renema, Shchesnovich, and Garcia-Patron.
Incidentally, if anything’s become clear over the last two weeks, it’s that I, the co-inventor of BosonSampling, am no longer any sort of expert on the subject’s literature!
Anyway, one source of uncertainty regarding the photon loss issue is that, as I said in my last post, the USTC experiment implemented a 2016 variant of BosonSampling called Gaussian BosonSampling (GBS)—and Jelmer tells me that the computational complexity of GBS in the presence of losses hasn’t yet been analyzed in the relevant regime, though there’s been work aiming in that direction. A second source of uncertainty is simply that the classical simulations work in a certain limit—namely, fixing the rate of noise and then letting the numbers of photons and modes go to infinity—but any real experiment has a fixed number of photons and modes (in USTC’s case, they’re ~50 and ~100 respectively). It wouldn’t do to reject USTC’s claim via a theoretical asymptotic argument that would equally well apply to any non-error-corrected quantum supremacy demonstration!
OK, but if an efficient classical simulation of lossy GBS experiments exists, then what is it? How does it work? It turns out that we have a plausible candidate for the answer to that, originating with a 2014 paper by Gil Kalai and Guy Kindler. Given a beamsplitter network, Kalai and Kindler considered an infinite hierarchy of better and better approximations to the BosonSampling distribution for that network. Roughly speaking, at the first level (k=1), one pretends that the photons are just classical distinguishable particles. At the second level (k=2), one correctly models quantum interference involving pairs of photons, but none of the higher-order interference. At the third level (k=3), one correctly models three-photon interference, and so on until k=n (where n is the total number of photons), when one has reproduced the original BosonSampling distribution. At least when k is small, the time needed to spoof outputs at the kth level of the hierarchy should grow like nk. As theoretical computer scientists, Kalai and Kindler didn’t care whether their hierarchy produced any physically realistic kind of noise, but later work, by Shchesnovich, Renema, and others, showed that (as it happens) it does.
In its original paper, the USTC team ruled out the possibility that the first, k=1 level of this hierarchy could explain its experimental results. More recently, in response to inquiries by Sergio, Gil, Jelmer, and others, Chaoyang tells me they’ve ruled out the possibility that the k=2 level can explain their results either. We’re now eagerly awaiting the answer for larger values of k.
Let me add that I owe Gil Kalai the following public mea culpa. While his objections to QC have often struck me as unmotivated and weird, in the case at hand, Gil’s 2014 work with Kindler is clearly helping drive the scientific discussion forward. In other words, at least with BosonSampling, it turns out that Gil put his finger precisely on a key issue. He did exactly what every QC skeptic should do, and what I’ve always implored the skeptics to do.
II. BosonSampling vs. Random Circuit Sampling: A Tale of HOG and CHOG and LXEB
There’s a broader question: why should skeptics of a BosonSampling experiment even have to think about messy details like the rate of photon losses? Why shouldn’t that be solely the experimenters’ job?
To understand what I mean, consider the situation with Random Circuit Sampling, the task Google demonstrated last year with 53 qubits. There, the Google team simply collected the output samples and fed them into a benchmark that they called “Linear Cross-Entropy” (LXEB), closely related to what Lijie Chen and I called “Heavy Output Generation” (HOG) in a 2017 paper. With suitable normalization, an ideal quantum computer would achieve an LXEB score of 2, while classical random guessing would achieve an LXEB score of 1. Crucially, according to a 2019 result by me and Sam Gunn, under a plausible (albeit strong) complexity assumption, no subexponential-time classical spoofing algorithm should be able to achieve an LXEB score that’s even slightly higher than 1. In its experiment, Google reported an LXEB score of about 1.002, with a confidence interval much smaller than 0.002. Hence: quantum supremacy (subject to our computational assumption), with no further need to know anything about the sources of noise in Google’s chip! (More explicitly, Boixo, Smelyansky, and Neven did a calculation in 2017 to show that the Kalai-Kindler type of spoofing strategy definitely isn’t going to work against RCS and Linear XEB, with no computational assumption needed.)
So then why couldn’t the USTC team do something analogous with BosonSampling? Well, they tried to. They defined a measure that they called “HOG,” although it’s different from my and Lijie Chen’s HOG, more similar to a cross-entropy. Following Jelmer, let me call their measure CHOG, where the C could stand for Chinese, Chaoyang’s, or Changed. They calculated the CHOG for their experimental samples, and showed that it exceeds the CHOG that you’d get from the k=1 and k=2 levels of the Kalai-Kindler hierarchy, as well as from various other spoofing strategies, thereby ruling those out as classical explanations for their results.
The trouble is this: unlike with Random Circuit Sampling and LXEB, with BosonSampling and CHOG, we know that there are fast classical algorithms that achieve better scores than the trivial algorithm, the algorithm that just picks samples at random. That follows from Kalai and Kindler’s work, and it even more simply follows from a 2013 paper by me and Arkhipov, entitled “BosonSampling Is Far From Uniform.” Worse yet, with BosonSampling, we currently have no analogue of my 2019 result with Sam Gunn: that is, a result that would tell us (under suitable complexity assumptions) the highest possible CHOG score that we expect any efficient classical algorithm to be able to get. And since we don’t know exactly where that ceiling is, we can’t tell the experimentalists exactly what target they need to surpass in order to claim quantum supremacy. Absent such definitive guidance from us, the experimentalists are left playing whac-a-mole against this possible classical spoofing strategy, and that one, and that one.
This is an issue that I and others were aware of for years, although the new experiment has certainly underscored it. Had I understood just how serious the USTC group was about scaling up BosonSampling, and fast, I might’ve given the issue some more attention!
III. Fock vs. Gaussian BosonSampling
Above, I mentioned another complication in understanding the USTC experiment: namely, their reliance on Gaussian BosonSampling (GBS) rather than Fock BosonSampling (FBS), sometimes also called Aaronson-Arkhipov BosonSampling (AABS). Since I gave this issue short shrift in my previous post, let me make up for it now.
In FBS, the initial state consists of either 0 or 1 photons in each input mode, like so: |1,…,1,0,…,0⟩. We then pass the photons through our beamsplitter network, and measure the number of photons in each output mode. The result is that the amplitude of each possible output configuration can be expressed as the permanent of some n×n matrix, where n is the total number of photons. It was interest in the permanent, which plays a central role in classical computational complexity, that led me and Arkhipov to study BosonSampling in the first place.
The trouble is, preparing initial states like |1,…,1,0,…,0⟩ turns out to be really hard. No one has yet build a source that reliably outputs one and only one photon at exactly a specified time. This led two experimental groups to propose an idea that, in a 2013 post on this blog, I named Scattershot BosonSampling (SBS). In SBS, you get to use the more readily available “Spontaneous Parametric Down-Conversion” (SPDC) photon sources, which output superpositions over different numbers of photons, of the form $$\sum_{n=0}^{\infty} \alpha_n |n \rangle |n \rangle, $$ where αn decreases exponentially with n. You then measure the left half of each entangled pair, hope to see exactly one photon, and are guaranteed that if you do, then there’s also exactly one photon in the right half. Crucially, one can show that, if Fock BosonSampling is hard to simulate approximately using a classical computer, then the Scattershot kind must be as well.
OK, so what’s Gaussian BosonSampling? It’s simply the generalization of SBS where, instead of SPDC states, our input can be an arbitrary “Gaussian state”: for those in the know, a state that’s exponential in some quadratic polynomial in the creation operators. If there are m modes, then such a state requires ~m2 independent parameters to specify. The quantum optics people have a much easier time creating these Gaussian states than they do creating single-photon Fock states.
While the amplitudes in FBS are given by permanents of matrices (and thus, the probabilities by the absolute squares of permanents), the probabilities in GBS are given by a more complicated matrix function called the Hafnian. Roughly speaking, while the permanent counts the number of perfect matchings in a bipartite graph, the Hafnian counts the number of perfect matchings in an arbitrary graph. The permanent and the Hafnian are both #P-complete. In the USTC paper, they talk about yet another matrix function called the “Torontonian,” which was invented two years ago. I gather that the Torontonian is just the modification of the Hafnian for the situation where you only have “threshold detectors” (which decide whether one or more photons are present in a given mode), rather than “number-resolving detectors” (which count how many photons are present).
If Gaussian BosonSampling includes Scattershot BosonSampling as a special case, and if Scattershot BosonSampling is at least as hard to simulate classically as the original BosonSampling, then you might hope that GBS would also be at least as hard to simulate classically as the original BosonSampling. Alas, this doesn’t follow. Why not? Because for all we know, a random GBS instance might be a lot easier than a random SBS instance. Just because permanents can be expressed using Hafnians, doesn’t mean that a random Hafnian is as hard as a random permanent.
Nevertheless, I think it’s very likely that the sort of analysis Arkhipov and I did back in 2011 could be mirrored in the Gaussian case. I.e., instead of starting with reasonable assumptions about the distribution and hardness of random permanents, and then concluding the classical hardness of approximate BosonSampling, one would start with reasonable assumptions about the distribution and hardness of random Hafnians (or “Torontonians”), and conclude the classical hardness of approximate GBS. But this is theoretical work that remains to be done!
IV. Application to Molecular Vibronic Spectra?
In 2014, Alan Aspuru-Guzik and collaborators put out a paper that made an amazing claim: namely that, contrary to what I and others had said, BosonSampling was not an intrinsically useless model of computation, good only for refuting QC skeptics like Gil Kalai! Instead, they said, a BosonSampling device (specifically, what would later be called a GBS device) could be directly applied to solve a practical problem in quantum chemistry. This is the computation of “molecular vibronic spectra,” also known as “Franck-Condon profiles,” whatever those are.
I never understood nearly enough about chemistry to evaluate this striking proposal, but I was always a bit skeptical of it, for the following reason. Nothing in the proposal seemed to take seriously that BosonSampling is a sampling task! A chemist would typically have some specific numbers that she wants to estimate, of which these “vibronic spectra” seemed to be an example. But while it’s often convenient to estimate physical quantities via Monte Carlo sampling over simulated observations of the physical system you care about, that’s not the only way to estimate physical quantities! And worryingly, in all the other examples we’d seen where BosonSampling could be used to estimate a number, the same number could also be estimated using one of several polynomial-time classical algorithms invented by Leonid Gurvits. So why should vibronic spectra be an exception?
After an email exchange with Alex Arkhipov, Juan Miguel Arrazola, Leonardo Novo, and Raul Garcia-Patron, I believe we finally got to the bottom of it, and the answer is: vibronic spectra are not an exception.
In terms of BosonSampling, the vibronic spectra task is simply to estimate the probability histogram of some weighted sum like $$ w_1 s_1 + \cdots + w_ m s_m, $$ where w1,…,wm are fixed real numbers, and (s1,…,sm) is a possible outcome of the BosonSampling experiment, si representing the number of photons observed in mode i. Alas, while it takes some work, it turns out that Gurvits’s classical algorithms can be adapted to estimate these histograms. Granted, running the actual BosonSampling experiment would provide slightly more detailed information—namely, some exact sampled values of $$ w_1 s_1 + \cdots + w_ m s_m, $$ rather than merely additive approximations to the values—but since we’d still need to sort those sampled values into coarse “bins” in order to compute a histogram, it’s not clear why that additional precision would ever be of chemical interest.
This is a pity, since if the vibronic spectra application had beaten what was doable classically, then it would’ve provided not merely a first practical use for BosonSampling, but also a lovely way to verify that a BosonSampling device was working as intended.
V. Application to Finding Dense Subgraphs?
A different potential application of Gaussian BosonSampling, first suggested by the Toronto-based startup Xanadu, is finding dense subgraphs in a graph. (Or at least, providing an initial seed to classical optimization methods that search for dense subgraphs.)
This is an NP-hard problem, so to say that I was skeptical of the proposal would be a gross understatement. Nevertheless, it turns out that there is a striking observation by the Xanadu team at the core of their proposal: namely that, given a graph G and a positive even integer k, a GBS device can be used to sample a random subgraph of G of size k, with probability proportional to the square of the number of perfect matchings in that subgraph. Cool, right? And potentially even useful, especially if the number of perfect matchings could serve as a rough indicator of the subgraph’s density! Alas, Xanadu’s Juan Miguel Arrazola himself recently told me that there’s a cubic-time classical algorithm for the same sampling task, so that the possible quantum speedup that one could get from GBS in this way is at most polynomial. The search for a useful application of BosonSampling continues!
And that’s all for now! I’m grateful to all the colleagues I talked to over the last couple weeks, including Alex Arkhipov, Juan Miguel Arrazola, Sergio Boixo, Raul Garcia-Patron, Leonid Gurvits, Gil Kalai, Chaoyang Lu, John Martinis, and Jelmer Renema, while obviously taking sole responsibility for any errors in the above. I look forward to a spirited discussion in the comments, and of course I’ll post updates as I learn more!
Follow
