Shtetl-Optimized

A group led by Jianwei Pan, based mainly in Hefei and Shanghai, announced today that it achieved BosonSampling with 40-70 detected photons—up to and beyond the limit where a classical supercomputer could feasibly verify the results. For more, see also Emily Conover’s piece in Science News, or Daniel Garisto’s in Scientific American, both of which I consulted on. (Full disclosure: I was one of the reviewers for the Pan group’s Science paper, and will be writing the Perspective article to accompany it.)

The new result follows the announcement of 14-photon BosonSampling by the same group a year ago. It represents the second time quantum supremacy has been reported, following Google’s celebrated announcement from last year, and the first time it’s been done using photonics rather than superconducting qubits.

As the co-inventor of BosonSampling (with Alex Arkhipov), obviously I’m gratified about this.

For anyone who regards it as boring or obvious, here and here is Gil Kalai, on this blog, telling me why BosonSampling would never scale beyond 8-10 photons. (He wrote that, if aliens forced us to try, then much like with the Ramsey number R(6,6), our only hope would be to attack the aliens.) Here’s Kalai making a similar prediction, on the impossibility of quantum supremacy by BosonSampling or any other means, in his plenary address to the International Congress of Mathematicians two years ago.

Even if we set aside the quantum computing skeptics, many distinguished colleagues told me they thought experimental BosonSampling was a dead end, because of photon losses and the staggering difficulty of synchronizing 50-100 single-photon sources. They said a convincing demonstration of quantum supremacy would have to await the arrival of quantum fault-tolerance—or at any rate, some hardware platform more robust than photonics. I always agreed that they might be right. Furthermore, even if 50-photon BosonSampling was possible, after Google reached the supremacy milestone first with superconducting qubits, it wasn’t clear that anyone was going to bother.

Obviously the new result isn’t dispositive. Maybe Gil Kalai really WON, BY A LOT, before Hugo Chávez hacked the Pan group’s computers to make it seem otherwise? (Sorry, couldn’t resist.) Nevertheless, as someone whose intellectual origins are close to pure math, it’s strange and exciting to find myself in a field where, once in a while, the world itself gets to weigh in on a theoretical disagreement.

Since excitement is best when paired with accurate understanding, please help yourself to the following FAQ, which I might add more to over the next couple days.

What is BosonSampling? You must be new here! In increasing order of difficulty, here’s an MIT News article from back in 2011, here’s the Wikipedia page, here are my PowerPoint slides, here are my lecture notes from Rio de Janeiro, here’s my original paper with Arkhipov…

What is quantum supremacy? Roughly, the use of a programmable or configurable quantum computer to solve some well-defined computational problem much faster than we know how to solve it with any existing classical computer. “Quantum supremacy,” a term coined by John Preskill in 2012, does not mean useful QC, or universal QC, or scalable QC, or fault-tolerant QC, all of which remain outstanding challenges. For more, see my Supreme Quantum Supremacy FAQ, or (e.g.) my recent Lytle Lecture for the University of Washington.

If Google already announced quantum supremacy a year ago, what’s the point of this new experiment? To me, at least, quantum supremacy seems important enough to do at least twice! Also, as I said, this represents the first demonstration that quantum supremacy is possible via photonics. Finally, as the authors point out, the new experiment has one big technical advantage over Google’s: namely, many more possible output states (~10³⁰ of them, rather than a mere ~9 quadrillion). This makes it infeasible to calculate the whole probability distribution over outputs and store it on a gigantic hard disk (after which one could easily generate as many samples as one wanted), which is what IBM proposed doing in its response to Google’s announcement.

Is BosonSampling a form of universal quantum computing? No, we don’t even think it can simulate universal classical computing! It’s designed for exactly one task: namely, demonstrating quantum supremacy and refuting Gil Kalai. It might have some other applications besides that, but if so, they’ll be icing on the cake. This is in contrast to Google’s Sycamore processor, which in principle is a universal quantum computer, just with a severe limit on the number of qubits (53) and how many layers of gates one can apply to them (about 20).

Is BosonSampling at least a step toward universal quantum computing? I think so! In 2000, Knill, Laflamme, and Milburn (KLM) famously showed that pure, non-interacting photons, passing through a network of beamsplitters, are capable of universal QC, provided we assume one extra thing: namely, the ability to measure the photons at intermediate times, and change which beamsplitters to apply to the remaining photons depending on the outcome. In other words, “BosonSampling plus adaptive measurements equals universality.” Basically, KLM is the holy grail that Pan’s and other experimental optics groups around the world have been working toward for 20 years, with BosonSampling just a more achievable pit stop along the way.

Are there any applications of BosonSampling? We don’t know yet. There are proposals in the literature to apply BosonSampling to quantum chemistry and other fields, but I’m not yet convinced that these proposals will yield real speedups over the best we can do with classical computers, for any task of practical interest that involves estimating specific numbers (as opposed to sampling tasks, where BosonSampling almost certainly does yield exponential speedups, but which are rarely the thing practitioners directly care about).

How hard is it to simulate BosonSampling on a classical computer? As far as we know today, the difficulty of simulating a “generic” BosonSampling experiment increases roughly like 2ⁿ, where n is the number of detected photons. It might be easier than that, particularly when noise and imperfections are taken into account; and at any rate it might be easier to spoof the statistical tests that one applies to verify the outputs. I and others managed to give some theoretical evidence against those possibilities, but just like with Google’s experiment, it’s conceivable that some future breakthrough will change the outlook and remove the case for quantum supremacy.

Do you have any amusing stories? When I refereed the Science paper, I asked why the authors directly verified the results of their experiment only for up to 26-30 photons, relying on plausible extrapolations beyond that. While directly verifying the results of n-photon BosonSampling takes ~2ⁿ time for any known classical algorithm, I said, surely it should be possible with existing computers to go up to n=40 or n=50? A couple weeks later, the authors responded, saying that they’d now verified their results up to n=40, but it burned $400,000 worth of supercomputer time so they decided to stop there. This was by far the most expensive referee report I ever wrote!

Also: when Covid first started, and facemasks were plentiful in China but almost impossible to get in the US, Chaoyang Lu, one of the authors of the new work and my sometime correspondent on the theory of BosonSampling, decided to mail me a box of 200 masks (I didn’t ask for it). I don’t think that influenced my later review, but was appreciated nonetheless.

Huge congratulations to the whole team for their accomplishment!

While a lot of pain is still ahead, this year I’m thankful that a dark chapter in American history might be finally drawing to a close. I’m thankful that the mRNA vaccines actually work. I’m thankful that my family has remained safe, and I’m thankful for all the essential workers who’ve kept our civilization running.

A few things:

1. Friend-of-the-blog Jelani Nelson asked me to advertise an important questionnaire for theoretical computer scientists, about what the future of STOC and FOCS should look like (for example, should they become all virtual?). It only takes 2 or 3 minutes to fill out (I just did).
   
2. Here’s a podcast that I recently did with UT Austin undergraduate Dwarkesh Patel. (As usual, I recommend 2x speed to compensate for my verbal tics.)
   
3. Feel free to use the comments on this post to talk about recent progress in quantum computing or computational complexity! Like, I dunno, a (sub)exponential black-box speedup for the adiabatic algorithm, or anti-concentration for log-depth random quantum circuits, or an improved shadow tomography procedure, or a quantum algorithm for nonlinear differential equations, or a barrier to proving strong 3-party parallel repetition, or equivalence of one-way functions and time-bounded Kolmogorov complexity, or turning any hard-on-average NP problem into one that’s guaranteed to have solutions.
   
4. It’s funny how quantum computing, P vs. NP, and so forth can come to feel like just an utterly mundane day job, not something anyone outside a small circle could possibly want to talk about while the fate of civilization hangs in the balance. Sometimes it takes my readers to remind me that not only are these topics what brought most of you here in the first place, they’re also awesome! So, I’ll mark that down as one more thing to be thankful for.

Hook ’em Hadamards!

If you’re a prospective PhD student: Apply here for the CS department (the deadline this year is December 15th), here for the physics department (the deadline is December 1st), or here for the ECE department (the deadline is 15th). GREs are not required this year because of covid. If you apply to CS and specify that you want to work with me, I’ll be sure to see your application. If you apply to physics or ECE, I won’t see your application, but once you arrive, I can sometimes supervise or co-supervise PhD students in other departments (or, of course, serve on their committees). In any case, everyone in the UT community is extremely welcome at our quantum information group meetings (which are now on Zoom, naturally, but depending on vaccine distribution, hopefully won’t be by the time you arrive!). Emailing me won’t make a difference. Admissions are very competitive, so apply broadly to maximize your chances.

If you’re a prospective postdoctoral fellow: By January 1, 2021, please email me a cover letter, your CV, and two or three of your best papers (links or attachments). Please also ask two recommenders to email me their letters by January 1. While my own work tends toward computational complexity, I’m open to all parts of theoretical quantum computing and information.

If you’re a prospective faculty member: Yes, faculty searches are still happening despite covid! Go here to apply for an opening in the CS department (which, in quantum computing, currently includes me and MIP*=RE superstar John Wright), or here to apply to the physics department (which, in quantum computing, currently includes Drew Potter, along with a world-class condensed matter group).

Update (Nov. 14): I now have a deluge of serious hosting offers—thank you so much, everyone! No need for more.

Since I’m now feeling better that the first authoritarian coup attempt in US history will probably sort itself out OK, here’s a real problem:

Nearly a score years ago, I created the Complexity Zoo, a procrastination project turned encyclopedia of complexity classes. Nearly half a score years ago, the Zoo moved to my former employer, the Institute for Quantum Computing in Waterloo, Canada, which graciously hosted it ever since. Alas, IQC has decided that it can no longer do so. The reason is new regulations in Ontario about the accessibility of websites, which the Zoo might be out of compliance with. My students and I were willing to look into what was needed—like, does the polynomial hierarchy need ramps between its levels or something? The best would be if we heard from actual blind or other disabled complexity enthusiasts about how we could improve their experience, rather than trying to parse bureaucratese from the Ontario government. But IQC informed us that in any case, they can’t deal with the potential liability and their decision is final. I thank them for hosting the Zoo for eight years.

Now I’m looking for a volunteer for a new host. The Zoo runs on the MediaWiki platform, which doesn’t work with my own hosting provider (Bluehost) but is apparently easy to set up if you, unlike me, are the kind of person who can do such things. The IQC folks kindly offered to help with the transfer; I and my students can help as well. It’s a small site with modest traffic. The main things I need are just assurances that you can host the site for a long time (“forever” or thereabouts), and that you or someone else in your organization will be reachable if the site goes down or if there are other problems. I own the complexityzoo.com domain and can redirect from there.

In return, you’ll get the immense prestige of hosting such a storied resource for theoretical computer science … plus free publicity for your cause or organization on Shtetl-Optimized, and the eternal gratitude of thousands of my readers.

Of course, if you’re into complexity theory, and you want to update or improve the Zoo while you’re at it, then so much the awesomer! It could use some updates, badly. But you don’t even need to know P from NP.

If you’re interested, leave a comment or shoot me an email. Thanks!!

Unrelated Announcement: I’ll once again be doing an Ask-Me-Anything session at the Q2B (“Quantum to Business”) conference, December 8-10. Other speakers include Umesh Vazirani, John Preskill, Jennifer Ouellette, Eric Schmidt, and many others. Since the conference will of course be virtual this year, registration is a lot cheaper than in previous years. Check it out! (Full disclosure: Q2B is organized by QC Ware, Inc., for which I’m a scientific advisor.)

The title of this post is not an allegory.

At 10am this morning, I had a previously-scheduled appointment with an oral surgeon to remove a large, hideous, occasionally painful growth on the inside of my lower lip. (I’d delayed getting it looked at for several months because of covid, but I no longer could.)

So right now I’m laying in bed at home, with gauze on my lips, dazed, hopped up on painkillers. I regret that things ever got to the point where this was needed. I believe, intellectually, that the surgeon executed about as competently as anyone could ask. But I still wish, if we’re being honest, that there hadn’t been quite this much pain in the surgery or in the recovery from it.

Again intellectually, I know that there’s still lots more pain in the days ahead. I’m not sure that whatever it was won’t just quickly grow back. And yet, I couldn’t be feeling more joy through my whole body with every one of these words that I write. At last I can honestly tell myself: the growth is gone.

Update (Sep. 17): Several people, here and elsewhere, wrote to tell me that while they completely agreed with my strategic and moral stance in this post, they think that it’s the ads of Republican Voters Against Trump, rather than the Lincoln Project, that have been most effective in changing Trump supporters’ minds. So please consider donating to RVAT instead or in addition! In fact, what the hell, I’ll match donations to RVAT up to $1000.

For the past few months, I’ve alternated between periods of debilitating depression and (thankfully) longer stretches when I’m more-or-less able to work. Triggers for my depressive episodes include reading social media, watching my 7-year daughter struggle with prolonged isolation, and (especially) contemplating the ongoing apocalypse in the American West, the hundreds of thousands of pointless covid deaths, and an election in 48 days that if I didn’t know such things were impossible in America would seem likely to produce a terrifying standoff as a despot and millions of his armed loyalists refuse to cede control. Meanwhile, catalysts for my relatively functional periods have included teaching my undergrad quantum information class, Zoom calls with my students, life on Venus?!? (my guess is no, but almost entirely due to priors), learning new math (fulfilling a decades-old goal, I’m finally working my way through Paul Cohen’s celebrated proof of the independence of the Continuum Hypothesis—more about that later!).

Of course, when you feel crushed by the weight of the world’s horribleness, it improves your mood to be able even just to prick the horribleness with a pin. So I was gratified that, in response to a previous post, Shtetl-Optimized readers contributed at least $3,000, the first $2,000 of which I matched, mostly to the Biden-Harris campaign but a little to the Lincoln Project.

Alas, a commenter was unhappy with the latter:

Lincoln Project? Really? … Pushing the Overton window rightward during a worldwide fascist dawn isn’t good. I have trouble understanding why even extremely smart people have trouble with this sort of thing.

Since this is actually important, I’d like to spend the rest of this post responding to it.

For me it’s simple.

What’s the goal right now? To defeat Trump. In the US right now, that’s the prerequisite to every other sane political goal.

What will it take to achieve that goal? Turnout, energizing the base, defending the election process … but also, if possible, persuading a sliver of Trump supporters in swing states to switch sides, or at least vote third party or abstain.

Who is actually effective at that goal? Well, no one knows for sure. But while I thought the Biden campaign had some semi-decent ads, the Lincoln Project’s best stuff seems better to me, just savagely good.

Why are they effective? The answer seems obvious: for the same reason why a jilted ex is a more dangerous foe than a stranger. If anyone understood how to deprogram a Republican from the Trump cult, who would it be: Alexandria Ocasio-Cortez, or a fellow Republican who successfully broke from the cult?

Do I agree with the Lincoln Republicans about most of the “normal” issues that Americans once argued about? Not at all. Do I hold them, in part, morally responsible for creating the preconditions to the current nightmare? Certainly.

And should any of that cause me to boycott them? Not in my moral universe. If Churchill and FDR could team up with Stalin, then surely we in the Resistance can temporarily ally ourselves with the rare Republicans who chose their stated principles over power when tested—their very rarity attesting to the nontriviality of their choice.

To my mind, turning one’s back on would-be allies, in a conflict whose stakes obviously overshadow what’s bad about those allies, is simultaneously one of the dumbest and the ugliest things that human beings can do. It abandons reason for moral purity and ends up achieving neither.

Another Update (Sep. 15): Sorry for the long delay; new post coming soon! To tide you over—or just to distract you from the darkness figuratively and literally engulfing our civilization—here’s a Fortune article about today’s announcement by IBM of its plans for the next few years in superconducting quantum computing, with some remarks from yours truly.

Another Update (Sep. 8): A reader wrote to let me know about a fundraiser for Denys Smirnov, a 2015 IMO gold medalist from Ukraine who needs an expensive bone marrow transplant to survive Hodgkin’s lymphoma. I just donated and I hope you’ll consider it too!

Update (Sep. 5): Here’s another quantum computing podcast I did, “Dunc Tank” with Duncan Gammie. Enjoy!

Thanks so much to Shtetl-Optimized readers, so far we’ve raised $1,371 for the Biden-Harris campaign and $225 for the Lincoln Project, which I intend to match for $3,192 total. If you’d like to donate by tonight (Thursday night), there’s still $404 to go!

Meanwhile, a mere three days after declaring my “new motto,” I’ve come up with a new new motto for this blog, hopefully a more cheerful one:

When civilization seems on the brink of collapse, sometimes there’s nothing left to talk about but maximal separations between randomized and quantum query complexity.

On that note, please enjoy my new one-hour podcast on Spotify (if that link doesn’t work, try this one) with Matthew Putman of Utility+. Alas, my umming and ahhing were more frequent than I now aim for, but that’s partly compensated for by Matthew’s excellent decision to speed up the audio. This was an unusually wide-ranging interview, covering everything from SlateStarCodex to quantum gravity to interdisciplinary conferences to the challenges of teaching quantum computing to 7-year-olds. I hope you like it!

Update (Sep 1): Thanks for the comments, everyone! As you can see, I further revised this blog’s header based on the feedback and on further reflection.

The Right could only kill me and everyone I know.
The Left is scarier; it could convince me that it was my fault!

(In case you missed it on the blog’s revised header, right below “Quantum computers aren’t just nondeterministic Turing machines” and “Hold the November US election by mail.” I added an exclamation point at the end to suggest a slightly comic delivery.)

Update: A friend expressed concern that, because my new motto appears to “blame both sides,” it might generate confusion about my sympathies or what I want to happen in November. So to eliminate all ambiguity: I hereby announce that I will match all reader donations made in the next 72 hours to either the Biden-Harris campaign or the Lincoln Project, up to a limit of $2,000. Honor system; just tell me in the comments what you donated.

1. Good news, everyone! Following years of requests, this blog finally supports rich HTML and basic TeX in comments. Also, the German spam that used to plague the blog (when JavaScript was disabled) is gone. For all this, I owe deep gratitude to a reader and volunteer named Filip Dimitrovski.
   
2. Filip refused to accept any payment for fixing this blog. Instead, he asked only one favor: namely, that I use my platform to raise public awareness about the plight of the MAOI antidepressant Nardil. Filip tells me that, while tens of thousands of people desperately need Nardil—no other antidepressant ever worked for them—it’s become increasingly unavailable because the pharma companies can no longer make money on it. He points me to a SlateStarCodex post from 2015 that explains the problem in more detail (anyone else miss SlateStarCodex?). More recent links about the worsening crisis here, here, and here.
   
3. Here’s a fantastic interview of Bill Gates by Steven Levy, about the coronavirus debacle in the US. Gates, who’s always been notoriously and strategically nonpartisan, is more explicit than I’ve ever seen him before in explaining how the Trump administration’s world-historic incompetence led to where we are.
   
4. Speaking of which, here’s another excellent article, this one in The American Interest, about the results of “wargames” trying to simulate what happens in the extremely likely event that Trump contests a loss of the November election. Notably, the article sets out six steps that could be taken over the next few months to decrease the chance of a crisis next to which all the previous crises of 2020 will pale.
   
5. A reader asked me to share a link to an algorithm competition, related to cryptographic “proofs of time,” that ends on August 31. Apparently, my having shared a link to a predecessor of this competition—at the request of friend-of-the-blog Bram Cohen—played a big role in attracting good entries.
   
6. Huge congratulations to my former PhD student Shalev Ben-David, as well as Eric Blais, for co-winning the FOCS’2020 Best Paper Award—along with two other papers—for highly unconventional work about a new minimax theorem for randomized algorithms. (Ben-David and Blais also have a second FOCS paper, which applies their award paper to get the first tight composition theorem for randomized query complexity. Here’s the full list of FOCS papers—lots of great stuff, for a conference that of course won’t physically convene!) Anyway, a central idea in Ben-David and Blais’s new work is to use proper scoring rules to measure the performance of randomized algorithms—algorithms that now make statements like “I’m 90% sure that this is a yes-input,” rather than just outputting a 1-bit guess. Notably, Shalev tells me that he learned about proper scoring rules by reading rationalist blogs. So next time you lament your untold hours sacrificed to that pastime, remind yourself of where it once led!
   
7. What have I been up to lately? Besides Busy Beaver, hanging out with my kids, and just trying to survive? Mostly giving a lot of Zoom lectures! For those interested, here’s a Q&A that I recently did on the past and present of quantum computing, hosted by Andris Ambainis in Latvia. It did feel a bit surreal when my “interviewer” asked me to explain how I got into quantum computing research, and my answer was basically: “well, as you know, Andris, a lot of it started when I got hold of your seminal paper back in 1999…”

Like many academics, I’ve now been regularly “attending” conferences and giving talks via Zoom for four months. Naturally, I’ve learned a lot about how to use this platform—one that, despite numerous quirks and flaws, actually works well enough that it could probably replace at least 2/3 of in-person talks and meetings after the covid crisis is over. But one particular lesson is so important that I thought I’d make a public service announcement of it. So without further ado:

Email the link.

You know, the thing like

https://us02web.zoom.us/jblahblah

that you actually click to get to the actual conversation. Definitely email the link to the speaker (!). But also email it to whomever said they plan to attend. Resend the link between a day and an hour in advance, so that it doesn’t get buried, but turns up right away when people search their inboxes. If possible, put the link in every single email about the meeting or lecture. Even if you already sent the link for previous iterations of the meeting and it hasn’t changed, send it again. Don’t assume people will find the link on the web. Don’t make them click through five other links or open an attached PDF for it. Don’t send ten emails that explain every possible detail of the meeting except how to get to it. Just email the link. That’s all. Thanks!