Tag Archives: DoingScience

Zoomplitudes Retrospective

During Zoomplitudes (my field’s big yearly conference, this year on Zoom) I didn’t have time to write a long blog post. I said a bit about the format, but didn’t get a chance to talk about the science. I figured this week I’d go back and give a few more of my impressions. As always, conference posts are a bit more technical than my usual posts, so regulars be warned!

The conference opened with a talk by Gavin Salam, there as an ambassador for LHC physics. Salam pointed out that, while a decent proportion of speakers at Amplitudes mention the LHC in their papers, that fraction has fallen over the years. (Another speaker jokingly wondered which of those mentions were just in the paper’s introduction.) He argued that there is still useful work for us, LHC measurements that will require serious amplitudes calculations to understand. He also brought up what seems like the most credible argument for a new, higher-energy collider: that there are important properties of the Higgs, in particular its interactions, that we still have not observed.

The next few talks hopefully warmed Salam’s heart, as they featured calculations for real-world particle physics. Nathaniel Craig and Yael Shadmi in particular covered the link between amplitudes and Standard Model Effective Field Theory (SMEFT), a method to systematically characterize corrections beyond the Standard Model. Shadmi’s talk struck me because the kind of work she described (building the SMEFT “amplitudes-style”, directly from observable information rather than more complicated proxies) is something I’d seen people speculate about for a while, but which hadn’t been done until quite recently. Now, several groups have managed it, and look like they’ve gotten essentially “all the way there”, rather than just partial results that only manage to replicate part of the SMEFT. Overall it’s much faster progress than I would have expected.

After Shadmi’s talk was a brace of talks on N=4 super Yang-Mills, featuring cosmic Galois theory and an impressively groan-worthy “origin story” joke. The final talk of the day, by Hofie Hannesdottir, covered work with some of my colleagues at the NBI. Due to coronavirus I hadn’t gotten to hear about this in person, so it was good to hear a talk on it, a blend of old methods and new priorities to better understand some old discoveries.

The next day focused on a topic that has grown in importance in our community, calculations for gravitational wave telescopes like LIGO. Several speakers focused on new methods for collisions of spinning objects, where a few different approaches are making good progress (Radu Roiban’s proposal to use higher-spin field theory was particularly interesting) but things still aren’t quite “production-ready”. The older, post-Newtonian method is still very much production-ready, as evidenced by Michele Levi’s talk that covered, among other topics, our recent collaboration. Julio Parra-Martinez discussed some interesting behavior shared by both supersymmetric and non-supersymmetric gravity theories. Thibault Damour had previously expressed doubts about use of amplitudes methods to answer this kind of question, and part of Parra-Martinez’s aim was to confirm the calculation with methods Damour would consider more reliable. Damour (who was actually in the audience, which I suspect would not have happened at an in-person conference) had already recanted some related doubts, but it’s not clear to me whether that extended to the results Parra-Martinez discussed (or whether Damour has stated the problem with his old analysis).

There were a few talks that day that didn’t relate to gravitational waves, though this might have been an accident, since both speakers also work on that topic. Zvi Bern’s talk linked to the previous day’s SMEFT discussion, with a calculation using amplitudes methods of direct relevance to SMEFT researchers. Clifford Cheung’s talk proposed a rather strange/fun idea, conformal symmetry in negative dimensions!

Wednesday was “amplituhedron day”, with a variety of talks on positive geometries and cluster algebras. Featured in several talks was “tropicalization“, a mathematical procedure that can simplify complicated geometries while still preserving essential features. Here, it was used to trim down infinite “alphabets” conjectured for some calculations into a finite set, and in doing so understand the origin of “square root letters”. The day ended with a talk by Nima Arkani-Hamed, who despite offering to bet that he could finish his talk within the half-hour slot took almost twice that. The organizers seemed to have planned for this, since there was one fewer talk that day, and as such the day ended at roughly the usual time regardless.

We also took probably the most unique conference photo I will ever appear in.

For lack of a better name, I’ll call Thursday’s theme “celestial”. The day included talks by cosmologists (including approaches using amplitudes-ish methods from Daniel Baumann and Charlotte Sleight, and a curiously un-amplitudes-related talk from Daniel Green), talks on “celestial amplitudes” (amplitudes viewed from the surface of an infinitely distant sphere), and various talks with some link to string theory. I’m including in that last category intersection theory, which has really become its own thing. This included a talk by Simon Caron-Huot about using intersection theory more directly in understanding Feynman integrals, and a talk by Sebastian Mizera using intersection theory to investigate how gravity is Yang-Mills squared. Both gave me a much better idea of the speakers’ goals. In Mizera’s case he’s aiming for something very ambitious. He wants to use intersection theory to figure out when and how one can “double-copy” theories, and might figure out why the procedure “got stuck” at five loops. The day ended with a talk by Pedro Vieira, who gave an extremely lucid and well-presented “blackboard-style” talk on bootstrapping amplitudes.

Friday was a grab-bag of topics. Samuel Abreu discussed an interesting calculation using the numerical unitarity method. It was notable in part because renormalization played a bigger role than it does in most amplitudes work, and in part because they now have a cool logo for their group’s software, Caravel. Claude Duhr and Ruth Britto gave a two-part talk on their work on a Feynman integral coaction. I’d had doubts about the diagrammatic coaction they had worked on in the past because it felt a bit ad-hoc. Now, they’re using intersection theory, and have a clean story that seems to tie everything together. Andrew McLeod talked about our work on a Feynman diagram Calabi-Yau “bestiary”, while Cristian Vergu had a more rigorous understanding of our “traintrack” integrals.

There are two key elements of a conference that are tricky to do on Zoom. You can’t do a conference dinner, so you can’t do the traditional joke-filled conference dinner speech. The end of the conference is also tricky: traditionally, this is when everyone applauds the organizers and the secretaries are given flowers. As chair for the last session, Lance Dixon stepped up to fill both gaps, with a closing speech that was both a touching tribute to the hard work of organizing the conference and a hilarious pile of in-jokes, including a participation award to Arkani-Hamed for his (unprecedented, as far as I’m aware) perfect attendance.

The Sum of Our Efforts

I got a new paper out last week, with Andrew McLeod, Henrik Munch, and Georgios Papathanasiou.

A while back, some collaborators and I found an interesting set of Feynman diagrams that we called “Omega”. These Omega diagrams were fun because they let us avoid one of the biggest limitations of particle physics: that we usually have to compute approximations, diagram by diagram, rather than finding an exact answer. For these Omegas, we figured out how to add all the infinite set of Omega diagrams up together, with no approximation.

One implication of this was that, in principle, we now knew the answer for each individual Omega diagram, far past what had been computed before. However, writing down these answers was easier said than done. After some wrangling, we got the answer for each diagram in terms of an infinite sum. But despite tinkering with it for a while, even our resident infinite sum expert Georgios Papathanasiou couldn’t quite sum them up.

Naturally, this made me think the sums would make a great Master’s project.

When Henrik Munch showed up looking for a project, Andrew McLeod and I gave him several options, but he settled on the infinite sums. Impressively, he ended up solving the problem in two different ways!

First, he found an old paper none of us had seen before, that gave a general method for solving that kind of infinite sum. When he realized that method was really annoying to program, he took the principle behind it, called telescoping, and came up with his own, simpler method, for our particular case.

Picture an old-timey folding telescope. It might be long when fully extended, but when you fold it up each piece fits inside the previous one, resulting in a much smaller object. Telescoping a sum has the same spirit. If each pair of terms in a sum “fit together” (if their difference is simple), you can rearrange them so that most of the difficulty “cancels out” and you’re left with a much simpler sum.

Henrik’s telescoping idea worked even better than expected. We found that we could do, not just the Omega sums, but other sums in particle physics as well. Infinite sums are a very well-studied field, so it was interesting to find something genuinely new.

The rest of us worked to generalize the result, to check the examples and to put it in context. But the core of the work was Henrik’s. I’m really proud of what he accomplished. If you’re looking for a PhD student, he’s on the market!

The Point of a Model

I’ve been reading more lately, partially for the obvious reasons. Mostly, I’ve been catching up on books everyone else already read.

One such book is Daniel Kahneman’s “Thinking, Fast and Slow”. With all the talk lately about cognitive biases, Kahneman’s account of his research on decision-making was quite familiar ground. The book turned out to more interesting as window into the culture of psychology research. While I had a working picture from psychologist friends in grad school, “Thinking, Fast and Slow” covered the other side, the perspective of a successful professor promoting his field.

Most of this wasn’t too surprising, but one passage struck me:

Several economists and psychologists have proposed models of decision making that are based on the emotions of regret and disappointment. It is fair to say that these models have had less influence than prospect theory, and the reason is instructive. The emotions of regret and disappointment are real, and decision makers surely anticipate these emotions when making their choices. The problem is that regret theories make few striking predictions that would distinguish them from prospect theory, which has the advantage of being simpler. The complexity of prospect theory was more acceptable in the competition with expected utility theory because it did predict observations that expected utility theory could not explain.

Richer and more realistic assumptions do not suffice to make a theory successful. Scientists use theories as a bag of working tools, and they will not take on the burden of a heavier bag unless the new tools are very useful. Prospect theory was accepted by many scholars not because it is “true” but because the concepts that it added to utility theory, notably the reference point and loss aversion, were worth the trouble; they yielded new predictions that turned out to be true. We were lucky.

Thinking Fast and Slow, page 288

Kahneman is contrasting three theories of decision making here: the old proposal that people try to maximize their expected utility (roughly, the benefit they get in future), his more complicated “prospect theory” that takes into account not only what benefits people get but their attachment to what they already have, and other more complicated models based on regret. His theory ended up more popular, both than the older theory and than the newer regret-based models.

Why did his theory win out? Apparently, not because it was the true one: as he says, people almost certainly do feel regret, and make decisions based on it. No, his theory won because it was more useful. It made new, surprising predictions, while being simpler and easier to use than the regret-based models.

This, a theory defeating another without being “more true”, might bug you. By itself, it doesn’t bug me. That’s because, as a physicist, I’m used to the idea that models should not just be true, but useful. If we want to test our theories against reality, we have a large number of “levels” of description to choose from. We can “zoom in” to quarks and gluons, or “zoom out” to look at atoms, or molecules, or polymers. We have to decide how much detail to include, and we have real pragmatic reasons for doing so: some details are just too small to measure!

It’s not clear Kahneman’s community was doing this, though. That is, it doesn’t seem like he’s saying that regret and disappointment are just “too small to be measured”. Instead, he’s saying that they don’t seem to predict much differently from prospect theory, and prospect theory is simpler to use.

Ok, we do that in physics too. We like working with simpler theories, when we have a good excuse. We’re just careful about it. When we can, we derive our simpler theories from more complicated ones, carving out complexity and estimating how much of a difference it would have made. Do this carefully, and we can treat black holes as if they were subatomic particles. When we can’t, we have what we call “phenomenological” models, models built up from observation and not from an underlying theory. We never take such models as the last word, though: a phenomenological model is always viewed as temporary, something to bridge a gap while we try to derive it from more basic physics.

Kahneman doesn’t seem to view prospect theory as temporary. It doesn’t sound like anyone is trying to derive it from regret theory, or to make regret theory easier to use, or to prove it always agrees with regret theory. Maybe they are, and Kahneman simply doesn’t think much of their efforts. Either way, it doesn’t sound like a major goal of the field.

That’s the part that bothered me. In physics, we can’t always hope to derive things from a more fundamental theory, some theories are as fundamental as we know. Psychology isn’t like that: any behavior people display has to be caused by what’s going on in their heads. What Kahneman seems to be saying here is that regret theory may well be closer to what’s going on in people’s heads, but he doesn’t care: it isn’t as useful.

And at that point, I have to ask: useful for what?

As a psychologist, isn’t your goal ultimately to answer that question? To find out “what’s going on in people’s heads”? Isn’t every model you build, every theory you propose, dedicated to that question?

And if not, what exactly is it “useful” for?

For technology? It’s true, “Thinking Fast and Slow” describes several groups Kahneman advised, most memorably the IDF. Is the advantage of prospect theory, then, its “usefulness”, that it leads to better advice for the IDF?

I don’t think that’s what Kahneman means, though. When he says “useful”, he doesn’t mean “useful for advice”. He means it’s good for giving researchers ideas, good for getting people talking. He means “useful for designing experiments”. He means “useful for writing papers”.

And this is when things start to sound worryingly familiar. Because if I’m accusing Kahneman’s community of giving up on finding the fundamental truth, just doing whatever they can to write more papers…well, that’s not an uncommon accusation in physics as well. If the people who spend their lives describing cognitive biases are really getting distracted like that, what chance does, say, string theory have?

I don’t know how seriously to take any of this. But it’s lurking there, in the back of my mind, that nasty, vicious, essential question: what are all of our models for?

Bonus quote, for the commenters to have fun with:

I have yet to meet a successful scientist who lacks the ability to exaggerate the importance of what he or she is doing, and I believe that someone who lacks a delusional sense of significance will wilt in the face of repeated experiences of multiple small failures and rare successes, the fate of most researchers.

Thinking Fast and Slow, page 264

Thoughts on Doing Science Remotely

In these times, I’m unusually lucky.

I’m a theoretical physicist. I don’t handle goods, or see customers. Other scientists need labs, or telescopes: I just need a computer and a pad of paper. As a postdoc, I don’t even teach. In the past, commenters have asked me why I don’t just work remotely. Why go to conferences, why even go to the office?

With COVID-19, we’re finding out.

First, the good: my colleagues at the Niels Bohr Institute have been hard at work keeping everyone connected. Our seminars have moved online, where we hold weekly Zoom seminars jointly with Iceland, Uppsala and Nordita. We have a “virtual coffee room”, a Zoom room that’s continuously open with “virtual coffee breaks” at 10 and 3:30 to encourage people to show up. We’re planning virtual colloquia, and even a virtual social night with Jackbox games.

Is it working? Partially.

The seminars are the strongest part. Remote seminars let us bring in speakers from all over the world (time zones permitting). They let one seminar serve the needs of several different institutes. Most of the basic things a seminar needs (slides, blackboards, ability to ask questions, ability to clap) are present on online platforms, particularly Zoom. And our seminar organizers had the bright idea to keep the Zoom room open after the talk, which allows the traditional “after seminar conversation with the speaker” for those who want it.

Still, the setup isn’t as good as it could be. If the audience turns off their cameras and mics, the speaker can feel like they’re giving a talk to an empty room. This isn’t just awkward, it makes the talk worse: speakers improve when they can “feel the room” and see what catches their audience’s interest. If the audience keeps their cameras or mics on instead, it takes a lot of bandwidth, and the speaker still can’t really feel the room. I don’t know if there’s a good solution here, but it’s worth working on.

The “virtual coffee room” is weaker. It was quite popular at first, but as time went on fewer and fewer people (myself included) showed up. In contrast, my wife’s friends at Waterloo do a daily cryptic crossword, and that seems to do quite well. What’s the difference? They have real crosswords, we don’t have real coffee.

I kid, but only a little. Coffee rooms and tea breaks work because of a core activity, a physical requirement that brings people together. We value them for their social role, but that role on its own isn’t enough to get us in the door. We need the excuse: the coffee, the tea, the cookies, the crossword. Without that shared structure, people just don’t show up.

Getting this kind of thing right is more important than it might seem. Social activities help us feel better, they help us feel less isolated. But more than that, they help us do science better.

That’s because science works, at least in part, through serendipity.

You might think of scientific collaboration as something we plan, and it can be sometimes. Sometimes we know exactly what we’re looking for: a precise calculation someone else can do, a question someone else can answer. Sometimes, though, we’re helped by chance. We have random conversations, different people in different situations, coffee breaks and conference dinners, and eventually someone brings up an idea we wouldn’t have thought of on our own.

Other times, chance helps by providing an excuse. I have a few questions rattling around in my head that I’d like to ask some of my field’s big-shots, but that don’t feel worth an email. I’ve been waiting to meet them at a conference instead. The advantage of those casual meetings is that they give an excuse for conversation: we have to talk about something, it might as well be my dumb question. Without that kind of causal contact, it feels a lot harder to broach low-stakes topics.

None of this is impossible to do remotely. But I think we need new technology (social or digital) to make it work well. Serendipity is easy to find in person, but social networks can imitate it. Log in to facebook or tumblr looking for your favorite content, and you face a pile of ongoing conversations. Looking through them, you naturally “run into” whatever your friends are talking about. I could see something similar for academia. Take something like the list of new papers on arXiv, then run a list of ongoing conversations next to it. When we check the arXiv each morning, we could see what our colleagues were talking about, and join in if we see something interesting. It would be a way to stay connected that would keep us together more, giving more incentive and structure beyond simple loneliness, and lead to the kind of accidental meetings that science craves. You could even graft conferences on to that system, talks in the middle with conversation threads on the side.

None of us know how long the pandemic will last, or how long we’ll be asked to work from home. But even afterwards, it’s worth thinking about the kind of infrastructure science needs to work remotely. Some ideas may still be valuable after all this is over.

4gravitons, Spinning Up

I had a new paper out last week, with Michèle Levi and Andrew McLeod. But to explain it, I’ll need to clarify something about our last paper.

Two weeks ago, I told you that Andrew and Michèle and I had written a paper, predicting what gravitational wave telescopes like LIGO see when black holes collide. You may remember that LIGO doesn’t just see colliding black holes: it sees colliding neutron stars too. So why didn’t we predict what happens when neutron stars collide?

Actually, we did. Our calculation doesn’t just apply to black holes. It applies to neutron stars too. And not just neutron stars: it applies to anything of roughly the right size and shape. Black holes, neutron stars, very large grapefruits…

LIGO’s next big discovery

That’s the magic of Effective Field Theory, the “zoom lens” of particle physics. Zoom out far enough, and any big, round object starts looking like a particle. Black holes, neutron stars, grapefruits, we can describe them all using the same math.

Ok, so we can describe both black holes and neutron stars. Can we tell the difference between them?

In our last calculation, no. In this one, yes!

Effective Field Theory isn’t just a zoom lens, it’s a controlled approximation. That means that when we “zoom out” we don’t just throw out anything “too small to see”. Instead, we approximate it, estimating how big of an effect it can have. Depending on how precise we want to be, we can include more and more of these approximated effects. If our estimates are good, we’ll include everything that matters, and get a good approximation for what we’re trying to observe.

At the precision of our last calculation, a black hole and a neutron star still look exactly the same. Our new calculation aims for a bit higher precision though. (For the experts: we’re at a higher order in spin.) The higher precision means that we can actually see the difference: our result changes for two colliding black holes versus two colliding grapefruits.

So does that mean I can tell you what happens when two neutron stars collide, according to our calculation? Actually, no. That’s not because we screwed up the calculation: it’s because some of the properties of neutron stars are unknown.

The Effective Field Theory of neutron stars has what we call “free parameters”, unknown variables. People have tried to estimate some of these (called “Love numbers” after the mathematician A. E. H. Love), but they depend on the details of how neutron stars work: what stuff they contain, how that stuff is shaped, and how it can move. To find them out, we probably can’t just calculate: we’ll have to measure, observe an actual neutron star collision and see what the numbers actually are.

That’s one of the purposes of gravitational wave telescopes. It’s not (as far as I know) something LIGO can measure. But future telescopes, with more precision, should be able to. By watching two colliding neutron stars and comparing to a high-precision calculation, physicists will better understand what those neutron stars are made of. In order to do that, they will need someone to do that high-precision calculation. And that’s why people like me are involved.

4gravitons Exchanges a Graviton

I had a new paper up last Friday with Michèle Levi and Andrew McLeod, on a topic I hadn’t worked on before: colliding black holes.

I am an “amplitudeologist”. I work on particle physics calculations, computing “scattering amplitudes” to find the probability that fundamental particles bounce off each other. This sounds like the farthest thing possible from black holes. Nevertheless, the two are tightly linked, through the magic of something called Effective Field Theory.

Effective Field Theory is a kind of “zoom knob” for particle physics. You “zoom out” to some chosen scale, and write down a theory that describes physics at that scale. Your theory won’t be a complete description: you’re ignoring everything that’s “too small to see”. It will, however, be an effective description: one that, at the scale you’re interested in, is effectively true.

Particle physicists usually use Effective Field Theory to go between different theories of particle physics, to zoom out from strings to quarks to protons and neutrons. But you can zoom out even further, all the way out to astronomical distances. Zoom out far enough, and even something as massive as a black hole looks like just another particle.

Just click the “zoom X10” button fifteen times, and you’re there!

In this picture, the force of gravity between black holes looks like particles (specifically, gravitons) going back and forth. With this picture, physicists can calculate what happens when two black holes collide with each other, making predictions that can be checked with new gravitational wave telescopes like LIGO.

Researchers have pushed this technique quite far. As the calculations get more and more precise (more and more “loops”), they have gotten more and more challenging. This is particularly true when the black holes are spinning, an extra wrinkle in the calculation that adds a surprising amount of complexity.

That’s where I came in. I can’t compete with the experts on black holes, but I certainly know a thing or two about complicated particle physics calculations. Amplitudeologists, like Andrew McLeod and me, have a grab-bag of tricks that make these kinds of calculations a lot easier. With Michèle Levi’s expertise working with spinning black holes in Effective Field Theory, we were able to combine our knowledge to push beyond the state of the art, to a new level of precision.

This project has been quite exciting for me, for a number of reasons. For one, it’s my first time working with gravitons: despite this blog’s name, I’d never published a paper on gravity before. For another, as my brother quipped when he heard about it, this is by far the most “applied” paper I’ve ever written. I mostly work with a theory called N=4 super Yang-Mills, a toy model we use to develop new techniques. This paper isn’t a toy model: the calculation we did should describe black holes out there in the sky, in the real world. There’s a decent chance someone will use this calculation to compare with actual data, from LIGO or a future telescope. That, in particular, is an absurdly exciting prospect.

Because this was such an applied calculation, it was an opportunity to explore the more applied part of my own field. We ended up using well-known techniques from that corner, but I look forward to doing something more inventive in future.

Understanding Is Translation

Kernighan’s Law states, “Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it.” People sometimes make a similar argument about philosophy of mind: “The attempt of the mind to analyze itself [is] an effort analogous to one who would lift himself by his own bootstraps.”

Both points operate on a shared kind of logic. They picture understanding something as modeling it in your mind, with every detail clear. If you’ve already used all your mind’s power to design code, you won’t be able to model when it goes wrong. And modeling your own mind is clearly nonsense, you would need an even larger mind to hold the model.

The trouble is, this isn’t really how understanding works. To understand something, you don’t need to hold a perfect model of it in your head. Instead, you translate it into something you can more easily work with. Like explanations, these translations can be different for different people.

To understand something, I need to know the algorithm behind it. I want to know how to calculate it, the pieces that go in and where they come from. I want to code it up, to test it out on odd cases and see how it behaves, to get a feel for what it can do.

Others need a more physical picture. They need to know where the particles are going, or how energy and momentum are conserved. They want entropy to be increased, action to be minimized, scales to make sense dimensionally.

Others in turn are more mathematical. They want to start with definitions and axioms. To understand something, they want to see it as an example of a broader class of thing, groups or algebras or categories, to fit it into a bigger picture.

Each of these are a kind of translation, turning something into code-speak or physics-speak or math-speak. They don’t require modeling every detail, but when done well they can still explain every detail.

So while yes, it is good practice not to write code that is too “smart”, and too hard to debug…it’s not impossible to debug your smartest code. And while you can’t hold an entire mind inside of yours, you don’t actually need to do that to understand the brain. In both cases, all you need is a translation.