Interview with Jeffrey Winicour, Richard Isaacson & Nigel Bishop

0:00 There we go, so the tape is on, and it's the 5th of November, about 2 in the afternoon, and I'm speaking with Nigel Bishop and Richard Ideson and Jeff Winiford, Roberta Bowman, and Bell. So I should probably explain briefly for your advantage that I'm sort of a former physicist who has been working on radiation reaction problems and relativity as it happens, and I'm in the process of converting to a historian of science. And actually what I'm doing at pretty much like sociology and I'm interested in recent and current work, theoretical work that's in one way related to the problem of gravitational wave detection and in particular for instance I'm interested in numerical work to study binary black vocal lessons and the aim to get templates for library. So I know from these papers that most of you have done work in this area. So I was interested first to get an idea of what's the current status of working in this area. I'm trying to remember when you were here the last time. Yes, it was early 1995. So where were we back in 1995? What was the last thing I said? But at the time, of course, we were mostly interested in earlier work, but the main thing that I remember at the end of that was that we'd be talking about the quadruple formula conferences, and you mentioned, you said that one thing that you thought that might be the outcome of current work, numerical relativity work in this area, would be that issues like quadruple forma in that would more or less become irrelevant. It wouldn't matter anymore, you know, what was a quadruple forma, what was a kind of At least when you get to the non, really not many of you in your perturbation description, you don't know where they were in the control condition. So hopefully someday, we're not there yet, it's been four years, but we haven't done that problem yet, but someday we'll be able to do the two-body problem and check the cortical form in some controlling regime, but then just be able to keep on tracing the orbits down to the, you know, all through the merger

2:30 and see how, see how it breaks down. Or see even the reason why it doesn't even make sense. But, but we're not there yet. I'm still trying to remember where we were back in 95. I think that was about the time the Grand Challenge was starting. No, but it was, in 95 we were, we just worked on the Scala-White equation, we hadn't at that stage really started on Geo, we were just talking about starting on Geo at the beginning of 95. So it was, really, we hadn't, so at that time we had no three-dimensional, we had no three-dimensional colors, there were no two-dimensional colors, no axisymmetric colors. It wasn't just, we were doing GR, we had done GR, so he had already finished an Axiosymmetra code back about that time, he got his PhD in 1904. This is Philip Papadopoulos. He's at Potsdam now. We would have done some things with cold. You know, we developed the cold, tested it, calibrated it, surely stable, did some trivial things with it, you know, some trivial waveforms. with pure vacuum space times, and I don't think we did much with black holes, but that's the time at which the Grand Challenge was started in this country, right about 9, 4, and then 5, and then 7, and then everything switched to 3D, so all of a sudden, maxi-symmetric codes were I think a lot of groups were in this position that people would just solve the axis symmetric problem of general physics not just our codes which are characteristic codes but the group at Cornell had a had a matter code Cauchy code the group at ncsa had another axisymmetric code vacuum code and they all and everyone would

5:00 have just done a lot of axisymmetric general relativity at that stage it wasn't for the grand challenge because the codes were all there and developed ready to know the physics of it And all of a sudden we got, because of this guy's instigation, the B&R bond that we're supposed to solve the binary black hole problem. This was the important thing to LIGO and you couldn't do that with an axisymmetric code. So we all, we all just stopped showing, wrapped up, wrapped up whatever there was left, you know, to milk in a few months out of axisymmetric systems and went on to developing food decodes. Some of the stuff, some of the stuff we did in axisymmetric systems and wrapping it up, I don't know if you, we talked about this, that was the science out of the kawana, reinterpreting how the event horizon belonged and you know, it's a bit of a business about topological censorship and rotating codes. Yeah, we didn't talk about it last time. What's that? We didn't talk about it last time. Yeah, so that was all some of the wrapping up things that happens, you know, with the axis and other stuff. But that didn't really require any new code development. That was just, you know, looking at the output and probably in post-processing It's a grand challenge. Yeah. It's a bad answer. Yeah. You know, and, uh, but I think it also showed what happens when you throw a lot of people together in a group, like the Grand Challenge, and forcing the work together, uh, that there are, besides the negative feed generated, there are some positive outcomes. And, you know, everyone looking at that problem, you know, from a different point of view, it because it wasn't clear at that stage before we all got together and thought about exactly what all that data meant but you know how you really what were the details of the pair of pants trousers picture with colliding about what it's always not all there what was the physics topological censorship issue how did that how did

7:30 that there on the toroidal horizons, you know, we found in rotating glass, that wasn't what we used to do either. So all these things became clarifying due to the fact that people were thrown together working on a grand challenge. Somewhere I think there's a quote from Hamming, isn't it, that the purpose of computation is insight, not numbers. Here's a case where they, you know, they have piles of numbers. Yeah, actually were able to get some brand new things that nobody had ever seen before out of it. So that was sort of an immediate, you know. Yeah, that was the early, uh, that was an early outcome of the branch balance. And then everyone, to some extent, went their own ways for a while, because, at least from point of view the other groups were all developing kosher codes but we were supposed to develop a characteristic code which we did in order to do the exterior problem so the idea was that the kosher codes would have a strong field of regions where the two like those would be would be located and then out around that region would be in some road to where the kosher solution would be matched to the characteristic solution and the characteristic solution would supply the boundary conditions the kosher solution would also carry off the radiation scry where you can do all the rigmarish stuff about waveform uh interpretation of it according to the theory so so you're having a false space line there's a good backification which is what like scry is on the grid right it's nice talking to this term So that, you know, so to some extent our group was working by, you know, family individually at that, you know, in the first few years of the Grand Challenge, just trying to develop a 3D characteristic code. And then the second stage of that effort, which is what we're involved in now, is matching the kosher code right okay so uh so as i just almost said the first part of that we fairly

10:00 must complete it although we've got we've got a 3d artistic code and it's a little bit calibrated and it's probably right now the you know the the best 3d code you know in existence including the I think we want to strip the kosher goals, at least temporarily, because we're able to deal, we're really able to describe any single black hole spacetime in that goal. And we've done a lot of problems of this sort, where we have to just go to a black hole and we evolve it stately for as long as we want, get the waveform. on you know uh you know completely general problems where the waveform is through polarization modes and been able to stop that so this is a this is a big step from going from where we were four years ago going on now that basically we you know we've jumped from the axe having the general axis symmetric problem solved and my characters flow now to having really the uh general single black hole problem yeah so and we'd love to be able to do the two black hole problem that way but you just can't do it that way but you need really to match to a coaching code to handle two black holes because one now quote can handle the natural male surface that's coming out from one black hole describe a single one but if you have two of them then you get pairs of you know two different families and no services that you can't handle with a single characteristic code so that's where we are now these guys are working and how it kind of matched to the ADM code, 3D ADM codes in the middle at the time period. I think that means that, you know, if we can do that, then we can stop doing this binary black hole code. And if the matching were done at this stage, right now with the present ADM codes, if we could get this matching accomplished, we would be able to start getting waveforms from two

12:30 black holes. Now, it would be hard to say how long you'd be able to evolve that system right now. But we know that we can evolve a single black hole for as long as we want. And it's just a question of how well this matching is going to work. The whole issue of boundary conditions for Cauchy evolution is a problem that goes well beyond general relativity because of all hyperbolic systems and fluid dynamics is probably the one that most work is done on in any other system there are more fluid dynamics than from general to this but long shot so that this is a rather uh live widely worked on problem and uh and there is important problem especially the nominee of problems where you can't choose a an analytic approximation of what the exterior region should look like in order to put on some world to boundary condition of the problem. It's a linear problem you can basically try to do some kind of a linear approximation. Approximation to the linear solution here is something you can't you probably can't solve exactly you can do some far field approximation but in the non-linear problem you can't even do that so best techniques that have been developed in computational fluid dynamics are all based upon linearizing the problem at least in the region outside of which you're you're doing the numerical evolution so they basically take the numerical evolution and treat that as a linear problem, as a nonlinear problem, a fully nonlinear problem, but then linearize the exterior. And that is still quite a task to solve that problem correctly, even making that linearized approximation to the exterior because you've got to make sure that waves coming from the nonlinear interior when the Hitcher boundary condition that you've constructed by this linearization don't get reflected back in but keep on going back straight out.

15:00 It turns out that the methods really necessary to do that are very difficult, not just numerically or computationally, but just conceptually. Mathematically, they're very deep methods. It's a global solution of a linear problem that's just an asymptotic one over our expansion type solution in order to do it right. It turns out you know the determinant problem, doing it that way. So, what we've done by this technique of matching onto a characteristic solution in the exterior, it's really created another approach to that problem where you don't have familiar ones. And it's an approach that hadn't been considered anywhere else in the literature, you know, even by the fluid dynamics of the literature. So it was something that was created within general activity as a new approach. And we worked this out first for the scale-of-wave problem, nonlinear scale-of-wave. say it takes an equation like box feed plus feed the water some potential now the potential in the coastal space and solve that problem by the characteristic matching technique and we were able to do that food 3d numerical solutions matching crochet solutions to that product with characteristic solutions to the outside and and we're able to you know compare this to exact solutions of the problem well we bought you know we found that there was no black reflection that everything everything converged to numerical work you worry about whether your solution converges to the analytic solution when your grid size shrinks to zero so that was all checked that we could get accurate and on top of that we found that these methods even are more efficient than the traditional coaching boundary conditions that people like the computation of fluid dynamics were applying problems like that that putting on this compassion this characteristic exterior at a given accuracy that you want to attain would lead to a more efficient way of computing

17:30 solution than the conventional techniques in the literature so we had something that you know had a couple pluses to it it was efficient and on top of that it gave you the right answer not the right answer you know so so this was a so this is a you know a big arguing point for the methodology at least to do this in general relativity but the general relativity problem is much more complicated because you're a scalar problem you're matching one field and in general activity you're matching a tensor field and you have to imagine a lot of derivatives of that field that have to be matched properly and so it's a much more complicated problem and uh as these guys will not contradict me on that statement i'm sure and it's also that in relativity and the way in the scale of why the coordinates are fixed you know what coordinate transformations are in relativity they're not nigel in fact he was one of the guys that got me investigating into doing this project originally we were doing characteristic evolution and we thought you know we thought that was a would be happy to solve that yeah you know we weren't really worried about what we'd use it for at the time we thought we didn't worry about that later but then i just came along with this idea that you should match it to a post evolution rather than trying to do an entire space time this way that you know it made more sense to try to try to do an entire space time like the binary black hole space time by matching you know breaking it up into pieces and solving each piece with the code that's appropriate to that piece and so uh so that that that's what we're uh percy and i think we're getting a symptom right now as i said of course because because the gravitational version of that problem has turned out to be much harder than I read a paper in 93 describing the theory of doing this, but now it sounds good. You mentioned solving the different parts based on the code appropriate to that. Are

20:00 the codes for a case like that very much integrated together or is it a case of, are the codes somehow run in parallel or is it more that they're going like you run one code to get the input for the next code? They run in parallel yeah the evolutions keep in step with each other so you update the co-sheet and then you update the characteristic and then so one you're updating right on a slant at 45 degrees and one like this but you're going out now like this one after another So basically, this whole infrastructure is there now, there's a code right in there. Somewhere in that room, I don't know what physical form, there's a code that does this. It has modules that consist of an ADM module, evolution module, and a characteristic evolution module, and another module which is a matching module that interfaces them. So this thing does exactly that. And you can put in data and evolve it, but just that it goes unstable. it so that it doesn't give a useful you know solution solution for various for a long time so it's the state is to stabilize in this code that is a nightmare right now a real challenge uh you know and you know to talk to enough people in america who are going to request a common nightmare so you come up with this wonderful algorithm at the intellectual level that you think it's been a solid problem. And then you go, you go and write a code, which isn't much fun. It's not as much fun as thinking of the art of it, but it's fairly straightforward. And then you debug it, which is straightforward, and then you come to the novel where it's just... Extremely painful. And then you come to this, after all that, you know, that, you find this thing just blows up on you. Because somehow the iteration scheme has some instability in it. And that's always instantly staged to get yourself through, if you're dealing with new equations. So at that stage when you're trying to stabilize, is that more like debugging or is it more

22:30 like having to tinker with the algorithm? It's more like tinkering with the algorithm. Once you're sure you've got to debug, you're never sure you've finished the debugging. so we try to set things up in a way in which we're still checking into the buggy, but at the same time there are ways to improve the algorithm. It's not really the, you know, it's not really the hot, you know, the macroscopic structure of the algorithm, right? It's these little things about how you pick your finite different stencils. Just a little thing about how you do an interpolation, which there might be a hundred different variations of choices you could make. Out of those, maybe only two out of a hundred will lead to a stable algorithm. It's just, you're in a problem with a lot of possibilities, and normally there aren't too many that work. Let me even step back from the details, there's been some reference to this great challenge, so let me say a few things about what that is and how it arose. The National Science Foundation, and particularly the Computer Science Directorate of the Foundation, was interested in high-performance computing, as was all the other government funding agencies. And there was a program to try and anticipate that there would be, you know, over a five-year period, tremendous step forward, a thousand, a factor of a thousand in advance, in any parameter in the computer speed, memory, you know. So, a bunch of government agencies were putting money into trying to get the scientific community moving to anticipate the uh these kinds of changes and trying out things and maybe providing feedback to the manufacturers about what kind of architectures might work um of course industry you know it's driven by a lot of things but at least having some some very large problems would would be helpful to them so various programs in the national science foundation One of the things I put on most of the things was to set up a competition to create grand

25:00 challenges in a whole variety of fields, physics, astronomy, biology, everything you can imagine, that would utilize the expected power and architectures of these Cubing and Terraforming machines. And there was a competition across all areas of science for, you know, what, what, what kind of, uh, project should be supported by the government, and, uh, there were, you know, several hundred proposals and then just a handful, uh, I don't know, I don't know, I've already had a dozen that went across all of these areas of science. Yeah. And so this amazing problem in general activity, the basic new body problem in the field, was something that captured the imagination and its view as something that would be well-matched to the expected advances in the computing side and well-matched to the frontiers of science. So it was selected out of this. And because this was funded by the computer science people and they had a strong interest in new computer architectures and tools and things other than just getting out of physics, they were, they really were delighted and they encouraged collaborations of computer scientists and the computer scientists were involved in developing new source principles to help these problems run on these large massive parallel machines. So that's another part of the effort. You don't see represented in this room, but it's I think the two centers really were Circus and Texas. Brown at Weston and Fox is really good at Circus.

27:30 I was going to ask, since I do have a group of American people and parents that have worked in that area, how does the work get divided up, especially in problems like this was for a large number of people working on it. I don't know if there's any set procedure for doing that, I know if you know, I can tell you how our group works, but I think if you go and just look at people in computational areas, the research they do, there's some people that are numerical people in applied mathematics, never write a program, never write a program, they just come up with algorithms. You know, they just work by themselves to come up with algorithms and try and think mathematical theorems about algorithms being stable. You know, they're called algorithm people. And then other people actually, you know, take algorithms and run programs. You know, try to get phenomenology out of them. You know, so they do simulations. You know, there's a whole gamut of how it could be done. So, I think in our case, we've more, you know, we've got a little bit of everything here, and every one of us, okay, but I'm not as, you know, I couldn't efficiently submit a program to these, you know, supercomputer sentences, say, Roberto or Bella, and they have the, you know, they have the expertise of that. So, they have the edge over me, they can do the theory. but I can't do the the queuing finding compiler errors this is a I don't know you should talk to these guys get their point of view of how they think that should be done now that's an interesting topic but they're more expert on it than I am you know as part of this grand challenge The main computational effort in developing the computational techniques, what we said these guys at Brown, at Texas, and a group under Fox in Syracuse developed was something

30:00 called DAG-H. This is the binary black code grand challenge contribution to computer science. You've heard of this? I've heard of it. Yeah. you know okay so that's this uh very complicated parallel processing type software and i don't know how many people in a grand challenge really know how it works we're all supposed to be using it eventually but how many people really know how it works uh maybe about one out of one out of four and it's not more than one out of four you know you know so that part you know the software part of it really gets some you know it's been put on especially how many how many people know it here they know what activity they're going to be part of the reason for that is that science foundation hasn't provided the computers that were expected to come at the end of this five year period.