Interview with Bangalore Sathyaprakash (2nd)

0:00 I think it looks like it's working fine, and it's the 3rd of May at 20 past 2 in the afternoon, the year 2000, and I'm speaking with Professor Sathya Bakesh. So, I think the main thing that I didn't have a chance to ask you about when we spoke the last time was about PADI approximants and their role in data analysis and I guess maybe you could just start by telling me what the main need for something like the paleo-approximance is for the diagnosis problem in the case of geo-enliven. Right. I think when we started looking at the chance of detecting binary in spiral signals, we realized that there's not a fantastic chance, especially with initial interferometers, of So we decided that we should do the best job, as far as filtering is concerned, making use of all the available theoretical tools. When we examine what's available, then it is the usual second post-Newtonian waveforms that were available. and if you use them, then our experience with test mass going around the Swastrad black hole showed that the post-Tutonian approximants wouldn't do very well. Let me clarify what I mean by that. We are talking about massive binary systems here of the order of 10 to 20 or 30 solo masses. In that case, the binary spends the last few seconds to milliseconds of its life right in the heart of the LIGO bandwidth or geo-bandwidth, where the detector is most sensitive, which means that our signal models will have to be pretty accurate in that region. You know, the signal is inspiring and finally ends up right in the best sensitivity part of LIGO and GEO. If it continues on, like in the neutron star, neutron star case, and ends up its life much later, at a much higher frequency, then it doesn't matter.

2:30 But for massive systems such as 20 solar mass objects, anywhere between 10 to 30, 20 being somewhere in the middle, then the system ends up its life right in the part where the detector is more sensitive and it is very close to the last stable orbit. So we want to have a good hang of what is the last stable orbit, when the signal is going to terminate and a good model for the evolution of the system at that stage. It turned out that the existing post-Otonian predictions were not very good when we compared those post-Newtonian predictions with the exact evolution in the case of a test particle orbiting a soft-shell black hole. So that led us to think and try to improve those models. And the theoretical foundation for that, we started developing, I think, about five years back. We looked at energy and flux functions, which are the formulas, which are the physical quantities which enter the phasing formula. They had certain properties which couldn't have been mimicked by a post-Newtonian series. For instance, both energy and flux have a pole at the light ring, which will never be mimicked by a post-Newtonian series. If you have post-Newtonian series for these two physical quantities, they will never catch a pole. So we thought that we have to have party approximants for these things, which are equivalent to these post-Hutonian approximants to the order at which we are exploring. But then they can catch pole because they are rational polynomials. So we were not the first. Other people also considered body approximants. We know from private communication they were looking at this. But then the crucial point was that if you are trying to approximate a quantity which has a pole and a branch cut as well, succeed with the body approximates so the crucial thing was to go and define more basic quantities both for flux and energy and then use them to construct new physical quantities and and then use those new physical quantities to arrive at new waveforms so that's in a just body approximates now the second part of

5:00 the answer is how much are we going to lose i mean if we don't use the body approximants but something like post newtonian approximants how much are we going to lose so i'm going to give you a rough number here if we if it's necessary we can go into the you know actual exact details i had a paper right on my table because i was working on it okay for massive systems which are the most important i think we will lose something like between 10-12% of the total signal-to-last ratio which when translated into number of events is something like 50% of the events. So if you have a chance of detecting let's say 10 events, you may detect only 5 and since we have a pretty low chance of detecting our own viewpoint in this is that we must maximize the chance of detecting, therefore use the best available models. What body approximants? Now, they can't be called just body approximants because there's something more that has gone into it, namely definition of new physical quantities. That's why we call them P approximants. So these P approximants will help us to gain that extra 10% of the signal-to-nose ratio, making it, you know, we can recover all the events that may be taking place within the span of G and Virgo and LIGO. That's what we're going to do. The... I don't know, apart from having heard you spoke about it before and having read, say, the numerical recipes chapter on patty approximants, I don't know much about them. I remember what Numerical Recipes kind of says is that there's something mysterious about the convergence of Paddy series that, you know, as it were, it somehow, in some way that you don't have guesses where the thing is going to converge to. And obviously, that's ideal in a situation where we don't actually know. Exactly. I completely agree with the statement that you made. In fact, many people ask the question, it converges, and your P-approximate series seem to converge, but how do you know that it converges to the right answer? There, I think, it's a matter of extrapolation

7:30 plus some arguments of continuity, you know. What we have done is to test this, in the case where we know the exact answer, namely the test particle orbiting a solar-chair black hole. In that case, we know the exact flux, we know the exact energy. So on the one hand we have an exact waveform and on the other hand we have a post-Lutonian waveform at various orders and P-approximate waveforms at various orders and we can compare them. There the results are fantastic. I mean you do see that the P-approximates give you very good overlaps with the exact waveforms. Once you go beyond 1.5 post-Lutonian order, you have very good results. whereas P approximates even when you go to 2nd post-Newtonian or even 11th post-Newtonian which is what is available at the moment you don't get any convergence that's the trouble with the post-Newtonian approximates now as far as equal mass binaries are concerned which is where we want to apply this as well as of course asymmetric masses in which case it's very good if you are talking about something like solar mass objects then I think the approximants are quite good because we have tested that in the eta tending to zero limit not test mass limit so in the equal in the equal mass case what we did was to introduce an arbitrary parameter in a model exact waveform I mean we model an exact we call something an exact waveform which uses all the information available and that's our exact waveform and in that exact waveform we try to change a parameter by 100% in two different directions. It's really what we are doing to be more precise is simply changing the location of the last stable orbit. In the exact model we change it by something like 100% on one side, 100% on the other side P approximates on top of that to make a prediction of what that is and they do it extremely well that's the point so we have checked the robustness going by varying our parameters a factor of 2 either way in our modeled exact waveform so in that

10:00 sense we have checked the robustness but I don't think anybody can say that we have converged the right waveform impossible i don't we don't make the claim either we have not made any claims saying that this is the exact waveform but it's most likely that it is the exact waveform so the next step therefore is to check now with more improved waveforms that will become available just to see how the convergence behaves. If we include now one more parameter, the third post-Newtonian and the 3.5 post-Newtonian waveforms should become available pretty soon. So that gives us a two-parameter family of waveforms which we can use to test our P-approximant models. So would you see the P-approximants as being suitable Or as it were, a first detection or more for subsequently when you feel you're confident that you're seeing something and then you can use these to pick out more signal? Yeah, but no, I think not the second. I think it's the first. I think as far as GO is concerned, we are very much going to use the P approximants to do the first detection itself. The main reason is really that, I mean, if it is going to reduce our signal-to-noise ratio by a factor of, let's say, 1.2 or 1.3, what are we talking about? We can accept something like a signal-to-noise ratio of 7 as something which is quite good. But now if you lose 10% of that, it may not sound so very bad. But I think the way we are setting up thresholds and doing this multiple hierarchical search is such that we may lose the signal in the first level of the hierarchy. And I think it's... I think the strategy will be something like this. Since we are going to be interested only in the high mass region, and in the high mass region, not many, many signals to be searched for, what we should do is to use both the approximate signals as well as the conventional post-Newtonian ones and maybe even more.

12:30 You know, you can just simply model you a signal like Tom Prince at the moment is thinking of something called stationary phase formalism, not stationary phase approximation to Fourier transform of chirps, but stationary phase formalism where he has got a phase in the Fourier domain with the usual post-Newtonian terms but also extra terms which is one can vary by hand and they don't correspond to any solutions to Einstein's equations at any post-Newtonian order but since we don't know the exact signal you may want to play around with that just to increase your parameter space of your manifold the volume of the manifold of your search think since it doesn't cost too much in the region to search for these binaries anyway we should perhaps use as many different models as possible but certainly P approximates and in GEO that's the strategy we are going to take we are going to use some of the post-Huttonian ones as well as the most accurate post-Huttonian obviously and the P approximates ones. I'll tell you the reason why I'm very hesitant to use just the post-Huttonian approximate ones they fluctuate much i mean if you use something at second post plutonium you're not guaranteed that 2.5 is very very good yeah in fact everyone within the data analysis community knows that even those 2.5 post plutonium is available we should be really using only second post plutonium it's because all the other physical quantities like energy and flux when you plot it for the 2.5 post plutonium they are very different from all other orders so I'm always worried about the non-convergence of the post-Newtonian you know, they can't really trust them too much whereas these few approximants are trustworthy as far as convergence is concerned but we don't know where they are converging to that's the problem with that and have you have you encountered people who've said well, you know, because one this, you know, isn't a suitable formalism for an initial detention? Yeah. Initially, I think I did come across people saying that. I've not heard of such statements of late,

15:00 except that people say, oh, you're only going to get another 10% improvement or 15% improvement. Why bother? Do you think we have to be so careful? of my argument is really that you know because of paucity of these events it's it's stupid not to use the best available tool at your at your disposal you want to maximize your chance of detection and therefore use the best available tool i mean there might be a little bit uncertainty as to whether it is converging if you ask me personally i mean i have no doubt that it is converging more towards the exact waveform. But the only reason why I have slight hesitation is that there are one or two examples in the usual theory of body approximants where it doesn't converge to the proper one. But most of the examples, including those from quantum electrodynamics calculating g-factor of the electron, they were done much before the experiments the value of g and they use this p approximants in quantum electrodynamics it has had a very great success i think because we know physical from physical reasons that there is a pole and from mathematical mathematical point of view p approximants are the best ones to catch the pole and because in the test mass case it has had such a great success and all these reasons you know force me to believe that these P approximates are much closer to the true waveforms than the post-Otonian ones but there are criticisms there were criticisms I don't know whether they still exist or not as far as the GEO is concerned we are going to use P approximates in our search within GEO I know that in the case of LIGO most of the data analysis work that's been done so far has been by theorists and my impression is that that's also true in Geo but are some of the experimentalists in Geo interested in data analysis in general or even specific things like which type of prophecy? Not to the detail that I'm going into as far as sources are concerned they're interested in different sorts of data analysis like detector diagnostics so we do time frequency searches let's say

17:30 and we go and show them a graph all the details about how the data was analyzed but they are not participating to the extent that they will come and you know tell okay please use this method to search for binary inspirals rather than something else so I think we take their input after we have done the data analysis we design the data analysis algorithms we go back to them and I think and then we go and improve our models data analysis models not the theoretical models so I think so far we haven't had any member of geo experimental groups directly involved in algorithm design that's the true statement and you think that's likely to change I know sometimes people say to me well at the moment the experimenters are too busy Maybe once they get the instrument done. Is that sort of the same situation? It's true, because there are people who are really interested in general relativity itself who would like to make the scientific use of the data. So in that case, they have to be participating in data analysis. So I believe people like Karsten Danceman or Jim Half will eventually be interested in all the details of the data analysis. But I think at the moment, they are so busy constructing the interferometer, so bothered. But once we start analyzing it, analyzing real data, that situation is going to change. And I think it should change. But I don't know whether it can change at the level of theoretical studies. I think it's very, very hard because the field is extremely technical. You know that as far as post-Newtonian gravitational waves are concerned, right? There are perhaps about half a dozen to a dozen people in the whole world who can do that kind of calculation or rather who are doing that kind of calculation not that who can do but who are doing that kind of calculation and and and to do that requires a lot of training to do such calculations requires a lot of training so it's unlikely they would participate at the level of giving you inputs into theoretical modeling of sources and how that can be used but i think as far as data analysis details are concerned. I'm sure they would be looking into that aspect as time goes on.

20:00 Like, you know, what filter did you use? How did you construct your filter function? How did you construct your power specter density? What windows did you use to estimate your transfer function? All these things will be questioned by experimentalists at a later stage, but Do you have the impression, actually, that the experimentalists do come to data analysis from a very different perspective, or look at data and signal and noise in a different way? Yeah, sometimes there could be gaps. And we have to make every effort to bridge that gap. I think last time I spoke I told you that we do meet with them but not very very frequently it has to increase so the language is different to give an example we use signal to noise ratios as simply numbers whereas they use dBs which are logarithmic scales and then they are talking about mostly instrumental artifacts that are affecting your data, that's what they're mostly interested in. Whereas you would like to pin down whether there is any external influence. As a theorist, you would like to know what else is affecting it from an external environment. I think there are slightly different perspectives in data analysis as to what they want to do. What the theorists, I mean experimentalists mostly want to do is to construct an ideal instrument. they want to go and find out okay what are the things you have found in your data strength tell me what kind of signals you have found they are not so much bothered about whether you have looked for a specific gravitational wave signature or not but rather they are interested in is there a burst of a certain kind or not if they find such a burst they want to remove it that's why I said their attitude at the moment is to diagnose the detector to understand the detector performance they are not so much bothered about searching for sources so that's the difference in perspective at the moment

22:30 as far as data analysis is concerned the more you can tell them about what are the different features you find the more happy they are Bernie when I was in Potsdam mentioned to me that Cardiff was just, you were just involved in releasing a set of data dealing with data acquisition, or a set of software dealing with data acquisition. And I guess I was curious to ask what's the function of the software? The data acquisition one? It's developed by Ian Taylor. What it does is to acquire the data from the various subsystems, and we are essentially using the LIGO data acquisition design, but tailored, I mean, but slightly tailored to the needs of Jio. For instance, we are not using this reflective memory system of LIGO, but rather we are just writing the data directly onto the disk. So it's a software that is written under Triana, which acquires the data from different locations, all the environmental channels, and writes it onto the local disks. And then it also makes the data available on a local area network for for detector people to carry out their own analysis, and they will be using Triana as a tool to analyze the data, and we are providing, we are writing software to do all the detector diagnostics, like they may want to cross-correlate two different channels, and they may want to look at whether certain kinds of signals are present in one of them, they may want to do covariance analysis or any sort of analysis that's needed for detected diagnostics, that's what we're developing under Triana at the moment. So is Triana basically a kind of a windowing system that allows you to...? It's a GUI, a graphic user interface. How it works is that there is a toolbox which contains several different units, which are basically, you know, program pieces written in Java. And then there is a workplace. You pull the units into the workplace, you know,

25:00 by just using your mouse, connect them together just like you connect your, let's say, filter boxes in a data analysis toolbox. And then you just click the first button and say run. and passes it through this network that you have created graphically. The advantage of that is that the person who is analyzing the data doesn't have to know anything about programming, doesn't have to know about algorithms. The algorithm description is there. I mean, on each unit, if you click a button, then the help file comes up and tells you what it does. And there are help files at different levels, and you can go to any level you want, and ultimately to the technical level. So it offers a graphical way of constructing a very complicated data analysis tool. The scientist only needs to know the scientific way the algorithms work. They don't need to know the programming. You know, correlation. There's a toolbox which says, let's say, correlation analysis. And you pull that toolbox and connect it to the data and you'll have to give it two inputs it'll correlate those two and output it onto a grapher or a file or whatever you like it's a graphical user interface and is it aimed primarily at experimenters or at theorists or at anyone at the moment it's primarily used by the experimenters right in the beginning when Triana was developed We were thinking that it will also act as the front end to our database and data analysis system, not database, but to our data analysis system. We thought that we will run all of our analysis. We will develop codes and we'll run all of our analysis by just this graphical user interface. But having gained experience with Triana now, it may still act as a managerial tool, but it may really not run the processes with full control. It might simply tell the status of various programs, but it will really not buffer the data from one unit to the other by itself. What we have found is that it can slow down quite heavily

27:30 because of this buffering that is required. When it goes from one program to another, you know that if you're just running a C program, all the data is there in the memory and then the program can simply switch from one subroutine to another without any problem. But here, if Triana has to control the thing, it has to buffer the data and pass it from one to another could take quite a long time so that's the reason why we don't think it's it's it is very efficient to work in a situation where there are very long pipelines so we may not be using it as a as a complete front end it could be still managing you know the various tasks that are running but at a very superficial level but we are still not completed that design we're still working on the Triana, the last bit of Triana that has to be written is being now written which is to make Triana available over the network you have several units on your workplace which are all connected up but one of the units might be running in Hanover, the other one might be running in Cardiff and yet another one in Albert Einstein Institute as far as the user is concerned It doesn't matter at all. I mean, it just gives you those connections. That's the feature we would like to have. For instance, we constructed this build system in Hanover just a couple of weeks back, right? That's the one that we will be using for binary-inspired searches, but it will be controlled by either AEI or here. So the process is running there, but the Triana is running here, and it has to show me all the status. So that part is missing, that, I think we will then start designing how Triana will interface with our data analysis system. So that design is not complete. So up to now, at any rate, it's looked as if it's not efficient to use it for running the actual data analysis? I don't think that it is very efficient to use it that way. So the main application at the moment would be more for what you mentioned just a while ago, that is the experimentalist who is looking at the data he's getting at and wanting to analyze it from the point of view of what's happening in the detector. That's right. As well as for

30:00 exploratory data analysis. In the sense I'm interested in seeing whether a particular algorithm works or not. And I don't want to write large C codes which are tailor made for that. So I just pull some units and see what happens to the data if I process it For such applications, it's okay. And I use it once in a while. Bernard Schultz uses it. There are not many people in geo-theory side who use Triana, apart from, of course, Ian Taylor, who is the main writer of this. But we are encouraging it more and more, especially because we have taken the responsibility to develop these tools for detector diagnostics, some of which are already written up. But one of my postdocs will be very seriously working on writing more tools. In course of time, I think, it'll be more and more used for exploratory data analysis. And do the geo-experimentors already use this system? Yeah, especially in Hanover and Glasgow. People have been using it. Not so much in Glasgow, I think. Yeah, Glasgow, they had a trial, but the real detector is only in Hanover, and that's where most of the use is really. I think it was also demonstrated LIGO but it did not pick up so much of popularity I think but Bernie mentioned to me that it's just been launched commercially yeah that's right so you see it also as being something that would be usable by a wider experimental community not just gravitational I'm not sure about other experimental community at the moment we are getting response from such a wide variety of people that I'm quite amazed why it has not picked up so much in the UK, it has picked up so much in America and there are people in the Navy who have shown interest, people in NASA who have shown interest and we have several downloads of I mean until recently it was a free software so there were several downloads several hundreds of downloads of the system but I don't know how experimentalists would pick it up all by themselves we've got to go and show how the system works and only then they will get attracted to it

32:30 the reason why those downloads have taken place is because Ian would go and do the demonstration in some place in a big conference and immediately we have several hundreds of downloads so it really happens only because we go and give talks and publicize it otherwise It really doesn't happen all by itself. With the commercial venture, we'll have to see whether that becomes any better. But I don't know whether experimentalists would like to just pick it up just to play. We'll have to see how that goes. We have no idea. But it is suitable. Having said that, it requires training. But I think it's suitable for a wide range of communities, including music, computer industry, particle physics, possibly. I don't know, really, at the moment. So just out of curiosity, will the commercial venture try to actually send people out to get to do the demonstrations and try to work out interested in that part of the plan? Yes, we do say that, but except that what happens is that the charges are much higher in that case. We can't send out people to train. right at the moment we are looking for people, new people to be employed at Cardiff for that purpose. You know, for us to train them and then they will perhaps go and train other people if there is a request. But that's not going to happen without much investment from the buying company, I guess. But I have a feeling that maybe just to popularize it, initially we had to go and just do the demos, free demos. Incidentally, I must just tell you one thing with regard to this commercialization. it turns out that Triana is the first software that is marketed electronically, right, by the web page in the whole country, in the whole of UK. We are the first one to go e-commerce in any British university. I mean, it's not any place, but it's the first British university that's putting up a software for e-selling or whatever, you know.

35:00 I had not known that incidentally. I came to know of that only because there were some people from the press on the inauguration day, apparently, I was not here, and apparently they told us, Bernard Schwartz and Ian Taylor, that this is the first university software that's made available on the web for, you know, credit card and other ways of buying it. I was quite amazed but it took us a long time it took us about a year to develop it I mean not the software but just the web installation, developing all the necessary software, it took us about a year and lazing with the bank lazing with another provider internet provider who gives you secure way of communication and what not yeah I can imagine there must have been a lot so is the main hope that the commercial side will finance further development of Kriyan not the group itself I don't know whether that's going to happen because it's still, it's owned by Cardiff University they are the proprietors I don't know whether the group can benefit greatly from it Yeah, but we would like to develop it further, I mean as a software we will invest a lot from the earnings into software developments but whether that can benefit other activities in the group it's very hard to say at the moment, whether it can help fund gravitational not clear to me at the moment. Yeah, I'll just leave it. Yeah. Well, what else? No, five past three. Five past, okay. Well, good, so we can just wrap it up quickly. Yeah. The, uh, to get back to the, uh, the, uh, you should have had a process or something, and then I was kind of, I was there. Incidentally, in your interviews, did you talk to other people about peer approximants, and did you hear any controversial remarks?

37:30 No, I haven't talked to them much in specific terms. It's just that I've, although I hope to do a bit more, for instance, I'm hoping to go to Paris at the end of the month, and I'm hoping to talk more about the people there, and then once I know more myself, sort of see what I hear from people. Thievo is one person who is interesting to talk to about this and his new work on what we have already started calling as BD waveforms, Bonanno and Damur waveforms. They have come up with a new approximation to the binary in spiral and plunge. And they believe that they now have the waveform complete up to, not just the last stable orbit, but up to the light ring. from 6M in the test mass language from 6M up to 3M and they think that now that can act as initial data for numerical relativity to evolve it until the quasi-normal modes so they think that they basically have the waveform up to the last point up to which when they detect it and I'm working on it at the moment trying to see how that compares with P approximates now So that's our next piece of work. So I guess that actually does lead into the sort of general area that I had in mind, that for a long time, I guess, the general idea that I think was in people's heads, I guess in my head at any rate, was that you were going to have something like post-Antonian approximations up to a certain point, and then presumably numerical relativity was going to take over beyond that. So the impression that I get is that more and more the putative, and it was always a little ambiguous, point at which there would be some changeover between the two is sort of moving further and further in. Into the, yeah, that's right. The analytic work is going further and further in the works, and the point at which the numerical work has to start goes in. Is that your impression also? And do you see it sort of extending to the point where there's very little left needed for the numerical? Yeah, I still think that there's a very big question mark as to what happens even from 3M until the two black holes form one black hole.

40:00 I think it's a very big question mark because there are two completely contradictory viewpoints on the subject, as you know very well. On the one hand, there are some people who think of radiation at that stage they call it the merger waveform or the coalescence waveform and they think that that could be significant compared to the inspiral phase itself so much so that that may be the first thing that we see some people think that on the other hand other people especially people who have worked on the close limit approximation seem to think that there's no dramatic thing that's going to happen there's going to be an inspiral and it'll smoothly join the quasi-normal modes. No big burst, nothing. So there are two schools which we think in entirely opposite ways and therefore we really need numerical relativity to resolve it. Well, I mean there could be some breakthrough in analytical work and somebody might come and say okay, this is what it is. But I see that it's very unlikely. Even in the latest work of Banano and Damo, it's again some approximation It's not an exact solution to a system of Hamiltonian system of equation. So I think numerical relativity will have to come for us to understand what happens in those last 100 milliseconds or 10 milliseconds. We do need it, I think. So even with the ingenuity, the analytical people have demonstrated getting closer in with post-Newtonian and other techniques, and then even with, I guess, the people who model the actual final merger as a perturbation of one black hole. That's right. There's still that little area sort of right at the end, just before that between those two, that you really would need a numerical unless something else. Yeah, that's my feeling. But now the happy thing is that numerical derivatives seem to think if you give them the data very close to the merger they will be able to evolve they think that they can handle it obviously if you reduce the time that they need to run for small enough exactly 10 milliseconds so they will run for 40 hours and give you 10 milliseconds worth of data but it's very very useful so well it'll be interesting to see

42:30 I guess it's 5 past now Thank you. Thank you.