Modélisation du syst`eme triple autour du pulsar radio PSR J0337+
Transcription
Modélisation du syst`eme triple autour du pulsar radio PSR J0337+
Master Science de la matière École Normale Supérieure de Lyon Université Claude Bernard Lyon I Stage 2013–2014 Guillaume Voisin M2 Physique Modélisation du système triple autour du pulsar radio PSR J0337+1755 Résumé : Ce stage avait pour but de parvenir à une description précise et stable au cours du temps des temps d’arrivée au radio-télescope de Nançay des impulsions du pulsar PSR J0337+1755 récemment découvert par Ransom et al. (2014). Ce pulsar est en effet en orbite relativement proche, inférieure à la taille d’un système planétaire comme le système solaire, avec deux naines blanches. De ce fait les intéractions gravitationnelles à trois corps doivent être prises en compte, numériquement, dans la reconstitution des orbites. Ce système est unique en son genre dans la mesure où il est à ce jour le seul système constitué de trois objets de masse stellaire comportant un pulsar radio milliseconde. Il est par conséquent possible de l’étudier en profitant de l’époustouflante précision offerte par cette ”horloge interne” au système qu’est le pulsar. Néanmoins, pour être efficace, le chronométrage de pulsar doit s’appuyer sur un modèle réaliste des délais de propagation de la lumière, modèle qui n’existe pas à ce jour pour un tel système triple. Mots clefs : pulsar, étoile à neutron, chronométrage, trois corps Stage encadré par : Lucas Guillemot lucas.guillemot@cnrs-orleans.fr / tél. (+33) 2 38 25 52 87 Ismaël Cognard icognard@cnrs-orleans.fr / tél. (+33) 2 38 25 79 08 Laboratoire de Physique et de Chimie de l’Environnement et de l’Espace 3A, Avenue de la Recherche Scientifique 45071 Orléans cedex 2 France http://lpce.cnrs-orleans.fr/ August 2, 2014 Remerciements Je remercie chaleureusement mes tuteurs Ismaël Cognard et Lucas Guillemot pour leur accueil et leur bienveillance. Je remercie également Jean-Matthias Grießmeier pour toutes les discussions enrichissantes tenues au cours de ces quatre mois à partager le même bureau ainsi que pour ses conseils, sans oublier les deux autres membres de l’équipe pulsar du LPC2E : Gilles Theureau et Kuo Liu. Merci aussi à Claire Revillet pour ses conseils et son assitance sur le plan informatique. Plus généralement je salue l’accueil du LPC2E, et toutes les personnes avec qui j’ai eu des échanges amicaux. I would like to thank Anne Archibald for her very appreciated help during my visit at ASTRON. Also I would like to thank Jorge Piekarewicz1 and Farrukh Fattoyev2 for answering my questions about neutron stars mechanical properties. Contents 1 Introduction 1 2 Introduction to the J0337+1755 system 2.1 Previous work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Orders of magnitude and accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 4 3 A full three-body timing model 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Computing the orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Retrieving the two-body results with a numerically integrated trajectory 3.2.2 Computation of the three-body orbits . . . . . . . . . . . . . . . . . . . 3.2.3 Numerical intricacies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 First order systematic delay : the Rømer delay . . . . . . . . . . . . . . . . . . 3.3.1 First order correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Second order correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Second order systematic delays . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.1 The Einstein delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.2 Tidal effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 5 6 6 6 8 9 10 10 10 10 13 4 Fitting to data : a very delicate optimization problem 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Starting solution and the choice of parameters . . . . . . 4.2.1 Initial solution . . . . . . . . . . . . . . . . . . . 4.2.2 Choice of fitting parameters . . . . . . . . . . . . 4.3 Minimization with the Minuit library . . . . . . . . . . . 4.4 The Markov-Chain-Monte-Carlo (MCMC) approach . . . 14 14 15 15 16 16 18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Conclusion 20 References 21 1 Jorge Piekarewicz, Department of Physics; Florida State University Tallahassee, FL 32306-4350 Farrukh Fattoyev, Department of Physics and Astronomy, Texas A&M University-Commerce, Commerce, Texas 75429-3011, USA ; Institute of Nuclear Physics, Tashkent 100214, Uzbekistan 2 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin Appendix A The gravitational two-body system defined by orbital elements 23 A.1 Relation between orbital elements and state vectors for a single body . . . . . . 23 A.2 The case of two bodies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 A.3 The mass function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Appendix B Jacobian of the Newtonian three-body differential system 24 Appendix C Likelihood function 24 3 Modélisation du système triple autour du pulsar radio PSR J0337+1755 1 Guillaume Voisin Introduction The first pulsar was discovered by Jocelyn Bell and Tony Hewish in 1968 (Hewish et al., 1968). Its periodic radio pulse was soon identified to be that of a highly magnetized neutron star (see in particular Gold (1968) and Gold (1969)), until then hypothetical objects proposed among others by Robert Oppenheimer in the 30’s. A neutron star is believed to be the remnant of a supernova, made of a neutron dominated matter with a density of the order of that of an atomic nucleus. Nevertheless, it is of macroscopic size with a radius of about ten kilometers, thus weighting about 1.5 Solar masses (just above the Chandrasekhar limit). The consequence is that these stars are part of the so-called compact objects, with a compactness parameter of about 0.5 (1 for a black hole, 10−6 for the Sun), in the vicinity of which the curvature of space-time cannot be neglected at any rate. The surroundings of the star are filled with a magnetosphere supported by a gigantic magnetic field that typically ranges from 108 Gauss for millisecond pulsars to 1015 Gauss3 for magnetars (Earth’s magnetic field is about 0.5 Gauss for comparison). It is widely believed to be the locus of the electromagnetic emissions of the star as first suggested Gold (1968). The detailed mechanisms are poorly understood but observation proved that it ranges over a very large spectrum from GHz in radio, observed with radio-telescopes, to gamma rays, observed recently with the Fermi-LAT satellite. Besides, this emission must be very localized in the magnetosphere, such that it emits in a narrow beam that we can see only once at every pulsar rotation if by chance it crosses our line of sight, just like a cosmic light house. These radiations are powered by the rotation of the pulsar that together with the magnetic field makes the star behave like a giant dynamo. The pulsar consequently slows down at a measurable pace, related to its magnetic field intensity and spin frequency. Pulsar spin periods range from a few milliseconds to several seconds, a ”normal” pulsar turning about himself in typically one second. The former are naturally called millisecond pulsars (MSPs). They form the most interesting family for pulsar timing, which consists in measuring with a high accuracy the times of arrival of the pulses on Earth. Indeed, their particularity is to have a lower magnetic field (108 Gauss as said above) and thus a very slow spin down rate. In other words they are extremely stable clocks which pulse can be predicted within a few hundred nanoseconds interval (Ransom, 2012). These pulsars had a rich life, they are sometimes referred to as recycled pulsars. Indeed, contrary to what their rapid spin frequencies suggest they are old pulsars, but that have lived with a companion, which makes a big difference ! The commonly accepted scenario is the following : in a binary one of the stars goes supernova, leaving a neutron star that slows down to become a normal, one-second pulsar in a hundred Myr while the companion continues its evolution during Gyrs. When the latter starts ascending the giant branch, it transfers angular momentum and mass to the neutron star via accretion, thus spinning up again the old neutron star. Accretion also significantly consumes the magnetic field, and we end up with an old, low magnetic field and so frequencystable, millisecond pulsar. These MSPs thus provide a very accurate internal probe of the system they belong to. This is the reason why they are sometimes called ”neutron star laboratories”, because it is one of the rare astrophysical precision measurements one can achieve. Moreover, the extreme conditions of gravitation and magnetic fields surrounding these stars make them unique candidates to test the limits of fundamental theories, and in particular the theory of gravitation. To do this one needs to take into account all the systematic effects, and in particular the orbital motion of the binary system, the delays due to deformations of space-time, interstellar dispersion, 3 1 Gauss = 10−4 Tesla. 1 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin the motion of the Earth, atmospheric delays etc... Models were developed since the 70’s, including an increasing number of systematics to an always higher order of approximation, in particular for the post-Newtonian description of orbital motion (Blandford and Teukolsky (1976) and Damour and Deruelle (1986) in particular). Nevertheless, MSPs always had only one companion star, usually a white dwarf, but that recently changed with the discovery of PSR J0337+1755 (Ransom et al., 2014). Indeed, this MSP has two companion white dwarves in rather close orbits. For the historical reasons I just gave no model was developed so far to deal with such a system. It was consequently impossible to time it to a high level of accuracy for the simple reason that every systematic was not taken into account. When the LPC2E team in Orléans began recording data from this system with the Nançay radio-telescope, they quickly faced that problem. That is the reason why they proposed an internship to deal with that issue. That supposed to make a new numerical model from scratch. In the next section, I present in more details the system we are dealing with and why it needs a specific modelization, in section 3 I present the model itself as I developed it so far, including orbital motion and some post-Newtonian corrections, then in section 4 I show how I could fit this model to the data at my disposal and the current, encouraging results that were obtained. 2 Introduction to the J0337+1755 system 2.1 Previous work The discovery of this system was initially published by Ransom et al. (2014). It is constituted of a radio pulsar with a spin frequency of 366Hz orbiting two white dwarves in hierarchical order : the inner white dwarf lies within 1 ls (light second) of the pulsar for an orbital period of 1.6 days while the outer is within 100 ls for an orbital period of about 327 days (see table ?? for more detailed parameters and error bars). Thus this system is quite compact and the interactions are rather strong compared to anything that was discovered before : the previously known three-body systems involved planetary mass objects, such as B1620−26, while here we have three stellar-mass objects in rather close orbits. A qualitative knowledge of the parameters of the system can be achieved thanks to the tools that already exist to deal with pulsars with planets. More specifically the so-called BTX model implemented in Tempo (Hobbs et al. (2006) and Edwards et al. (2006)) was made to deal with planets in the approximation that they do not cross interact. It is actually good enough, replacing planets by white dwarves, to keep track of the phase of the pulsar at plus or minus half a turn as shown in figure 1. Nevertheless this is not enough to fold the raw data on a long period of time, that is piling up consecutive pulse profiles to make an averaged one with higher accuracy, since this demands to predict the next pulse with a below-microsecond accuracy4 . This is the main goal of this internship to be able to do that. Besides that, the fact that there are three bodies with nonnegligible interactions allows to solve some degeneracies in the orbital parameters and thus to extract more information about the masses or the inclination of the orbits than is possible with an ordinary binary system, as we will see in section 4.2. One may also want to take advantage of the incomparable regularity of pulsars and the unique configuration of this system to test fundamental principles of general relativity such as the strong equivalence principle. This is what we are looking forward. 4 The data is recorded with a binning of 250ns which sets the higher precision limit 2 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin Figure 1: Residuals of the BTX model applied to the last-to-date Nançay data, that is the difference between the time of arrivals (TOAs) predicted by the model and the measured times. i Barycenter Direction of ascending node Earth Earth Figure 2: Sketch of the orbits as fitted in Ransom et al. (2014), not to scale. The neutron star is the smallest of the bodies but the heaviest so has a smaller amplitude of motion. Together with the closest (red) white dwarf they form the inner system. To a good approximation this one can be considered as a body orbiting the outer (green) white dwarf to form the outer system. To achieve these goals it is necessary to solve the exact three-body equations of motion since the gravitational coupling is too strong to keep the desired level of accuracy (see below) over years of data. This is the object of sections 3. Additionally one needs to take into account 3 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin systematic effects such as Roemer, Einstein or Shapiro delays (see section 3.3 and 3.4). During this work, we took contact with Anne Archibald5 who works on the fitting model used in Ransom et al. (2014) and started a collaboration realized by a visit in late June. We then agreed that the numerical codes would remain fully independent from each other so as to be able to confirm results. Indeed, on each side this computation revealed to be delicate and sometimes tricky because of the very high level of accuracy to achieve. 2.2 Orders of magnitude and accuracy The input we have is a set of 5019 times of arrival (TOAs) of the pulsar radio pulse received at the Nançay decimeter telescope6 between August 2013 and July 2014. We corrected them from all the systematics related to the Solar system and the interstellar medium using the standard tools provided in Tempo (Hobbs et al. (2006) and Edwards et al. (2006)) : mainly Solar-system orbital motions and interstellar dispersion. Consequently we will from now on only consider barycentric time of arrivals at infinite frequency, that is pulses as if they arrived directly at the Solar-system barycenter, considered a Galilean frame, at infinite energy (so without being dispersed at all). These TOAs are recorded with an accuracy of about 3µs. They are given in modified Julian days7 (MJD), for example the first one is : 56,492.297,192,640,59 MJD. Given that there are 86,400 s in one MJD we soon realize that the accuracy of our TOAs lies in the last digit. From a numerical point of view, such a number is a float coded on 64 bits (double in C or Fortran) and we readily understand that numerical round-off will be an issue. The solution will be to use 80-bits-C long doubles for most computation (see section 3.2.3) to avoid loosing the physical accuracy in numerical round-off. From the point of view of the pulsar itself, it means that its relative location must be known with an accuracy of a few hundred meters at any time, as compared to the outer orbit wideness of about 1 AU8 , and this from an estimated distance of more than 1kpc9 ! In table 2.2, I summarized the relative weight of interactions in several known 3-body systems in order to get a broader picture of the system. We see that the interactions in the system J0337 clearly follow a hierarchy that somewhat lies in between the Earth-Moon-Sun and the Sun-Jupiter-Saturn hierarchies. Note that these can be treated with a very reasonable accuracy by perturbation methods, but those will intrinsically fail to give the amazing accuracy needed here over long time periods. Another reason not to use perturbation methods is that post-Newtonian effects are likely to appear in this system, such as precession of the periastron, that would make them even harder to derive. Indeed, contrary to anything in the Solar system, we here deal with compact objects only. The compactness parameter is given by the ratio of Schwarzschild radius Rs = 2GM/c2 (Misner and Wheeler, 1973) of the object over its radius, and is about 10−4 for a white dwarf to 0.5 for a neutron star. Post-Newtonian effects are likely to appear for two reasons : if the speed is a reasonable fraction of the speed of light or if the gravitational field is sufficiently large. The speed of the pulsar is about 10−4 c (a few tens of kilometers per seconds) which is enough to yield a significant Einstein effect (see section 3.4.1) while the relative strength of the gravitational field can be determined looking at RS /r where r is the distance between the two bodies. If this ratio is close to zero then gravity tends to be Newtonian (Will, 2014). Here it is about 10−5 5 Netherlands Institute for Radio Astronomy (ASTRON), Dwingeloo, The Netherlands Radiotélescope décimétrique de Nançay, Observatoire de Paris, route de Souesmes, 18330 Nançay, France 7 The length of a modified Julian day is by definition always equal to 86400 seconds 8 1AU = 1.496 · 1011 m 9 1kpc ' 3 · 1019 m 6 4 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Energy 0 - 1 Energy 0 - 2 Energy 1 - 2 Force 0 - 1 Force 0 - 2 Force 1 - 2 Earth - Moon - Sun 1/10 1 1/300 1 1/100 1/240 Sun - Jupiter - Saturn 1 1/6 1/3000 1 1/12 1/100,000 to 1/10,000 Guillaume Voisin Pulsar - Inner WD - Outer WD 1 1/230 1/1000 1 1/450 1/3100 Table 1: Summary of the relative weights of gravitational energy and force terms in different known triple systems. Numbers 0,1,2 refer to the bodies of a given system in the same order as they are mentioned. WD stands for White Dwarf. for the inner white dwarf in the field of the neutron star. Hence, both because of speed and gravity this system is in a weak field regime where a first order post-Newtonian description should be sufficient (Will, 2014). For comparison, for Earth around Sun, RS /r⊕ ' 10−9 and v⊕ /c ' 10−4 . 3 3.1 A full three-body timing model Introduction As I tried to show in the previous section we need to compute an exact solution, at least to desired accuracy, at all times. For a three-body system it is usually well accepted that no analytical solution exist to date. This is not exact since Karl F. Sundman published such a solution in 1909 (see Henkel (2001) for an historical approach and Sundman (1913)). Nevertheless this solution is based on slowly converging series and thus is of little interest for actual computation, as compared to purely numerical approaches. However the numerical way is not without difficulties as we shall see in section 3.2.3. Once the motion of the pulsar and its two companion white dwarves is computed we need to model the times of arrivals of the radio pulses. This involves taking into account intrinsic spinning parameters of the pulsar as well as systematic delays due to geometry (Roemer delay), post-Newtonian effects (Shapiro delay, Einstein delay...) or tidal deformations. We will consider that the Roemer delay is first-order and the others second-order meaning that the former is at a 100 s level while the latter are at most a few hundreds of microseconds. In the proper frame of the pulsar one just computes the phase, or number of turns N , with a Taylor expansion to second order (Blandford and Teukolsky, 1976) : N (τ ) = N (τ0 ) + f (τ0 )(τ − τ0 ) + 1 df (τ0 )(τ − τ0 )2 2 dτ (1) Where τ is the proper time of the pulsar. The timing model must then be expressed in the Solar-system-barycenter frame. Now delays come in : N (ta ) = N (t0 ) + f (t0 )(ta − ∆t(1) − ∆t(2) − τ0 ) + 1 df (t0 )(ta − ∆t(1) − ∆t(2) − τ0 )2 2 dt (2) Where ta is the time of arrival, ∆t(1) and ∆t(2) are respectively the first and second order systematic delays and ta − ∆t(1) − ∆t(2) = τre is the retarded proper time of emission of 5 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin the pulsar. Indeed we cannot know what was the actual time of emission unless we know the distance between the pulsar and us. Fortunately, although this parameter is often poorly measured, it is not necessary to analysis. We will thus assume that all our results are retarded by an unknown constant delay. However the system could undergo a relative radial motion with respect to the Solar system. Ongoing observations with the VLBA10 telescope will address that matter (Ransom et al., 2014), but as long as this relative proper motion is constant (no acceleration), it essentially amounts to a linear drift of the phase which is equivalent to a slight shift in spin frequency : consequently our model will be just as efficient but with an effective spin period different from the intrinsic one. τ0 is our retarded proper time of reference at which are taken all the orbital and spin parameters. Since coordinates systems can be choosen arbitrarily, I decided for convenience to take τ0 = t0 the first Nançay TOA : 56,492.297,192,640,59 MJD. 3.2 3.2.1 Computing the orbits Retrieving the two-body results with a numerically integrated trajectory The aim here was to reproduce the residuals of programs such as Tempo for known binary pulsars but, instead of using the well known analytical expressions for the Newtonian two-body problem, I replaced it by a fully numerical approach. The numerical integrator used was the routine odeint of the scipy library11 . It implements a predictor-corrector algorithm (Press et al., 1992) using double floats (64 bits). The equations were cast into a dimensionless form using the typical scales of the problem. Typically : dq1 dτ = q˙1 dq2 = q˙2 dτ −q2 dq˙1 (3) = −M2 kqq1−q k 3 1 2k dτ −q1 k dq˙2 = −M1 kqq2−q dτ k3 1 2 Where, if r and t stand for the dimensional position and time, P is the orbital period, a may be the semi-major axis of one of the bodies if they are similar (which is the case for similar masses) or an intermediate value otherwise, M is the mass of the Sun, and G is the gravitational constant, I define : a3 M P 2 G τ = t/P q = r/a q̇ = ṙP/a k = (4) (5) (6) (7) Thus all variables and interaction terms should be of the order of one. For a predictorcorrector scheme one needs to provide the Jacobian as well but we shall see that in the next section, for the more general three-body case. 3.2.2 Computation of the three-body orbits We shall have here exactly the same approach as before, but we will somewhat generalize the notations. The adimensionned state vector Q gathers all the eighteen components of the 10 11 Very Large Base Array Scientific Python : http://www.scipy.org/ 6 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin J1455+3330 Residuals 15 10 Residuals (µs) 5 0 5 10 15 55800 56000 56200 56400 56600 Barycentric time of arrival (Julian days) 56800 Figure 3: Residuals obtained for the pulsar J1455+3330. The weighted standard deviation is 2.81µs while with tempo we have 3.55µs. The error bars for each TOA, though not represented here for the sake of clarity, are mostly between 2 and 3 µs and thus are compatible with a null residual. problem (9 components of position, and 9 in the tangential velocity space). Schematically : Q = (q1 , q2 , q3 , q˙1 , q˙2 , q˙3 ). The differential system becomes : k dQ = F (Q) dτ (8) Where Qi≤9 = F (Qi+9 ) and F (Q{I 0 ,J 0 ,K 0 } ) = −MJ QI − QJ QI − QK and circular permutations of {I, J, K}. (9) 3 − MK kQI − qJ k kQI − QK k3 I is the position space of the first body and I 0 its associated tangential space, i.e. the velocity space of the first body. Hence QI must be understood as ”the projection of Q on I” or equivalently ”the 3-vector (Q1 , Q2 , Q3 ) = q1 ”. As for J and K they refer respectively to the spaces of the second and third body. ∂Fi From there one can also derive the Jacobian of the system Jij = ∂Q , needed by some j integration schemes such as the Scipy predictor-corrector, or to perform perturbations of an initial solution. This Jacobian is worked out in appendix B 7 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin Figure 4: A three body system with the characteristics given in table 3, as well as numerical errors on first integrals of motion. L is the component of the angular momentum directed toward the observer, the impulsion P and the Barycenter numerical errors are represented in the plane of sky. 3.2.3 Numerical intricacies Once the mathematical problem is posed one needs to integrate numerically. Naturally, the easiest way would be to to use again the predictor-corrector integrator provided with Scipy 12 . This proves to work well if one integrates over no more than a few weeks but beyond that roundoff errors become more important than a few microseconds which our observational accuracy as previously stated. The reason for that lies in the numerical accuracy of the variables : Python defines floats on 64 bits giving a mantissa as long as 16 digits in decimal base. One will easily check that to keep an accuracy at the microsecond level for Roemer delays, it is necessary to compute the location of the pulsar with about 100m precision. Given that we are dealing with distances as large as 1011 m, it means a relative accuracy of 109 and thus numerical round-off errors have only 6 decimal places to go before reaching our desired level of accuracy. This is easily reached since the number of evaluations of the right hand side of the differential equation 9 reaches a few millions. As a consequence we cannot do otherwise than using another type of variable and a library that can deal with it. It is the case of the C++ library BOOST 13 which provides numerous differential integrators as well as multi-type support and a Python interfacing library. The type used will be the native C++ long double which encodes floats on 80 bits on most processors : it is usually the longest type supported in hardware, that keeps it fast14 , and it provides a twenty-digits mantissa that proves sufficient for our purpose. The numerical integrator chosen follows a Bulirsch-Stoer scheme, since it is famous for its high accuracy (Press et al., 1992). 12 See note page 6. Boost : http://www.boost.org 14 More bits would have to be handled through a specific software which would be much slower. 13 8 Modélisation du système triple autour du pulsar radio PSR J0337+1755 3.3 Guillaume Voisin First order systematic delay : the Rømer delay All of this is fine but one needs to give himself/herself a frame to perform his/her computations. Since it is by far preferable to have a Galilean frame it is reasonable to use integrals of motion for this purpose, since these will remain constant with respect to such a frame. More specifically we’ll use the invariant plane given by nH = H/ kHk where H is the total angular momemtum of the system. Moreover, the unit vector n going from the solar system barycenter to the pulsar system barycenter can be considered as constant in time given the distances at stake (though this might need to be corrected in case of high proper motion). Then it proves useful to take the intersection between the invariant plane and the plane of the sky (the line of ascending ×nH node) : na = knn ×n Hk From those vectors, we can draw two others to form two direct basis, n3 = na × nH and n03 = na ×n , such that eventually we have two frames RH = (na , nH , n3 ) and R = (na , n , n03 ). The first one will be useful as a local frame to carry out analytical calculations helping to see what is going on, while the second one is going to be used to compute numerically the Rømer delay as seen from the barycenter of the Solar system. Let’s remark that these two are related by a rotation of angle i = (n , nH ) ∈ [0, π[. Now let’s compute a general formula for the so-called barycentric time of arrival at infinite frequency (Blandford and Teukolsky (1976)). This is the time of arrival of the pulsar light in the Solar-system barycenter, assumed to be a good approximation to a Galilean frame, corrected from any local and interstellar delays such as the dispersion of the Earth atmosphere or the dispersion of the intergalactic medium. ta = tem + ttravel (10) And, neglecting for now relativistic effects : distance(P, b ) (11) c P is the pulsar location, b the solar-system barycenter and bP shall be that of the pulsar kbP −P k system. To first order in kb ≡ bbPbPP we end up with : −bP k ttravel = ttravel = (b bP + n · P )/c (12) Further we shall get rid of the term in b bP above, since it is hard to measure and does not provide any essential information as far as we neglect the effect of proper motion ( even so we would only need its derivative). Thus we redefine ta as : ta = tem + n · P/c (13) The only difference being that any function of tem , such as the frequency of the pulsar and its derivative, will be a retarded function of the ”True” tem : the state of the pulsar is known at an undetermined epoch in the past . We now need to parametrize the position of the pulsar P with respect to its barycenter. In RH spherical coordinates seem appropriate, especially since in the case of the triple system, the orbits are almost coplanar (see Ransom et al. (2014)). We define : θ = (nH , P ); φ = (na , P − nH · P ); rP = kP k (14) Thus, the Rømer delay is : ∆R (tem ) = n · P/c = rP /c (cos(i) cos(θ) + sin(i) sin(θ) sin(φ)) In R , it simply stands as ∆R (tem ) = y/c where y is the position component along n . 9 (15) Modélisation du système triple autour du pulsar radio PSR J0337+1755 3.3.1 Guillaume Voisin First order correction Since the only thing we measure is ta we would need to solve ta − ∆R (ta − ∆R (tem )) = tem to know the Rømer delay at tem . Fortunately the Rømer delay is about a few seconds while the characteristic time scale T of the orbits is at least a day. Thus we will use a first order correction in = ∆R (ta )/T (the time derivative hereafter yields a 1/T in order of magnitude). This is the approach used by Damour and Deruelle (1986) or Blandford and Teukolsky (1976) : d∆R (16) ∆R (tem ) = ∆R (ta ) − ∆R (ta ) dt ta In R this gives : ∆R (tem ) = y/c(1 − 3.3.2 y0 ). c Second order correction For most applications i.e. for most binary pulsars, a first order correction is sufficient. However, it is certainly a careful practice to check that against at least the following order. Moreover, we are using microsecond accurate data here, and it is likely that the second order will eventually matter for high accuracy applications. Indeed, if we take T to be the inner orbit period and ∆R to be at most the time needed to travel across the outer orbit, be 100 seconds, ' 7 · 10−4 and 2 ' 5 · 10−7 . Then the first order yields at most a 0.1s correction while the second order yields a few microseconds and the next order would not contribute to more than tens on nanoseconds, well below our observational accuracy. Let’s remark that in the traditional case of a binary pulsar, it is the same orbit that generates both the biggest delay and the time scale of orbital motion, which since light travels much faster than the pulsar end up in an already very accurate formula at first order. For instance, in the case of the binary millisecond pulsar J1455+3330 we already mentioned, this correction amounts to at most 160µs while the second order falls below one nanosecond. The expansion of Rømer delay to second order in ∆R (ta )/T reads : 1 2 00 0 2 0 (17) ∆R (te ) = ∆R (ta ) − ∆R (ta )∆R |ta + ∆R (ta )∆R |ta + ∆R (ta ) ∆R |ta + ◦(2 ) 2 In R , ∆00R |ta = y 00 /c. Computationally y 00 is picked on-the-fly from the right-hand-side of the differential system during the integration of the equations. 3.4 3.4.1 Second order systematic delays The Einstein delay As I mentioned in the introduction of this section (3.1), from the pulsar viewpoint, namely in its proper time, the only variation of its phase occurs quadratically. However, we need to turn it into our own time to model observations, and that is the so-called Einstein effect. Actually we shall simplify the problem and rather express equation 1 in the Solar-system barycenter (SSB) since other programs like Tempo can convert from an Earth-based-observatory frame with a lot of refinements. Let’s precise what the proper frame of the pulsar is : it is a frame following the speed of the barycenter of the pulsar and undergoing every external gravitational fields. By symmetry the pulsar gravitational field does not play any role, and it is good because we don’t want it : it would only contribute to a constant deformation of space-time that would not be detectable. To perform computation, we are going to put ourselves in the first post-Newtonian approximation (Will, 2014). Namely, if one considers a scale given by ∼ v 2 /c2 ∼ U/c2 where U is 10 Modélisation du système triple autour du pulsar radio PSR J0337+1755 80 a) Delay (µs) 60 20 0 20 40 60 b) First order 72.77 +5.649e4 c) Delay (µs) Delay (s) 40 5000 4000 3000 2000 1000 0 1000 2000 3000 40002.30 0.1 0.0 0.1 0.2 0.3 0.4 0.5 2.30 Guillaume Voisin Roemer delay 80 56492.3 56599.4 56707.7 Time since first toa (MJD) Second order 37.75 +5.649e4 Figure 5: Second Order Correction of Rømer delay for the fitted parameters given in table 3. a) is the complete delay to second order, small variations can be distinguished ; b) Zoom on the first order correction, an envelope with a longer time scale can be seen until the first half pseudo-period ; c) Zoom on the second order correction to Rømer delay, here again an envelope can be distinguished plus some specific patterns the external gravitational potential and v the speed of the pulsar with respect to the SSB, the line element is given by : dτ 2 = (1 − 2U )dt2 − (1 + 2U )dr2 /c2 + ◦() Where t is the proper time of the SSB and U = From what it is straightforward that : Gmi |ri −rp | + (18) Gmo |ro −rp | dτ v2 = 1 − U/c2 − 2 + ◦() dt 2c The Einstein delay to the first post-Newtonian order comes out as : Z v2 ∆E = t − τ = U/c2 + 2 dt 2c (19) (20) That being said, it is clear from this formula that this delay will increase to infinity. That is normal, times are not passing at the same speed. If the integrand where to remain constant it would just be a constant deformation of time, as before, and the only effect would be to change the apparent spin frequency : we get the ”real” spin period, as it is from the pulsar viewpoint. A much more interesting effect is due to the variations of the integrand with time. It can be seen just by removing the linear component in equation 20 : Z v2 v2 δ∆E = U/c2 + 2 − hU/c2 + 2 idt (21) 2c 2c What remain are the variations due to orbital motion. It thus has two pseudo-periods corresponding to PI and PO , as can be seen on figure 6.a. Hereafter, we shall always implicitly remove this linear component. 11 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin b) a) c)c) Figure 6: a) Einstein delay for the parameters drawn from 8-month data (green dots) of J0337 at Nança, b) Component due to the outer white dwarf during the same time as in a), c) Zoom on the component due to the coupling between the outer and inner speeds. The pseudo-period is that of the inner orbit, about 1.6 days. We can get more insight by decomposing equation 20 in terms of inner, outer and coupled components. Namely : U = Ui + Uo 2 2 v 2 = vp/I + vI2 + vI · vp/I (22) (23) Where Ui and Uo are respectively the gravitational potential of the inner and the outer companion, vI is the speed of the inner system barycenter and vp/I is the velocity of the pulsar with respect to that barycenter. Hence we expect that those components will show the same periodicity as the sub-systems they belong to. As we are only interested in variations it is clear that eccentricity will play a major role, since a circular orbit would offer neither variation of speed nor of potential. One can estimate these variations to the first non-null order in eccentricity e, assuming the motion follows momentarily its osculating orbits : 1 Ûi eI cos(2ωI t) + ◦(e2I ) 2 1 = Ûo eO cos(2ωo t) + ◦(e2O ) 2 = v̂I2 e2O cos(2ωI t) + ◦(e2o ) 2 = v̂p/I e2I cos(2ωo t) + ◦(e2I ) δUi = (24) δUo (25) δvI2 2 δvp/I (26) (27) Where the hat means ”value at periastron” and the ω = 2π/P are the instantaneous pulsations of the orbits. Let’s put numbers (see table 3 or Ransom et al. (2014)). The inner white dwarf (WD) being the lightest with a small eccentricity ( 10−4 ), it scarcely contributes a few microseconds. The comparatively more eccentric and larger orbit of the outer WD makes it 12 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin contributes hundreds of microseconds (see figure 6.b). Looking at velocities gives similar results except that the outer motion contributes only at a microsecond level. The coupling term, however, shows a particular behaviour which is by definition specific to a three-body system. Indeed, the full amplitude of the term plays a role as the pulsar goes forward and backward along the outer orbit. Moreover we expect properties of both sub-systems : a 2 fast pseudo-period Pi (and not PI /2 as the vp/I term), an envelope of pseudo-period PO /2 due to the outer eccentricity and a drift of pseudo-period PO because of the misalignment of the two speeds at the end of each inner period, as we can see in figure 6.c. Not only the Einstein correction makes our model more accurate, but it also separates otherwise highly correlated parameters. More specifically, we saw in section 3.3 that the Rømer delay is mostly proportional to a sin i which makes the inclination angle hardly possible to work out independently from the semi-major axis. With the Einstein delay, we can pull apart these two parameters since it does not depend on i while it heavily depends on a. Moreover, we may expect that this will help to determine masses more accurately : masses are somewhat correlated with period and semi-major axes by the mass function (see section 4.2 and appendix A.3) and here these parameters come in differently, with a pseudo period often divided by two. The Einstein correction is thus a major effect, with characteristics specific to a three-body system, that needs to be tackled to achieve a high accuracy model for this system. 3.4.2 Tidal effects Tidal effects can in principle severely affect both the orbital parameters of the system and the intrinsic parameters of the neutron star. Nevertheless, we shall give orders of magnitude to demonstrate that such effects are negligible here. To do so we will focus on the strongest interaction, between the inner companion and the neutron star, and consider negligible the effect of the outer companion. We shall first consider how it might change the spin parameters. Indeed, if the star flattens perpendicularly to its spinning axis ~z, its radius orthogonal to ~z will become larger and the star will spin down as a consequence of conservation of angular momentum. Here we will assume that the orbital axis of the companion is aligned with the spin axis of the pulsar, since this is the case for which the effect is maximized and we do not actually know there relative orientations. Conservation of angular momentum of the neutron star in this configuration just reads : LNS = ωNS INS (28) R Where ωNS is the spin pulsation and INS = NS (z 2 − r2 /3dV ) is the associated moment of inertia. In case of a perturbation of order we would have : LNS = ωNS INS + δωNS INS + ωNS δINS + ◦() (29) Where first order terms must cancel, yielding : δωNS = −ωNS δINS INS (30) Now we need to estimate the δINS involved by the companion. If the star was Newtonian, we R would have INS = NS ρ(~r)(x2 + y 2 )dV , with ρ the local density, and the deformation induced by the external gravitational field of the white dwarf would be computed using Hooke’s law, as it can be found in standard textbooks such as Beutler. Here we need to be more careful for two reasons : the neutron star matter is in the strong field regime and its equation of state is highly uncertain. In the case of a weak gravitational perturbation (the distance from the 13 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin source is much greater than its Shwartzschild radius) given by the quadrupolar tidal field Eij , we may however compute the quadrupole moment of the star Qij through the so-called tidal polarizability λ (Hinderer (2008), Thorne (1998), Fattoyev et al. (2013)) : δQij = −λEij The quadrupole moment of the star δQij tends to δQij (31) −→ Newtonian δρ(~r)(xi yj − ~r2 /3)dV in the Newtonian limit, while the external quadrupolar tidal field Eij tends to the Hessian of 2V the external gravitational potential V : Eij → ∂x∂i ∂x j . In our case we shall use these limits. Assuming that the external tidal field is mainly along the axis x between the two bodies, and the neutron star remains axisymmetric around z we can relate the variation of the moment of inertia to the quadrupole moment as: δINS = δIzz = 3Qxx + 3δIxx (32) The tidal polarizability needs much more cumbersome computations : one needs to integrate the Tollman-Oppenheimer-Volkoff system along with a given equation of state and that cannot usually be achieved analytically. Therefore, we will restrain ourselves to the results obtained by Hinderer (2008) and Fattoyev et al. (2013) for a range of polytropic equations of state, considered as realistic, and take the largest. For a star of radius R = 10km this is about λ ' 3 · 1029 m2 s2 · kg−3 . Let’s now put everything together. It is clear that for a circular orbit the external field would remain constant and the effect would result in a constant shift irrelevant to observation. For a small eccentricity e however, the distance varies by a small amount ed0 from the distance of closest approach d0 , such that the quadrupolar field Exx changes by an amount δExx ' 3eExx during one orbital period. During this period δIxx remains constant and so cancels in equation 32. Thus the shift in spin frequency is equal to : λExx ωNS (33) δωNS ' 9e INS Finally we put some conservative numbers in this equation. ωNS = 325Hz, e = 10−3 , INS = 2π MNS R2 ' 1.2 · 1038 kgm2 (this is for a homogeneous density, but in a neutron star most 15 G ' 9.9 · 10−7 N/m where of the mass lies in the core, so this is conservative) and Exx = MWD d30 we took MWD = 0.2M , MNS = 1.5M and d0 = 1 light-second. It yields : δωNS ' 7 · 10−15 Hz (34) From what we easily conclude that this effect cannot be observed. Another effect would be the torque of the white dwarf on the neutron star tidal deformation. This one could be more important since it would be cumulative over time due to dissipation of the spinning energy (while the previous one was only periodic over an orbital period). However Thorne (1998) reminds us that the neutron matter has a very low viscosity, actually neutron stars are believed to be partly made of a superfluid, and therefore the deformation is well aligned with the field at any time, thus yielding no significant torque. 4 4.1 Fitting to data : a very delicate optimization problem Introduction Our purpose is to fit to data from Nançay the model described by equation 2. The data are TOAs ti with Gaussian uncertainties characterized by their standard deviations σi . If we take a 14 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin short enough period we can also infer how many turns the pulsar did in between two TOAs so as to work out a set Ni of turn numbers. Thus we may just fit N (ti , {θk }) to Ni , where {θk } are the parameters. We need to work out the likelihood of the Ni s from that of the ti s. If one neglects the frequency derivative term as well as delays in equation 2 this is fairly straightforward : one will end up with a Gaussian of standard deviation σi0 = f σi and an amplitude corrected by 1/f . The density of likelihood for the whole set of TOAs thus reads : Y i p(Ni )dNi ∝ Y1 (N (ti , {θk }) − Ni ))2 exp dNi f 2σi0 i (35) The point is that this approximation makes things very simple while we miss very little : as I point out in appendix C, this should not make any significant difference unless we had decades of data. We now have to maximize the likelihood function so as to find the most probable set of parameters that describe the system. In practice we will rather minimize the absolute value of the logarithm, since it is much easier to handle numerically. Given the complexity of the problem we must find an appropriate set of parameters to make optimization easier as well as an appropriate starting solution. Then we need to do the actual minimization with a library called Minuit (James and Winkler, 2004), developed at CERN for 40 years to solve multivariate minimization problems. Finally, we shall find the uncertainties of the parameters we found using a Markov-Chain-Monte-Carlo program called emcee (Goodman and Weare, 2010) that returns a sample of the probability distribution of each parameter. 4.2 4.2.1 Starting solution and the choice of parameters Initial solution Since we still are in a weak regime of gravity, it is possible to use a zeroth-order approximation of the orbits, in other words to assume there are no interactions in between the companions. Such a model is already implemented in the Tempo software under the name of BTX, as previously stated. It was initially designed to take into account the effect of planets turning around pulsars. Though, planets are much lighter than white dwarfs and the BTX model will neither give a reliable solution on a long time period nor an accurate one, but it shall be a good initial guess. Moreover, it proves accurate enough to keep track of the number of turns about itself the pulsar did, and thus provides the N (ti ) needed for the fit. The BTX model does only describe a superposition of two-body motions, involving the neutron star and each of its companions. Such a motion is usually described by so-called orbital elements. In the case of a bounded, elliptic orbit these sum up as : the semi-major axis a, the eccentricity of the orbit e, the orbital period P , the time of passage to the periastron tp , the longitude of the periastron with respect to the line of ascending nodes ω and the angular inclination with respect to the plane of the sky i. The line of ascending nodes is characterized by the intersection of the plane of the ellipse with the plane of the sky. The angle ω is taken with respect to the point of the ellipse where the pulsar is moving away from Earth. These definitions can be found in any celestial mechanics textbook such as Beutler. In the case of the BTX, Tempo will output two sets of orbital elements, one for the inner system and another for the outer system, as well as the spin characteristics of the pulsar f and f 0 . Masses can be determined by taking the root of the so-called mass function (Lorimer and Kramer) f (m, M ) = 0 when one of the masses m or M is known. f (m, M ) = m3 − 4π 2 a3 (m + M )2 = 0 GP 2 15 (36) Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin In the BTX, the mass mp of the pulsar is arbitrarily set to 1.35M , from which one gets the mass of the inner companion using f (mi , mp ) and the mass of the outer companion with f (mo , mi + mp ). More details on the mass function are given in appendix A. 4.2.2 Choice of fitting parameters Since we are using an initial solution defined in terms of orbital elements, it is quite natural to remain with them. Indeed there is a bi-univoque relation between a set of six orbital elements and the position and speed of a body at a given time (Beutler). In the general case they describe the so-called osculating orbit to the actual motion, namely its tangential Keplerian orbit. But the best reason to keep this particular set of parameters is that it avoids degeneracies. Indeed, a naive approach to the problem suggests to fit on a total of six parameters, position and velocity, for the motion of each body, two intrinsic parameters of the pulsar that are its spin period and derivative, and three masses. Eventually we count 23 unknown parameters ! Fortunately, the laws of mechanics can help us to lower it down. If we assume the three bodies constitute an isolated gravitational system, then it has 10 constants of motion among which six we can arbitrarily choose : the center of mass which we set to zero and the impulsion which can also be set to zero as a beginning (proper motion is seldom detectable). Eventually we remain with 17 unknowns. From our previous discussion of the BTX model, we have 12 orbital elements, the pulsar spin period and its derivative. We still need two more parameters reflecting the interactions between the companions that were neglected so far. These can be the masses of the white dwarfs since the mass function is not valid anymore. Indeed, the mass function arises from Kepler’s third law which is only valid for the two-body motion. However this could be quite a hazardous choice of parameters since they come in very symmetric ways in the equations and would likely yield high covariances. Hence I suggest to rewrite equation 38 as follow : 4π 2 a3 GP 2 m3 =µ (m + M )2 (37) In the two-body case, µ = 1 and we get back to equation 38. In the three-body case, however, we obtain two new parameters µpi and µIo 15 . We retrieve masses as before by solving a generalized mass function : 4π 2 a3 (m + M )2 = 0 (38) fµ (m, M ) = m3 − µ GP 2 This parametrization allows to have a direct idea of how the system departs from a zerothorder, double two-body system such as the BTX while reasonably avoiding high covariances. Table 2 summarizes the set of parameters presented in this section. 4.3 Minimization with the Minuit library Minuit is the name of a library that contains several minimizers. The most famous is called Migrad, and is mainly the one I used. Migrad is a variable metric minimizer (DAVIDON, 1991). The so-called metric is actually nothing more than the Hessian H of the the function f of variables x to minimize. It roughly follows the scheme below : • Compute the local Hessian. 15 I will use index i to point out the inner companion but index I for the inner system, pulsar and inner companion. 16 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Names of intrinsics Pulsar spin frequency Pulsar spin frequency derivative Pulsar mass Mass coefficient Mass coefficient Intrinsic f f0 mp µip µIo Inner eI ap Ωp TI PI iI Outer eO ao Ωo TO PO iO Guillaume Voisin Names of orbitals Eccentricity Semi-major axis Longitude of periastron Periastron epoch Orbital period Inclination of the orbit Table 2: Summary of the parameters of the system. Capital indices are used when the parameter applies to the whole (inner or outer) system, while lower-case indexes apply when it is specific to one of the bodies (in this case the parameter of the other body is a combination of the given orbital elements). • Perform a Newton-Raphson step from the current position xc to evalutate the position xmin of the local minimum : xmin = xc + H −1 gc where gc is the current gradient. This step is exact for a purely quadratic function. • Evaluate the difference between f (xmin ) and its quadratic approximation. If inferior to an arbitrary limit, it stops, otherwise the metric is modified and a new step is attempted. There are different ways to update the metric, Minuit uses the scheme proposed by Fletcher (1970), that uses the local gradients and avoids an expensive computation of the full Hessian. As we can see, the accurate computation of derivatives is a critical point in Migrad. Since it proceeds with finite differences, numerical accuracy may be, and was, an issue. Another point is due to the high number of parameters. It implies a high number of local minima, or more likely there might be a local minimum for a given set of parameters that disappears because of covariances as soon as one allows to fit on a larger set of parameters. Alternatively, fitting all the parameters at the same time can fail simply because the parameter space is too big and the minimizer gets stuck with an unphysical solution. For example it may start from position where two variables are very covariant, and thus hardly sees a minimum, even though in the real solution these parameters are not that much correlated. In this case, a solution is to start by fixing one of the parameters and to proceed in two successive fits. To be honest, there is a fair portion of try and miss in this kind of search. However, it is very helpful to develop tools to diagnose why a fit may have failed. I mostly developed two such tools. One is a tracker, that picks and records every step that Migrad does during a fit. It proved particularly useful to understand numerical accuracy issues. For example if numerical steps in the likelihood function occur for a variation of eccentricity of order 10−14 and that Migrad tries to compute a gradient by doing two such steps, it might yield a catastrophe, even though we are well below our desired physical accuracy. The other tool that proved helpful was the use of fake pulsars. Here is the principle : take data as well as some plausible set of parameters and, instead of fitting parameters to data, fit TOAs to parameters. Then just change the parameters by some random very little amount that gives a posterior probability similar to the one we have with the actual data. Once you have it, just apply the same fitting strategy to both fake and real pulsar. The former will tell you with certainty whether you got closer to the solution or not (you can have a better probability but got further away from the real solution), and thus allow you to reject or keep the solution of the latter. On figure 7 I show the current status of the fit. The weighted standard deviation of these 17 Modélisation du système triple autour du pulsar radio PSR J0337+1755 100 Guillaume Voisin Time residuals 80 Residuals (µs) 60 40 20 0 20 40 60 56450 56500 56550 56600 56650 Times of arrival (MJD) 56700 56750 Figure 7: Best timing residuals obtained so far including Rømer and Einstein delays. It includes 3063 TOAs from Nançay spaning over 200 days. residuals is 23µs. The parameters corresponding to this fit are given in table 3. One of the current limitations for this fit is that the time span does not cover the full outer orbit, thus making it difficult to resolve precisely. This might be provided by sharing data with the Ransom et al. (2014) team, in the framework of the collaboration with Anne Archibald. Now that we have a candidate solution, we want to get the uncertainties on it. This is the purpose of the next section. 4.4 The Markov-Chain-Monte-Carlo (MCMC) approach The MCMC algorithms are not really minimizers but rather statistical samplers. Indeed they output a sample of the probability distribution of each parameter. On the one hand they can be used to fit a model to data, since in this case we are looking for the most probable model, but they often reveal themselves to be pretty slow at it and thus should only be used when everything else fails. On the other hand, the statistical samples can be used to get realistic error bars on each parameter, which is not possible with minimizers like Minuit that implicitly assume that the model is quadratic (this is usually true locally but it can severely bias the global probability distribution). As a consequence we will use MCMC to get the error bars of our model. To be a bit more specific, let’s introduce the posterior probability density of having a given set of parameters Θ = {θi } for a given set of data D, p(Θ|D), the likelihood of D given Θ, p(D|Θ), and the prior probability p(Θ). We then have the simple bayesian relation : p(D|Θ) × p(Θ) = p(Θ|D) × p(D) (39) The prior is usually flat, there is no reason to privilege a set of parameters over another, and will rather be used to restrict the search over an arbitrary parameter-space volume. The likelihood is known from section 4.1. Hence we know the posterior probability density up to a constant factor p(D) which all the Metropolis algorithm needs (see Diaconis (2009) or Foreman-Mackey et al. (2013)). 18 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin 6.9 6.93 6.94 6.95 6.96 7 3.5 2 3.5 5 3 3.5 0 3 3.5 5 40 3.5 −4 3e.5O (x 3.5103.5) 25 30 35 40 Indeed the algorithm starts with an arbitrary probability p(Θ0 |Θ) to go from a set to another. Then it successively tries to jump to new parameter sets at each time adjusting p(Θ0 |Θ) until it becomes stationary. When it is so, the fundamental theorem of Markov chains (see Diaconis (2009) for a rigorous mathematical treatment) tells us that the chain has converged and that from now on any sample of Θs drawn from the successive trials of the chain will follow a distribution equal to p(Θ|D) (To be mathematically exact the sample must of course be infinite). This algorithm is very robust : the only assumption is the simple connectedness of the parameter space, the likelihood and a fortiori the posterior probability do not even need to be analytical. Moreover, it will theoretically explore completely the allowed volume and find every local minimum. In practice, this would require a real random generator and an infinite time ahead, which is why it is preferable to already have a good solution to start with. Moreover, although random evaluations are slow, it is easily parallelized by running several chains at the same time (a hundred in our case). One can even use parallel tempering, which consists in running chains with broader likelihood (actually it enlarges the standard deviations of the Gaussian, just like a temperature coefficient would do, hence the name of the algorithm), that will spend less time in local minima and go through parameter space faster, before eventually sharing this information with other, more accurate, chains (Earl and Deem, 2005). eI (x10−2 ) eO (x10−4 ) Figure 8: Statistical distribution of the inner and outer eccentricities as well as their correlation plot. We can see that this last plot is roughly isotropic, and so that the statistical correlation is low, which is the case for all outer parameters with respect to inner parameters, as one might expect. The blue lines show where the fitted value with Minuit is. In practice I used a program called emcee implemented by Foreman-Mackey et al. (2013) featuring the affine invariant scheme16 of Goodman and Weare (2010). A run of this program thus gives all the statistical information one might need about each parameter, including correlations with other parameters, as illustrated in figure 8. The final results for the first 3063 TOAs from Nançay are given in table 3 with there error bars estimated at a 90% confidence level 17 .S 16 Affine invariant scheme : such a scheme sees linear combinations of parameters as a same class of equivalence. For instance, it yields the same results whatever a length was scaled in micrometers or in light years. It is thus very robust in that matter. 17 For example, x = 1.230+6.7 −4.5 means that, if the real value xr is above x it has 90% of being between 1.23 and 1.297, and respectively if it is below. 19 Modélisation du système triple autour du pulsar radio PSR J0337+1755 f (s−1 ) f 0 (s−2 ) Mp (M ) µip µIo 365.95332+1.4 −2.2 −14 −4.443+0.73 −1.1 · 10 +6.1 1.4325−4.7 1.0000+1.4 −1.9 1.00000+2.6 −2.9 eI ap (ls) Ωp (rad) TI (MJD) PI (days) iI (rad) −4 6.9569952+2.7 −3.2 · 10 1.2178+2.5 −3.4 1.6447+6.7 −8.4 55917.5+2.0 −7.2 1.6294+4.9 −5.7 1.5483+9.6 −7.2 Guillaume Voisin −2 eO 3.53150+1.3 −1.2 · 10 aI (ls) 74.677+5.1 −5.0 ΩI (rad) 1.6708+4.1 −3.5 +3.3 TO (MJD) 56317.21−2.6 +2.0 PO (days) 327.262.0 iO (rad) 1.5709+5.5 −7.9 Table 3: This table shows the best fitted parameters for the 3063 first TOAs from the Nançay decimetric telescope, for the system J0337+1755, with their error bars. The errors are given for the last digit at a 90% confidence level. (ls stands for light-second) 5 Conclusion The principal aim of this work was to improve the model used to analyse the raw data from the Nançy radio-telescope. A look at the timing residuals of the initial model (figure 1) compared to the residuals given by the model developed during this internship (figure 7) suggests that this goal was somewhat achieved. Indeed the residuals were decreased by almost two orders of magnitude and are now stable over long periods. To achieve this, a model was realized from scratch with comparison to the existing binary models when possible. Post-Newtonian effects such as the Einstein correction were taken into account and proved to matter at a high level. From a more technical point of view, numerical round-off errors were systematically taken into account and tested against. It involves a set of tools : the main frame is written in Python with an underlying layer of Cython 18 for all the computations where accuracy and speed are critical, iMinuit and emcee libraries are in charge of the minimization operations, and the whole interoperates with a C++ code featuring the library Boost for numerical integration of the differential equations of motion. From a user point of view, it presents itself as a fully documented Python class that can be forked from a Subversion19 repository hosted at LPC2E. Though this model and code is not yet complete. A better accuracy can undoubtedly be achieved and must be if one wants to consider tests of the strong equivalence principle. It will need to include more post-Newtonian effects, in particular in the equations of motion which are currently merely Newtonian. Indeed, orders of magnitude clearly suggest that periastron advance is to be considered in this system. For the sake of completeness, one should also work out the Shapiro delay20 even though this one is not likely to matter at a very significant level in this system, as deduced both from orders of magnitude and according to the more advanced work of Anne Archibald. Moreover, a greater number of times of arrival covering a larger period of time will be needed to perfectly resolve all the parameters. This point might eventually be solved in the frame of the collaboration with Anne Archibald, by sharing data, and might also be critical to achieve the most accurate possible test of the strong equivalence principle. In the next weeks, Ismaël Cognard and Lucas Guillemot will try to use the model developed during this internship to improve the accuracy of the times of arrival recorded at Nançay, with the hope of decreasing the uncertainty by an order of magnitude, down to a few hundred nanoseconds. 18 Cython : C-extension to Python that allows to use C-static-type definitions and optimisations. http://www.cython.org/ 19 Subversion is a centralized version control system. https://subversion.apache.org/ 20 Shapiro delay : delay added by the gravitational field of a companion when the light passes close to it on its way to a telescope. 20 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin References Scipy. URL http://www.scipy.org/. Gerhard Beutler. Methods of Celestial Mechanics: Volume I: Physical, Mathematical, and Numerical Principles. Springer, 1st softcover edition without CD-ROM of original hardcover edition. edition edition. ISBN 9783540407492. Roger Blandford and Saul A. Teukolsky. Arrival-time analysis for a pulsar in a binary system. The Astrophysical Journal, 1976. Thibault Damour and Nathalie Deruelle. General relativistic celestial mechanics of binary systems. II. the post-newtonian timing formula. 44:263–292, 1986. William DAVIDON. VARIABLE METRIC METHOD FOR MINIMIZATION. 1991. Persi Diaconis. The markov chain monte carlo revolution. Bulletin of the American Mathematical Society, 2009. David J. Earl and Michael W. Deem. Parallel tempering: Theory, applications, and new perspectives. 7(23):3910, 2005. ISSN 1463-9076, 1463-9084. doi: 10.1039/b509983h. URL http://arxiv.org/abs/physics/0508111. R. T. Edwards, G. B. Hobbs, and R. N. Manchester. TEMPO2, a new pulsar timing package - II. The timing model and precision estimates. Monthly Notices of the Royal Astronomical Society, 372:1549–1574, November 2006. doi: 10.1111/j.1365-2966.2006.10870.x. F. J. Fattoyev, J. Carvajal, W. G. Newton, and Bao-An Li. Constraining the high-density behavior of nuclear symmetry energy with the tidal polarizability of neutron stars. Physical Review C, 87(1), January 2013. ISSN 0556-2813, 1089-490X. doi: 10.1103/PhysRevC.87. 015806. URL http://arxiv.org/abs/1210.3402. arXiv: 1210.3402. R. Fletcher. A new approach to variable metric algorithms. The Computer Journal, 13(3): 317–322, January 1970. ISSN 0010-4620, 1460-2067. doi: 10.1093/comjnl/13.3.317. URL http://comjnl.oxfordjournals.org/content/13/3/317. D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman. emcee: The MCMC Hammer. Publications of the Astronomical Society of the Pacific, 125:306–312, March 2013. doi: 10. 1086/670067. T. Gold. Rotating neutron stars as origin of pulsating radio sources. Nature, 218(5143):731–&, 1968. ISSN 0028-0836. doi: 10.1038/218731a0. WOS:A1968B164900007. Thomas Gold. Rotating neutron stars and the nature of pulsars. Nature, 221(5175):25–27, January 1969. doi: 10.1038/221025a0. URL http://www.nature.com/nature/journal/ v221/n5175/abs/221025a0.html. Jonathan M. Goodman and Jonathan Weare. Ensemble samplers with affine invariance. Communications in Applied Mathematics and Computational Science, 2010. Malte Henkel. Sur la solution de sundman du problème des trois corps. Philosophia Scientiae, 5(2):161–184, 2001. ISSN 1281-2463. URL https://eudml.org/doc/103658. 21 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin A. Hewish, S. J. Bell, J. D. H. Pilkington, P. F. Scott, and R. A. Collins. Observation of a rapidly pulsating radio source. Nature, 217(5130):709–713, February 1968. ISSN 0028-0836. doi: 10.1038/217709a0. Tanja Hinderer. Tidal love numbers of neutron stars. The Astrophysical Journal, 677(2): 1216–1220, April 2008. ISSN 0004-637X, 1538-4357. doi: 10.1086/533487. URL http: //arxiv.org/abs/0711.2420. arXiv: 0711.2420. G. B. Hobbs, R. T. Edwards, and R. N. Manchester. TEMPO2, a new pulsar-timing package I. An overview. Monthly Notices of the Royal Astronomical Society, 369:655–672, June 2006. doi: 10.1111/j.1365-2966.2006.10302.x. Fred James and Matthias Winkler. Minuit user’s guide. CERN, Geneva, 2004. URL http: //www.ftp.uni-erlangen.de/pub/mirrors/gentoo/distfiles/mnusersguide.pdf. D. R. Lorimer and M. Kramer. Handbook of Pulsar Astronomy. Cambridge University Press. ISBN 9780521828239. Charles W. Misner and John Archibald Wheeler. Gravitation. W. H. Freeman, September 1973. ISBN 9780716703440. W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical recipes in FORTRAN. The art of scientific computing. 1992. S. M. Ransom, I. H. Stairs, A. M. Archibald, J. W. T. Hessels, D. L. Kaplan, M. H. van Kerkwijk, J. Boyles, A. T. Deller, S. Chatterjee, A. Schechtman-Rook, A. Berndsen, R. S. Lynch, D. R. Lorimer, C. Karako-Argaman, V. M. Kaspi, V. I. Kondratiev, M. A. McLaughlin, J. van Leeuwen, R. Rosen, M. S. E. Roberts, and K. Stovall. A millisecond pulsar in a stellar triple system. Nature, 505:520–524, January 2014. doi: 10.1038/nature12917. Scott M. Ransom. Pulsars are cool. seriously. Proceedings of the International Astronomical Union, 8(S291):3–10, 2012. URL http://journals.cambridge.org/abstract_ S1743921312023046. Karl F. Sundman. Mémoire sur le problème des trois corps. Acta Mathematica, 36(1):105– 179, December 1913. ISSN 0001-5962, 1871-2509. doi: 10.1007/BF02422379. URL http: //link.springer.com/article/10.1007/BF02422379. Kip Thorne. Tidal stabilization of rigidly rotating, fully relativistic neutron stars. Physical Review D, 58(12), 1998. doi: 10.1103/PhysRevD.58.124031. Clifford M. Will. The confrontation between general relativity and experiment. Living Reviews in Relativity, 17, 2014. ISSN 1433-8351. doi: 10.12942/lrr-2014-4. URL http://relativity. livingreviews.org/Articles/lrr-2014-4/. 22 Modélisation du système triple autour du pulsar radio PSR J0337+1755 A A.1 Guillaume Voisin The gravitational two-body system defined by orbital elements Relation between orbital elements and state vectors for a single body There is a bi-univoque relation between between state vectors (position and velocity) and orbital elements. In the frame we may call R0 made of the semi-major axis, the semi-minor axis and a third vector orthogonal to the plane of the ellipse (say the angular momentum ~h this relation stands as follow : ~r = a(1 − e cos E) (cos(v), sin(v), 0)0 √ 2π a d~r 2 = − sin E, 1 − e cos E, 0 dt (1 − e cos E) P 0 (40) (41) Where E is the eccentric anomaly and v the true anomaly which can be defined using the following relations : t − tp (42) E − e sin E = 2π P Which is known as Kepler’s equation. It cannot be solved analytically so we use a fixed-point method to solve it to desired accuracy (see for example Beutler). Then one needs two more relations which merely account for the geometrical relation between E and v : √ 1 − e2 sin E (43) cos v = 1 − e cos E cos E − e sin v = (44) 1 − e cos E Let’s remark that we did not use two parameters : i and ω. These allow to orientate the frame in space. In this case, one will need to perform two rotations on 40 : one around ~h of angle ω such that the first component be along the line of ascending node n~a , and a second around n~a of angle i such that the third component be along the line of sight n~ . A.2 The case of two bodies In the case of two bodies one only needs to add one parameter, which is the mass of one of the bodies, the second one being computed using Kepler’s third law (see below). The reason why one does not need to add as many parameters as there are new initial conditions to the equations of motion is that six integrals of motion can be set arbitrarily, thus providing six relations between the initial conditions. Those integrals are the three components of the center of mass and of the impulsion of the system. Remark that this assumes an isolated system. A.3 The mass function The mass function f (m, M ) is simply the expression of Kepler’s third law, which stands that the mass m of a body involved into a Keplerian two-body orbit is related to the period P, the semi-major axis a, the tilt angle i and the mass M of the other object by finding the root of f : 4π 2 (a sin i)3 f (m, M ) = (m sin i)3 − (m + M )2 = 0 (45) GP 2 23 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin Where G is the gravitational constant. This is a third order polynom in m which can be efficiently solved by standard root finding algorithms, such as those implemented in Scipy (sci). We clearly see here that the single-body motion is equivalent to the two-body case with one infinite mass. B Jacobian of the Newtonian three-body differential system Following the notations of section 3.2.2, it is possible to map the subsets I, I 0 , J, J 0 , K, K 0 to the component indices of Q : I I0 J J0 K K0 → → → → → → [1 : 3] [10 : 12] [4 : 6] [13 : 15] [6 : 9] [16 : 18] From that we can compute the Jacobian as Jij = ∂FL = ∂Q0M ∂FI0 ∂QI ∂FI0 ∂QJ ∂FI0 ∂QK ∂Fi ∂Qj (46) (47) (48) (49) (50) (51) : δLM where L, M ∈ {I, J, K} = −MJ JacI (fg (QI , QJ )) − MK JacI (fg (QI , QK )) = −MJ JacJ (fg (QI , QJ )) and circular permutations of {I, J, K} = −MK JacK (fg (QI , QK )) Here I used the gravitational force function fg (q1 , q2 ) = respect to the 3-vector q1 , reads : q1 −q2 kq1 −q2 k3 (52) the Jacobian of which, with ∂fg i δij 3 (q1 − q2 )i (q1 − q2 )j = 3 − ∂q1 i 2 kq1 − q2 k kq1 − q2 k5 (53) And : JacI (fg (QI , QJ )) = −JacJ (fg (QI , QJ )) = −JacI (fg (QJ , QI )) = JacJ (fg (QJ , QI )) C (54) Likelihood function Given a Gaussian likelihood density for a toa t, the likelihood for the associated turn number N is given by straightforward probabilities : Z Z ∀∆t, p(t)dt = p0 (N )dN (55) ∆t ∆N (∆t) dN 0 dt ⇔ p(t)dt = p (N ) (56) dt 24 Modélisation du système triple autour du pulsar radio PSR J0337+1755 Guillaume Voisin Hence one has to compute the derivative of the phase with respect to time. Though it is clear that delays will contribute to second order since their characteristic period of variation is the orbital period (let’s take the shortest PI ) and they are bounded to a hundred seconds at most (in the case of the Rømer delay). Their derivatives will consequently be bounded and small compared to the main contribution : dN = f + f 0 (t − t0 ) + (delay terms) dt (57) In computations, we try to minimize the logarithm of the likelihood which to first order in − t0 ) ' 10−11 (t − t0 ) reads : f0 (t f f0 log(p(N )) ' log(p(t)) − log(f ) − (t − t0 ) f (58) The last term is the only one new compared to equation 35. It can become important since it is proportional to time. Also we must remind ourselves that the previous formula must be summed up for all toas. So assuming we can observe 5000 TOAs a year for n years how long will it be before it changes the log of the total likelihood by, say, 0.001 (which in practice not significant for such a number of TOAs) ? 5000n f0 × 365n > 0.001 ⇒ n > 7 f (59) Compared to a purely Gaussian likelihood this implies a shift of about 10−7 turn per turn or 1ns per TOA, well below the experimental accuracy. 25