Draft version November 11, 2019Preprint typeset using LATEX style emulateapj v. 01/23/15

THE SLOWEST SPINNING X-RAY PULSAR IN AN EXTRAGALACTIC GLOBULAR CLUSTER

Ivan Yu. Zolotukhin1,2,3, Matteo Bachetti4, Nicola Sartore5, Igor V. Chilingarian6,2, Natalie A. Webb5,1

1 Universite de Toulouse; UPS-OMP, IRAP, 9 avenue du Colonel Roche, BP 44346, F-31028 Toulouse Cedex 4, France2 Sternberg Astronomical Institute, Moscow State University, Universitetskij pr., 13, 119992, Moscow, Russia

3 Special Astrophysical Observatory of the Russian Academy of Sciences, Nizhnij Arkhyz 369167, Russia4 INAF/Osservatorio astronomico di Cagliari, via della Scienza 5, I-09047 Selargius, Italy

5 CNRS, IRAP, 9 avenue du Colonel Roche, BP 44346, F-31028 Toulouse Cedex 4, France and6 Smithsonian Astrophysical Observatory, 60 Garden St. MS09, Cambridge, MA, 02138, USA

Draft version November 11, 2019

ABSTRACT

Neutron stars are thought to be born rapidly rotating and then exhibit a phase of a rotation-poweredpulsations as they slow down to 1–10 s periods. The significant population of millisecond pulsarsobserved in our Galaxy is explained by the recycling concept: during an epoch of accretion from adonor star in a binary system, the neutron star is spun up to millisecond periods. However, onlya few pulsars are observed during this recycling process, with relatively high rotational frequencies.Here we report the detection of an X-ray pulsar with Pspin = 1.20 s in the globular cluster B091D inthe Andromeda galaxy, the slowest pulsar ever found in a globular cluster. This bright (up-to 30%of the Eddington luminosity), high spin-up rate pulsar, persistent over the 12 years of observations,must have started accreting less than 1 Myr ago and has not yet had time to accelerate to hundredsof Hz. The neutron star in this unique wide binary with an orbital period Porb = 30.5 h in a 12 Gyrold, metal rich star cluster, accretes from a low mass, slightly evolved post-main sequence companion.We argue that we are witnessing a neutron star produced in an accretion induced collapse event thatsubsequently exchanged its low mass donor in a binary by a ∼0.8M� star – a viable scenario in amassive dense globular cluster like B091D with high global and specific stellar encounter rates. Thisintensively accreting non-recycled X-ray pulsar provides therefore a long-sought missing piece in thestandard pulsar recycling picture.Keywords: X-rays: binaries — pulsars: individual (3XMM J004301.4+413017, CXO

J004301.4+413016, XB091D) — globular clusters: individual (B091D, Bol D91, M31GCJ004301+413017) — galaxies: individual (M31) — astronomical databases: miscellaneous— virtual observatory tools

1. INTRODUCTION

Around 2000 pulsars are known, where the majority ofthese are ’regular’ pulsars which have pulse periods be-tween tens of miliseconds to approximately a second, andmagnetic field strengths of ∼1012 G (Manchester et al.2005). These pulsars show a general spin down to longerperiods. However, a few hundred of these pulsars showmuch shorter periods, of the order a millisecond, alongwith lower magnetic fields of ∼108 G. It is believed thatthese millisecond pulsars (MSPs) are the descendants ofneutron stars/pulsars found in X-ray binaries. Accretiononto a neutron star from a close companion is believedto transfer angular momentum to the neutron star, spin-ning it up to periods of milliseconds (Alpar et al. 1982;Radhakrishnan & Srinivasan 1982). This is strongly sup-ported by both the discovery of a MSP in an X-raybinary system (SAX 1808.4-3658, Wijnands & van derKlis 1998) and the presence of kilo-Hertz Quasi-PeriodicOscillations in many LMXBs, which have been foundto show millisecond pulsation periods (see van der Klis1998, and references therein). More recently, Archibaldet al. (2009) showed that the previously accreting mil-lisecond pulsar FIRST J102347.67+003841.2 had ceasedto accrete and radio pulsations could subsequently beobserved, thus supporting the ’recycled’ pulsar idea.

Many of the known MSPs are found in Galactic globu-lar clusters, where as of the end of 2015, almost 140 MSPs

have been detected1. Globular clusters are dense spher-ical systems of ∼104–106 old stars (e.g. Henon 1961).Their old age implies that they should also contain manycompact objects (e.g. Hut et al. 1992). Stellar encoun-ters, which are extremely rare in lower density regions,can occur in globular clusters on time-scales comparablewith or less than the age of the Universe. This wouldindicate that many Galactic globular cluster stars haveundergone at least one encounter in its lifetime. En-counters between stars is one way in which binaries canbe produced. The encounter rate (Γ) due to tidal cap-ture (Fabian et al. 1975) is proportional to the encountercross-section, the relative velocity of the stars and thenumber density of stars in the cluster (core). Both pri-mordial binary systems and those formed due to encoun-ters should exist in globular clusters due to the denseenvironments, but encounters between a binary and ei-ther a single star or a binary system would more readilyoccur as the cross sections are significantly larger, thusincreasing the likelihood of an encounter. This explainsthe large number of recycled pulsars that we observe inGCs.

However, a small number (6) have periods greater than0.1 s and these pulsars have not yet been fully recycled.The longest of these, B 1718-19 in NGC 6342, has a pe-riod of 1.004 s and a magnetic field of ∼ 1012 G, typical

1 http://www.naic.edu/~pfreire/GCpsr.html

arX

iv:1

602.

0519

1v1

[as

tro-

ph.H

E]

16

Feb

2016

2 Zolotukhin et al.

of a ’regular’ pulsar. It appears to be a young pulsar,with a characteristic age of 1 × 107 years (Lyne et al.1993). Lyne et al. (1993) propose that this young pulsaroriginated either from a collision between an old neutronstar and a cluster star, or that a white dwarf accretingfrom a companion underwent an accretion induced col-lapse (Michel 1987).

Following the recent detection of coherent pulsationsfrom an ultra-luminous X-ray source (ULX) in the M82galaxy ('3.5 Mpc away) using NuSTAR data (Bachettiet al. 2014), which showed that this bright source was infact a neutron star, we started to search archive XMM-Newton data to find similar sources, in order to addressquestions such as how can such super-Eddington lumi-nosities be possible in a neutron star (see e.g. Mushtukovet al. 2015; King & Lasota 2016, for further discussion).

In this paper we describe the analysis that made itpossible to detect a 1.2 s pulsar in the low-mass X-raybinary 3XMM J004301.4+413017, associated with theglobular cluster B091D from the Revised Bologna Cat-alog of M31 globular clusters (RBC V.5; Galleti et al.2004), using public data. For brevity we denote thepulsar XB091D after its host globular cluster designa-tion. This is the first X-ray pulsar detected in M31, andit also has the longest period among all known pulsars(rotation-powered and accreting) in globular clusters, be-ing more than an order of magnitude slower than the re-cycled accreting pulsars from the two GCs NGC 6440 andTerzan 5 (Altamirano et al. 2010; Papitto et al. 2011).We note that this pulsar has also recently been detectedby Esposito et al. (2016), but these authors interpret thenature of this source quite differently. In this paper, wediscuss possible evolutionary scenarios that may producesuch slowly rotating neutron stars in globular clusters.

The content is organized as follows: Section 2 brieflydescribes the XMM-Newton photon database (which willbe described in detail elsewhere) and the dataset usedfor the initial pulsation detection, as well as the XMM-Newton pulsar factory analysis methods which resultedin the automated detection of the pulsed X-ray emission;Section 3 covers the manual blind search for pulsed emis-sion in all available XMM-Newton data for this sourceand the determination of source’s orbital parameters;Section 4 and 5 summarize our findings on the timingand spectral properties of this source which are then dis-cussed in Section 6 where we also argue on the possibleorigin and evolution of this system.

2. PULSATIONS SEARCH DATA AND METHODS

2.1. Photon database

For this study we created a database of all photonsregistered by the EPIC/pn detector of the XMM-Newtonsatellite and operated in imaging mode during the 7781observations that took place between 2000 and mid 2013.These are the same observations that were used to com-pile the 3XMM-DR5 catalog (Rosen et al. 2015). Thephoton data were taken from the event lists called PIEVLIfiles publicly available from the XMM-Newton sciencearchive2. These are science-ready data products thatcome from the pipeline run by the XMM-Newton Sur-vey Science Centre (XMM SSC)3. Each file is a binary

2 http://xmm.esac.esa.int/xsa/3 http://xmmssc.irap.omp.eu

table with one row per event that contains the follow-ing information: (a) event time (satellite clock), (b) rawCCD pixel and projected sky coordinates relative to thenominal pointing position, (c) corrected and uncorrectedevent energy; (d) event quality flag; (e) CCD number.Each PIEVLI file also contains good time intervals de-termined separately for each pn CCD by the XMM SSCpipeline software. We point out that it is not necessary toretrieve the complete XMM-Newton observation archive(ODF archive) for the massive scale timing analysis withthe XMM-Newton.

These event lists represent the lowest and most de-tailed data level we use for initial timing studies. Theseare accompanied by higher level data in the form of the3XMM-DR5 catalog of X-ray sources and their individ-ual observations acting essentially as an index for thenavigation within the photon database. The connectionbetween these two very different sets of information isachieved by means of the known transformation betweenpixel coordinates of the events and world coordinates ofthe X-ray sources from the 3XMM-DR5 catalog. Weused the World Coordinate System (WCS) transforma-tion data available for every EPIC/pn exposure in thePIEVLI file. In this fashion one can make event extrac-tion for any detection from the 3XMM-DR5 catalog usingits world coordinates, right ascension and declination.

The catalog of X-ray sources was cross-matched withother catalogs of astrophysical objects in order to findpossible counterparts and determine the source type orlocation within a galaxy. This gives us the possibility toeasily extract arbitrary photon lists of e.g. all photonscoming from the M31 galaxy, or all photons from knownmagnetars.

The last stage before being able to launch our timinganalysis codes over the sequence of extracted photon listsis the barycentric correction, i.e. correcting the eventtimes to the Solar System barycenter. For this purposewe used the XMM-Newton orbit files fed to the barycentask from the XMM-Newton SAS version 13.5 softwareand the source positions from the 3XMM-DR5 catalog.

We note that it is now possible to access this pho-ton database as well as the 3XMM-DR5 catalog datathrough the convenient web interface4 that we developedwhile working on the 3XMM-DR5 catalog compilationand this project. In particular, one can extract barycen-tered photons in arbitrary regions from the observationlevel event lists using nothing but a web browser. Moredetails on this website for the quick-look science analysisof the XMM-Newton data will be presented in a separatepaper (Zolotukhin et al., in prep.)

2.2. Pulsar factory analysis software

The photon database provides a way to easily extractcalibrated and barycentered event lists for arbitrary setsof astrophysical sources observed with the EPIC/pn cam-era on board the XMM-Newton satellite. We developedan analysis software aimed at finding coherent pulsationsin the photon database, optimized for high throughput.

It consists of few analysis layers. First, for each eventlist it produces the power density spectrum (PDS) adapt-ing the time binning so that the Nyquist frequency is wellabove 2 kHz and the total number of bins between gaps

4 http://xmm-catalog.irap.omp.eu

A slow luminous extragalactic X-ray pulsar 3

3 arcsec

DSS: 48-inch Schmidt

10 arcmin

30 arcsec

HST ACS (F555W+F814W)

10 pc

Figure 1. The position of the XB091D and its host globular cluster B091D in the Andromeda galaxy in several scales – using DigitizedSky Survey and HST ACS colored images. On the zoomed inset in the bottom left corner 1σ X-ray positional uncertainties from Chandra(red circle) and XMM-Newton (blue circle) catalogs are overplotted.

in the light curve is factorable with small prime numbersfor better performance with the numpy FFT algorithm.In the PDS the algorithm searches for peaks exceedingthe 99 per cent detection level, including the numberof trials, following the standard rules from Leahy et al.(1983). If such peaks are found, it launches the Z2

n test(Buccheri et al. 1983) in the vicinity of their frequencies.If a peak from the PDS is confirmed with the Z2

n test,an accelerated search with the PRESTO code (Ransom2001) is launched to constrain further or discard this pul-sation frequency candidate. As the final step of the auto-mated procedure the pulsar factory checks this frequencyin other observations of this source from the 3XMM-DR5catalog. If the detection is highly significant (>nσ, wheren can vary between runs) or the same frequency is foundin other detections, an operator is notified.

We first tested the XMM-Newton pulsar factory codeusing XMM-Newton data on a set of known X-ray brightmagnetars from the McGill Online Magnetar Catalog 5

(Olausen & Kaspi 2014). After achieving the stable de-tection of known coherent pulsation periods in this auto-mated regime, we launched a larger period search on theunstudied sample of X-ray sources in the catalog. TheM31 galaxy was an obvious choice of survey region for itslarge number of known X-ray sources and its relativelylow distance, and was included in our first survey.

3. DETECTION AND ORBITAL PARAMETERS

During the M31 test run, the automated XMM-Newton pulsar factory algorithm detected pulsations,at about the same period, in 3 observations (ObsIDs:

5 http://www.physics.mcgill.ca/~pulsar/magnetar/main.html

Table 1Best orbital solution found in this paper.

Parameter Value

Porb 1.27101304(16) dTasc MJD 56104.791(26)a sin i/c 2.89(13) l-sece <0.003

Note. — Brackets indicate 1-σ uncertainties, as re-turned by TEMPO2. The upper limit on the eccentricityis the maximum uncertainty returned by the ELL1 model,for values of the eccentricity always consistent with 0.

0112570101, 0505720301, P0650560301) of the source3XMM J004301.4+4130176. The Z2

n test triggered bythe detection confirmed the candidate. These detectionsare fully reproducible online from the source’s event ex-traction pages, e.g. for observation 0112570101: http://xmm-catalog.irap.omp.eu/pievli/101125701010068.

In order to look for the detected pulsation in more Ob-sIDs, we ran an accelerated search with PRESTO in allObsIDs containing the source. The very high values ofperiod derivative required by the accelerated search, theclear improvement of detection significance when addinga second derivative in the search, and the shape of thetrack in the phaseogram shown by PRESTO, pointedtowards the presence of orbital modulation. Largely fol-lowing the same procedure described in Bachetti et al.(2014), we cut the two longest ObsIDs into chunks, 10–30 ks long, and ran an accelerated search with custom-made software, searching the solution in the ν − ν plane

6 See the source web page at the 3XMM-DR5 catalog website:http://xmm-catalog.irap.omp.eu/source/201125706010086

4 Zolotukhin et al.

Table 2XMM-Newton observations used for timing analysis in this study and its main results.

ObsID Obs. start Exposure Pspin S / d.o.f. Pulse depthUTC (s) (s)

0112570101 2002-01-06 18:45:45 64317 1.203898(49) 8.7 0.33 ± 0.07

0405320701 2006-12-31 14:24:50 15918 1.203731(11) 4.1 0.28 ± 0.080405320901 2007-02-05 03:44:24 16914

0505720201 2007-12-29 13:42:13 27541 1.203738(10) 7.3 0.20 ± 0.050505720301 2008-01-08 07:01:05 272190505720401 2008-01-18 15:11:47 228170505720501 2008-01-27 22:28:21 21818

0551690201 2008-12-30 03:27:52 21916 1.203662(8) 5.0 0.23 ± 0.050551690301 2009-01-09 06:19:54 219180551690501 2009-01-27 07:23:03 219120551690601 2009-02-04 13:21:03 26917

0600660201 2009-12-28 12:42:54 18820 1.203675(13) 4.2 0.25 ± 0.080600660501 2010-01-25 02:39:14 19715

0650560301 2011-01-04 18:10:16 33415 1.203651(8) 9.1 0.31 ± 0.060650560401 2011-01-15 00:16:57 243160650560601 2011-02-03 23:58:12 23918

0674210201 2011-12-28 01:07:36 19034 1.203634(13) 4.9 0.35 ± 0.090674210301 2012-01-07 02:47:01 154330674210401 2012-01-15 15:00:38 199160674210501 2012-01-21 12:22:03 17317

0690600401 2012-06-26 06:29:43 122355 1.203698(4) 20.6 0.23 ± 0.030700380501 2012-07-28 15:16:27 119140700380601 2012-08-08 23:08:08 23916

0701981201 2013-02-08 22:19:55 23918 1.20373(13) 7.0 0.45 ± 0.14

Note. — Available observations were split into 9 blocks which were analysed with theassumption that NS spin period Pspin does not change much within them. S / d.o.f. isthe statistical significance of the obtained solution as defined in the Appendix A.1. Pulsedepth (also referred as pulsed fraction) is defined as quantity A there as well. Uncertaintiesrepresent 1σ confidence interval.

(where ν indicates the pulse frequency) that yielded thehighest Z2

2 statistics.Inside the long ObsIDs (0112570101, 0690600401) the

best-solution frequency and frequency derivative clearlyfollowed a sinusoidal law with a period between 1 and 2days, as expected from orbital modulation. We fitted si-multaneously the two values of frequency and frequencyderivative in ObsID 0690600401 with the expected varia-tion due to orbital motion, and obtained a first estimateof the orbital parameters (Porb ∼ 1.2 d, a sin i/c ∼ 2.60 l-sec, Tasc ≈ MJD56104.789, where a sin i/c is the pro-jected semi-major axis and Tasc the time of passagethrough the ascending node). Starting from the firstrough estimate of these parameters and, by trial-and-error, trying to align the pulses in the phaseogram in thechunks first, then calculating TOAs with a custom imple-mentation of the fftfit method (Taylor 1992) and usingTEMPO2 to fit an orbital solution with the ELL1 model7,we reached a solution valid to ObsID 0690600401. Wethen applied the solution to the other long ObsID, re-fining the orbital parameters so that every residual or-bital modulation was eliminated by assuming a constantspin through the observation. The solution found in this

7 This model is appropriate for quasi-circular orbits, seehttp://www.atnf.csiro.au/people/pulsar/tempo/ref_man_sections/binary.txt

way is the following: Porb ≈ 1.2695 d, a sin i/c ≈ 2.886,Tasc ≈ 56104.7907. Then, we addressed the ObsIDs inbetween, using “quantized” values of the orbital periodthat conserved the ascending node passages close to thetwo long ObsIDs. For every value, we calculated thescatter that it produced on the TOAs and selected theone that produced the lowest scatter. The eccentricityfitted by ELL1 was always consistent with 0 at the ∼ 2σlevel, with an uncertainty of 0.002–0.003. We use thislast number as an upper limit on the eccentricity. Thefull solution is in Table 1.

We used this solution to look for pulsations in the re-maining ObsIDs and refine the estimate. In the nextsection we describe this procedure.

4. REFINED TIMING ANALYSIS

Except a few cases (ObsIDs: 0112570101, 0650560301,0690600401), all other individual observations yield poorphoton statistics in order to determine the pulsation pe-riod and the pulse shape with enough statistical signifi-cance for detailed interpretation.

Therefore, we attempted the search of coherent pul-sations by combining several datasets spanning 2 to 5months in different years of observations by assumingthat the period did not change among individual obser-vations within each block. First, we corrected all thephoton arrival times using the orbital elements of the bi-

A slow luminous extragalactic X-ray pulsar 5

nary system reported above. Then we used the periodgrid search using the S statistics with regularization (seeAppendix for details) leaving the orbital phase as an ad-ditional free parameter.

Then, for each year we started with the first obser-vation typically taken in late December or early Jan-uary and then started adding observations checking thatthe S statistics around the probable period, accordingto the increase in the exposure time suggesting that thepulsations are still coherent. If the S statistics (Leahyet al. 1983) did not increase or decrease when we addedthe additional observation, we concluded that the pe-riod changed significantly and started a new block. Theresults of our timing analysis are provided in Table 2and the blocks of observations used for coherent searchesare separated with horizontal lines. We estimated un-certainties of the period measurements analytically fromthe photon statistics, exposure time, and the pulse prop-erties as explained in the Appendix. In Fig. 2 we providethe 9 recovered pulse profiles between 2002 and 2012.

We note that when correcting photon arrival timesusing the orbital solution of Esposito et al. (2016),we were only able to get sufficient pulsations signifi-cance in three observations (0112570101, 0650560301,0690600401), whereas in others the signal was clearly suboptimal. At the same time the orbital solution obtainedin this study allows to significantly detect pulsations inall observations listed in Table 2. Though these two or-bital parameter estimates agree on the order of magni-tude with each other, the solution presented here is moreprecise.

This analysis of combined datasets reveals a clear spin-up trend observed in XB091D which is otherwise hardto be detected in individual observations, see Fig. 3.Despite the significant deviations and probably P signchange (note e.g. a period increase in the next to lastdataset comprising observations from Jun to Aug 2012),it is most likely that the neutron star spins up. On av-erage spin-up rate amounts to P ≈ −5.7 × 10−13 s s−1

if we consider all 9 period estimates obtained, or P ≈−8.5 × 10−13 s s−1 if we reject 2 most recent datasetswith outlying period estimate.

5. SPECTRAL ANALYSIS

We considered all imaging observations in whichXB091D was in the f.o.v. of the EPIC instruments. Datareduction was performed with the standard epproc andemproc pipelines coming with the XMM-Newton ScienceAnalysis Software (SAS v.14) and using the latest cali-bration files. From the raw event lists we selected highenergy photons (>10 keV for the pn and between 10and 12 keV for the MOS) and built light curves with50 s binning, in order to identify and filter out time in-tervals affected by high particle background. Thresholdcount rates were set at 0.4 ct s−1 and 0.35 ct s−1 for thepn and MOS, respectively. We then used the interac-tive xmmselect task from the SAS to extract source andbackground spectra from the filtered event lists, start-ing with ObsID 0690600401 which has the largest countsstatistics.

In this observation the source lies in the f.o.v. ofall three EPIC cameras. We extracted source countsfrom a circular region of 25 arcsec radius, while for

Figure 2. Recovered pulse profiles of XB091D obtained from thesearch for coherent pulsations using the regularized S statistics in9 combined datasets listed in Table 2. The shift along the Y axisis arbitrary and is made for clarity.

the background we selected events from a nearby regionfree of sources within the same CCD. We applied the(FLAG==0)&&(PATTERN==0) filtering options during se-lection of pn events, in order to have the spectrum of thehighest possible quality. For MOS counts we used thestandard filtering flags, (#XMMEA EM)&&(PATTERN<=12).The extracted spectra were then re-binned in order tohave at least 40 and 25 counts per energy bin for the pnand MOS, respectively.

Spectral fitting was performed with XSPEC v12.9 (Ar-naud 1996). The spectrum of XB091D can be de-scribed by an absorbed power law with exponential cut-off, wabs(cflux*cutoffpl) in XSPEC. The spectrum

6 Zolotukhin et al.

51000 52000 53000 54000 55000 56000 57000Epoch, MJD

0.0001

0.0002

0.0003

0.0004

0.0005

0.0006

0.0007

Spin period − 1.2033, s

2000 2002 2004 2006 2008 2010 2012 2014Year

Figure 3. Evolution of neutron star spin period. Spin period esti-mates obtained from individual observations are displayed in lightcolor. They fluctuate strongly, however an uncertainty of individ-ual measurement is large due to the poor photon statistics. Blackdots correspond to the spin periods estimated from 9 combineddatasets listed in Table 2. Their uncertainties are computed as perAppendix A.1. The dashed gray line is a linear fit to the period es-timates from combined datasets. Dashed cyan line is the same butwith outliers rejected (without 2 most recent period estimates).

is hard, with a photon index Γ = 0.20 ± 0.5, a cut-off energy Ecut = 4.6 ± 0.4 keV, and low absorption,nH = 3.79×1020 cm−2, obtained from PN spectrum onlyand kept fixed at that value because of unstable resultsin the simultaneous pn+MOS fit. The best-fit columndensity is in broad agreement with the expected valuefrom the Dickey & Lockman (1990) in the direction ofM31. The reduced chi square of the fit is 1.16 for 310 de-grees of freedom. We present a plot of the best-fit foldedspectrum and model for ObsID 0690600401 in Fig. 4.

We then analyzed all other observations, discard-ing those data sets where the source extraction re-gion overlapped with CCD gaps or columns of badpixels, and applying less stringent filtering options,(FLAG==0)&&(PATTERN<=4). In any case, given the lowercount statistics, for all but ObsID 0700380601 a simplerabsorbed power law model, wabs(cflux*powerlaw), issufficient to fit the spectra adequately. We used the mul-tiplicative component cflux, which returns the source’sflux directly as a parameter of the fit, to charachterizevariations with the epoch of the unabsorbed 0.3−10 keVband luminosity, and of the hardness ratio HR, whereHR = Flux5−10 keV/Flux0.3−5 keV, assuming a distance toM31 of 752 kpc and isotropic emission (Fig. 5). Intringu-ingly, the shape of the spectrum seems to be related withthe luminosity of the source, where the harder spectraoccur at higher luminosities, see Fig. 6.

6. DISCUSSION

We first estimate the chance association probability ofthis X-ray source with the globular cluster B091D pro-jection without being physically associated, as Espositoet al. (2016) favor an interpretation with the pulsar notassociated with the GC. However, there are 38 GCs (in-cluding candidates) between 14 and 16 arcmin projecteddistance from M31’s center in the Bologna catalog whichyields their density 6.0× 10−5 GCs per square arcsec at

Figure 4. Best-fit X-ray spectrum and folded model of XB091Dduring ObsID 0690600401. pn data are black, MOS1 are green andMOS2 are red. See main text for details.

this angular distance from galaxy center. This gives a lowprobability (1.2× 10−2) of having a GC within 1 arcsecof one of the 70 brightest X-ray sources from the 3XMM-DR5 catalog in M31. In fact, only 4 of the 70 brightestX-ray sources lie in this annulus. Therefore, their den-sity of 6.3 × 10−6 per square arcsec is translated to theprobability of only 7.5 × 10−4 of having a bright X-raysource within 1 arcsec of any of the globular clusters. Itis therefore highly likely that this X-ray source belongsto B091D globular cluster. Given the probability of co-incidence, we assume below that the source belongs tothe GC, and do not discuss alternative interpretations.

The detection of pulsations at ∼ 0.83 Hz secures theidentification of the source with a spinning neutron star.This is the first object of this class identified in the An-dromeda Galaxy, and one of the most distant pulsarsobserved to date. Also given that XB091D resides in-side the globular cluster it spins slower than any knownsources in a GC (see Introduction) and about ten timesslower than the slowest accreting pulsar in a globularcluster known previously, the ∼ 11 Hz IGR J17480-2466in Terzan 5 (Papitto et al. 2011). Its orbital period of30.5 hr is also the longest known for accreting globularcluster binaries (Heinke 2010).

Given the maximum observed X-ray luminosity of thesystem LX = 1.2 × 1038 erg s−1 (see Fig. 5) which isclose to Eddington, one can determine the correspond-ing mass accretion rate assuming that bolometric lu-minosity does not greatly exceed the X-ray luminosity:

A slow luminous extragalactic X-ray pulsar 7

Figure 5. (Top) Unabsorbed 0.3–10 keV luminosity of XB091D,in units of 1038 erg/s, versus the epoch of the observation.(Bottom) Ratio of the unabsorbed fluxes estimated in the 5 − 10keV and 0.3− 5 keV bands, respectively. Error bars correspond to1σ confidence limits.

Figure 6. Hardness ratio versus the unabsorbed 0.3–10 keV lumi-nosity of XB091D. Error bars correspond to 1σ confidence limits.

M = LXRNS

GMNS= 1.0 × 10−8 M� yr−1. We note that

the observed X-ray variability of a factor ' 3 − 4 (seeFig. 5) excludes significant contribution from other un-resolved X-ray sources residing in the same host globularcluster. We cannot completely rule out the possibilitythat we observe 2 superimposed X-ray sources residingin the same globular cluster, though the probability ofthis coincidence is extremely low, for the reasons statedabove. In case of the superposition we can only refer tothe pulsed part of the flux (typically 20 per cent, see Ta-ble 2) as originating from this X-ray source. This wouldadjust our calculation of the accretion rate and the mag-netic field below by a small factor, without affecting theimportant conclusions.

For a neutron star spinning with a 1.2 s period,the corotation radius is RC = (GMP 2/4π2)1/3 '1890M

1/31.4 km, where M1.4 is the NS mass in units of

1.4M�.We employ same approach as in Papitto et al. (2011)

to estimate the lower and upper limits for the magneticfield of the neutron star given that it persistently accretesmatter from the companion. It is reasonable to assume inthis case that the accretion disk is truncated at the radiusRin between the radius of the neutron star RNS and the

corotation radius RC: RNS < Rin ∼< RC = 1890M1/31.4 .

The inner disk radius can be approximately defined (e.g.Ghosh & Lamb 1978) from the balance between kineticenergy of the falling gas and the magnetic energy of

the NS magnetosphere: Rin ' 160M1/71.4 R

−2/76 L

−2/737 µ

4/728 ,

where R6 is the neutron star radius in units of 10 km,L37 is the accretion luminosity in units of 1037 erg s−1

and µ28 is the magnetic dipole moment of the NS inunits of 1028 G cm3. After trivial transformations weobtain: 0.008L

1/237max

< µ28 ∼< 75L1/237min

We substitute the

minimum (L37min= 3.5) and maximum (L37max

= 12.0)observed X-ray luminosity into the resulting expressionand obtain: 2.7× 108 G < B ∼< 1.4× 1012 G. This upperlimit corresponds to the case of equilibrium rotation ofthe neutron star when the accretion disk is truncated atcorotation radius and the falling matter does not trans-fer any angular momentum to the neutron star. If thesystem is indeed in equilibrium, the derivative of the NSspin period P would be fluctuating in sign with smalltypical values. This is indeed what happens if one anal-yses individual observations of XB091D (see Fig. 3, lightgray). Assuming coherent rotation between adjacent ob-servations separated by few months we were able to sig-nificantly improve the statistics and constrain period tomuch better precision which allowed us to detect con-stant average spin-up of the neutron star.

We find a spin-up rate ν ≈ 4.0–5.9×10−13 Hz s−1, onlya factor of two–three less than that of IGR J17480−2446which possesses a spin frequency derivative of ν ≈ 1.2×10−12 Hz s−1 (Cavecchi et al. 2011; Patruno et al. 2012).Such a significant spin-up rate clearly indicates thatXB091D is not in equilibrium rotation and its magneticfield estimate should be lower than the equilibrium valueof B ' 1.4 × 1012 G. IGR J17480−2446’s spin-up rateagrees well with the accretion of disk matter angular mo-mentum given the observed luminosity 2–7×1037 erg s−1

and independent estimates of the inner radius of accre-tion disk from the quasi-periodic oscillations (Papitto

8 Zolotukhin et al.

et al. 2011). From simple considerations a torque exertedon the neutron star by accreted material moving in a Ke-plerian disk is N = 2πIν = M

√GMRin, where I is the

NS moment of inertia usually assumed to be 1045 g cm2.Therefore we could roughly estimate the inner disk ra-dius to be Rin ' 30 km and corresponding magneticfield B ' 5×108 G computed for the minimum observedaccretion rate M = 2.6 × 10−9 M� yr−1. This simplecomputation yields results which are very similar to theones for IGR J17480−2446: Rin ∼< 20 km, B ' 7×108 G

(Papitto et al. 2011). The observed spin-up rate is notcompatible with this simple model for the maximum ob-served accretion rate M = 1.0 × 10−8 M� yr−1 as itgives Rin below 10 km.

It is not clear to what extent our P estimate is in-fluenced by the pulsar’s acceleration along the line ofsight al in the globular cluster gravitational field, whichadds a term al/c to P /P . However, the largest known

absolute value of P /P from pulsars in globular clus-ters (measured in B1718−19 in NGC 6342) amounts to

P /P ≈ 1×10−15 s−1 and can be used as the upper limitof al/c for pulsars in globular clusters. We therefore donot expect it to influence our spin-up rate estimate much.

XB091D exhibits rather small luminosity changes dur-ing the observed 11 years span (see Fig. 5) which in-creases to 12 years if we consider an ObsID 0112570601from the end of 2000 at very large off-axis angle whichhowever shows very similar flux in the 3XMM-DR5 cat-alog. Together with its X-ray spectrum characteristic tolow-mass X-ray binaries (see Fig. 4) it allows to classify itas a persistent LMXB. Assuming 100 per cent duty cycleand persistent accretion at the observed rate, it becomespossible to estimate time to spin up XB091D to millisec-ond period, P/P ' 50–100 × 103 yr. Extrapolating tothe past, it would take XB091D only 1 Myr or even lessto spin up from 10 s period to its observed value.

6.1. Properties of host globular cluster B091D andcomparison with Terzan 5

In order to compare B091D to Terzan 5, the host clus-ter of the slowest known accreting X-ray pulsar (Papittoet al. 2011), we used archival imaging and measurementsof stellar velocity dispersion and stellar populations ofboth clusters from literature. Knowing the cluster struc-tural parameters (e.g. the central stellar density ρ0 andthe core radius rc) and central stellar velocity dispersionsσ0 in cluster cores, we can directly compare the numbersof LMXB formation events by tidal capture using the re-lation for encounters rate from Verbunt & Hut (1987):

Γ ∝ ρ20r

3c/σ0 (1)

Terzan 5 is a massive (M = 2 · 106M�) very metalrich cluster ([Fe/H]∼0 dex) hosting a rich population ofX-ray binaries. It possesses multiple stellar populations(Ferraro et al. 2009) so it is believed to be a nucleus of adwarf galaxy heavily stripped by the Milky Way. It has asmall core radius (5.9 arcsec = 0.24 pc), a very high cen-tral stellar density ρ0 ∼ 1–4×106M� pc−3 and a tidal ra-dius of ≈ 6.7 pc (Lanzoni et al. 2010). Its central velocitydispersion is estimated to be σ0 = 12.7 km s−1 (Gnedinet al. 2002). Terzan 5 has one of the densest known stel-

lar cores among globular clusters in our Galaxy and itis likely near the point of core collapse (Djorgovski &King 1986) implying that the close binary formation bytidal capturing and three-body interactions (Cohn et al.2002).

B091D was included in the sample of M31 GCs withstellar velocity dispersions measured from high resolutionoptical spectra (Strader et al. 2011). However, their dy-namical models are based on ground based images and,therefore, might be inaccurate.

We downloaded high resolution Hubble Space Tele-scope optical images for B091D from the Barbara A.Mikulski Archive for Space Telescopes8 obtained inthe framework of the HST GO program 10273 “Accu-rately Mapping M31’s Microlensing Population” (P.I.:A. Crotts). The cluster is located close to the edge of thefield of view on two single exposures in the F555W (ex-posure time texp = 151 sec) and F814W (texp = 457 sec)filters obtained with the HST Advanced Camera for Sur-veys Wide Field Camera. Several pixels in the F814Wimage close to the cluster core are saturated. We gener-ated HST point-spread functions in the two filters usingthe tinytim software (Krist et al. 2011) and then ranthe galfit 2-dimensional image fitting code (Peng et al.2002) and fitted King (1966) profiles into B091D imagesmasking saturated pixels in the F814W image.

We obtained the following parameters for the F555Wprofile: core radius rc = 0.110 arcsec= 0.42pc, trun-cation radius rt = 7.1 arcsec=27 pc, central surfacebrightness µ0,555 = 13.61 mag arcsec−2 or I0 = 1.31 ·105L� pc−2 that corresponds to ρ0 ≈ 8 ·105M� pc−3, el-lipticity e = 0.92. Uncertainties on the structural param-eters are an order of 5–7 per cent. The parameters ob-tained from fitting the F814W image (rc = 0.15 arcsec,rt = 6.3 arcsec, µ0,814 = 13.44 mag arcsec−2, e = 0.89)are consistent within uncertainties with those obtainedfrom F555W except the core radius probably affected bysaturated pixels. We note, that our rc value is a factor of4 smaller than the half-light radius rhb = 1.5 pc reportedby Peacock et al. (2010) even though for moderately andhighly concentrated systems such as B091D they shouldbe almost identical. Likely, this illustrates the superiorquality of HST data we used for our analysis.

Following Richstone & Tremaine (1986) and convert-ing into proper units, we estimate the V band dy-namical mass-to-light ratio of B091D as: (M/L)V ≈333σ2

0/(rcI0), where σ0 is a central projected stellar ve-locity dispersion in km s−1, rc is a core radius in pc, I0is a central surface brightness in L� pc−2. The aperturecorrection for the observed value σ = 18.6± 1.0 km s−1

(Strader et al. 2011) obtained by the integration ofthe King model yields σ0 = 21.0 ± 1.3 km s−1, hence(M/L)

V,dyn = 2.6 ± 0.4(M�/L�)V or Mdyn = (9.6 ±1.5) · 105M� assuming V = 15.39 mag.

Substituting these values in Eq. 1, we estimate the ra-tio of the stellar encounter rates in Terzan 5 and B091Dto be: ΓTerzan 5/ΓB091D ∼ 0.7± 0.3 that means that theLMXB formation by capture in the core of B091D is 1.5times more likely than in Terzan 5, one of the densestglobular clusters in our Galaxy which also possesses therichest population of X-ray sources observed in a globular

8 http://mast.stsci.edi/

A slow luminous extragalactic X-ray pulsar 9

cluster (Heinke et al. 2006). As encounter number Γ wasshown to correlate with the numbers of X-ray binariesin globular clusters (Pooley et al. 2003) we can there-fore expect B091D to be quite prolific globular clusterhaving X-ray binaries population similar or larger thanthat of Terzan 5 and continuously forming new systemsat present epoch.

Caldwell et al. (2011) reported old stellar population(t ≈ 12 Gyr) and a metallicity [Fe/H]= −0.70 dex forB091D making it a representative of “red” metal-richglobular clusters. For the Kroupa (2002) stellar initialmass function, these parameters correspond to the stel-lar mass-to-light ratio (M/L)∗,V = 2.6(M�/L�). Thisremarkable agreement with the dynamical (M/L)

V,dynsuggests that B091D did not experience stellar mass lossdue to tidal stripping by M31 after dynamical relaxationand stellar mass segregation if it had been born with theKroupa IMF.

With an estimated dynamical relaxation timescale oftrelax ≈ 1 Gyr, B091D is not expected to have undergonemass segregation (when low mass stars must migrate out-wards and massive stars including neutron stars go intothe core) because it occurs at tss ≈ 15trelax (Baumgardtet al. 2002). Consequently, encounters with M31 wouldhave caused equal degree of tidal stripping for stars ofall masses. This makes it similar to substantially moremassive ultracompact dwarf galaxies (Drinkwater et al.2003; Chilingarian et al. 2011) at least some of whichwere formed via tidal stripping and still contain cen-tral black holes from their progenitors (Seth et al. 2014),and to compact elliptical galaxies formed via tidal strip-ping (Chilingarian et al. 2009; Chilingarian & Zolotukhin2015). Keeping in mind that Terzan 5 hosting anotherslow X-ray pulsar is also believed to be a stripped dwarfgalaxy nucleus, this raises a question whether some spe-cific conditions in nuclear star clusters favor the forma-tion of LMXBs. It seems natural that GCs with severalstellar populations favor wider set of evolutionary pathsfor LMXBs than single population GCs where the poolof donor stars is more limited.

We also compare another dynamical parameter forGCs, the encounter rate for a single binary, γ ∝ ρ0/σ0

(Verbunt 2003; Verbunt & Freire 2014). It is expectedthat a higher γ indicates a higher rate of exchange en-counters in a globular cluster and therefore higher ob-served frequency of exchange encounter products – suchas isolated pulsars, slow young neutron stars and otherkinds of exotic objects believed to be formed by the dis-ruption of X-ray binaries. The lifetime of a binary untilthe next encounter which increases their chances to getdisrupted or to exchange companion star is proportionalto 1/γ. In the units of the reference globular cluster M4from Verbunt & Freire (2014), γB091D = 14.6 γM4. Thisputs B091D in the top 5 GCs of our Galaxy by this pa-rameter.

6.2. System age and formation scenarios

Caldwell et al. (2011) estimated the age of B091D glob-ular cluster to be 12 Gyr. Generally, there exists few evo-lutionary sequences of binaries formed more than 10 Gyrago that start an accretion episode of required intensityat the present time and hence could explain the origin ofXB091D. For instance, in Podsiadlowski et al. (2002) bi-

naries with initial mass of a secondary star M2 between1.0 and 1.2 M�, initial orbital period P between 0.5 and100 days, exhibit accretion episodes of order ' 100 Myrin duration after ' 10–12 Gyr of evolution, reaching peakaccretion rate Mpeak ' few×10−8 M� yr−1 with average

accretion rate ˙〈M〉 ' few × 10−9 M� yr−1. This showsthat scenario of primordial origin of XB091D when thesystem formed at early epochs of its host globular clusteraround 12 Gyr ago and started accretion episode whichwe observe today very recently, is not forbidden by theevolution theories of isolated binary systems.

But this scenario, however, poses difficulty to explainhow the neutron star which formed 12 Gyr ago kept itsmagnetic field of at least 2.7× 108 G. Neutron stars areborn with magnetic field B ' 1013–1014 G (Faucher-Giguere & Kaspi 2006) which is expected to decay withthe characteristic time that spans from few × 106 (e.g.Narayan & Ostriker 1990) to few× 108 years (e.g. Bhat-tacharya et al. 1992). Even assuming the most con-servative magnetic field characteristic decay time of ∼>100 Myr the primordial neutron star origin does not lookfeasible. Moreover, as we show in the Section 6.1 due tothe dynamical properties of B091D globular cluster it ishighly unlikely that a primordial binary system does notget disrupted for such a long time in such high-densitycluster with high dynamical interactions rate.

We therefore need to evoke a formation scenario thatis capable of producing a neutron star in an old glob-ular cluster. One such possibility that looks promisingis an accretion-induced collapse (AIC) when a massive(' 1.2M) ONeMg white dwarf (WD) accretes matterfrom a companion until it reaches the Chandrasekharlimit M = 1.44M�. AIC is anticipated to be responsi-ble for the population of slow isolated pulsars with highmagnetic fields in GCs (e.g. Lyne et al. 1996; Bretonet al. 2007) and to be the origin of some slow accretingX-ray pulsars in the field such as 4U 1626−67 (e.g. Yun-gelson et al. 2002). Ivanova et al. (2008) claim that ina typical globular cluster during 9.5−12.5 Gyr produc-tion of LMXBs from AIC is two to three times moreefficient than any other dynamical formation channel,such as physical collisions, tidal captures and binary ex-changes. A cluster like Terzan 5 (and therefore very sim-ilar B091D) is thought to produce from AIC 9.35± 1.20LMXBs per Gyr at ages 11 ± 1.5 Gyr (Ivanova et al.2008) which is not negligible even considering the shortlifetime of such binaries.

Importantly, AIC produces NSs with systematicallylarger magnetic fields as it favors more magnetic pro-genitors: it is known that magnetic WDs typically havehigher masses than nonmagnetic WDs (Wickramasinghe& Ferrario 2005). Therefore there is no difficulty in ex-plaining large possible value of the magnetic field B '1.4 × 1012 G even if the NS formed a few hundred Myrago. At the same time, our rough estimate of the mag-netic field B ' 5×108 G is rather small which hints at thesignificant age of the neutron star. It is clear that the cur-rently observed intensive accretion M ' few× 10−9 M�yr−1 which is known to bury magnetic field, could notcause such significant magnetic field decay as this accre-tion epoch appears to have started very recently, mostlikely less than 1 Myr ago – unless the characteristic de-

10 Zolotukhin et al.

cay time is much shorter than it is expected or currentrapid spin-up epoch was preceded by the prolonged equi-librium accretion epoch.

On the basis of existing data it is not possible to distin-guish AIC neutron star from the one formed as the resultof coalescence of double white dwarfs (merger-inducedcollapse, MIC), another possible scenario of the NS ori-gin in XB091D. Characteristic property of a NS formedin AIC event is its low mass, MNS ' 1.26 M�. Wedoubt however that it is possible to constrain the neu-tron star mass in this system to support or discard anAIC hypothesis. No optical identification and spectro-scopic observations of companion star is possible withcurrent generation of astronomical instrumentation of asource in globular cluster at M31 distance.

MIC however produces a single neutron star whichtherefore needs to acquire a donor star later by tidal cap-ture – a process with relatively small cross-section com-pared to e.g. binary star interactions. In Ivanova et al.(2008) simulations LMXBs produced from tidal capturechannel make ≈ 10 per cent of AIC LMXBs in a glob-ular cluster like Terzan 5. Moreover tidal capture of asecondary tends to produce more compact systems withperiods inferior to 1 day (Verbunt 2003), so it seems lessplausible scenario for the origin of XB091D.

There are several reasons to suspect that we currentlyobserve accretion episode powered by a new donor afterdynamical exchange event took place with the originalbinary that hosted AIC, despite that in Ivanova et al.(2008) this channel gives only 20 per cent LMXBs com-pared to AIC. During AIC white dwarf loses roughly0.2M� in the form of binding energy and probably somemass in a supernova shell, which makes the binary or-bit wider therefore detaching binary and halting masstransfer. Time between AIC and resumption of masstransfer in the ultra-compact system with white dwarfdonor is ' 108 yr (Verbunt et al. 1990), though it obvi-ously strongly depends on donor properties – for exampleon the presence of magnetic braking that brings the sec-ondary into contact with its Roche lobe, or the rate ofthe secondary star radius increase due to its nuclear evo-lution. On the other hand, the donor loses a significantfraction of its mass to power AIC event and its nuclearevolution can be slowed down. Prolonged epochs withoutaccretion increases the chances for the binary exchangeinside a globular cluster with a high specific encounterrate γ. Finally, there is the incompatibility between thesignificant neutron star age (estimated from its magneticfield) and the very short duration of the observed accre-tion epoch indicated by the significant period derivative.It is therefore highly unlikely that the current accretionepisode caused the NS magnetic field decay to its ob-served value. This implies that either the neutron starhad accretion episodes before and we observe a secondaryrecycling, or there is prolonged period of time of order1 Gyr or more between the NS formation and the cur-rent accretion epoch so that magnetic field decayed to itspresent value.

The mass loss of the original donor to trigger theAIC event should be within 0.2–0.3 M� range. Fromour orbital solution we estimate the donor mass func-tion to be 0.0160, in agreement with Esposito et al.(2016). This is translated to the minimum donor massof M2 = 0.36M� in case of edge-on system with incli-

nation angle i = 90 deg. The lack of X-ray eclipses inthe longest observation which covers all orbital periodmeans we can constrain the donor mass a little furtheras the system’s inclination is then less than ' 70 deg:M2 > 0.38M�. For a random distribution of inclinationangles, one has the 90 per cent a priori probability ofobserving a binary system at an angle i > 26 deg. Forthe observed mass function, this inclination correspondsto M2 = 1.04M�. Therefore the 90 per cent confidenceinterval for the donor mass is 0.38 ≤ M2 ≤ 1.04M�. Infact for a 12 Gyr old globular cluster B091D the main se-quence turn-off mass is 0.8M�. All stars within B091Dwith mass ∼> 1.0 M� must have evolved to red giantsand even become white dwarfs. So the reasonable up-per limit on the donor mass is M2 ' 0.8–0.9 M� whichcorresponds to the low-mass sub-giants and stars leavingthe main sequence.

Using Eggleton (1983) formula and considering limitsfor the mass ratio (M1 = 1.25M�, M2 = 0.9M� andM1 = 2.0M�, M2 = 0.4M�) it is easy to estimate thesize of the Roche lobe for the companion star: 1.64 ≤RL2

≤ 2.24R�. Therefore, to fill its Roche lobe thedonor must be an evolved star which has recently leftthe main sequence and has its radius increased to about2R�. The current high accretion rate can be understoodas driven by the nuclear evolution of the binary withrapidly increasing radius.

It is very unlikely that a low-mass star with M2 '0.4–0.5M� (e.g. donor that powered AIC event) canreach 1.6R� radius. This could however be influencedby the X-ray irradiation of the donor surface but its ef-fect is not well understood presently. For the most likelyvalue of the donor mass M2 = 0.8M� the inclination ofthe orbital plane is ' 30 deg and the orbital separationis 6.27–6.96R� (0.029–0.032 AU). This makes XB091Dthe widest known accreting binary system in a globularcluster. It is known that tidal capture usually producessystematically more compact systems with orbital peri-ods of less than 1 day whereas exchange encounters favorwide binaries (e.g. Verbunt 2003). Therefore we have anAIC NS that exchanged its companion rather than a MICNS that tidally captured a donor star.

The pulsar recycling theory (Bhattacharya & van denHeuvel 1991) assumes that after rotation-powered phaseof classical pulsar finishing with Pspin ≈ 1–10 s neu-tron stars can be spun up to Pspin ≈ 1–10 ms by anaccreting donor in binary system. XB091D very well fitsinto this picture being observed at the earliest stages ofits accretion spin-up phase. Whereas very similar pul-sar IGR J17480−2466 from Terzan 5 represents a mildlyrecycled system, XB091D is a missing example of non-recycled neutron star which nonetheless is accreting veryintensively.

7. CONCLUSIONS

We report an independent detection of a luminous(LX = 3–12 × 1037 erg s−1) accreting X-ray pulsar inthe Andromeda galaxy in the public data of 38 ob-servations obtained by the XMM-Newton observatorybetween 2000 and 2013. In 13 observations we de-tected 15 to 30 per cent pulsed emission with periodof 1.2 s. Our analysis is fully reproducible online usingthe XMM-Newton photon database available at http://xmm-catalog.irap.omp.eu.

A slow luminous extragalactic X-ray pulsar 11

We demonstrate that this X-ray binary is associatedwith a massive core collapsed globular cluster B091Dwith the age of 12 Gyr possessing a high stellar encounterrate. Therefore this system is a very unusual example ofa non-recycled pulsar intensively accreting. At 1.2 s itsneutron star spins 10 times slower than the former slow-est known X-ray pulsar in globular clusters – a mildlyrecycled system IGR J17480−2446 in Terzan 5.

From the X-ray timing analysis we estimate the binarysystem orbital parameters, including its orbital period of30.5 h. By combining several adjacent datasets in orderto increase the photon statistics, we obtain a phase con-nected solution in 9 extended periods of time over the11 years baseline and confidently detect and measure theneutron star’s spin-up rate of P ' −8.5 × 10−13 s s−1.From this number, we estimate that the accretion on-set happened less than 1 Myr ago, because it will onlytake ≈ 105 yr for this system to become a conventionalmillisecond pulsar.

A significant spin-up rate matches that expected fromthe angular momentum accretion of Keplerian disk withthe accretion rate M ' 3 × 10−9 M� yr−1 if the in-ner boundary of the disk is at Rin ' 30 km. Fromthe fact that the neutron star in XB091D is not in theequilibrium rotation we estimate its magnetic field to beB ' 5×108 G. This value is large enough to rule out theprimordial origin of the neutron star, but nevertheless itis highly unlikely that it decayed to its present value inthe course of a very short (∼< 1 Myr) currently observedaccretion epoch.

In a 12 Gyr old globular cluster, a neutron star haslikely formed 1–2 Gyr ago as a result of an accretion-or merger-induced collapse, the first one being preferredon the basis of recent simulations (Ivanova et al. 2008).From the orbital separation and the donor Roche lobesize, also keeping in mind that the system has been per-sistently accreting over the past 12 years, we concludethat the secondary must be a slightly evolved low-massstar with the mass close to the main sequence turn-offfor a 12 Gyr old globular cluster M2 ' 0.8M�.

It is also very likely that after the AIC event the systemexperienced another exchange interaction and the youngneutron star captured another low-mass donor star whichstarted to overflow its Roche lobe very recently, less than1 Myr ago. This seemingly exotic formation scenario isin line with the measured properties of XB091D and cor-responds to the expectations that follow from the globalproperties of the host globular cluster, namely its highencounter rate for a single binary γ, a predicted indicatorof the frequency of exotic systems that form via exchangeencounters. XB091D is the first accreting non-recycledX-ray pulsar which completes the picture of pulsar recy-cling.

ACKNOWLEDGMENTS

This work is based on observations obtained withXMM-Newton, an ESA science mission with instrumentsand contributions directly funded by ESA Member Statesand the USA (NASA). This research has made useof the VizieR catalogue access tool, CDS, Strasbourg,France. Part of the plots were produced using Veusz byJeremy Sanders. Authors are grateful to citizen scien-tists M. Chernyshov, A. Sergeev, and A. Timirgazin for

their help with the development of the XMM-Newton cat-alog website http://xmm-catalog.irap.omp.eu usedthroughout this study. IZ acknowledges the support bythe Russian Scientific Foundation grant 14-50-00043 forthe data processing and grant 14-12-00146 for the timinganalysis. IC and IZ acknowledge the joint RFBR/CNRSgrant 15-52-15050 supporting the Russian–French col-laboration on the archival and Virtual Observatory re-search, the RFBR grant 15-32-21062 and the president ofthe Russian Federation grant MD-7355.2015.2 support-ing the studies of globular clusters and compact stellarsystems. The work of NS was supported by the FrenchSpace Agency CNES through the CNRS. MB was sup-ported by the Sardinian Region through a fundamentalresearch grant under Regional Law 7th. Part of the de-tection chain was adapted from the software library forX-ray timing MaLTPyNT (Bachetti 2015).

REFERENCES

Alpar, M. A., Cheng, A. F., Ruderman, M. A., & Shaham, J.1982, Nature, 300, 728

Altamirano, D., Patruno, A., Heinke, C. O., et al. 2010, ApJL,712, L58

Archibald, A. M., Stairs, I. H., Ransom, S. M., et al. 2009,Science, 324, 1411

Arnaud, K. A. 1996, in Astronomical Society of the PacificConference Series, Vol. 101, Astronomical Data AnalysisSoftware and Systems V, ed. G. H. Jacoby & J. Barnes, 17

Bachetti, M. 2015, Astrophysics Source Code Library, recordascl:1502.021

Bachetti, M., Harrison, F. A., Walton, D. J., et al. 2014, Nature,514, 202

Baumgardt, H., Hut, P., & Heggie, D. C. 2002, MNRAS, 336,1069

Bhattacharya, D., & van den Heuvel, E. P. J. 1991, Phys. Rep.,203, 1

Bhattacharya, D., Wijers, R. A. M. J., Hartman, J. W., &Verbunt, F. 1992, A&A, 254, 198

Breton, R. P., Roberts, M. S. E., Ransom, S. M., et al. 2007,ApJ, 661, 1073

Buccheri, R., Bennett, K., Bignami, G. F., et al. 1983, A&A, 128,245

Caldwell, N., Schiavon, R., Morrison, H., Rose, J. A., & Harding,P. 2011, AJ, 141, 61

Cavecchi, Y., Patruno, A., Haskell, B., et al. 2011, ApJ, 740, L8Chilingarian, I., Cayatte, V., Revaz, Y., et al. 2009, Science, 326,

1379Chilingarian, I., & Zolotukhin, I. 2015, Science, 348, 418Chilingarian, I. V., Mieske, S., Hilker, M., & Infante, L. 2011,

MNRAS, 412, 1627Cohn, H. N., Lugger, P. M., Grindlay, J. E., & Edmonds, P. D.

2002, ApJ, 571, 818Dickey, J. M., & Lockman, F. J. 1990, ARA&A, 28, 215Djorgovski, S., & King, I. R. 1986, ApJ, 305, L61Drinkwater, M. J., Gregg, M. D., Hilker, M., et al. 2003, Nature,

423, 519Eggleton, P. P. 1983, ApJ, 268, 368Esposito, P., Israel, G. L., Belfiore, A., et al. 2016, MNRAS, 457,

L5Fabian, A. C., Pringle, J. E., & Rees, M. J. 1975, MNRAS, 172,

15Faucher-Giguere, C.-A., & Kaspi, V. M. 2006, ApJ, 643, 332Ferraro, F. R., Beccari, G., Dalessandro, E., et al. 2009, Nature,

462, 1028Galleti, S., Federici, L., Bellazzini, M., Fusi Pecci, F., & Macrina,

S. 2004, A&A, 416, 917Ghosh, P., & Lamb, F. K. 1978, ApJ, 223, L83Gnedin, O. Y., Zhao, H., Pringle, J. E., et al. 2002, ApJ, 568, L23Heinke, C. O. 2010, in American Institute of Physics Conference

Series, Vol. 1314, American Institute of Physics ConferenceSeries, ed. V. Kalogera & M. van der Sluys, 135–142

12 Zolotukhin et al.

Heinke, C. O., Wijnands, R., Cohn, H. N., et al. 2006, ApJ, 651,1098

Henon, M. 1961, Annales d’Astrophysique, 24, 369Hut, P., McMillan, S., Goodman, J., et al. 1992, PASP, 104, 981Ivanova, N., Heinke, C. O., Rasio, F. A., Belczynski, K., &

Fregeau, J. M. 2008, MNRAS, 386, 553King, A., & Lasota, J.-P. 2016, arXiv, arXiv:1601.03738King, I. R. 1966, AJ, 71, 64Krist, J. E., Hook, R. N., & Stoehr, F. 2011, in Society of

Photo-Optical Instrumentation Engineers (SPIE) ConferenceSeries, Vol. 8127, Society of Photo-Optical InstrumentationEngineers (SPIE) Conference Series, 0

Kroupa, P. 2002, Science, 295, 82Lanzoni, B., Ferraro, F. R., Dalessandro, E., et al. 2010, ApJ,

717, 653Leahy, D. A. 1987, A&A, 180, 275Leahy, D. A., Darbro, W., Elsner, R. F., et al. 1983, ApJ, 266, 160Lyne, A. G., Biggs, J. D., Harrison, P. A., & Bailes, M. 1993,

Nature, 361, 47Lyne, A. G., Manchester, R. N., & D’Amico, N. 1996, ApJ, 460,

L41Manchester, R. N., Hobbs, G. B., Teoh, A., & Hobbs, M. 2005,

AJ, 129, 1993Michel, F. C. 1987, Nature, 329, 310Mushtukov, A. A., Suleimanov, V. F., Tsygankov, S. S., &

Poutanen, J. 2015, MNRAS, 447, 1847Narayan, R., & Ostriker, J. P. 1990, ApJ, 352, 222Olausen, S. A., & Kaspi, V. M. 2014, ApJS, 212, 6Papitto, A., D’Aı, A., Motta, S., et al. 2011, A&A, 526, L3Patruno, A., Alpar, M. A., van der Klis, M., & van den Heuvel,

E. P. J. 2012, ApJ, 752, 33Peacock, M. B., Maccarone, T. J., Knigge, C., et al. 2010,

MNRAS, 402, 803

Peng, C. Y., Ho, L. C., Impey, C. D., & Rix, H.-W. 2002, AJ,124, 266

Podsiadlowski, P., Rappaport, S., & Pfahl, E. D. 2002, ApJ, 565,1107

Pooley, D., Lewin, W. H. G., Anderson, S. F., et al. 2003, ApJ,591, L131

Radhakrishnan, V., & Srinivasan, G. 1982, Current Science, 51,1096

Ransom, S. M. 2001, PhD thesis, Harvard UniversityRichstone, D. O., & Tremaine, S. 1986, AJ, 92, 72Rosen, S. R., Webb, N. A., Watson, M. G., et al. 2015, ArXiv

e-prints, arXiv:1504.07051Seth, A. C., van den Bosch, R., Mieske, S., et al. 2014, Nature,

513, 398Strader, J., Caldwell, N., & Seth, A. C. 2011, AJ, 142, 8Taylor, J. H. 1992, Philosophical Transactions: Physical Sciences

and Engineering, 341, 117van der Klis, M. 1998, Advances in Space Research, 22, 925Verbunt, F. 2003, in Astronomical Society of the Pacific

Conference Series, Vol. 296, New Horizons in Globular ClusterAstronomy, ed. G. Piotto, G. Meylan, S. G. Djorgovski, &M. Riello, 245

Verbunt, F., & Freire, P. C. C. 2014, A&A, 561, A11Verbunt, F., & Hut, P. 1987, in IAU Symposium, Vol. 125, The

Origin and Evolution of Neutron Stars, ed. D. J. Helfand &J.-H. Huang, 187

Verbunt, F., Wijers, R. A. M. J., & Burm, H. M. G. 1990, A&A,234, 195

Wickramasinghe, D. T., & Ferrario, L. 2005, MNRAS, 356, 1576Wijnands, R., & van der Klis, M. 1998, Nature, 394, 344Yungelson, L. R., Nelemans, G., & van den Heuvel, E. P. J. 2002,

A&A, 388, 546

APPENDIX

IMPROVED PERIOD SEARCH AND ERROR ANALYSIS FOR THE EPOCH FOLDING TECHNIQUE

Error analysis for the epoch folding of a single observations

Here we follow Leahy et al. (1983); Leahy (1987) in order to perform the period search using the epoch folding tech-nique and make an additional step and estimate the period determination uncertainty from statistical considerations.

Let us consider a harmonic signal with a period P over some constant background so that the pulse shape is expressedas:

f(t) = a+ b sin t

a = Nγ/T (A1)

where b is the pulse amplitude and a is the background value that we estimate from the total number of photons Nγregistered during the total exposure time T . Here we assume that the exposure filling factor is 100 per cent, i.e. nogaps took place during the exposure time due to e.g. soft proton flares.

If we now perform the epoch folding with a slightly different period P + ∆P,∆P � P , it will cause the phase shiftof the last pulse:

∆φ =2πT∆P

P 2(A2)

Hereafter, we take the continuum limit and replace all sums with integrals. The discretization does not change thefinal results much because the reduction in the method sensitivity is only 3.3 per cent for n = 10 bins per phase and0.8 per cent for n = 20 used by us here (Leahy 1987). Then we can estimate the phase smeared folded pulse shape asa function of ∆φ:

f1(t,∆φ) =1

∆φ

∫ ∆φ

0

(a+ b sin(t− τ))dτ =

a+2b

∆φsin

∆φ

2sin(t− ∆φ

2) (A3)

A slow luminous extragalactic X-ray pulsar 13

Now we can compute the S statistics (Leahy et al. 1983) as

S(∆φ) =

∫ 2π

0

(f1(t,∆φ)− a)2

f1(t,∆φ)dt ≈ 1

a

∫ 2π

0

(f1(t,∆φ)− a)2dt =

2πb2

a∆φ2(1− cos ∆φ) =

4πb2

a∆φ2sin2 ∆φ

2(A4)

For the simplicity of the computation, here we assume that the pulse is shallow (e.g. b � a) and hence we take 1/aoutside the integral. More general analytical calculation is bulky and does not change the result very much because itdepends weakly on the b/a ratio as ∼

√1− (b/a)2.

Now keeping in mind that S is in fact the χ2 statistics, we can estimate the period uncertainty by solving theequation S(∆φ) = S(0)− 1. The Taylor expansion of Eq. A4 to the 4-th power on ∆φ yields:

πb2

a(1− ∆φ2

12) =

πb2

a− 1 (A5)

Solving it for ∆φ and introducing the pulse depth A = b/a (as in Leahy 1987):

∆P =

√3

π3

P 2

A√NγT

(A6)

Analytic formulation for the epoch folding of coherent pulsations in two observations

Now let us consider a harmonic signal with a period P observed in two observations with exposure times T and T/n(without the loss of generality we take a real number n ≥ 1). The second observation starts at the moment mT (m isa real number, m ≥ 1). Then, the folded phase shape resulting from the sum of the two observations becomes:

f1(t,∆φ,m, n) =1

∆φ(

∫ ∆φ

0

(a+ b sin(t− τ))dτ +

∫ (m+1/n)∆φ

m∆φ

(a+ b sin(t− τ))dτ) (A7)

Here the second integral has a multiplier (1/∆φ) rather than (n/∆φ) because it will contribute as (1/n) to the totalpulse. Omitting bulky computations and trigonometric transformations, the S statistics computed in the shallow pulseapproximation (b� a, see above) becomes:

S(∆φ,m, n) =4πb2

a∆φ2(sin2 ∆φ

2+ sin2 ∆φ

2n+ 2 sin

∆φ

2sin

∆φ

2ncos(

∆φ

2− ∆φ

2n−m∆φ)) (A8)

This expression is non-negative for any ∆φ and it has several properties of interest for our analysis. It is virtuallyidentical to the light pattern formed by the double slit diffraction of a coherent source on the slits of unequal widths.Adding the second dataset separated from the first one in time introduces the modulation of the S statistics from asingle observation (Eq. A4) with the amplitude (1 + 1/n)2 and the high frequency m/2π (see Fig. 7). Then, dependingon the statistics defined by the total number of registered photons, the 1σ confidence region for the period P mayeither shrink into a single modulated peak so that the period uncertainty ∆P will be given by the equation similar toEq. A6:

∆Ppeak =

√3

π3

P 2√

1 + 1/n

A√NγT (m+ 1/n)

(A9)

or cover several secondary peaks in which case a single value of ∆P becomes meaningless because the real P valuemay reside in one of several secondary peaks in the confidence region. The multiple solution situation occurs when thephase shift ∆φ that corresponds to the period uncertainty ∆P from Eq. A4 (with T corrected to the total exposuretime Ttot = T (1 + 1/n)) exceeds the phase distance between secondary peaks: ∆φ > 2π/m assuming m� 1. Hence,we can estimate the number of secondary peaks nsec within the upper envelope of the S distribution by equating thephase shift to 2πnsec/m:

nsec = 2b√

3

π3

m

A√ac (A10)

This quantity depends on the count rate a = Nγ/Ttot and not on the actual counts. This conclusion looks counter-intuitive at first, however, it is trivially explained by the fact that we work in phase space. Therefore, when the pulsestatistics grows because of the increased exposure time, the main peak becomes narrower in terms of ∆P and if wekeep the separation between observations constant, m will decrease. In the case of XB091D a ≈ 0.1 and A ≈ 0.3, wewill get ∼ 3m secondary peaks on either side of the primary one if we combine two observations. Another importantproperty of Eq. A8 is that one can use the measured modulation amplitude of the S statistics for the period search intwo observations in order to estimate the possible period variation. If the observed modulation stays close to the valuepredicted by Eq. A8 plus the stochastic component Snoise that a χ2

n−1 statistics with mean (n− 1) (Leahy 1987), wecan conclude that the period change between two observations is undetectable.

14 Zolotukhin et al.

Figure 7. Examples of the S statistics for the period search (without the stochastic component) for one observation (black line) witha total exposure T shown as black solid line and for two observations (red line) with exposure times T1 = T and T2 = T/4 with thesecond observations starting at the moment Ts = 12T . Shaded areas show examples of confidence regions: for the red curve it containstwo secondary peaks on either side of the primary one.

All the computations presented above can be trivially generalized to the case of Nobs > 2. However, all expressionsbecome very lengthy and difficult to understand. The envelope size for the confidence region that might still containmultiple narrow peaks (1σ) can be estimated as:

∆Pmult =

√3

π3

P 2

A√NγtotTtot

(A11)

where Ttot indicates the total good time interval (GTI) and Nγtot is the total number of detected photons.

Regularization of the pulse shape

We can improve the period search procedure by assuming the pulse shape to be smooth. This idea is similar toregularization techniques used in the image/signal reconstruction and therefore we can exploit similar mathematicalmethods. As long as the period is determined by searching the maximum of the S statistics, we introduce thepenalization factor that depends on the pulse shape p(φ) as:

Sreg = S(1 + k

∫ 2π

0

(d2p(φ)

dφ2)2dφ)1/2 (A12)

Here k is a positive coefficient that defines the degree of regularization. Numerically, the second derivative is computedfor a discrete pulse shape as 3

2 (box(p, 3)−p) where box(p, 3) denotes the boxcar smoothing of the pulse p with a windowof 3 pixels.

Applying the regularization to real data results in much smoother pulse shapes while the periods always stay withinthe uncertainties predicted from the S statistics or the analytic formulations provided above (Eq. A11). We stressthat we do not smooth the actual pulse shape but rather bias the period search statistically towards smoother shapes.In Fig. 8 (left panel) we show an example of the recovered pulse shape from the dataset obtained in January of 2002(ObsID: 0112570101) with and without regularization using k = (2/3)2. The right panel shows the S statistics without(blue) and with (green) regularization obtained from the folded epoch period search in 20 phase bins. Therefore, theS statistics includes a stochastic contribution with the variance < Srand >= 19. It is clear from the plot thatthe regularized solution lies within the 1σ confidence area, i.e. the period values are consistent within statisticaluncertainties, however, the recovered pulse shape looks substantially smoother in the regularized case.

A slow luminous extragalactic X-ray pulsar 15

Figure 8. The pulse profile recovered from the observation collected on 2002-01-06 (ObsID: 0112570101). (Left) The yellow and red linesare pulse shapes obtained with and without regularization respectively. (Right) The blue and green lines represent the S statistics versusspin period for the search with and without regularization respectively.