1: \section*{Materials and methods}
2:
3: In the following, we describe several details of the discovery and follow-up observations of \psr\ as well as statistical arguments relating to origin in a globular cluster or a triple star system.
4:
5: \subsection*{Discovery}
6:
7: The survey \cite{cfl+06} collects data during 5-min pointings for each of seven positions on the sky using the Arecibo L-band Feed Array (ALFA) receiver and digital spectrometers covering 100-MHz of bandwidth centered at 1400\,MHz. The recorded data have 256 frequency channels and a time resolution of 64\,$\mu$s. We initially detected \psr\ in September 2006 using a search pipeline based on the PRESTO suite of pulsar analysis software\cite{rem02,presto}. The basic steps of the pipeline are similar to those of the ``quicklook pipeline'' \cite{cfl+06}, except that we process the data at full resolution in both time and radio frequency, we perform extensive radio-frequency interference excision in the time and frequency domains, and we look specifically for compact binary pulsars using linear ``acceleration'' searches \cite{lk05}. To remove the dispersive effects of the interstellar medium, the data were first dedispersed for 1272 trial dispersion measures (DMs) between 0$-$1003\,pc\,cm$^{-3}$. The DM trial values are spaced such that the pulse smearing due to interstellar dispersion is $<$1\,ms for all DMs $<$600\,pc\,cm$^{-3}$, thereby giving us unprecedented sensitivity to millisecond pulsars (MSPs) in the disk of our Galaxy. We conduct standard periodicity and acceleration searches by taking the Fast Fourier Transform (FFT) of each dedispersed time series using incoherent harmonic summing of up to 16 harmonics. In addition, single-pulse searches of the data were carried out which were sensitive to transient events in the time domain with widths in the range 64\,$\mu$s to 0.1\,s
8:
9: \subsection*{Timing}
10:
11: Once a pulsar is discovered it is characterized by its position, spin parameters and any orbital parameters. Using the search data these parameters are poorly constrained, but they can be measured to much greater levels of precision by fitting a timing model of the pulsar's rotation to the data. Integer numbers of pulsar rotations are fitted between the times-of-arrival of the observed pulses; in such a ``phase connected solution'' every rotation of the pulsar is accounted for.
12:
13: Follow-up timing observations were carried out with Arecibo at $\sim$1.5\,GHz on 22 days between September 2006 and January 2007 and at $\sim$2.1\,GHz for several days between January 2008 and April 2008, with the Westerbork Synthesis Radio Telescope (WSRT) at $\sim$1.4\,GHz on 14 days between April and December 2007, and with the Green Bank Telescope (GBT) primarily at $\sim$2.1\,GHz between December 2006 and April 2008. The Arecibo observations used the seven beam ALFA receiver and a Wideband Arecibo Pulsar Processor (WAPP) usually in the same configuration as used in the survey. The GBT observations made use of the Green Bank Pulsar Spigot \cite{kel+05}, a digital correlator which synthesized either 768 or 1536 frequency channels covering 600\,MHz of bandwidth and sampled every 81.92\,$\mu$s. Individual GBT observations were made at 5 and 9\,GHz with the Spigot in a mode with 1024 frequency channels recorded over 800\,MHz of bandwidth and sampled every 81.92\,$\mu$s. WSRT observations used the PUMa II data acquisition system \cite{kss08} to record eight 20\,MHz wide dual polarization bands. Each band was subsequently analyzed offline by forming a 64-channel coherent filterbank leading to a final time resolution of 16.8\,$\mu$s.
14:
15: The timing model parameters shown in Table 1 of the main article were derived from the best-fit model to the pulse arrival times we obtained with the above telescopes and instrumentation. A timing model using a simple Keplerian orbit provides a very poor fit to the data (reduced-$\chi^2$ of 29.5 for 327 degrees of freedom [DOF]). The addition of an orbital periastron advance ($\dot\omega$) dramatically improves the fit (reduced-$\chi^2$ of 2.12 for 326 DOF), giving $\dot\omega = 2.46(2)$$\times$10$^{-4}$\,deg\,yr$^{-1}$, but systematic trends remain in the timing residuals. Given the orientation of the orbit, these remaining trends are most likely explained by the Shapiro delay: the delay should have an amplitude of tens of $\mu$s, an order of magnitude larger than our arrival time measurement precision, even though most of that signal will be covariant with the projected semi-major axis, $a\sin i$. Using the ``DDGR'' model \cite{tw89,dd86}, which assumes that general relativity correctly describes the dynamics of the system with Keplerian orbit parameters plus the total system mass M$_{\rm tot}$ and the mass of the companion star M$_2$, the fit is further improved (reduced-$\chi^2$ = 1.13 for 325 DOF) and results in no visible systematics in the residuals (Fig.~1). In Fig.~S1, we show a $\chi^2$ contour map of the mass constraints from the DDGR model.
16:
17: In order to confirm that this timing is consistent with general relativity, the ``DD'' model \cite{dd85,dd86}, a theory independent relativistic description of the orbital parameters, was used separately fitting for the orbital periastron advance, the companion mass and sine of the inclination angle ($\sin$i). In this fit (reduced-$\chi^2$ = 1.12 for 324 DOF) all of the post-Keplerian parameters which are not specific to general relativity are consistent with the parameters from the general relativity specific DDGR model at the 1-2\,$\sigma$ level. The timing model parameters for the DD model are show in Table~S1.
18:
19: \subsection*{Search for a pulsar companion}
20:
21: Given the possibility that this is a double neutron star system, and that the companion is potentially a pulsar, we searched for pulsations from the companion using several of the early observations of the system taken at Arecibo. As well as searching each observation independently, we summed power spectra from the individual observations to increase sensitivity. We dedispersed each observation at the known DM of \psr. We then removed the deleterious orbital effects of the system prior to incoherently summing the power spectra from FFTs. The orbital accelerations were calculated and removed by noting that the orbital period and eccentricity of the companion's orbit are the same as for \psr, but the longitude of periastron is different by 180\degrees. The only unknown parameter in this calculation is the ratio of the semi-major axes of the pulsar and companion which is the inverse of the ratio of their masses. A search over a series of trial projected semi-major axes for the companion was therefore used, ranging from 60 to 185 light-seconds. This encompasses the range of well-measured neutron star masses \cite{lp07}. Finally the summed FFT was searched for candidates using the techniques described above. The search yielded no significant candidates.
22:
23: We set an upper limit on pulsed emission at 1.4\,GHz of 20\,$\mu$Jy for a period of 2\,ms and 9\,$\mu$Jy for a period of 200\,ms, both at a DM of 297\,pc\,cm$^{-3}$. These upper limits correspond to 1.4\,GHz radio luminosities of 0.8 and 0.4~mJy~kpc$^2$ respectively. Only two percent of all radio pulsars currently known have luminosities below the latter value. Our non-detection therefore excludes most of the observable radio pulsar population as a companion, but does not of course rule out a neutron star that is unfavorably beamed away from our line of sight or rotating too slowly to be observed as a pulsar.
24:
25: \subsection*{Optical and infra-red analysis}
26:
27: The field of PSR~J1903+0327 was observed with the Near InfraRed Imager (NIRI) on the Gemini North telescope on July 24, 2007. Dithered series of 6\,second exposures were obtained in the $J$, $H$ and $K_\mathrm{S}$ filters, amounting to a total exposure time of 10\,minutes in each filter. All images were corrected for dark current and flatfielded using skyflats. The non-uniform sky distribution was removed by substracting a sky frame constructed of the unregistered science images to remove the contributions of stars. After these corrections, the images taken with the same filter were registered and averaged.
28:
29: Astrometry was done relative to the 2MASS catalogue \cite{scs+06}, as the small field of view of NIRI ($2\amin\times2\amin$) contains no astrometric standards from the UCAC2 catalogue \cite{zuz+04}. A total of 71 2MASS stars coincided with the registered and average $K_\mathrm{S}$-band image and 51 of these were not saturated and appeared steller and unblended. After iteratively removing 9 outliers, the final astrometric calibration has rms residuals of $0\farcs086$ in right ascension and $0\farcs094$ in declination. A comparison of the positions of 281 UCAC2 standards coinciding with 2MASS stars within $10\amin$ from the position of PSR~J1903+0327 shows no significant shift between the positions in both catalogues.
30:
31: Taking the rms uncertainty on both coordinates as a measure for the 1-$\sigma$ uncertainty in the astrometric calibration, we find a single star inside the $0\farcs13$ error circle on the pulsar position at $\alpha_{\rm{J2000}}=19^{\rm{h}}03^{\rm{m}}05\fs796(6)$ and $\delta_{\rm{J2000}}= 03\degrees27\amin19\farcs19(9)$, where the uncertainty on the position is the quadratic sum of the uncertainty in the astrometry and the intrinsic positional uncertainty of the star in the image ($0\farcs02$ in each coordinate). This position if offset from the pulsar position by $0\farcs034(88)$ in right ascension and $0\farcs032(96)$ in declination.
32:
33: Photometry was also done relative to the 2MASS catalogue. Magnitude offsets were determined and outliers were iteratively removed, leaving 30 to 40 stars for the photometric calibration. The calibrations showed no significant dependence with star color. We find that the object in the error circle has $J=19.22(9)$, $H=18.41(10)$ and $K_{\rm{S}}=18.03(9)$.
34:
35: To estimate the reddening towards PSR~J1903+0327, we used red clump stars to trace the reddening as a function of distance \cite{lcgh02}. Again using the 2MASS catalogue, we selected 1400 stars within $5\amin$ from PSR~J1903+0327 and determined the $J-K_\mathrm{S}$ colour of the red clump stars at different $K_\mathrm{S}$ magnitudes using the formalism of \cite{dv06}. At the DM-distance of PSR~J1903+0327, $d=6.4(16)$\,kpc (assuming a 25\% uncertainty), we constrain the reddening at $A_V=4.9(11)$. For this distance and reddening, the intrinsic magnitudes of the star inside the error circle are $M_J=3.8(11)$, $M_H=3.5(10)$ and $M_{K_{\mathrm{S}}}=3.5(10)$.
36:
37: Though the metallicity and age of the star are unknown, stars of solar metallicity, ages of 1\,Gyr or less and masses between approximately 0.8 and 1.3\,M$_\odot$ will have similar absolute magnitudes \cite{gbbg00}. Stars older than 1\,Gyr will have evolved and will match the absolute magnitudes for lower masses. At 10\,Gyr and solar metallicity the mass range decreases to approximately 0.75 to 1.1\,M$_\odot$. For a mass of 1.05\,M$_\odot$ as determined from the radio timing, a solar metallicity star will match the observed absolute magnitudes if the age is less than 6\,Gyr.
38:
39: At the apparent magnitudes, the star is bright enough for optical or infrared spectroscopy to determine radial velocity variations which would unambiguously confirm the star as the binary companion to PSR~J1903+0327.
40:
41: \subsection*{Globular cluster calculations}
42:
43: While \psr\ is unprecedented in the Galactic plane several MSPs in eccentric binaries have been seen in globular clusters (GCs). GCs are known to be excellent breeding grounds for MSPs, with formation rates per unit mass $\sim$100 times that in the disk \cite{fpr75,Ver87}. \psr\ may have formed in a GC which was later disrupted, or escaped from an existing GC $\sim 10^9$~yr ago such that they are no longer in the same part of the sky. Many GCs may have been completely disrupted via gravitational interactions with the disk and bulge over the course of many orbits in the Galactic potential. As a result, their stars become part of the Galactic spheroid population (an approximately spherical distribution of older stars distributed in an extension of the central Galactic bulge, with a diameter of $\sim$10\,kpc). Estimates suggest that over half the spheroid mass could have come from such disrupted GCs \cite{go97}. Such disruption should require several Galactic orbits of the GC, the periods of which are $\sim$0.5$-$1.0 Gyr. However, in the particular case of \psr, we are constrained by the relatively small characteristic age of the system. Since a characteristic age is usually an upper limit on the true age of a radio pulsar \cite{lk05}, any model involving GC disruption requires a recycled pulsar to exchange its companion and acquire an eccentric orbit and its parent GC to then be totally disrupted in less than $\sim$2\,Gyr. This is challenging given the orbital periods of GCs.
44:
45: By comparing the densities of the spheroid and disk we can estimate the probability of \psr\ being found in the disk but actually being associated with the spheroid population. Using estimates of the stellar densities at low Galactic latitudes of both the spheroid \cite{rrc00} and disk \cite{Chab01} populations, we find $\rho_{\rm spheroid}/\rho_{\rm disk}=4.7\times10^{-5}{\rm\ M_\odot\ pc^{-3}}/4.3\times10^{-2}{\rm\ M_\odot\ pc^{-3}} \sim 1\times10^{-3}$. Assuming that half of the spheroid mass originated in disrupted GCs, the $\la$2\,Gyr age of \psr\ implies that $\la$20\% of that mass could have come from GCs disrupted since the recycling of the pulsar. Furthermore, if we assume that those GCs produced highly eccentric binary MSPs $\sim$10$^3$ times more efficiently per unit mass than the disk, we can crudely estimate that the probability that \psr\ came from a disrupted GC is $\la$10\%.
46:
47: We can make a similar estimate of the probability that \psr\ was ejected from a GC. The number of GC-ejected \psr-like systems should be roughly the ratio of the GC system mass (10$^7$\,M$_{\odot}$) to the disk mass (10$^{11}$\,M$_{\odot}$) times the 10$^3$ formation efficiency factor for highly eccentric binary MSPs times the $\sim$50\% fraction of pulsars which are eventually ejected from GCs \cite{ihr+07}. The resulting $\sim$5\% probability again indicates that some GC formation mechanism for \psr\ is plausible.
48:
49: \subsection*{Hierarchical triple system calculations}
50:
51: Hierarchical stellar triple systems are common, and such a system may explain the eccentricity we observe in \psr. To date no pulsars have been found in other stellar triples. Hierarchical triples contain an inner binary (in this case the MSP and the $\sim$1.05\,M$_{\odot}$ companion seen in timing at the 95-day orbit) with a third object orbiting at greater distance (in this case, the main sequence star observed in the infra-red). Here we assess the plausibility of forming such an end state.
52:
53: Constraints on the formation of the inner binary are: (i) to form a 2.15-ms MSP there must have been a long period of stable mass transfer to spin up the neutron star; (ii) to produce the wide 95-day orbit the white dwarf progenitor cannot be too massive, to avoid common-envelope (CE) evolution and spiral-in; (iii) the star must have been massive enough to leave a $\sim$1.05\,M$_{\odot}$ white dwarf after final mass transfer \cite{tvs00}. We therefore consider an initial system containing a central binary formed by a 8\,M$_{\odot}$ primary; a secondary that is just below CE mass, we shall here assume 3\,M$_{\odot}$; and a 0.9\,M$_{\odot}$ third star.
54:
55: A survey of the multiplicity of massive stars in the Orion Nebula \cite{pbh+99} suggests that 80\% of spectroscopic binaries with OB-star primaries have a third visual companion. These systems must survive the supernova that creates the pulsar: the inner binary will remain bound given suitable mass-loss and kick parameters \cite{Hill83}. The outer companion can remain bound \cite{py06} if its orbital velocity is comparable to the new system velocity of the inner binary (a few tens of km\,s$^{-1}$, \cite{mhth05}). The multiplicity survey \cite{pbh+99} finds $\sim$5-10\% of third companions are close-in enough to remain bound in the supernova. The semi-major axis of this outer orbit may then increase as mass is lost from the system during the pulsar spin-up phase \cite{spv97}. In this hierarchical triple the outer companion cannot unbind the inner binary, but its orbit evolves towards a Kozai-cycle resonance with the inner-binary relativistic periastron advance period\cite{fkr00}. Our above rough estimates of the various fractions combine to a roughly 4\% chance that a given evolved NS-WD binary is part of a hierarchical triple. As there are currently $\sim$50 NS-WD systems known in Galactic-disk binaries, this simple estimate is consistent with the idea that \psr\ is the first in such a triple. We do note that this formation scenario still faces the difficulty that thus far no other such systems with both a highly recycled pulsar and a wide-orbit very massive WD companion have been observed.
56:
57: Two other probability arguments for the detection of this system are not directly related to the formation plausibility. First, for a system with masses 1.74, 1.05 and 0.9\,M$_{\odot}$, a relatively high inclination of the third star's orbit $i > 45$\degrees and some eccentricity in the outer orbit, the inner-binary eccentricity is higher than $e=0.44$ for about 20\% of the oscillation time \cite{fkr00}. Second, for the line-of-sight acceleration from the third star to be less than our measured $\dot{P}$ this third star must currently be close to plane of the sky, the a priori probability of which is $\sim$5\%.
58:
59: \bibliography{1903}
60: \bibliographystyle{Science}
61:
62: \clearpage
63:
64: \begin{table*}
65: \begin{center}
66: \begin{tabular}{l l}
67: \hline
68: \multicolumn{2}{c}{Timing Parameters Using the DD Model} \\
69: \hline
70: Right Ascension (J2000) \dotfill & 19$^{\rm h}\;$03$^{\rm m}\;$05$\fs$79368(3) \\
71: Declination (J2000) \dotfill & 03$\degrees\;$27$\amin\;$19$\farcs$2220(11) \\
72: Spin Period (ms) \dotfill & 2.1499123298435(3) \\
73: Period Derivative (s/s) \dotfill & 1.879(2)$\times$10$^{-20}$ \\
74: Dispersion Measure (pc cm$^{-3}$) \dotfill & 297.537(7) \\
75: Epoch of Period (MJD) \dotfill & 54280.0 \\
76: Orbital Period (days) \dotfill & 95.1741176(2) \\
77: Projected Semi-Major Axis (lt-s) \dotfill & 105.593459(8) \\
78: Eccentricity \dotfill & 0.43667838(5) \\
79: Longitude of Periastron \dotfill & 141.652477(3)$\degrees$ \\
80: Epoch of Periastron (MJD) \dotfill & 54063.8402310(6) \\
81: Advance of Periastron (deg yr$^{-1}$) \dotfill & 2.46(3)$\times$10$^{-4}$\\
82: Sine of Orbital Inclination \dotfill & 0.966(10)\\
83: Companion Mass, M$_2$ (M$_\odot$) \dotfill & 1.3(2)\\
84: \hline
85: \end{tabular}
86:
87: \caption{Measured and derived parameters for \psr. The timing parameters were measured using the DE405 Solar System ephemerides \cite{DE405} and the ``DD'' timing model which uses a theory independent relativisitc model to describe the parameters of the binary system \cite{dd85,dd86}. A total of 342 pulse arrival times measured between MJDs 53990 and 54568 were fit. The numbers in parentheses are twice the TEMPO-reported 1-$\sigma$ uncertainties in the least significant digit or digits quoted as determined by a bootstrap error analysis.}
88:
89: \end{center}
90: \end{table*}
91:
92: \begin{figure}[hb]
93: \begin{center}
94: \includegraphics[height=5in,angle=270]{sfig1.ps}
95: \end{center}
96: \caption{A $\chi^2$-map showing the 1 and 2-$\sigma$ confidence regions for the total system mass (M$_{\rm tot}$) and companion mass (M$_2$) based on the timing solution presented in Table~1 after doubling the nominal TEMPO errors. Those errors were within 20\% of the bootstrap error estimates. The grey region in the lower right portion of the figure is excluded by the mass function. The grey lines show inclinations $i$ of 70, 75, 80, and 85$\degrees$. For \psr, M$_{\rm tot}$ is best constrained from the measurement of the relativistic advance of periastron $\dot\omega$ while M$_2$ is constrained by a measurement of $\sin i$ via the Shapiro delay (see Fig.~1).}
97: \end{figure}
98: