betaFeedback

WorksPhysics and Astronomy

APOLLO-Timing: Annealed Phase - Orbit Lattice Linking across Observations for Binary Pulsar Timing

Abstract

Sparse observations of binary pulsars can leave the orbit and the pulse count across observing gaps uncertain. We present APOLLO-Timing, a Bayesian phase-orbit lattice method that fits timing parameters together with integer pulse offsets between observing sessions. To validate candidate phase connections, we refit the timing model at fixed pulse numbers using the Python pulsar timing package PINT. We demonstrate the method on PSR J1912-1215, discovered in the CRAFTS survey. A preliminary Keplerian fit to variations in the apparent spin period and its derivative did not phase-connect 96 times of arrival (TOAs) spanning 1438.07 days in 23 sessions. Starting from its position, spin frequency F0F_0, and first time derivative F1F_1 estimated in that fit, with dispersion measure fixed, APOLLO-Timing recovers the orbit and a unique best-supported phase-connected solution without supplied orbital parameters. A repeat full-data blind search recovers the same pulse counts in 914.0 s, including the final fit. Setting the initial F1F_1 to zero in the 17-, 18-, and 19-session subsets also recovers the reference path. The solution gives Pb=17.526P_b=17.526 days, x=23.448x=23.448 light-seconds, e=2.65×10−5e=2.65\times10^{-5}, χ2=188.01χ^2=188.01, and a weighted rms residual of 5.08 μs5.08\,μ\mathrm{s}. For a pulsar mass of 1.4 M⊙1.4\,M_\odot and inclination of 60∘60^\circ, the inferred companion mass is 0.67 M⊙0.67\,M_\odot, favoring a CO/ONeMg white dwarf.

INTRODUCTION

Pulsars are rapidly rotating, highly magnetized neutron stars. Their extremely high densities, strong gravitational fields, and intense magnetic environments make them natural laboratories for studying extreme physical processes. Due to their remarkably stable rotational properties, the pulsed signals from pulsars serve as precise celestial clocks. Through long-term monitoring of pulse arrival times and the development of coherent timing models, pulsar timing enables accurate characterization of physical properties such as spin, astrometric position, proper motion, and binary orbital evolution. This high-precision measurement technique has yielded significant results across a broad range of fields. In particular, high-precision timing of binary pulsars provides a unique method for testing gravitational theories ([35, 10, 18]) and allows the study of general relativistic effects, including orbital decay, periastron advance, and geodetic precession ([8, 3, 17, 36, 5, 21, 41]). Additionally, pulsar timing facilitates precise measurements of the masses of neutron stars and their companions in binary systems, providing critical observational constraints on the equation of state of high-density nuclear matter (Özel & Freire [24]). Millisecond pulsars (MSPs), which exhibit extremely stable rotation and ultra-high timing precision, can be combined into pulsar timing arrays (PTAs). By monitoring the arrival-time residuals of a set of MSPs over extended periods, PTAs address fundamental issues such as neutron star rotational stability, interstellar medium perturbations, and solar-system ephemeris errors. Moreover, PTAs represent one of the most important observational tools for detecting nanohertz-frequency gravitational waves (Verbiest et al. [37]). In recent years, several PTA collaborations—including NANOGrav, EPTA/InPTA, PPTA, and CPTA—have successively reported evidence for a nanohertz gravitational-wave background ([1, 9, 28] (Xu et al. 2023)). All of these applications ultimately rely on a phase-connected timing solution, in which the integer number of pulsar rotations is tracked unambiguously across the entire observing span.

With the operation of new-generation, high-sensitivity radio telescopes such as FAST and MeerKAT, along with their large-scale survey projects, the sample of known pulsars has been growing rapidly. In particular, the FAST Galactic Plane Pulsar Snapshot survey (GPPS) and the Commensal Radio Astronomy FAST Survey (CRAFTS) has discovered a large number of new pulsars, significantly expanding the sample sizes of normal pulsars, millisecond pulsars, and RRATs ([13, 12, 7, 22, 41 b] (Xu et al. 2025)). Surveys such as MeerKAT/TRAPUM are also continuously discovering new pulsars in environments including extragalactic systems and globular clusters, providing richer observational samples for studying pulsar populations, formation and evolution, and extreme physics ([29, 27]).

Sparse follow-up observations often leave uncertain pulse counts across long gaps. Manual timing connects observations incrementally, refitting the spin and orbit as the counts are revised. Freire & Ridolfi [11] automated this process by testing integer-turn assignments and comparing their timing fits. The Algorithmic Pulsar Timer (APT; Phillips & Ransom [25]) automates phase connection for isolated pulsars, using F-tests and parameter covariances to select observations, add parameters, and compare models. The Algorithmic Pulsar Timer for Binaries (APTB; Taylor et al. [34]) extends this approach to binary parameters and pulse counts, using F-tests and the approximately quadratic variation of reduced χ2\chi^2 near the best phase wrap.

Here, we present APOLLO-Timing (Annealed Phase–Orbit Lattice Linking across Observations), a Bayesian method for jointly determining the orbit and rotation count of a binary pulsar. The phase variation within an observing session carries information about spin evolution and orbital Doppler modulation even when the pulse count between sessions is unknown. Requiring all sessions to follow the same spin evolution and orbital light-travel-time delay combines constraints that may be inconclusive in individual observations. We group TOAs into sessions, separate the measured phase variation from an unknown integer offset for each session, and infer the offsets together with the timing parameters. Uncertainties in the common parameters correlate the session offsets. Within each orbital-cycle cell, marginalization over the local continuous-parameter uncertainty gives a lattice distribution over candidate pulse paths. We retain several paths and follow them as the added phase noise is reduced to zero. PINT (Luo et al. [19]) provides the physical phase model throughout the search and performs the final fits with the retained pulse numbers held fixed. These fits and the associated consistency checks constitute the certification described in Section 2.

We test APOLLO-Timing on PSR J1912−1215, starting with the position, F0F_0, and F1F_1, with DM fixed and no orbital parameters supplied. We then remove sessions in reverse chronological order and compare the recovered pulse paths with the full-data solution. This comparison examines how the time span and orbital-phase coverage affect the ability to distinguish competing rotation counts. We measure the wall time from blind orbital initialization through the final timing fit, since the orbit and phase connection must be reassessed as follow-up observations accumulate. Such repeated searches are part of the timing follow-up of discoveries from large radio pulsar surveys.

The APOLLO-Timing source code is publicly available at https://github.com/flagtiya/apollo_timing.

Section 2 describes the methodological framework of APOLLO-Timing. Section 3 presents its application to PSR J1912−1215 and discusses the nature and possible evolutionary history of the system. Section 4 summarizes the main conclusions of this study.

METHODS

Pulsar timing involves comparing measured TOAs with predictions generated by a timing model that incorporates the pulsar’s rotation, astrometric and propagation effects, and, for binary pulsars, orbital motion. TOAs are estimated by matching observed, integrated pulse profiles to a template with a defined fiducial phase. A phase-connected timing solution consistently accounts for the integer number of rotations between observations by using a single timing model throughout the entire data span.

Uncertainty in the spin and orbital parameters can leave the number of rotations across an observing gap undetermined. Several combinations of pulse counts and timing parameters may fit sparse observations. APOLLO-Timing searches these combinations jointly, using phase changes within observations and offsets between them.

After accounting for Solar-System and binary delays, the timing model expands the spin phase about a reference epoch:

ϕspin(t)=ϕ0+F0(t−t0)+12F1(t−t0)2+16F2(t−t0)3+⋯(1)\phi_{\mathrm{spin}}(t)=\phi_0+F_0(t-t_0)+\frac{1}{2}F_1(t-t_0)^2+\frac{1}{6}F_2(t-t_0)^3+\cdots \tag*{(1)}

Here ϕ0\phi_0 is the phase at t0t_0 (PEPOCH), and F0F_0, F1F_1, and F2F_2 are the spin frequency and its first two time derivatives. Phase is measured in rotations and time differences in seconds. Let θ\theta denote the spin, astrometric, and orbital parameters. With the pulse-profile reference set to phase zero, a phase-connected model assigns each TOA an integer pulse number NiN_i satisfying

ϕ(ti;θ)≃Ni.(2)\phi(t_i;\theta)\simeq N_i. \tag*{(2)}

The difference between the model phase and assigned pulse number is

Δϕi=ϕ(ti;θ)−Ni,(3)\Delta\phi_i=\phi(t_i;\theta)-N_i, \tag*{(3)}

or, in time units,

Ri=ΔϕiF0.(4)R_i=\frac{\Delta\phi_i}{F_0}. \tag*{(4)}

For a given pulse-number assignment, we fit θ\theta by weighted least squares. APOLLO-Timing searches the allowed assignments together with the timing parameters. PINT evaluates the model phases and their parameter dependence throughout the search and performs the final fixed-count fits.

Figure 1 illustrates the main steps with synthetic TOAs from a circular binary. The example combines local PINT fits with a reduced phase model for the joint pulse-count calculation. We selected a noise realization with competing orbital aliases and searched periods of 6–14 days and offsets of ±1000\pm1000 turns. The plotted weights come from profile fits on this finite grid and do not include marginalization over timing-parameter uncertainty.

Figure 1. Phase connection of synthetic TOAs from an 8.500-day circular orbit. Each session contains six TOAs over two hours. The added phase uncertainty τ\tau, measured in turns, is reduced to zero during the search. (a) Phases after subtracting each session’s mean, with arrows marking observing gaps. (b) Separate period ranges near Path A (blue) and Path B (red). Dashed curves use local PINT fits, and solid curves refit continuous parameters at fixed pulse counts. A and B share a score reference within each method. (c) Session 2 weights summed over the other offsets. (d) Integer choices near A and B, relative to Session 3/4 offsets (−2-2, 1) and (266, 545). Gray outlines and black squares mark retained projections at τ=0.9\tau= 0.9 and τ=0\tau= 0 turns. Session 2 is projected out. (e) Weights during annealing, with 311 paths above the final 0.01 cutoff. (f) Earlier fitted TOAs (filled circles) and independent day-40 TOAs (open diamonds, shaded region). The spin and orbital parameters remain fixed at their earlier fits, with a free mean phase for the new session. Later error bars include measurement and conditional prediction uncertainties. Both paths fit the earlier observations. The later session remains consistent with the injected Path A and gives large residuals for Path B, distinguishing the two connections.

Phase Information within Observations

TOAs within an observation usually constrain relative pulse counts more tightly than TOAs separated by a long gap. We group TOAs by observing epoch, combining closely spaced scans into a session using a time-gap threshold. A count error across a gap shifts every pulse number in a session by the same integer, leaving their differences unchanged. We write the unknowns as

{θ,m,ν},(5)\{\theta,\mathbf{m},\boldsymbol{\nu}\}, \tag*{(5)}

where θ\theta contains the fitted timing parameters, m\mathbf{m} the within-session count corrections, and ν\boldsymbol{\nu} the whole-session integer offsets. For SS sessions,

ν=(n1,n2,…,nS−1)T.(6)\boldsymbol{\nu}=(n_1,n_2,\ldots,n_{S-1})^{\mathsf{T}}. \tag*{(6)}

Setting n0=0n_0=0 fixes the arbitrary count origin: a common integer shift leaves all pulse-number differences unchanged. We call a complete assignment a pulse-number path. For TOA ii in session ss,

Ks,i=Ks,i(0)+ms,i+ns.(7)K_{s,i}=K_{s,i}^{(0)}+m_{s,i}+n_s. \tag*{(7)}

The input model supplies the reference count Ks,i(0)K_{s,i}^{(0)}, to which we add the within-session correction ms,im_{s,i} and the whole-session offset nsn_s. Here Ks,iK_{s,i} is the session-indexed form of NiN_i. We retain alternative within-session counts when needed.

Within-session phase changes constrain the model before nsn_s is known. Let Φs(θ)\boldsymbol{\Phi}_s(\theta) contain the model phases. For a trial within-session count assignment, the residual vector is

rs=Φs(θ)−Ks(0)−ms.(8)\mathbf{r}_s=\boldsymbol{\Phi}_s(\theta)-\mathbf{K}_s^{(0)}-\mathbf{m}_s. \tag*{(8)}

For Gaussian TOA errors, the inverse phase-error covariance Ws\mathbf{W}_s gives the measurement weights. We separate the weighted mean offset asa_s from the remaining residuals qs\mathbf{q}_s:

as=1sTWsrs1sTWs1s,(9)a_s=\frac{\mathbf{1}_s^{\mathsf{T}}\mathbf{W}_s\mathbf{r}_s}{\mathbf{1}_s^{\mathsf{T}}\mathbf{W}_s\mathbf{1}_s}, \tag*{(9)}
qs=Ws1/2Psrs.\mathbf{q}_s=\mathbf{W}_s^{1/2}\mathbf{P}_s\mathbf{r}_s.

Here 1s\mathbf{1}_s and Is\mathbf{I}_s are the all-ones vector and identity matrix, with dimensions set by the number of TOAs in session ss. The projection

Ps=Is−1s(1sTWs1s)−11sTWs(10)\mathbf{P}_s=\mathbf{I}_s-\mathbf{1}_s\left(\mathbf{1}_s^{\mathsf{T}}\mathbf{W}_s\mathbf{1}_s\right)^{-1}\mathbf{1}_s^{\mathsf{T}}\mathbf{W}_s \tag*{(10)}

removes the common phase offset, as a separate JUMP does for a group of TOAs in DRACULA or APTB [11, 34]. The remaining slope and curvature constrain the timing

model. We obtain the offset from the weighted mean and scale the residuals by Ws1/2\mathbf{W}_{s}^{1/2}, so their squared sum accounts for the measurement uncertainties.

To connect sessions, we compare their mean offsets with session 0: d=(a1−a0,…,aS−1−a0)T\mathbf{d} = (a_1-a_0,\ldots,a_{S-1}-a_0)^{\mathsf{T}}. With vs=(1sTWs1s)−1v_s = (\mathbf{1}_{s}^{\mathsf{T}}\mathbf{W}_s\mathbf{1}_s)^{-1}, the covariance of these differences is

V=diag⁡(v1,…,vS−1)+v011T,(11)\mathbf{V} = \operatorname{diag}(v_1,\ldots,v_{S-1}) + v_0\mathbf{1}\mathbf{1}^{\mathsf{T}}, \tag*{(11)}

where 1\mathbf{1} has length S−1S-1. The shared uncertain reference produces the off-diagonal entries. A trial model must fit both qs\mathbf{q}_s and the relative mean offsets d\mathbf{d}, with covariance V\mathbf{V}.

Trial Orbits and the Orbital-Period Search

When no initial orbital solution is available, or only a subset of its parameters is known, we estimate the missing orbital parameters from phase changes within observing sessions. The resulting trial orbits are evaluated with the selected PINT binary model.

The orbital period PbP_b sets the rate of phase advance, and the projected semimajor axis x=apsin⁡i/cx = a_p \sin i/c sets the Roemer-delay amplitude. Here apa_p is the pulsar’s semimajor axis about the binary barycentre, ii is inclination, and ee is eccentricity. The parameter file denotes xx by A1A_1, in light-travel time. For nearly circular orbits, ELL1 uses TASCT_{\mathrm{ASC}} (TASC), ϵ1=esin⁡ω\epsilon_1 = e\sin\omega (EPS1), and ϵ2=ecos⁡ω\epsilon_2 = e\cos\omega (EPS2) to describe orbital phase and eccentricity corrections. These coordinates avoid the poorly determined periastron epoch T0T_0 and longitude ω\omega at small eccentricity. ELL1H and ELL1k use the same coordinates. BT, DD, DDS, DDGR, DDH, and DDK use T0T_0, ee, ω\omega. BTpiecewise also fits offsets between orbital segments.

For a trial orbital period, an initial parameter set for an ELL1 model may contain only

θc=(α,δ,F0,F1,tepoch,Pb).(12)\boldsymbol{\theta}_c = (\alpha,\delta,F_0,F_1,t_{\mathrm{epoch}},P_b). \tag*{(12)}

Here α\alpha, δ\delta are right ascension and declination, and tepocht_{\mathrm{epoch}} is the spin reference epoch. We calculate phases without the binary delay and select a densely sampled observing session. Its closely spaced TOAs trace the residual phase during the observation, constraining its slope and any measurable curvature. A circular orbit contributes sine and cosine terms, whose changes we fit relative to the first TOA. Let wiw_i be the phase modulo one rotation and t0t_0 the session’s first TOA. For

λi=2π(ti−t0)/Pb,\lambda_i = 2\pi(t_i-t_0)/P_b,
Di=(sin⁡λi−sin⁡λ0,cos⁡λi−cos⁡λ0).(13)\mathbf{D}_i = (\sin\lambda_i-\sin\lambda_0,\cos\lambda_i-\cos\lambda_0). \tag*{(13)}

The coefficients c=(u,v)T\mathbf{c} = (u,v)^{\mathsf{T}} give the orbital phase term. We fit them jointly with the within-session count corrections k\mathbf{k}, weighted by the phase errors σϕ,i\sigma_{\phi,i}. An upper bound xmax⁡x_{\max} restricts the amplitude to ∣F0∣xmax⁡|F_0|x_{\max}:

p(c,k∣w,Pb)∝1(∥c∥2≤∣F0∣xmax⁡)×∏iexp⁡[−(wi+Dic−ki)22σϕ,i2].(14)p(\mathbf{c},\mathbf{k}\mid\mathbf{w},P_b) \propto\mathbf{1}(\lVert\mathbf{c}\rVert_2 \le|F_0|x_{\max}) \times\prod_i \exp\left[-\frac{(w_i+\mathbf{D}_i\mathbf{c}-k_i)^2}{2\sigma_{\phi,i}^2}\right]. \tag*{(14)}

The indicator 1(⋅)\mathbf{1}(\cdot) equals one when the bound holds and zero otherwise. We increase xmax⁡x_{\max} through a specified sequence, then refine candidates with a second harmonic and a local F0F_0 correction. The coefficients are converted to A1,TASCA_1,T_{\mathrm{ASC}}, EPS1, EPS2 and checked with the PINT ELL1 delay. These trial orbits seed the phase-connection search.

If an initial solution provides only some of the orbital parameters, we estimate the missing values in the appropriate Keplerian coordinates:

qKep={(A1,TASC,ϵ1,ϵ2),ELL1 models(A1,T0,e,ω),eccentric models.(15)q_{\mathrm{Kep}} = \begin{cases} (A_1,T_{\mathrm{ASC}},\epsilon_1,\epsilon_2), & \text{ELL1 models} \\ (A_1,T_0,e,\omega), & \text{eccentric models} \end{cases}. \tag*{(15)}

For eccentric orbits, we search a finite grid in eccentricity, orbital phase, and periastron longitude. Compatible within-session counts are retained, and the specified PINT model refines the trial orbits. Multiple starts reduce dependence on a single coarse fit. Initialization supplies the basic orbit without adding F2F_2, P˙b\dot{P}_b, x˙\dot{x}, ϵ˙1\dot{\epsilon}_1, or ϵ˙2\dot{\epsilon}_2 automatically. Complete initial parameter sets retain their specified fit flags.

We scale parameter changes by their phase effects to search spin, position, and orbit together. PINT evaluates the physical timing model at each trial. Let TspanT_{\mathrm{span}} be the observing span and trt_r the search reference epoch. With the light-travel time across one astronomical unit, tAU=499.004783836 st_{\mathrm{AU}} = 499.004783836\,\mathrm{s}, the spin and position coordinates are

Fr=F0+F1(tr−PEPOCH) s,F_r = F_0 + F_1(t_r-\mathrm{PEPOCH})\,\mathrm{s},
κspin=12F1(Tspan)2,\kappa_{\mathrm{spin}} = \frac{1}{2}F_1(T_{\mathrm{span}})^2,
ξE=FrtAUcos⁡δ0 Δα,\xi_E = F_rt_{\mathrm{AU}}\cos\delta_0\,\Delta\alpha,
ξN=FrtAU Δδ.(16)\xi_N = F_rt_{\mathrm{AU}}\,\Delta\delta. \tag*{(16)}

To the order shown, FrF_r is the frequency at trt_r and κspin\kappa_{\mathrm{spin}} is the quadratic phase scale across the span. The subscript ss denotes seconds. Position offsets Δα,Δδ\Delta\alpha,\Delta\delta from (α0,δ0)(\alpha_0,\delta_0) are in radians, and ξE,ξN\xi_E,\xi_N give their phase scales through the annual Roemer delay. The time dependence also depends on source direction and Earth’s position. For ecliptic coordinates, the longitude offset carries a factor cos⁡β0\cos\beta_0, where β0\beta_0 is the input latitude.

Sparse observations may leave the number of elapsed orbits uncertain. We use the orbital cycle count CorbC_{\mathrm{orb}} across the data and the phase at trt_r:

Corb=TspanPb,ϕr=wrap⁡[2πtr−TorbPb].(17)\begin{aligned} C_{\mathrm{orb}} &= \frac{T_{\mathrm{span}}}{P_b},\\ \phi_r &= \operatorname{wrap}\left[2\pi\frac{t_r-T_{\mathrm{orb}}}{P_b}\right]. \tag*{(17)} \end{aligned}

Here TorbT_{\mathrm{orb}} is TASCT_{\mathrm{ASC}} for ELL1, ELL1H, and ELL1k and T0T_0 for eccentric models, and wrap places phase in one 2π2\pi interval. Changing CorbC_{\mathrm{orb}} by one adds or removes one orbit over TspanT_{\mathrm{span}}. Sparse observations can admit several such aliases. We search finite intervals in CorbC_{\mathrm{orb}}, termed orbital-cycle cells, with a specified maximum width.

The other orbital coordinates describe delay amplitude and shape. For ELL1, aELL1=(FrA1,FrA1EPS1,FrA1EPS2)\mathbf{a}_{\mathrm{ELL1}}=(F_r A_1,F_r A_1\mathrm{EPS1},F_r A_1\mathrm{EPS2}) gives the main term and eccentricity corrections in phase units. For eccentric models, aecc\mathbf{a}_{\mathrm{ecc}} contains arcsin⁡e\arcsin\sqrt{e} and sine–cosine coordinates for amplitude and periastron longitude at trt_r. A fitted OMDOT is scaled to the periastron angle accumulated over TspanT_{\mathrm{span}}. We collect these coordinates as

qM={(Corb,ϕr,Fr,aELL1,…),ELL1 models(Corb,ϕr,Fr,aecc,…),eccentric models(18)\mathbf{q}_M = \begin{cases} (C_{\mathrm{orb}},\phi_r,F_r,\mathbf{a}_{\mathrm{ELL1}},\ldots), & \text{ELL1 models}\\ (C_{\mathrm{orb}},\phi_r,F_r,\mathbf{a}_{\mathrm{ecc}},\ldots), & \text{eccentric models} \end{cases} \tag*{(18)}

For models with orbital segments, we append the fitted segment offsets bpw\mathbf{b}_{\mathrm{pw}} to the eccentric-model coordinates. The omitted entries are other free timing parameters. Scaling by the allowed ranges gives the dimensionless vector z\mathbf{z}, which we convert to physical parameters for each PINT evaluation. The allowed parameters and integer offsets define the search range, or support. The priors specify the relative weights within these bounds. The numbers of starts and retained paths are finite.

Joint Fits of Timing Parameters and Pulse Numbers

For each trial orbit, we seek pulse counts that connect all sessions with one timing model. Let y\mathbf{y} collect the within-session residuals qs\mathbf{q}_s. Near a trial model, the first-order response of y\mathbf{y} and the inter-session offsets to δz\delta\mathbf{z} is

y(z+δz)≃y0+Aδz,d(z+δz)≃d0+Bδz.(19)\begin{aligned} \mathbf{y}(\mathbf{z}+\delta\mathbf{z}) &\simeq\mathbf{y}_0+\mathbf{A}\delta\mathbf{z},\\ \mathbf{d}(\mathbf{z}+\delta\mathbf{z}) &\simeq\mathbf{d}_0+\mathbf{B}\delta\mathbf{z}. \tag*{(19)} \end{aligned}

Each column of A\mathbf{A} or B\mathbf{B} gives the phase response to a small parameter change, evaluated with PINT. This linear approximation allows timing uncertainty and pulse counts to be compared jointly.

A connection must fit the within-session phase changes and the differences between session means. Let DD denote the TOAs and cc the orbital interval. Combining the Gaussian likelihood and priors, with added phase uncertainty τ\tau (§ 2.4), gives

p(z,m,ν,c∣D,τ)∝p(c) p(z∣c)×p(m,ν∣c)exp⁡(−12Qτ),Qτ=∑s∥qs(z,ms)∥2+eTVτ−1e,e=d(z,m)−ν.(20)\begin{aligned} p(\mathbf{z},\mathbf{m},\boldsymbol{\nu},c\mid D,\tau) &\propto p(c)\,p(\mathbf{z}\mid c)\\ &\quad\times p(\mathbf{m},\boldsymbol{\nu}\mid c)\exp\left(-\frac{1}{2}Q_\tau\right),\\ Q_\tau&=\sum_s\left\lVert\mathbf{q}_s(\mathbf{z},\mathbf{m}_s)\right\rVert^2+\mathbf{e}^{\mathsf{T}}\mathbf{V}_\tau^{-1}\mathbf{e},\\ \mathbf{e}&=\mathbf{d}(\mathbf{z},\mathbf{m})-\boldsymbol{\nu}. \tag*{(20)} \end{aligned}

The weighted squared phase differences QτQ_\tau play the role of a timing χ2\chi^2. The two terms measure the within-session residuals and the differences between predicted and integer session offsets, including correlated errors. The priors weight the orbital interval, timing parameters, and counts. For an isolated pulsar, cc is fixed because there is no orbital-period search.

Within each orbital interval, starts cover a finite grid in cycle count and phase and include the interval centre. We also use any supplied parameter-prior component means that lie in the interval. Bounded PINT fits refine these starts. Around each resulting model, we adopt the local Gaussian prior δz∼N(0,Λ)\delta\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{\Lambda}), whose covariance Λ\mathbf{\Lambda} sets the parameter widths and correlations. The within-session data give

Σy=(Λ−1+ATA)−1,μy=−ΣyATy0.(21)\begin{aligned} \mathbf{\Sigma}_y&=\left(\mathbf{\Lambda}^{-1}+\mathbf{A}^{\mathsf{T}}\mathbf{A}\right)^{-1},\\ \boldsymbol{\mu}_y&=-\mathbf{\Sigma}_y\mathbf{A}^{\mathsf{T}}\mathbf{y}_0. \tag*{(21)} \end{aligned}

The within-session fit leaves uncertainty in the timing parameters. Marginalizing over their local distribution gives the predicted mean c\mathbf{c} and covariance Sτ\mathbf{S}_\tau of the session offsets:

c=d0+Bμy,Sτ=Vτ+BΣyBT.(22)\begin{aligned} \mathbf{c}&=\mathbf{d}_0+\mathbf{B}\boldsymbol{\mu}_y,\\ \mathbf{S}_\tau&=\mathbf{V}_\tau+\mathbf{B}\mathbf{\Sigma}_y\mathbf{B}^{\mathsf{T}}. \tag*{(22)} \end{aligned}

The contribution BΣyBT\mathbf{B}\mathbf{\Sigma}_y\mathbf{B}^{\mathsf{T}} propagates the shared timing-parameter uncertainty and correlates the offsets. Measurement errors and added search uncertainty enter through Vτ\mathbf{V}_\tau, with V0=V\mathbf{V}_0=\mathbf{V}. In Sτ\mathbf{S}_\tau, diagonal entries are variances and off-diagonal entries are covariances. The offset vector c\mathbf{c} is distinct from the initialization’s harmonic coefficients.

At fixed orbital interval and within-session assignment, equal priors on allowed integer paths give the local Gaussian distribution

p(ν∣y,d,τ)∝exp⁡[−12(ν−c)TSτ−1(ν−c)].(23)p(\boldsymbol{\nu}\mid\mathbf{y},\mathbf{d},\tau)\propto\exp\left[-\frac{1}{2}(\boldsymbol{\nu}-\mathbf{c})^{\mathsf{T}}\mathbf{S}_\tau^{-1}(\boldsymbol{\nu}-\mathbf{c})\right]. \tag*{(23)}

Each integer vector assigns a set of session pulse counts. Increasing one component by one adds a rotation to that session and moves to a neighboring lattice point. We evaluate complete vectors with the covariance of the predicted offsets, accounting for correlations from the shared timing parameters. Their uncertainty-weighted distance from the prediction is

Qk,τ=(νk−c)TSτ−1(νk−c).(24)Q_{k,\tau} = (\boldsymbol{\nu}_k-\mathbf{c})^{\mathrm{T}}\mathbf{S}_{\tau}^{-1}(\boldsymbol{\nu}_k-\mathbf{c}). \tag*{(24)}

For path kk, Qk,τQ_{k,\tau} is the squared Mahalanobis distance, including offset uncertainties and correlations. Smaller values give larger local weights. We enumerate a finite set of paths at each start in every orbital interval. To assess contributions from omitted paths near a trial model, we sum the first KK weights:

ZZ,K(τ)=1(2π)(S−1)/2∣Sτ∣1/2∑k=1Kexp⁡(−Qk,τ2).(25)Z_{\mathbb{Z},K}(\tau) = \frac{1}{(2\pi)^{(S-1)/2}\lvert\mathbf{S}_{\tau}\rvert^{1/2}} \sum_{k=1}^{K} \exp\left(-\frac{Q_{k,\tau}}{2}\right). \tag*{(25)}

This Gaussian-normalized sum is the truncated lattice mass. At fixed τ\tau, we increase KK: continued growth indicates omitted local weight, while stabilization means additional paths contribute little near this model. The check cannot exclude distant timing solutions or orbital aliases. § 3.1.1 reports the convergence diagnostics.

Refining Candidate Connections

An inaccurate trial spin or orbit can make a viable pulse assignment fit poorly. To avoid rejecting it before refitting, we initially allow extra uncertainty in session offsets and reduce it through τ0>τ1>⋯>τL=0\tau_0 > \tau_1 > \cdots> \tau_L = 0, in rotations. At each stage,

Vτ=V+τ2(IS−1+1S−11S−1T).(26)\mathbf{V}_{\tau} = \mathbf{V} + \tau^2\left(\mathbf{I}_{S-1} + \mathbf{1}_{S-1}\mathbf{1}_{S-1}^{\mathrm{T}}\right). \tag*{(26)}

Here IS−1\mathbf{I}_{S-1} is the identity matrix and 1S−1\mathbf{1}_{S-1} the all-ones vector. The added term represents independent session phase variances τ2\tau^2, whose differences from the shared reference have off-diagonal covariances. This search uncertainty leaves the within-session weights unchanged. At τ=0\tau= 0, only the measurement covariance V\mathbf{V} remains.

At each stage, bounded fits update the timing parameters and integer enumeration updates the path weights. This annealing procedure carries a finite set of the highest-weight paths forward as the added uncertainty decreases. It uses deterministic fits and enumeration without Markov chain Monte Carlo (MCMC) sampling. Paths consistent at τ=0\tau= 0 proceed to full-model fits.

Fits with the Full Timing Model

For each retained path, we reconstruct Ks,icert=Ks,i(0)+ms,i+nsK^{\mathrm{cert}}_{s,i} = K^{(0)}_{s,i} + m_{s,i} + n_s and save the pulse numbers in the timing file. Starting from the τ=0\tau= 0 parameters, PINT fits the full model by weighted least squares with pulse-number tracking enabled. All pulse numbers remain fixed.

Fit flags come from the input and initialization. At fixed pulse numbers, we compare binary models and additional timing terms using the Bayesian information criterion (BIC), their joint significance relative to the simpler model, and parameter bounds. Adding F2F_2, secular orbital derivatives, or post-Keplerian terms requires a fit improvement large enough to offset the BIC penalty.

For each path, we record parameters, residuals, χ2\chi^2, reduced χ2\chi^2 (per degree of freedom), and weighted rms. Reloading solution.par and solution.tim into PINT must reproduce the fit. Acceptance requires parameters within bounds, reduced χ2<20\chi^2 < 20 before and after reloading, and ∣Δχ2∣≤0.01\lvert\Delta\chi^2\rvert\le0.01. The separate target criterion is reduced χ2<10\chi^2 < 10. These PINT certification checks assess fit quality and file consistency. They cannot independently establish the pulse count.

Solution Selection and Search Limits

Among accepted paths, we report the one with the smallest fixed-count PINT χ2\chi^2 for the same model and TOA errors. Approximate path weights guide search retention, while the final comparison uses the full timing fits. Fit quality, accepted-path counts, and search limits are reported separately.

Coverage within each orbital interval is limited by the finite starting grid, integer enumeration, and path retention. Connections may remain undiscovered even when the local weight sum converges. We record the orbital intervals, candidate limits, phase-uncertainty schedule, and fit diagnostics. § 3.1.1 tests sensitivity to selected numerical settings and the initial spin model.

The 24 scans form 23 sessions because the first day’s two scans are grouped together. We test shorter spans by successively removing the latest session, retaining all TOAs in the first 23 down to the first 12 sessions. After each search, we compare counts with the full-data reference, allowing a common integer shift. Table 2 reports accepted paths, reference matches, and original wall times. § 3.1.1 gives the initialization and search-setting tests.

SessionsTOAsOutcomeAccepted pathsReferenceWall time (s)
2396Reference recovered11922.9
2292Reference recovered11892.4
2188Reference recovered11834.2
2084Reference recovered11787.8
1980Reference recovered11771.1
1876Reference recovered11757.9
1772Reference recovered11660.7
1668Reference plus one other path21681.2
1564Reference recovered11668.6
1460Different accepted path20650.4
1356Different accepted path10597.9
1252Different accepted path20620.8

Table 2. Blind recovery for chronological session subsets of PSR J1912−1215, using PEPOCH = MJD 59368. “Accepted paths” counts distinct assignments covering all input TOAs and passing the fit-quality and reload checks. “Reference” counts matches to the full-data assignment. A common integer shift is allowed in both counts. The second 19-session fit failed the checks and is excluded. Wall times were measured on the same single-threaded host.

RESULTS

We applied APOLLO-Timing to the newly discovered binary pulsar PSR J1912−1215, for which no phase-connected timing solution has previously been published. We describe the blind search and the dependence of phase recovery on session coverage below, followed by the implications of the timing solution for the companion and the possible evolutionary history of the system.

Figure 2. Timing residuals obtained from the preliminary timing solution for PSR J1912−1215.

Application of APOLLO-Timing to PSR J1912−1215

PSR J1912−1215 was discovered by the CRAFTS survey (?). The FAST observations used here span 1438.07 days and include 24 observing scans, which form 23 observing sessions under the APOLLO-Timing session definition. Each FAST scan lasted 5 minutes and recorded full Stokes parameters with a sampling time of 49.152 μs\mu\mathrm{s}. Additionally, to calibrate the polarization of this pulsar, a 40-second noise-diode injection was performed at the start of each observing scan. Figure 3 shows the polarization profile of PSR J1912−1215.

Initial follow-up observations showed significant epoch-to-epoch variations in the apparent spin period of PSR J1912−1215 and its derivative, consistent with radial-velocity and acceleration changes in a binary orbit. We fitted a preliminary Keplerian model with the Java-based tool newfitorbit1, estimating the orbital period PbP_b, periastron epoch T0T_0, projected semimajor axis xx, and spin-frequency derivative F˙1\dot{F}_1. The residuals from this initial timing solution (Figure 2) show that it does not phase-connect all observations.

For the blind APOLLO-Timing calculation, we retained the position, F0F_0, and preliminary F1F_1, fixed DM, and removed the binary component and orbital parameters. The search obtained its starting orbits from the within-session phase changes before connecting the observations. We repeated this complete blind calculation with PEPOCH set to MJD 60604 for the solution reported in Table 1. It recovered the same relative pulse counts as the original full-data run used for the chronological-prefix experiment.

Pulsar namePSR J1912−1215
Measured Parameters
R.A., α\alpha (hh:mm:ss)19:12:34.9733(2)
Decl., δ\delta (dd:mm:ss)−12:15:21.9180(2)
Pulse frequency, ν\nu (s−1^{-1})189.047239445402(8)
Frequency derivative, ν˙\dot{\nu} (s−2^{-2})−3.024(2) × 10−16^{-16}
Dispersion measure, DM (pc cm−3^{-3})39.495(1)
Faraday rotation measure, RM (rad m−2^{-2})−31(26)
Orbital period, PbP_b (days)17.526125680(7)
Epoch of ascending node, TASC (MJD)60606.87637480(14)
Projected semimajor axis, xx (lt-s)23.4476454(6)
EPS1−2.558(6) × 10−5^{-5}
EPS26.76(6) × 10−6^{-6}
Binary modelELL1
Fitting Parameters
First TOA (MJD)59367.8
Last TOA (MJD)60805.9
Timing epoch (MJD)60604.0
Data span (yr)3.94
Number of TOAs96
Solar system ephemeris modelDE421
PINT χ2\chi^2188.01
Reduced χ2\chi^22.19
Weighted RMS residual (μs\mu\mathrm{s})5.08
APOLLO wall time (s)914.0
Maximum RSS (MiB)255
UNITSTDB
Derived Parameters
PP (s)0.0052896831656132(2)
P˙\dot{P} (s s−1^{-1})8.462(5) × 10−21^{-21}
Galactic longitude, ll (°)24.290
Galactic latitude, bb (°)−10.189
DM distance, dd (kpc)
YMW161.27
NE20251.67
Characteristic age, τc\tau_c (Myr)9.90 × 103^3
Surface magnetic field, BsB_s (G)2.14 × 108^8
Spin-down luminosity, E˙\dot{E} (erg s−1^{-1})2.26 × 1033^{33}

Table 1. Best-fitting Timing Parameters for PSR J1912−1215.

Note: The timing solution and resource measurements are from a repeat blind run with PEPOCH set to MJD 60604. Table 2 reports the original session-prefix runs. The DM was determined by dividing the full observing band into two sub-bands of equal bandwidth but different central frequencies, and then fitting these sub-bands. The timing fit held DM fixed at 39.48 pc cm−3^{-3}. The rotation measure (RM) was measured using the rmfit tool from the PSRCHIVE software package [15].

APOLLO-Timing searched the orbital-period intervals and correlated session offsets. One candidate reached zero added uncertainty and passed the final PINT fit. The solution (Table 1) has χ2=188.01\chi^2 = 188.01, reduced χ2=2.19\chi^2 = 2.19, and a weighted rms residual of 5.08 μs\mu\mathrm{s} with the reported TOA errors. Figure 4 shows the residuals against time and orbital phase. No systematic orbital-phase trend, including a Shapiro-delay signature, is apparent. The YMW16 (Yao et al. [40]) and NE2025 (Ocker & Cordes [23]) electron-density models give DM-based distances of 1.27 and 1.67 kpc.

Post-fit residuals for PSR J1912−1215 shown against MJD and orbital phase

Figure 4. The post-fit residuals for PSR J1912−1215. The left and right panels show the residuals vs. MJD and orbital phase, respectively.

Figure 3. Integrated pulse profiles for PSR J1912−1215. The black line in the lower panel of each subplot represents the total intensity, and the red line and blue line represent the linear polarization intensity and circular polarization intensity respectively. From the plot, it can be seen that the peak intensity is normalized to 1.

Table 2 follows recovery as the latest sessions are removed. It counts accepted paths and matches to the full-data reference separately, allowing a common integer shift. Fits that fail the reduced-χ2\chi^2 or reload checks are excluded from the accepted count. The 15-, 17-, 18-, and 20- to 23-session runs retained only the reference. The 16-session run also retained one alternative. The second 19-session fit had different counts but failed the checks. At 12 to 14 sessions, accepted paths differed from the reference. Each removal changes the baseline and orbital-phase coverage, as well as the finite candidate set retained by the search.

Search Stability and Alternative Pulse Paths

To examine the extra path at 16 sessions, we retained its pulse numbers for the first 15 sessions (64 TOAs) and refitted the timing parameters. The result had χ2=153.60137\chi^2 = 153.60137, reduced χ2=2.84447\chi^2 = 2.84447, and reload difference ∣Δχ2∣=2.38×10−5\lvert\Delta\chi^2\rvert= 2.38 \times10^{-5}. All parameters stayed within the original 15-session bounds, and the path passed the same acceptance checks while remaining distinct from the reference after a common shift. Thus an acceptable alternative existed even though the original 15-session search retained only the reference. This fixed-path test leaves the blind-search counts and times in Table 2 unchanged.

Fifteen additional runs of the 17- to 19-session subsets tested a wider phase grid, larger path-retention and enumeration limits, a zero starting F1F_1, and a wider F1F_1 range. All selected the reference as their best fit. The coarse accumulated path weight did not converge in any run, leaving search completeness unverified. These tests assess the stability of the preferred path under the tested settings. Independent fixed-path posterior weights and evidence ratios were not calculated.

Starting at F1=0F_1 = 0 recovered the reference at 17, 18, and 19 sessions. With a wider F1F_1 range, the 18-session search also found an accepted alternative with χ2=498.39\chi^2 = 498.39, reduced χ2=7.55\chi^2 = 7.55, and reload Δχ2=0.0027\Delta\chi^2 = 0.0027. Its F1=+2.39×10−13 Hz s−1F_1 = +2.39 \times10^{-13}\ \mathrm{Hz}\,\mathrm{s}^{-1} and eight other fitted parameters lie within the original bounds. It is poorer than the reference by Δχ2=373.87\Delta\chi^2 = 373.87. For the same Gaussian error model, its maximum likelihood is 6.5×10−826.5 \times10^{-82} times that of the reference. This ratio compares the maximum likelihoods without integrating over timing parameters or including prior path probabilities. The acceptance thresholds admit this much poorer fit in the 18-session subset. The full 96-TOA result is unchanged.

Parameter bounds set the coordinate scales and local Gaussian prior covariance, so widening them changes both the search and the effective prior. We imposed neither F1<0F_1 < 0 nor a stellar-companion-mass restriction on PbP_b and xx. Such astrophysical priors require justification and sensitivity tests, especially over a short timing baseline. They should not be chosen merely to remove alternatives. Here, fit quality favors the reference over the alternative found with the wider F1F_1 range.

Origin of the system

Based on the timing results we have obtained, we can infer the mass range of the companion star using the mass function.

f(mp,mc)M⊙=(mcsin⁡i)3M2M⊙=4π2x3T⊙Pb2,(27)\frac{f(m_{\mathrm{p}},m_{\mathrm{c}})}{M_\odot} = \frac{(m_{\mathrm{c}}\sin i)^3}{M^2M_\odot} = \frac{4\pi^2 x^3}{T_\odot P_b^2}, \tag*{(27)}

where M=mp+mcM = m_{\mathrm{p}} + m_{\mathrm{c}} is the total mass of the binary system, with mpm_{\mathrm{p}} and mcm_{\mathrm{c}} denoting the pulsar and companion masses, respectively. The constant T⊙≡GM⊙/c3=4.925490947 μsT_\odot\equiv GM_\odot/c^3 = 4.925490947\ \mu\mathrm{s} represents the solar mass expressed in time units, where GG is the gravitational constant and cc is the speed of light. Here ff, mpm_{\mathrm{p}}, mcm_{\mathrm{c}}, and MM are dimensional masses. The quantities xx and PbP_b must be expressed in seconds when T⊙T_\odot is inserted in seconds. Assuming a canonical pulsar mass of mp=1.4 M⊙m_p = 1.4\,M_\odot and an orbital inclination angle of i=60∘i = 60^\circ, we obtain a companion mass of mc=0.67 M⊙m_c = 0.67\,M_\odot for PSR J1912−1215 (the corresponding minimum mass at i=90∘i = 90^\circ is 0.56 M⊙0.56\,M_\odot). The inferred intermediate companion mass falls within the white dwarf mass gap identified by Shamohammadi et al. (2023), specifically the transitional region between low-mass helium white dwarfs (<0.46 M⊙< 0.46\,M_\odot) and typical massive CO white dwarfs (0.70.7–0.9 M⊙0.9\,M_\odot). A precise measurement of the companion mass would make this system a crucial test case for determining whether this apparent mass gap is genuine.

Figure 5 shows orbital eccentricity (left panel) and median companion mass (right panel) as functions of orbital period for known binary systems. The left panel reveals that systems containing helium white dwarf (He WD) and carbon-oxygen/oxygen-neon-magnesium white dwarf (CO/ONeMg WD) companions exhibit broad and overlapping eccentricity distributions, with both populations showing an overall increase in orbital eccentricity as a function of orbital period in the log–log plane [26, 16]. He WD systems approximately follow the theoretical relation e∝Pbe \propto P_b, whereas CO/ONeMg WD systems display greater scatter. For systems with relatively massive CO/ONeMg WD companions, Cohen et al. [6] found an empirical relation of approximately e∝Pb3/2e \propto P_b^{3/2}. The right panel demonstrates that the He WD systems are broadly consistent with the numerical calculations of Tauris & Savonije [33]. Based on the companion mass of PSR J1912−1215 and its position in the diagram, its companion is more likely to be a CO/ONeMg WD.

Figure 5. The left and right panels display the relationships between orbital period and, respectively, eccentricity and median companion mass for known binary systems. The data are sourced from the ATNF Pulsar Catalogue (version 2.8.1, [14, 20]). The dashed curve in the right panel represents the numerical results for MSP–He WD systems from Tauris & Savonije [33].

According to the standard evolutionary model, the initially more massive star in a millisecond pulsar binary first evolves into a neutron star. It is subsequently ”recycled” to millisecond spin periods by accreting matter and angular momentum from its evolving companion, a process accompanied by X-ray emission [2, 4]. Low-mass X-ray binaries generally undergo prolonged mass transfer, producing pulsars with spin periods of a few milliseconds, whereas intermediate-mass X-ray binaries evolve more rapidly, limiting the duration of accretion and typically producing pulsars with spin periods of tens of milliseconds.

PSR J1614−2230 consists of a fully recycled millisecond pulsar with a spin period of only 3.15 ms and a massive CO WD companion. Evolutionary simulations by Tauris et al. [31] (2012) suggest that the system most likely originated from an intermediate-mass X-ray binary with an initial donor mass of approximately 4.5 M⊙4.5\,M_\odot. The relatively long-lived and stable Case A Roche-lobe overflow, especially its final phase AB lasting approximately 10 Myr, allowed the neutron star to accrete enough mass and angular momentum to be spun up to a period of 3.15 ms. Combined analyses of the binary and spin evolution rule out the Case C common-envelope channel because it cannot provide sufficient accretion to fully recycle the neutron star, whereas the Case BB channel cannot release enough orbital energy to eject the donor envelope. Therefore, Case A evolution is considered essential for the formation of this system.

PSR J1912−1215 may share the intermediate-mass X-ray binary origin proposed for PSR J1614−2230. Long-lived, stable Case A Roche-lobe overflow could have supplied enough angular momentum to produce its 5.29-ms spin. Testing this possibility requires a detailed evolutionary calculation. Longer timing may detect Shapiro delay (Shamohammadi et al. 2023), constraining companion mass and inclination and, through the measured mass function, both component masses. Other post-Keplerian effects are less accessible in this wide, nearly circular orbit (Pb≃17.5 dP_b \simeq17.5\,\mathrm{d}): periastron advance and Einstein delay are small, and gravitational-wave damping of PbP_b is negligible over 3.9 yr. Shapiro delay offers the most promising mass constraint. Version 2.8.1 of the ATNF Pulsar Catalogue lists only about 20 such systems outside globular clusters.

Discussion and Conclusions

APOLLO-Timing uses a common spin and orbital model to connect observations whose relative pulse counts are unknown. Orbital delays restrict the possible rotation counts across gaps, and those counts constrain the orbit through the accumulated phase. The phase–orbit lattice retains these coupled alternatives while the continuous timing parameters are inferred. Combining the within-session phase measurements in this way allows separated observations to constrain an orbit before a coherent ephemeris has been established. Candidates are followed to zero added phase noise and checked with the full timing model at fixed pulse numbers.

APOLLO-Timing connected 96 TOAs of PSR J1912−1215 over 3.94 yr, grouped from 24 scans into 23 sessions, without supplied orbital parameters or a binary component. The full data yielded a unique best-supported solution with χ2=188.01\chi^2 = 188.01, reduced χ2=2.19\chi^2 = 2.19, and weighted rms 5.08 μ\mus. Shorter spans gave nonmonotonic recovery: the 15-, 17-, 18-, and 20- to 23-session runs accepted only the reference, the 16-session run accepted an alternative, and the second 19-session fit failed the checks. At 12–14 sessions, accepted paths differed from the reference despite smaller residuals. Removing sessions changes both the baseline and orbital-phase coverage, so these tests define no universal minimum observation count. Small residuals alone cannot identify the pulse count.

The 5.29-ms pulsar has Pb=17.526P_b = 17.526 days, a projected semimajor axis of 23.448 lt-s, and e≃2.65×10−5e \simeq2.65 \times10^{-5}. For a pulsar mass of 1.4 M⊙1.4\,M_\odot and inclination of 60∘60^\circ, the inferred 0.67 M⊙0.67\,M_\odot companion favors a CO/ONeMg white dwarf. It lies between typical He white dwarfs (<0.46 M⊙< 0.46\,M_\odot) and massive CO white dwarfs (0.70.7–0.9 M⊙0.9\,M_\odot), making a direct mass measurement relevant to the proposed white-dwarf mass gap. Stable Case A mass transfer from an intermediate-mass X-ray binary is a possible origin. Shapiro-delay constraints on the component masses could test this interpretation.

For the solution in Table 1, the complete 23-session calculation, from blind orbital initialization through PINT certification, required 914.0 s (15.2 min) and a maximum resident memory of 255 MiB. The calculations ran on a Ryzen 9 9950X3D host, with eight CPUs visible to the execution environment and single-threaded numerical libraries. The original chronological-prefix experiment in Table 2 took 597.9–922.9 s per run. The computational cost for other systems will depend on their observing coverage and the search support.

The runtime measured for J1912−1215 makes repeated blind searches feasible as timing observations accumulate. In the follow-up of large radio pulsar surveys with single-dish telescopes and interferometric arrays, such searches can test whether new observations resolve competing orbital and rotation-count solutions. The resulting timing solutions yield mass functions and eccentricities that constrain companion properties and the mass-transfer histories of newly discovered binary systems.

Acknowledgments

We thank the researchers and students in the pulsar group at Xinjiang Astronomical Observatory, and Zhichen Pan and his students at the National Astronomical Observatories, Chinese Academy of Sciences, for testing APOLLO-Timing in preparation for its open-source release.

This work is supported by the National Natural Science Foundation of China Grants (No. 12288102, 12273100), the Specialized Research Fund for State Key Laboratory of Radio Astronomy and Technology (SDQZ-KF-A2615), the Tianshan Talents Program (No. 2023TSYCTD0013), Tianshan Talents Innovation Team, the Tianchi Talents Program (Young Doctor Program) and the CAS project (No. JZHKYJPT-2021-06). W. M. Yan is supported by the Tianshan Talents Program (No. 2024TSYCCX0073) and Technology Program of Xinjiang Uygur Autonomous Region (No. 2022A03013-4). J. T. Bai is supported by the National Natural Science Foundation of China Grants (No. 12603156), and the Startup Research Fund of Henan Academy of Science under Grants (No. 20251845001). This work is also supported by the National Natural Science Foundation of China (Nos. 11988101 and T2241020) and the Tianshan Talents Program (No. 2022TSYCLJ0059).

This work made use of the data from the Five-hundred-meter Aperture Spherical radio Telescope (FAST; https://cstr.cn/31116.02.FAST). FAST is a Chinese national megascience facility, operated by the National Astronomical Observatories, Chinese Academy of Sciences.

References

  1. [1]Agazie, G., Anumarlapudi, A., Archibald, A. M., et al. 2023, ApJL, 951, L8, doi: 10.3847/2041-8213/acdac6DOI
  2. [2]Alpar, M. A., Cheng, A. F., Ruderman, M. A., & Shaham, J. 1982, Nature, 300, 728, doi: 10.1038/300728a0DOI
  3. [3]Barker, B. M., & O’Connell, R. F. 1975, ApJL, 199, L25, doi: 10.1086/181840DOI
  4. [4]Bhattacharya, D., & van den Heuvel, E. P. J. 1991, PhR, 203, 1, doi: 10.1016/0370-1573(91)90064-SDOI
  5. [5]Cameron, A. D., Bailes, M., Champion, D. J., et al. 2023, MNRAS, 523, 5064, doi: 10.1093/mnras/stad1712DOI
  6. [6]Cohen, Y., Ginzburg, S., Levy, M., Bar Shalom, T., & Siman Tov, Y. 2024, MNRAS, 534, 455, doi: 10.1093/mnras/stae2136DOI
  7. [7]Cruces, M., Champion, D. J., Li, D., et al. 2021, MNRAS, 508, 300, doi: 10.1093/mnras/stab2540DOI
  8. [8]Damour, T., & Ruffini, R. 1974, Academie des Sciences Paris Comptes Rendus Serie Sciences Mathematiques, 279, 971
  9. [9]EPTA Collaboration, InPTA Collaboration, Antoniadis, J., et al. 2023, A&A, 678, A50, doi: 10.1051/0004-6361/202346844
  10. [10]Fonseca, E., Stairs, I. H., & Thorsett, S. E. 2014, ApJ, 787, 82, doi: 10.1088/0004-637X/787/1/82DOI
  11. [11]Freire, P. C. C., & Ridolfi, A. 2018, MNRAS, 476, 4794, doi: 10.1093/mnras/sty524DOI
  12. [12]Gao, S.-J., Shao, Y.-X., & Li, X.-D. 2026, ApJ, 997, 210, doi: 10.3847/1538-4357/ae3243DOI
  13. [13]Han, J. L., Zhou, D. J., Wang, C., et al. 2025, Research in Astronomy and Astrophysics, 25, 014001, doi: 10.1088/1674-4527/ada3b7DOI
  14. [14]Hobbs, G., Manchester, R., Teoh, A., & Hobbs, M. 2004, in IAU Symposium, Vol. 218, Young Neutron Stars and Their Environments, ed. F. Camilo & B. M. Gaensler, 139, doi: 10.48550/arXiv.astro-ph/0309219DOI
  15. [15]Hotan, A. W., van Straten, W., & Manchester, R. N. 2004, PASA, 21, 302, doi: 10.1071/AS04022DOI
  16. [16]Hui, C. Y., Wu, K., Han, Q., Kong, A. K. H., & Tam, P. H. T. 2018, ApJ, 864, 30, doi: 10.3847/1538-4357/aad5ecDOI
  17. [17]Kramer, M. 1998, ApJ, 509, 856, doi: 10.1086/306535DOI
  18. [18]Kramer, M., Stairs, I. H., Manchester, R. N., et al. 2021, Physical Review X, 11, 041050, doi: 10.1103/PhysRevX.11.041050DOI
  19. [19]Luo, J., Ransom, S., Demorest, P., et al. 2021, ApJ, 911, 45, doi: 10.3847/1538-4357/abe62fDOI
  20. [20]Manchester, R. N., Hobbs, G. B., Teoh, A., & Hobbs, M. 2005, AJ, 129, 1993, doi: 10.1086/428488DOI
  21. [21]Meng, L., Zhu, W., Kramer, M., et al. 2024, The Astrophysical Journal, 966, 46, doi: 10.3847/1538-4357/ad381cDOI
  22. [22]Miao, C. C., Zhu, W. W., Li, D., et al. 2023, MNRAS, 518, 1672, doi: 10.1093/mnras/stac1305DOI
  23. [23]Ocker, S. K., & Cordes, J. M. 2026, ApJ, 1002, 3, doi: 10.3847/1538-4357/ae5825DOI
  24. [24]Özel, F., & Freire, P. 2016, ARA&A, 54, 401, doi: 10.1146/annurev-astro-081915-023322DOI
  25. [25]Phillips, C., & Ransom, S. 2022, AJ, 163, 84, doi: 10.3847/1538-3881/ac403eDOI
  26. [26]Phinney, E. S. 1992, Philosophical Transactions of the Royal Society of London Series A, 341, 39, doi: 10.1098/rsta.1992.0084DOI
  27. [27]Prayag, V., Levin, L., Geyer, M., et al. 2026, MNRAS, 545, staf2193, doi: 10.1093/mnras/staf2193DOI
  28. [28]Reardon, D. J., Zic, A., Shannon, R. M., et al. 2023, ApJL, 951, L6, doi: 10.3847/2041-8213/acdd02DOI
  29. [29]Sengar, R., Anumarlapudi, A., Kaplan, D. L., et al. 2026, ApJ, 1001, 119, doi: 10.3847/1538-4357/ae4872DOI
  30. [31]Tauris, T. M., Langer, N., & Kramer, M. 2011, MNRAS, 416, 2130, doi: 10.1111/j.1365-2966.2011.19189.xDOI
  31. [33]Tauris, T. M., & Savonije, G. J. 1999, A&A, 350, 928, doi: 10.48550/arXiv.astro-ph/9909147DOI
  32. [34]Taylor, J., Ransom, S., & Padmanabh, P. V. 2024, ApJ, 964, 128, doi: 10.3847/1538-4357/ad1ce9DOI
  33. [35]Taylor, J. H., & Weisberg, J. M. 1989, ApJ, 345, 434, doi: 10.1086/167917DOI
  34. [36]Venkatraman Krishnan, V., Bailes, M., van Straten, W., et al. 2019, ApJL, 873, L15, doi: 10.3847/2041-8213/ab0a03
  35. [37]Verbiest, J. P. W., Vigeland, S. J., Porayko, N. K., Chen, S., & Reardon, D. J. 2024, Results in Physics, 61, 107719, doi: 10.1016/j.rinp.2024.107719DOI
  36. [40]Yao, J. M., Manchester, R. N., & Wang, N. 2017, ApJ, 835, 29, doi: 10.3847/1538-4357/835/1/29DOI
  37. [41]Zhao, D., Wang, N., Yuan, J. P. 2024a, ApJL, 964, L7, doi: 10.3847/2041-8213/ad2fb3DOI

Paper details

vlnc:W_dbazzep9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents