betaFeedback

WorksPhysics and Astronomy

Gamma-Ray Pulsar Discovery Using a Lattice-Based Method

Abstract

Pulsars are difficult to discover via γγ-ray pulsations because typically only one photon is observed per ∼108{\sim}10^8 rotations. We report the first γγ-ray pulsar discovered with an extension of a recently proposed algorithm, which uses efficient tools for identifying short vectors in high-dimensional lattices instead of exhaustively searching for periodic signals. PSR J1622−-7204, an isolated pulsar spinning at 261 Hz, was found in 15 years of Fermi-LAT data using about ten CPU hours. In contrast, the previous state-of-the-art (a semicoherent search) failed after ∼600{\sim}600 CPU years. This new approach may enable the discovery of many new γγ-ray pulsars, including in binary systems, a previously intractable problem without detailed prior knowledge of the orbital parameters.

Introduction

The Fermi Large Area Telescope (LAT) [1] records γ\gamma-ray photon arrival times with microsecond precision. Pulsations have been detected from more than 300 pulsars—rapidly rotating, highly magnetized neutron stars [2]—including the fastest Galactic pulsar [3], the most compact Galactic pulsar binary [4], and the heaviest known neutron stars [5], which constrain the nuclear equation of state [6]. Millisecond pulsars can also form a γ\gamma-ray pulsar timing array that probes the gravitational-wave background without radio-specific systematic uncertainties [7].

Yet, among the thousands of sources in the fourth Fermi-LAT source catalog (4FGL) [8], hundreds with pulsar-like spectra and properties remain unidentified after both radio and γ\gamma-ray pulsation searches. Searching these for γ\gamma-ray pulsations is computationally challenging. With an effective collecting area of ∼1 m2\sim1\ \mathrm{m}^{2}, the LAT detects fewer than ∼103\sim10^{3} photons from a faint pulsar over its 18-year mission, during which the pulsar rotates 10910^{9} to 101110^{11} times. If the detector had 10810^{8} times the collecting area, or the source were 10410^{4} times closer to us, then most rotations would yield at least one photon, making the rotation period easy to measure. Instead, we observe only a tiny, random subset of photons, contaminated by background. This makes the problem very difficult.

There are additional complications. As the pulsar radiates energy and angular momentum, its rotation slows. Furthermore, the pulse arrival times are modified by the LAT’s motion around the Solar System. Accounting for these effects in a brute-force search requires an intractable number (∼1022\sim10^{22}) of trials [9].

Existing methods [9, 10] trade sensitivity for a ∼5\sim5-order-of-magnitude reduction in search points, yet still require substantial computing resources, such as those of the Einstein@Home volunteer distributed computing project. Nevertheless, this approach has enabled the discovery of more than eighty pulsars [2, 11, 12, 13, 14, 15, 16].

Recently, Gazith, Pearlman, and Zackay proposed a new method [17], which in practice reduces the computational cost of a search by a further 5 to 6 orders of magnitude. This is achieved by transforming it into a classic geometry problem: finding the shortest nonzero vector in a lattice. A lattice is a regularly spaced grid of points: all integer linear combinations of a set of linearly independent vectors.

Here, we report an extension of this algorithm and its first success. We hunted for pulsations from the unassociated source 4FGL J1622.2−7202 [8], whose spectral and other properties have long suggested it may host a pulsar. We found a new isolated millisecond pulsar J1622−7204 in ≈1%\approx1\% of our trial runs, each taking under 5 min on one CPU core. This is by far the least computationally expensive successful blind γ\gamma-ray pulsar search to date. Previous Einstein@Home searches (using the setup of [16]) and TRAPUM radio searches [18, 19] both failed to detect pulsations from PSR J1622−7204.

**Method—**We use an improved version of the lattice-based pulsar search presented in [17]. First, we describe

Published by the American Physical Society under the terms of the Creative Commons Attribution 4.0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI. Open access publication funded by the Max Planck Society. how the pulsar search is turned into a lattice problem and discuss our modifications.

*Input data—*The inputs for the lattice algorithm are the photon arrival times (PATs) at the satellite of nn photons detected by the Fermi-LAT (with n=80n = 80 in our search) from a small sky region around a target source, in this case 4FGL J1622.2−7202. The ordered PATs are denoted by t1<t2<⋯<tnt_1 < t_2 < \cdots< t_n. We also need the photon weights wi∈[0,1]w_i \in[0,1] for i=1,…,ni = 1,\ldots,n, which are the (prior) probability that the iith photon was emitted by the source, rather than the background [20]. The wiw_i are derived from the photon energies and directions, using a spectral-spatial model of the time-averaged γ\gamma-ray sky.

The PATs are affected by factors both related and unrelated to the pulsar’s motion. The largest arise from the LAT’s motion around Earth, and the Earth’s motion around the Sun. Over a year, these perturb the satellite-to-pulsar distance, creating annual PAT variations of up to ≈500\approx500 s (Sun-to-Earth light travel time). To account for these, we compute the times τ\tau at which photons would arrive at the Solar System barycenter (SSB) if emitted from the pulsar simultaneously with the observed ones. [Barycentric dynamical time (TDB) τ\tau is defined in the International Celestial Reference System (ICRS) frame, a standard inertial coordinate system with origin at the SSB, and runs at the same average rate as Global Positioning System time at the Earth geoid].

*Pulsar rotation phase model—*The relationship between the PATs tt and τ\tau depends upon the (unknown) direction to the pulsar, given by its right ascension α\alpha and declination δ\delta in the ICRS frame:

α=α0+Δα,δ=δ0+Δδ.(1)\alpha= \alpha_0 + \Delta\alpha,\qquad\delta= \delta_0 + \Delta\delta. \tag*{(1)}

The nominal direction α0\alpha_0, δ0\delta_0 in Table I is a source localization estimate using 14 years of 4FGL data [21]. The arrival times at the SSB are

τ(t,α,δ)=τ(t,α0,δ0)+Δα∂τ∂α(t,α0,δ0)+Δδ∂τ∂δ(t,α0,δ0),(2)\tau(t,\alpha,\delta) = \tau(t,\alpha_0,\delta_0) + \Delta\alpha\frac{\partial\tau}{\partial\alpha}(t,\alpha_0,\delta_0) + \Delta\delta\frac{\partial\tau}{\partial\delta}(t,\alpha_0,\delta_0), \tag*{(2)}

where we linearize in the sky offsets Δα\Delta\alpha, Δδ\Delta\delta. For each PAT, we compute τ(ti,α0,δ0)\tau(t_i,\alpha_0,\delta_0) and its partial derivatives using PINT [22, 23, 24], which uses the location of the LAT within the Solar System [25]. We set τ=0\tau= 0 at the reference epoch, modified Julian date (MJD) 57662.23910470847.

We assume that the pulsar is isolated, moving at constant velocity with respect to the SSB, and slowly radiating rotational energy and angular momentum. Its rotation phase is described by a quadratic spin-down

Φ(τ)≡ϕ0+fτ+12f˙τ2,(3)\Phi(\tau) \equiv\phi_0 + f\tau+ \frac{1}{2}\dot{f}\tau^2, \tag*{(3)}

increasing by one for every full rotation. At time zero, ϕ0∈[−0.5,0.5)\phi_0 \in[-0.5,0.5) is the initial phase, ff is the pulsar spin frequency, and f˙\dot{f} (typically <0< 0) is the spin-down rate. Last, we assume that photons are emitted in pulses centered about the integer values of Φ\Phi.

*Pulsar search problem—*Our model implies that PATs lie close to times τ\tau for which Φ(τ)\Phi(\tau) is integer. From those PATs, we want to find the 5-vector

p=[ϕ0,f,f˙,fΔα,fΔδ](4)\mathbf{p} = [\phi_0,f,\dot{f},f\Delta\alpha,f\Delta\delta] \tag*{(4)}

containing the correct values of the three phase-model parameters and the two (scaled) sky direction offsets. As explained above, this is difficult because we do not know the number of rotations ki∈Zk_i \in\mathbb{Z} associated with the iith photon. With some caveats described later, the search problem is solved by minimizing the (squared) residual phases φi≡Φ[τ(ti,α,δ)]−ki\varphi_i \equiv\Phi[\tau(t_i,\alpha,\delta)] - k_i. If kik_i is the closest integer to Φ[τ(ti,α,δ)]\Phi[\tau(t_i,\alpha,\delta)], then φi∈[−0.5,0.5]\varphi_i \in[-0.5,0.5]. Once the integer vector k=[k1,…,kn]\mathbf{k} = [k_1,\ldots,k_n] is known, it is easy to find p\mathbf{p}. Note that in practice the PATs are widely separated, so physical solutions must have increasing kk, meaning increasing rotation counts k1<k2<⋯<knk_1 < k_2 < \cdots< k_n.

We express the search problem in linear form

v≡uB,(5)\mathbf{v} \equiv uB, \tag*{(5)}

where the full-rank (n+5)×(n+5)(n+5)\times(n+5) matrix

B≡[−In0Sη](6)B \equiv\begin{bmatrix} -I_n & 0 \\ S & \eta\end{bmatrix} \tag*{(6)}

is computed at the start of the search, and the row vector u=[k p]\mathbf{u} = [\mathbf{k}\ \mathbf{p}] contains the n+5n+5 unknowns. (Notation: [xy][xy] concatenates the nn columns of xx before the 5 columns of yy.) Here, InI_n is the n×nn\times n identity matrix, SS is the 5×n5\times n matrix

S≡[1⋯1τ(t1,α0,δ0)⋯τ(tn,α0,δ0)12τ2(t1,α0,δ0)⋯12τ2(tn,α0,δ0)∂τ∂α(t1,α0,δ0)⋯∂τ∂α(tn,α0,δ0)∂τ∂δ(t1,α0,δ0)⋯∂τ∂δ(tn,α0,δ0)],(7)S \equiv\begin{bmatrix} 1 & \cdots& 1 \\ \tau(t_1,\alpha_0,\delta_0) & \cdots& \tau(t_n,\alpha_0,\delta_0) \\ \frac{1}{2}\tau^2(t_1,\alpha_0,\delta_0) & \cdots& \frac{1}{2}\tau^2(t_n,\alpha_0,\delta_0) \\ \frac{\partial\tau}{\partial\alpha}(t_1,\alpha_0,\delta_0) & \cdots& \frac{\partial\tau}{\partial\alpha}(t_n,\alpha_0,\delta_0) \\ \frac{\partial\tau}{\partial\delta}(t_1,\alpha_0,\delta_0) & \cdots& \frac{\partial\tau}{\partial\delta}(t_n,\alpha_0,\delta_0) \end{bmatrix}, \tag*{(7)}

0\mathbf{0} is a zero matrix of appropriate shape [in (6), n×5n \times5], and η\eta is the 5×55\times5 diagonal matrix

η≡diag⁡(σσϕ0,σσf,σσf˙,σσfΔα,σσfΔδ).(8)\eta\equiv\operatorname{diag}\left(\frac{\sigma}{\sigma_{\phi_0}},\frac{\sigma}{\sigma_f},\frac{\sigma}{\sigma_{\dot{f}}},\frac{\sigma}{\sigma_{f\Delta\alpha}},\frac{\sigma}{\sigma_{f\Delta\delta}}\right). \tag*{(8)}

The subscripted σ\sigma’s in Table I are the widths of zero-mean Gaussian priors on the search parameters p\mathbf{p} in (4); σ=0.1\sigma= 0.1 rotations is the assumed Gaussian pulse width. The priors σfΔα,σfΔδ\sigma_{f\Delta\alpha},\sigma_{f\Delta\delta} define a sky patch slightly larger than the 4FGL 95% confidence radius for α0,δ0\alpha_0,\delta_0 at 300 Hz.

Search parameters.
PATs used for latticenn80
Nominal right ascension (RA)α0\alpha_016^{\mathrm h}22^{\mathrm m}12\fs72
Nominal declination (DEC)δ0\delta_0-72^\circ03'57\farcs6
Assumed pulse width (duty cycle)σ\sigma0.1
Initial phase prior widthσϕ0\sigma_{\phi_0}1
Spin frequency prior widthσf\sigma_f300 Hz
Spin-down prior widthσf˙\sigma_{\dot f}10−14 Hz s−110^{-14}\ \mathrm{Hz\,s}^{-1}
Scaled RA prior widthσfΔα\sigma_{f\Delta\alpha}(0^{\mathrm h}0^{\mathrm m}12\fs0)_{{\sigma_f}}
Scaled DEC prior widthσfΔδ\sigma_{f\Delta\delta}(0^\circ03'00\farcs0)_{{\sigma_f}}

Table 1.

We now show that via (5), the row vector u\mathbf{u} containing the correct pulsar parameters and rotation counts produces a row vector v\mathbf{v} with relatively short length ∥v∥\lVert\mathbf{v}\rVert. This transforms the pulsar search problem into a form similar to the classic “shortest vector problem” (SVP) in lattice theory.

Consider the first entry v1v_1 of (5), dropping higher orders in Δα,Δδ\Delta\alpha,\Delta\delta, and assuming ∣f˙∣\lvert\dot f\rvert is small enough that ∣τf˙(Δα∂τ/∂α+Δδ∂τ/∂δ)∣∼∣f˙∣(5500 d)(500 s)(1×rad−1)(3 arcmin)∼∣f˙∣108 s/Hz≪1\lvert\tau\dot f(\Delta\alpha\partial\tau/\partial\alpha+\Delta\delta\partial\tau/\partial\delta)\rvert\sim\lvert\dot f\rvert(5500\ \mathrm d)(500\ \mathrm s)(1\times\mathrm{rad}^{-1})(3\ \mathrm{arcmin})\sim\lvert\dot f\rvert10^8\ \mathrm{s/Hz}\ll1. Inspection of (6) and (7) gives v1=−k1+Φ[τ(t1,α,δ)]v_1=-k_1+\Phi[\tau(t_1,\alpha,\delta)]. For the correct u=[k p]\mathbf{u}=[\mathbf{k}\ \mathbf{p}], this gives v1/σ=ϕ1/σv_1/\sigma=\phi_1/\sigma: the normalized residual phase of the first photon, of order unity. This also holds for the next n−1n-1 components (photons).

Next, consider the (n+1)(n+1)th entry of (5). Inspection of (6) and (8) gives vn+1/σ=ϕ0/σϕ0v_{n+1}/\sigma=\phi_0/\sigma_{\phi_0}. This “penalty” for the first search parameter ϕ0\phi_0 is of order unity if ∣ϕ0∣≲σϕ0\lvert\phi_0\rvert\lesssim\sigma_{\phi_0}, i.e., if ϕ0\phi_0 lies within the “prior range” of Table I. This also holds for the final four components (search parameters).

Thus, dividing ∥v∥2\lVert\mathbf{v}\rVert^2 by σ2\sigma^2,

∥v∥2σ2=ϕ02σϕ02+f2σf2+f˙2σf˙2+f2Δα2σfΔα2+f2Δδ2σfΔδ2+∑i=1nϕi2σ2.(9)\frac{\lVert\mathbf{v}\rVert^2}{\sigma^2} = \frac{\phi_0^2}{\sigma_{\phi_0}^2} +\frac{f^2}{\sigma_f^2} +\frac{\dot f^2}{\sigma_{\dot f}^2} +\frac{f^2\Delta\alpha^2}{\sigma_{f\Delta\alpha}^2} +\frac{f^2\Delta\delta^2}{\sigma_{f\Delta\delta}^2} +\sum_{i=1}^{n}\frac{\phi_i^2}{\sigma^2}. \tag*{(9)}

This will be ≈n+5\approx n+5 for the correct integer rotation counts k\mathbf{k} and pulsar parameters p\mathbf{p}, and much larger for the wrong values, provided that the pulse profile has width ≈σ\approx\sigma. Thus, the solution to the search problem corresponds to parameters u\mathbf{u} that lead to “short” vectors v\mathbf{v}.

If the phase residuals and model parameters are assumed to have been drawn from independent zero-mean Gaussians with the widths in Table I, then −∥v∥2/2σ2-\lVert\mathbf{v}\rVert^2/2\sigma^2 is the logarithm of the posterior probability [17]. Short vectors therefore correspond to solutions with high posterior probability.

The algorithm presented in an earlier work [17] discretizes the five continuous real parameters p\mathbf{p}, transforming the search into an (n+5)(n+5)-dimensional SVP. We improve this with an additional step.

Lattice formulation

Below, we solve for the five real parameters p\mathbf{p} of u=[k p]\mathbf{u}=[\mathbf{k}\ \mathbf{p}], reducing the dimension from n+5n+5 to nn. The remaining unknowns k\mathbf{k} are then found by solving the SVP. From (5) and (6), for fixed k\mathbf{k} one has

v=uB=[−k 0]+pW,(10)\mathbf{v}=\mathbf{u}B=[-\mathbf{k}\ \mathbf{0}]+\mathbf{p}W, \tag*{(10)}

where W≡[S η]W\equiv[S\ \eta] is the last five rows of BB. As p\mathbf{p} varies, v\mathbf{v} traces out a 5-plane in Rn+5\mathbb{R}^{n+5}. The shortest v\mathbf{v} is obtained when v\mathbf{v} is orthogonal to that 5-plane. Thus, the vector p=pmin⁡\mathbf{p}=\mathbf{p}_{\min} that minimizes ∥v∥\lVert\mathbf{v}\rVert satisfies the five equations

vWT=([−k 0]+pmin⁡W)WT=0,(11)\mathbf{v}W^{\mathsf T}=\left([-\mathbf{k}\ \mathbf{0}]+\mathbf{p}_{\min}W\right)W^{\mathsf T}=0, \tag*{(11)}

since the dot product between v\mathbf{v} in (10) and any vectors in the 5-plane (here, the five columns of the transpose WTW^{\mathsf T}) must vanish. Multiplying (11) on the right by (WWT)−1(WW^{\mathsf T})^{-1} and moving the [−k 0][-\mathbf{k}\ \mathbf{0}] term to the rhs gives

pmin⁡=[k 0]WT(WWT)−1.(12)\mathbf{p}_{\min}=[\mathbf{k}\ \mathbf{0}]W^{\mathsf T}(WW^{\mathsf T})^{-1}. \tag*{(12)}

Define the (n+5)×(n+5)(n+5)\times(n+5) symmetric matrix

P≡−In+5+WT(WWT)−1W,(13)P\equiv-I_{n+5}+W^{\mathsf T}(WW^{\mathsf T})^{-1}W, \tag*{(13)}

noting that −P-P is a projection operator, since P2=−PP^2=-P. Substituting pmin⁡\mathbf{p}_{\min} from (12) into (10) gives

v=[k 0]P=kP⊥,(14)\mathbf{v}=[\mathbf{k}\ \mathbf{0}]P=\mathbf{k}P^\perp, \tag*{(14)}

where P⊥P^\perp is obtained from PP by removing the last five rows.

The set of vectors v=kP⊥\mathbf{v}=\mathbf{k}P^\perp with integer k∈Zn\mathbf{k}\in\mathbb{Z}^n forms a lattice with generator matrix P⊥P^\perp. Finding a nonzero integer k\mathbf{k} that minimizes the length of v\mathbf{v} is called the SVP. The matrix P⊥P^\perp is rectangular: n×(n+5)n\times(n+5). Equivalent square generator matrices may be found from (14), which implies ∥v∥2=kQkT\lVert\mathbf{v}\rVert^2=\mathbf{k}Q\mathbf{k}^{\mathsf T}. The symmetric positive-definite n×nn\times n Gram matrix Q≡P⊥(P⊥)TQ\equiv P^\perp(P^\perp)^{\mathsf T} has a unique positive definite symmetric square-root MM satisfying M2=QM^2=Q. From ∥v∥2=∥kM∥2\lVert\mathbf{v}\rVert^2=\lVert\mathbf{k}M\rVert^2, it follows that MM is an n×nn\times n generator matrix for the (possibly rotated) lattice generated by P⊥P^\perp. Alternatively, Cholesky decomposition produces a square lower triangular lattice generator matrix LL, with Q=LLTQ=LL^{\mathsf T} implying that ∥v∥2=∥kL∥2\lVert\mathbf{v}\rVert^2=\lVert\mathbf{k}L\rVert^2.

With this formulation, we do not search over the five parameters p\mathbf{p}. Instead, we search for rotation count vectors k\mathbf{k} that yield short vectors in the equivalent lattices generated by PP, MM, or LL. The corresponding p\mathbf{p} follow from (12).

Our approach differs from [17] in that we analytically extremize p\mathbf{p} and eliminate it, thereby reducing the dimension. In contrast, [17] works with the lattice generated by BB, after rescaling p\mathbf{p} so that its discrete integer values are used. This requires an inverse rescaling of the rows of WW in order to keep the product pW\mathbf{p}W in (10) invariant. The rescaled rows of WW have very small Euclidean norms, leading to a large number of short but unphysical lattice vectors with small integers in (the rescaled) p\mathbf{p} and k=0\mathbf{k}=\mathbf{0}. This issue is resolved using projections [different than our −P-\mathbf{P} defined by (13)]. In [17], Sec. 4 and the Appendix, the blocks SS and η\boldsymbol{\eta} of WW are orthogonalized separately, thereby changing the subspace spanned by the last five rows of BB, which is then projected out. We are still studying the advantages and disadvantages of the lattice that [17] would produce, which differs from the one presented here.

*Short vectors versus shortest vectors—*In practice, the shortest vectors v\boldsymbol{v} in the lattice come via (14) from rotation counts kk that are incorrect “solutions” to the search problem. Many of these kk have parameters pp via (12) that are unphysical. Examples are all kik_i being zero, except for one or a few being ±1\pm1, or nonmonotonic kk, or ∣f∣≲1\lvert f\rvert\lesssim1 mHz. The large trials factor also means that spurious short vectors may arise by chance. Furthermore, our data contain spurious background PATs, and most realistic pulsars have pulse profiles that are more complicated than our assumed Gaussian shape, lengthening v\boldsymbol{v} for the correct pp (e.g., the pulse profile of the pulsar we discovered has two peaks). Similar nonsolutions appear in [17] and we address this the same way, i.e., by generating many short vectors and identifying the correct solution among them. We are currently optimizing the method for more complicated pulse profiles, but defer this discussion to future work.

The correct solution pp is identified by testing the different possible solutions using the HH statistic [20, 26], normalized as in [13], Eqs. (7)–(11). The HH statistic is computed for the highest-weighted 1000 PATs, using the residual phases determined by pp. If H>150H > 150, then the parameters pp are recorded as a potential solution. The corresponding false alarm probability (for many independent random PATs) is p≈exp⁡(−0.4H)≈10−26p \approx\exp(-0.4 H) \approx10^{-26}, whose inverse is larger than the number of independent points in the parameter space (∼1022\sim10^{22}), making the trials-corrected significance similar to the standard 4.1σ4.1\sigma detection threshold [27].

*Implementation—*Several algorithms and libraries have been developed to solve the classic SVP, but we require many short vectors. We use a recent record holder in the SVP challenge, the G6K library [28]. This employs sieving, which repeatedly forms sums and differences of lattice vectors to produce shorter ones.

G6K requires integer generators, so we multiply LL by 103010^{30} and round to the nearest integers, following [17]. While one could sieve this basis directly, it is more efficient to first apply block Korkine-Zolotarev reduction and then invoke G6K’s “pump-up” procedure [28]. This begins by projecting the lattice orthogonally to the first n/2n/2 basis vectors and sieving the resulting n/2n/2-dimensional lattice. The projection dimension is then increased by adding one additional basis vector, after which projection and sieving are repeated. This continues until the full nn-dimensional lattice is reached. Throughout, G6K maintains and enlarges a database of short vectors. At the end, these are transformed back to the original basis, yielding many k\boldsymbol{k} with short kL\boldsymbol{k}L.

Vectors in database318 199
Unphysical vectors (nonincreasing kk)1434
Equivalent vectors removed126 244
Solutions (HH statistic H>150H > 150)1 (H≈231H \approx231)
Shortest vector norm∥v∥=0.099 98\lVert\boldsymbol{v}\rVert= 0.099\,98
Shortest physical vector norm∥v∥=1.093 36\lVert\boldsymbol{v}\rVert= 1.093\,36
Longest vector norm∥v∥=1.311 85\lVert\boldsymbol{v}\rVert= 1.311\,85
Solution vector norm∥v∥=1.212 15\lVert\boldsymbol{v}\rVert= 1.212\,15
Database vectors shorter than solution6827

Table 2. Vectors obtained from sieving with G6K.

We postprocess the vectors by flipping the signs of any kk with decreasing values, removing any unphysical vectors, and also removing any equivalent entries, where kk and k′k' differ by the same integer. Then, we examine the HH statistic as described above. A notebook reproducing our method is available in [29].

*Dataset—*We selected photons detected before MJD 60284 from the unassociated source 4FGL J1622.7−7202 [8]. We used Pass 8 P8R3 Fermi-LAT data [30, 31] and the P8R3_SOURCE_V2 instrument response functions, selecting SOURCE-class photons with energies above 100 MeV, and used the energy-dependent angular and point spread function event-type cuts described in [8]. We computed probability weights wiw_i with gtsrcprob [32], using the 4FGL-DR4 model [21] for the spectra and positions of the target and nearby sources, and the corresponding interstellar (gll_iem_v07.fits) and isotropic (iso_P8R3_SOURCE_V2_v1.txt) diffuse emission models [33].

*Multiple trials and random photon subsets—*The computational cost of the sieving process grows exponentially with nn. While including more photons should increase sensitivity, the algorithm is very sensitive to outlier PATs with large residual phases φi\varphi_i (background photons). To mitigate these effects, we start with the highest-weighted 120 photons, pick random subsets of size n=80n = 80 from these, and then carry out the lattice-based analysis described above. This is an extension of the method suggested in [17]. A search using a single random subset of PATs takes about 90 s to run on a single CPU core of an otherwise idle AMD EPYC 7351P 16-Core processor (this extends to ∼5\sim5 min under full load).

*First discovery—*The pulsar PSR J1622−7204 was found the first time that we tried our algorithm on 20 different random subsets. The search parameters are given in Table I. By chance, this trial used PATs containing only a few outliers. For other subsets containing a larger number of outliers, the algorithm failed to find the pulsar.

We got lucky with the 20 trials, of which one was successful. Based on 1000 trials, we estimate that the probability of successfully detecting this pulsar in a single subset of 80 PATs is p≈1%p \approx1\%. With NN trials, the probability of finding the pulsar at least once is 1−(1−p)N1-(1-p)^N, which is 18% for 20 trials, and 50% for 70 trials. A robust detection with 95% success probability is achieved for 300 trials.

Table II contains information about the shortest G6K vectors of the successful trial. The set of short vectors is illustrated in Figure 1 in the plane of vector norm versus HH statistic. The two groups of nonsolution vectors, the “physical” and “unphysical” ones, appear in all of our searches, independent of the source properties or its pulse profile. Even after the unphysical vectors are removed, the solution is not the shortest vector. However, its HH statistic H=231H = 231 has a pp-value p≈exp⁡(−0.4H)≈10−40p \approx\exp(-0.4H) \approx10^{-40} [20]. This is highly significant, since 1/p1/p is much larger than the number of independent trial points in parameter space, ∼1022\sim10^{22}.

Density of short vectors in the plane of vector norm versus H statistic

Figure 1. Density of the 318 199 short vectors found by G6K in the successful search, i.e., for one subset of 80 photons. The (correct) solution is circled. Vectors (blue) with HH statistic larger than the solution are unphysical: they have frequencies f<10−7f < 10^{-7} Hz and their kk are not increasing. The remaining vectors (red) have increasing kk and frequencies 33 mHz <f<2< f < 2 kHz.

Pulsar parameters—The physical properties of PSR J1622−7204 are summarized in Table III. These parameters and their uncertainties were determined from the 15 905 photons with weights wi>0.01w_i > 0.01, using the γ\gamma-ray pulsar timing method described in [34] to explore the likelihood surface in the neighborhood of the physical lattice vector with the highest HH statistic. These parameters and derived physical properties are typical for an isolated γ\gamma-ray millisecond pulsar (see, e.g., Fig. 2 of Ref. [2]). The uncertainties are relatively small because of the long data span of ∼15\sim15 years. Figure 2 shows the pulse profile and the 15 905 weighted photon phases over time. The effective number of photons from the pulsar is ∑iwi≈828\sum_i w_i \approx828, and the expected squared pulsation signal-to-noise ratio is ∑iwi2≈193\sum_i w_i^2 \approx193.

Units and conventions
Time coordinate τ\tauTDB
Reference epoch τ=0\tau= 0MJD 57662.23910470847
Solar System ephemeris modelDE440
Data spanMJD 54 694 to 60 284
Inferred parameters of PSR J1622−7204
Spin frequency ff260.896 208 070 72(2) Hz
Frequency derivative f˙\dot{f}−1.5845(4)×10−15-1.5845(4) \times10^{-15} Hz/s
Right ascension (RA) α\alpha16h22m16.553(2)16^{\mathrm{h}}22^{\mathrm{m}}16.553(2)
Declination (DEC) δ\delta−72∘04′25.27′′(1)-72^\circ04^\prime25.27^{\prime\prime}(1)
Proper motion in RA, μαcos⁡δ\mu_\alpha\cos\delta−2.9(1.6)-2.9(1.6) mas/yr
Proper motion in DEC, μδ\mu_\delta−4.7(2.1)-4.7(2.1) mas/yr
Derived parameters of PSR J1622−7204
Spin period3.832 941 871 385 6(3) ms
Spin-period derivative2.3279(6)×10−202.3279(6) \times10^{-20}
Characteristic agea^{a}2.6×1092.6 \times10^{9} yr
Surface magnetic fielda^{a}3.0×1083.0 \times10^{8} G
Light-cylinder magnetic fielda^{a}4.9×1044.9 \times10^{4} G
Spin-down powera^{a}1.6×10341.6 \times10^{34} erg/s

Table III. Properties of PSR J1622−7204. a^{a}These parameters are derived from the observed spin-period derivative, which includes contributions from the Shklovskii effect and Galactic acceleration relative to the SSB. We cannot correct for these (unknown) distance-dependent effects, but both are small and approximately cancel, leaving a net correction smaller than ±6%\pm6\%.

Weighted photon phases and integrated pulse profile

Figure 2. Bottom: weighted photon phases for the best-fit timing solution, showing two rotations. The left-axis star marks τ=0\tau= 0, the reference epoch of Table III. The rotational phase Φrot(τ)≡[Φ(τ)−Φ0] mod 1\Phi_{\mathrm{rot}}(\tau) \equiv[\Phi(\tau) - \Phi_0] \bmod1 with Φ0≈0.38\Phi_0 \approx0.38 is offset from (3) so that Φrot(0)=0\Phi_{\mathrm{rot}}(0) = 0. The 80 successful trial PATs are circled in red. Top: integrated pulse profile, with dark and light blue showing 1σ/2σ1\sigma/2\sigma posteriors. The dashed red line estimates the nonpulsed background. The assumed σ=0.1\sigma= 0.1 Gaussian pulse profile (dash-dot) is centered around integer Φ(τ)\Phi(\tau), marked by vertical lines, where the lattice algorithm identified peaks.

Conclusions—We have shown that lattice-based pulsar searches can discover previously undetected pulsars using 5 to 6 orders of magnitude less computing time than prior methods. The new method will be an important tool in future γ\gamma-ray pulsar searches.

PSR J1622−7204 is an isolated millisecond pulsar. An exciting prospect is that the lattice-based algorithm can also be used for blind searches for pulsars in binary systems. These are computationally infeasible with previous methods [9]. The Fermi-LAT source catalog [8] contains hundreds of unassociated sources whose spectral shape and temporal variability suggest that they might be pulsars [35], but searches for isolated pulsars have failed. Many of these could be γ\gamma-ray pulsars in binary systems. Searching for binary pulsars requires modeling the additional PAT delays due to the pulsar’s orbital motion, which produces additional rows in (7) and (8). A blind search for a pulsar in a binary system should have a computation cost ∼106\sim10^6 larger than the isolated search reported here [17], since it requires linearization around ∼106\sim10^6 binary parameter sets, each defining a different lattice. This should be feasible with Einstein@Home or similar resources. To compensate for the increased trials factor and search volume, we estimate that the threshold on HH increases by ∼71\sim71. We do not expect this to be an issue in practice, since many of the potential sources have enough high-weight photons to result in a much larger HH statistic.

Acknowledgments—We thank Toby Burnett for identifying 4FGL J1622.2−7202 as a promising pulsar candidate source for our searches, and M. A. Papa for helpful comments on the manuscript. B. Z. and D. G. are supported by the Minerva Foundation with funding from the Federal

German Ministry for Education and Research. The Fermi-LAT Collaboration acknowledges support for LAT development, operation, and data analysis from NASA and the U.S. Department of Energy; CEA/Irfu and IN2P3/CNRS in France; ASI and INFN in Italy; MEXT, KEK, and JAXA in Japan; and the K. A. Wallenberg Foundation, the Swedish Research Council, and the Swedish National Space Board. Science analysis support in the operations phase from INAF (Italy) and CNES (France) is also gratefully acknowledged. That work was performed in part under DOE Contract No. DE-AC02-76SF00515.

Data availability—The code to reproduce our result is available in [29] and includes a notebook implementing the lattice-based search as well as a notebook performing the parameter estimation. The resulting parameter file data/J1622-7204_shoogle_results.par (also included in [29]) can be loaded into PINT together with the PATs to produce a pulse profile compatible with Figure 2 as well as an HH statistic of ∼345\sim345.

References

  1. [1]W. B. Atwood, A. A. Abdo, M. Ackermann, W. Althouse, B. Anderson, M. Axelsson, L. Baldini, J. Ballet, D. L. Band, G. Barbiellini et al., The Large Area Telescope on the Fermi Gamma-ray Space Telescope mission, Astrophys. J. 697, 1071 (2009).
  2. [2]D. A. Smith, S. Abdollahi, M. Ajello, M. Bailes, L. Baldini, J. Ballet, M. G. Baring, C. Bassa, J. Becerra Gonzalez, R. Bellazzini et al., The third Fermi Large Area Telescope catalog of gamma-ray pulsars, Astrophys. J. 958, 191 (2023).DOI
  3. [3]C. G. Bassa, Z. Pleunis, J. W. T. Hessels, E. C. Ferrara, R. P. Breton, N. V. Guseinskaia, V. I. Kondratiev, S. Sanidas, L. Nieder, C. J. Clark et al., LOFAR discovery of the fastest-spinning millisecond pulsar in the Galactic field, Astrophys. J. Lett. 846, L20 (2017).DOI
  4. [4]L. Nieder, C. J. Clark, D. Kandel, R. W. Romani, C. G. Bassa, B. Allen, A. Ashok, I. Cognard, H. Fehrmann, P. Freire et al., Discovery of a gamma-ray black widow pulsar by GPU-accelerated Einstein@Home, Astrophys. J. Lett. 902, L46 (2020).
  5. [5]R. W. Romani, M. Belznezay, A. V. Filippenko, T. G. Brink, and W. Zheng, PSR J0952-0607: Tightening a record-high neutron star mass, Astrophys. J. 996, 101 (2026).arxiv.org/abs/2512.05099
  6. [6]F. Özel and P. Freire, Masses, radii, and the equation of state of neutron stars, Annu. Rev. Astron. Astrophys. 54, 401 (2016).DOI
  7. [7]M. Ajello, W. B. Atwood, L. Baldini, J. Ballet, G. Barbiellini, D. Bastieri, R. Bellazzini, A. Berretta, B. Bhattacharyya et al. (FERMI-LAT Collaboration), A gamma-ray pulsar timing array constrains the nanoHertz gravitational wave background, Science 376, 521 (2022).DOI
  8. [8]S. Abdollahi, F. Acero, M. Ackermann, M. Ajello, W. B. Atwood, M. Axelsson, L. Baldini, J. Ballet, G. Barbiellini, D. Bastieri et al., Fermi Large Area Telescope fourth source catalog, Astrophys. J. Suppl. Ser. 247, 33 (2020).DOI
  9. [9]L. Nieder, B. Allen, C. J. Clark, and H. J. Pletsch, Exploiting orbital constraints from optical data to detect binary gamma-ray pulsars, Astrophys. J. 901, 156 (2020).DOI
  10. [10]W. B. Atwood, M. Ziegler, R. P. Johnson, and B. M. Baughman, A time-differencing technique for detecting radio-quiet gamma-ray pulsars, Astrophys. J. 652, L49 (2006).DOI
  11. [11]A. A. Abdo, M. Ackermann, M. Ajello, B. Anderson, W. B. Atwood, M. Axelsson, L. Baldini, J. Ballet, G. Barbiellini, M. G. Baring et al., Detection of 16 gamma-ray pulsars through blind frequency searches using the Fermi LAT, Science 325, 840 (2009).DOI
  12. [12]P. M. Parkinson, M. Dormody, M. Ziegler, P. S. Ray, A. A. Abdo, J. Ballet, M. G. Baring, A. Belfiore, T. H. Burnett, G. A. Caliandro et al., Eight γ-ray pulsars discovered in blind frequency searches of Fermi LAT data, Astrophys. J. 725, 571 (2010).
  13. [13]H. J. Pletsch, L. Guillemot, B. Allen, M. Kramer, C. Aulbert, H. Fehrmann, P. S. Ray, E. D. Barr, A. Belfiore, F. Camilo et al., Discovery of nine gamma-ray pulsars in Fermi Large Area Telescope data using a new blind search method, Astrophys. J. 744, 105 (2012).DOI
  14. [14]H. J. Pletsch, L. Guillemot, B. Allen, D. Anderson, C. Aulbert, O. Bock, D. J. Champion, H. B. Eggenstein, H. Fehrmann, D. Hammer et al., Einstein@Home discovery of four young gamma-ray pulsars in Fermi LAT data, Astrophys. J. Lett. 779, L11 (2013).
  15. [15]C. J. Clark et al., The Einstein@Home gamma-ray pulsar survey. I. Search methods, sensitivity and discovery of new young gamma-ray pulsars, Astrophys. J. 834, 106 (2017).DOI
  16. [16]C. J. Clark, M. Di Mauro, J. Wu, B. Allen, O. Behnke, H. B. Eggenstein, B. Machenschalk, L. Nieder, P. M. Saz Parkinson, A. Ashok et al., Einstein@Home searches for gamma-ray pulsars in the inner Galaxy, Astrophys. J. 994, 149 (2025).DOI
  17. [17]D. Gazith, A. B. Pearlman, and B. Zackay, Recovering pulsar periodicity from time-of-arrival data by finding the shortest vector in a lattice, Astrophys. J. 979, 48 (2025).arxiv.org/abs/2402.07228
  18. [18]C. J. Clark, R. P. Breton, E. D. Barr, M. Burgay, T. Thongmeearkom, L. Nieder, S. Buchner, B. Stappers, M. Kramer, W. Becker et al., The TRAPUM L-band survey for pulsars in Fermi-LAT gamma-ray sources, Mon. Not. R. Astron. Soc. 519, 5590 (2023).
  19. [19]T. Thongmeearkom, C. J. Clark, R. P. Breton, M. Burgay, L. Nieder, O. G. Dodgson, B. McGloughlin, E. D. Barr, S. Buchner, B. W. Stappers et al., Pulsar discoveries from the TRAPUM UHF survey of Fermi-LAT sources, Mon. Not. R. Astron. Soc. 547, stag406 (2026).DOI
  20. [20]M. Kerr, Improving sensitivity to weak pulsations with photon probability weighting, Astrophys. J. 732, 38 (2011).DOI
  21. [21]J. Ballet, P. Bruel, T. H. Burnett, and B. Lott (Fermi-LAT Collaboration), Fermi Large Area Telescope fourth source catalog data release 4 (4FGL-DR4), arXiv:2307.12546.
  22. [22]J. Luo, S. Ransom, P. Demorest, P. S. Ray, A. Archibald, M. Kerr, R. J. Jennings, M. Bachetti, R. van Haasteren, C. A. Champagne et al., PINT: A modern software package for pulsar timing, Astrophys. J. 911, 45 (2021).
  23. [23]A. Subhanian, D. L. Kaplan, A. M. Archibald, J. Luo, P. S. Ray, T. T. Pennucci, S. M. Ransom, G. Agazie, W. Fiore, B. Larsen et al., PINT: Maximum-likelihood estimation of pulsar timing noise parameters, Astrophys. J. 971, 150 (2024).
  24. [24]J. Luo, S. Ransom, P. Demorest, R. van Haasteren, P. Ray, K. Stovall, M. Bachetti, A. Archibald, M. Kerr, J. Colen et al., PINT: High-precision pulsar timing analysis package, Astrophysics Source Code Library, record ascl:1902.007 (2019), https://ascl.net/1902.007.
  25. [25]Fermi-LAT Collaboration, Fermi data & products, https://fermi.gsfc.nasa.gov/ssc/data/ (2026), accessed: 2026-2-8.
  26. [26]O. C. de Jager, J. W. H. Swanepoel, and B. C. Raubenheimer, A powerful test for weak periodic signals with unknown light curve shape in sparse data, Astron. Astrophys. 221, 180 (1989), https://ui.adsabs.harvard.edu/abs/1989A%26A...221..180D/abstract.
  27. [27]D. A. Smith, P. Bruel, I. Cognard, A. D. Cameron, F. Camilo, S. Dai, L. Guillemot, T. J. Johnson, S. Johnston, M. J. Keith et al., Searching a thousand radio pulsars for gamma-ray emission, Astrophys. J. 871, 78 (2019).
  28. [28]M. R. Albrecht, L. Ducas, G. Herold, E. Kirshanova, E. W. Postlethwaite, and M. Stevens, The General Sieve Kernel and new records in lattice reduction, in Advances in Cryptology—EUROCRYPT 2019, edited by Y. Ishai and V. Rijmen (Springer International Publishing, Cham, 2019), pp. 717–746, 10.1007/978-3-030-17656-3_25.DOI
  29. [29]See Supplemental Material at http://link.aps.org/supplemental/10.1103/vlt2-xzgk for an example implementation of our method and the data and notebook for reproducing our timing solution.
  30. [30]W. Atwood, A. Albert, L. Baldini, M. Tinevella, J. Bregeon, M. Pesce-Rollins, C. Sgrò, P. Bruel, E. Charles, A. Drlica-Wagner et al., Pass 8: Toward the full realization of the Fermi-LAT scientific potential, in Proceedings of the 4th Fermi Symposium, Monterey, California, 2012, edited by T. J. Brandt, N. Omodei, and C. Wilson-Hodge, eConf C121028 (2013), pp. 8–13, arXiv:1303.3514.arxiv.org/abs/1303.3514
  31. [31]P. Bruel, T. H. Burnett, S. W. Digel, G. Jóhannesson, N. Omodei, and M. Wood, Fermi-LAT improved Pass 8 event selection, arXiv:1810.11394.arxiv.org/abs/1810.11394
  32. [32]Fermi-LAT Collaboration, Fermtools, v2.2.0, https://github.com/fermi-lat/ScienceTools/releases/tag/Fermitools-2.2.0, https://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/ (2022).
  33. [33]Fermi-LAT Collaboration, Fermi LAT background models, https://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html (2024), accessed: 2026-1-12.
  34. [34]C. J. Clark, S. Valtonina, L. Nieder, and R. van Haasteren, Timing gamma-ray pulsars using Gibbs sampling, Astron. Astrophys. 709, A215 (2026).
  35. [35]D. V. Malyshev, Towards resolving the Galactic center GeV excess with millisecond-pulsars-like sources using machine learning, Phys. Rev. D 111, 043033 (2025).DOI

Paper details

vlnc:W_7zg43tht
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

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