Bayesian Timing Pipeline Vela

• Bayesian Timing Pipeline Vela is a modular, probabilistic framework that fuses deterministic timing models with explicit stochastic noise modeling to extract astrophysical parameters from pulsar data.
• It employs full non-linear implementations and analytic marginalization over nuisance parameters, enabling robust inference of glitch characteristics and rotational dynamics.
• The pipeline, realized in software like Vela.jl, leverages evidence-based model selection and advanced sampling techniques to compare competing hypotheses in pulsar timing and gravitational-wave searches.

Searching arXiv for relevant papers on Bayesian timing pipelines and Vela-related pulsar timing. “Bayesian Timing Pipeline Vela” denotes a family of Bayesian inference workflows centered on pulsar timing and related signal-analysis tasks in which timing data, timing-model structure, and explicit probabilistic noise modeling are combined to estimate physical parameters, compare models, and propagate uncertainty in a statistically coherent manner. In the Vela context, the term refers most directly to two related but distinct lines of work: Bayesian pulsar-timing software named Vela or Vela.jl for single-pulsar timing and noise analysis (Susobhanan, 2024, Susobhanan, 15 May 2025), and Bayesian, timing-driven analyses of the Vela pulsar itself, including glitch modeling, internal-structure inference, and gravitational-wave searches triggered or constrained by radio timing (Collaboration et al., 2010, Montoli et al., 2020, Shannon et al., 2016). Across these settings, the common architecture is a timing-informed probabilistic pipeline in which deterministic phase or residual models are embedded in a likelihood, nuisance parameters and stochastic processes are marginalized or sampled, and evidences or Bayes factors are used for detection or model selection.

1. Definition and scope

In its most software-specific sense, Vela.jl is “a Bayesian pulsar timing and noise analysis package whose distinctive feature is a full, non-linear implementation of the wideband timing paradigm” and, more generally, “an efficient, modular, easy-to-use Bayesian pulsar timing and noise analysis package written in Julia” (Susobhanan, 15 May 2025, Susobhanan, 2024). It is designed for joint Bayesian inference on deterministic timing parameters and stochastic noise processes, with a Python interface, pyvela, exposing the timing engine to external samplers (Susobhanan, 15 May 2025).

In a broader methodological sense, a Bayesian timing pipeline associated with Vela comprises a structured sequence of operations: ingestion of timing or detector data, construction of a deterministic timing or signal model, definition of priors, evaluation of a likelihood under explicit noise assumptions, posterior sampling or evidence computation, and model comparison. This architecture appears in multiple Vela-related applications. In a gravitational-wave ringdown search triggered by the August 2006 Vela glitch, radio timing defines the glitch epoch, the signal start-time prior, and the on/off-source segmentation used for Bayesian model selection (Collaboration et al., 2010). In Bayesian modeling of the 2016 Vela glitch, pulse-to-pulse timing residuals are fitted with a three-component rotational model plus a magnetospheric term, and nested sampling is used to infer superfluid moments of inertia, couplings, and glitch timescales (Montoli et al., 2020). In long-baseline phase-coherent timing of Vela over 21 years, a Bayesian framework simultaneously models glitches and steep red timing noise and performs evidence-based model selection among glitch-recovery models (Shannon et al., 2016).

This suggests a unifying characterization: a Bayesian Timing Pipeline Vela is not a single algorithm but a class of pipelines in which timing observables are treated as the primary carrier of astrophysical information and are analyzed with explicit probabilistic structure. Depending on the application, the outputs may be posterior distributions for timing parameters, glitch parameters, neutron-star interior parameters, gravitational-wave signal amplitudes, or theory-level couplings in modified gravity (Vaglio et al., 2 May 2026).

2. Core probabilistic formulation

The core data model in Bayesian pulsar timing can be written in residual form. In the PTA-oriented Bayesian analysis pipeline for continuous gravitational waves, the timing residuals for a pulsar are modeled as

$\delta t = M\,\delta\boldsymbol{\xi} + n + s,$

where $M$ is the timing-model design matrix, $\delta\boldsymbol{\xi}$ are offsets from the best-fit timing parameters, $n$ is stochastic noise, and $s$ is the signal contribution (Ellis, 2013). After analytic marginalization over timing-model offsets, inference proceeds in the subspace orthogonal to the timing model via the projection matrix $G$, yielding a Gaussian likelihood in the projected residuals (Ellis, 2013). Vela.jl adopts an analogous full non-linear timing philosophy rather than relying exclusively on linearized approximations (Susobhanan, 2024).

In narrowband timing, Vela describes topocentric arrival times through a sum of physical delays,

$$t_{\text{arr}} = t_{\text{em}} + \Delta_{\text{B}} + \Delta_{\text{ltt}} + \Delta_{\text{DM}} + \Delta_{\text{scatter}} + \Delta_{\text{GW}} + \Delta_{\odot} + \Delta_{\text{clock}} + \Delta_{\text{jump}} + \mathcal{N}_{\text{R}} + \mathcal{N}_{\text{jitter}} + \dots,$$

and models rotational phase as

$$\phi = \phi_0 + \sum_{n=0}^{n_F} \frac{1}{(n+1)!} F_n (t_{\text{em}} - t_0)^{n+1} + \phi_{\text{glitch}} + \phi_{\text{SN}} + \dots$$

with timing residual

$r = \frac{\phi - \mathcal{N}[\phi]}{F}.$

The Gaussian likelihood then has the standard form

$$\ln L = -\frac{1}{2} r^T C^{-1} r - \frac{1}{2} \ln \det C$$

with covariance $M$0 containing white and correlated noise terms (Susobhanan, 2024).

In the wideband paradigm, each observing epoch yields both a TOA and a DM measurement. Vela.jl constructs a combined residual vector $M$1 and evaluates

$M$2

for the white-noise case, or an analytically marginalized form with

$M$3

when Gaussian-process components are present (Susobhanan, 15 May 2025).

In gravitational-wave timing-triggered analyses, the same Bayesian structure appears in a different data representation. In the 2006 Vela glitch ringdown search, time-frequency spectrogram power is modeled with non-central $M$4 likelihoods under the signal-present hypothesis and central $M$5 likelihoods under the noise-only hypothesis, and evidences are computed for a coherent signal model $M$6, a noise model $M$7, and an incoherent transient model $M$8 (Collaboration et al., 2010). The detection statistic is the odds ratio

$M$9

This use of timing to define priors and trigger windows, followed by Bayesian model selection, is an especially direct instance of a timing pipeline (Collaboration et al., 2010).

3. Software architecture of Vela and Vela.jl

Vela.jl is organized as a modular timing-and-noise engine with Julia as the numerical core and Python as the orchestration layer (Susobhanan, 2024, Susobhanan, 15 May 2025). The Python binding pyvela exposes a high-level SPNTA object intended to integrate with external samplers such as emcee (Susobhanan, 15 May 2025). One-time operations such as parsing par and tim files, clock corrections, and Solar-system ephemerides are delegated to PINT through pyvela (Susobhanan, 2024).

The central abstractions are TOA or WidebandTOA data objects, a TimingModel composed of ordered Component objects, a Kernel implementing the likelihood algebra, and a ParamHandler and prior layer for mapping between internal and physical parameter spaces (Susobhanan, 2024, Susobhanan, 15 May 2025). Components encode spin, astrometry, binary motion, Solar-system propagation, DM evolution, jumps, and stochastic or quasi-stochastic processes such as red spin noise or DM noise. In wideband mode, the same abstractions are extended so that both timing residuals and DM residuals are modeled simultaneously (Susobhanan, 15 May 2025).

Two kernel structures are emphasized. WhiteNoiseKernel evaluates the diagonal-covariance likelihood when only white noise is present. WoodburyKernel handles low-rank-plus-diagonal covariance,

$\delta\boldsymbol{\xi}$0

and exploits the Woodbury identity for efficient likelihood evaluation in the presence of Gaussian-process components or analytically marginalized linear parameters (Susobhanan, 15 May 2025). In narrowband timing, specialized kernels also treat ECORR-like epoch-correlated white noise with block structure, enabling $\delta\boldsymbol{\xi}$1 evaluation under simplifying assumptions (Susobhanan, 2024).

A notable design decision is the commitment to the full non-linear timing model. Many PTA frameworks linearize around a best-fit solution and analytically marginalize timing perturbations. Vela instead computes timing phases directly at each likelihood evaluation, except where explicit analytic marginalization over selected linear parameters is introduced as an optional acceleration (Susobhanan, 2024). This places it conceptually closer to full non-linear Bayesian timing codes while retaining modularity and sampler-agnostic usage.

The same software stack has already been used in production-level Bayesian timing applications. A 2026 analysis of PSR J1738+0333 employs the Bayesian timing pipeline Vela with a standard GR timing model, realistic white and red noise, and a subsequent theory-level resampling procedure to constrain Einstein-æther gravity, showing that the package is not limited to narrowband timing demonstrations (Vaglio et al., 2 May 2026).

4. Noise processes and inference strategy

Noise modeling is a defining feature of Bayesian timing pipelines because timing residuals are typically dominated by a mixture of radiometer noise, pulse jitter, red spin noise, DM noise, instrumental offsets, and sometimes profile-domain systematics. Vela.jl incorporates this through a decomposition of the covariance into white and correlated components (Susobhanan, 2024, Susobhanan, 15 May 2025).

White-noise variances in narrowband timing are written as

$\delta\boldsymbol{\xi}$2

where $\delta\boldsymbol{\xi}$3 is EFAC and $\delta\boldsymbol{\xi}$4 is EQUAD (Susobhanan, 2024). In wideband timing the analogous scaling becomes

$\delta\boldsymbol{\xi}$5

for TOA and DM uncertainties respectively (Susobhanan, 15 May 2025). Epoch-correlated narrowband jitter is treated through ECORR-like covariance blocks (Susobhanan, 2024).

Correlated processes are represented either through explicit Fourier-series components or through Gaussian processes whose coefficients have power-law priors. For red spin noise and DM noise, Vela uses Fourier-domain basis expansions with Gaussian priors on coefficients and power-law spectral hyperparameters such as TNREDAMP, TNREDGAM, TNDMAMP, and TNDMGAM (Susobhanan, 15 May 2025). To avoid Neal’s funnel geometry in sampling these GP models, the software rescales coefficients so that the transformed amplitudes are standard normal and independent of the spectral hyperparameters (Susobhanan, 2024).

In long-term Bayesian timing of the Vela pulsar itself, the timing noise is found to be well described by a steep power-law process with $\delta\boldsymbol{\xi}$6, and the evidence favors a stationary process independent of glitches (Shannon et al., 2016). Espinoza et al. show that steep red timing noise can generate step-like residual structures that mimic microglitches, which implies that a Bayesian Vela pipeline should treat red noise and glitch models as competing or joint explanatory components rather than assigning every short-timescale irregularity to a discrete glitch (Espinoza et al., 2020).

Inference engines differ by application. Vela.jl demonstrations use emcee for posterior sampling in both narrowband and wideband modes (Susobhanan, 2024, Susobhanan, 15 May 2025). The 2016 Vela glitch analysis uses dynesty via Bilby to obtain both posteriors and evidences (Montoli et al., 2020). The 21-year phase-coherent timing analysis of Vela uses TEMPONest with PolyChord nested sampling for high-dimensional posterior exploration and evidence computation (Shannon et al., 2016). In all cases, the common pattern is posterior sampling over nonlinear physical parameters with either explicit or analytic treatment of linear nuisance sectors.

5. Vela pulsar as a timing-pipeline testbed

The Vela pulsar is an especially demanding object for Bayesian timing because it combines high-quality timing information with strong rotational irregularities, substantial glitch activity, and potentially significant pulse-profile variability. Several strands of work illustrate distinct roles played by Bayesian timing pipelines in this system.

In the 21-year phase-coherent timing study, eight large glitches are modeled simultaneously with a power-law red-noise process. The preferred glitch model contains permanent frequency steps and two exponentially decaying transient frequency components, with shared timescales of approximately 25 d and 1300 d, and does not require permanent $\delta\boldsymbol{\xi}$7 steps (Shannon et al., 2016). This is a canonical example of a Bayesian timing pipeline resolving deterministic glitch recovery from stochastic timing noise.

In the pulse-to-pulse 2016 Vela glitch analysis, a three-component rotational model is written as

$\delta\boldsymbol{\xi}$8

leading to an analytic two-timescale solution for the observable glitch-induced deviation

$\delta\boldsymbol{\xi}$9

A magnetospheric residual-jump component is added to account for a transient pre-glitch residual increase, and nested sampling is used to infer moment-of-inertia fractions, coupling rates, lags, glitch rise and relaxation times, and evidence for overshoot (Montoli et al., 2020). The model comparison between a broad prior on superfluid reservoir fractions and a crust-limited prior yields $n$0 in favor of the broad model, supporting core participation in the glitch (Montoli et al., 2020).

Another line of work uses long-term glitch activity rather than individual glitch waveforms. Chamel derives the activity parameter

$n$1

and relates it to the superfluid and entrainment moments of inertia via

$n$2

For Vela, $n$3, and with strong crustal entrainment this implies

$n$4

Although this paper uses deterministic inequalities rather than a full Bayesian pipeline, it explicitly points toward a Bayesian implementation in which timing-derived $n$5 becomes a likelihood factor or prior on internal-structure parameters (Chamel, 2016).

The Vela pulse profile itself is also variable. Palfreyman et al. report width variations with quasi-periodicities of $n$6, $n$7, and $n$8 d, along with width jumps associated with microglitches and changes in bright-pulse rates (Palfreyman et al., 2016). A plausible implication is that any high-precision Bayesian timing pipeline for Vela should eventually move beyond fixed-profile TOA extraction and include profile-evolution or profile-domain components, especially when attempting sub-$n$9s precision or interpretation of subtle timing noise.

6. Timing-informed gravitational-wave analyses

A distinctive feature of Vela-related Bayesian pipelines is that timing information can be used not only to infer rotational and internal parameters, but also to drive targeted gravitational-wave searches.

In the search for gravitational waves associated with the August 2006 Vela glitch, radio timing from HartRAO is used to fit pre- and post-glitch spin models with TEMPO2 and to infer the glitch epoch by intersecting these models (Collaboration et al., 2010). The result,

$s$0

or GPS $s$1 s, directly defines the on-source window: $s$2 a 120 s interval centered on the glitch (Collaboration et al., 2010). The signal start time prior is then

$s$3

This is the clearest possible example of timing driving a Bayesian search.

The target signal is a damped $s$4-mode ringdown with

$s$5

and one dominant $s$6 spherical-harmonic component at a time (Collaboration et al., 2010). The pipeline constructs spectrograms from H1 and H2 data, models power with non-central $s$7 likelihoods, compares a coherent signal model against both Gaussian noise and an incoherent transient model, and calibrates the Bayesian odds empirically using 161 off-source trials (Collaboration et al., 2010). The on-source result $s$8 is consistent with background, so the pipeline proceeds to Bayesian upper limits, obtaining 90% credible bounds on intrinsic strain amplitudes from $s$9 to $G$0, depending on $G$1, and energy limits from $G$2 to $G$3 erg (Collaboration et al., 2010).

The broader lesson is methodological: a Bayesian timing pipeline can serve as a trigger-and-prior engine for multimessenger searches. Radio timing determines the event epoch, prior support, and background calibration strategy, while the Bayesian machinery handles coherence, nuisance models, and upper-limit construction (Collaboration et al., 2010).

A different but conceptually related timing-driven GW application is the Bayesian continuous-wave search toward Vela in Virgo VSR2. There the radio ephemeris provides an exact phase model used for heterodyning, and Bayesian inference is then carried out for the amplitude parameters $G$4 with a Student-$G$5-like likelihood after marginalization over segment-wise noise variance (Collaboration et al., 2011). The resulting 95% Bayesian upper limit of $G$6 for known orientation falls below the spin-down limit (Collaboration et al., 2011). Although the paper does not use the term “pipeline” in the software sense of Vela.jl, it is a mature timing-based Bayesian analysis architecture in which electromagnetic timing effectively supplies the phase model and parameter-space collapse (Collaboration et al., 2011).

7. Applications beyond Vela and implications

Vela.jl is not restricted to the Vela pulsar. It has been demonstrated on NANOGrav 12.5-year wideband data for PSR J1923+2515, where the model includes spin, astrometry, DM and DM1, wideband DM jumps, time jumps, EFAC, EQUAD, DMEFAC, spin noise, and DM noise, with a wideband likelihood evaluated on 119 TOA+DM measurements (Susobhanan, 15 May 2025). It has also been used on PSR J1738+0333 to obtain the timing posterior later propagated into Einstein-æther constraints, with 25,054 ToAs from Arecibo, Green Bank, Nançay, Parkes, and Westerbork (Vaglio et al., 2 May 2026). This suggests that the software component of Bayesian Timing Pipeline Vela is best understood as a general-purpose Bayesian timing engine whose name happens to coincide with the astrophysical object that motivated several methodological case studies.

The PTA continuous-wave pipeline of Babak et al. provides an instructive analogue. It organizes inference into search, characterization, and evaluation phases, using Adaptive Metropolis, parallel tempering, and thermodynamic integration to search for single SMBHB signals in pulsar timing arrays (Ellis, 2013). Although not a Vela-branded codebase, its three-layer architecture closely matches what later Vela-related pipelines implement in other contexts: fast localization of posterior maxima, posterior mapping, and evidence-based model comparison (Ellis, 2013). A plausible implication is that “Bayesian Timing Pipeline Vela” can be understood as a member of a broader class of modular Bayesian timing systems spanning radio pulsar timing, targeted GW searches, and theory tests.

Several recurrent design principles emerge across these applications. First, timing information is often the most constraining prior available, whether as a phase model, a glitch epoch, or a structural observable such as glitch activity. Second, explicit stochastic noise modeling is indispensable, especially for young pulsars such as Vela where red noise and glitch recovery can be strongly covariant (Shannon et al., 2016, Espinoza et al., 2020). Third, model comparison is not optional: crust-only versus core-participating glitch reservoirs, one- versus two-timescale glitch recoveries, coherent GW signals versus incoherent transients, and GR versus alternative-gravity interpretations all require evidences or Bayes factors rather than point estimation alone (Montoli et al., 2020, Collaboration et al., 2010, Vaglio et al., 2 May 2026).

A common misconception is that a Bayesian timing pipeline is merely a sampler wrapped around a standard pulsar timing model. The Vela literature indicates a broader conception. In practice, the pipeline includes data-model design, parameterization choices that regularize degeneracies, analytic marginalization of nuisance sectors, empirical calibration where needed, and hierarchical or multimessenger propagation of posteriors into physical constraints. In the Vela glitch problem, for example, the distinction between a permanent $G$7 step and a long exponential decay is not a matter of optimizer choice but of the model family itself, and evidence calculations show that the latter is preferred (Shannon et al., 2016).

Another misconception is that Bayesian timing pipelines are inherently tied to fixed-profile TOA analysis. Vela.jl’s wideband implementation shows that the Bayesian formalism naturally extends to joint TOA+DM measurements (Susobhanan, 15 May 2025). The observed pulse-profile evolution of Vela suggests that future pipelines may need to move still further toward profile-domain likelihoods when profile changes become astrophysically relevant or precision-limiting (Palfreyman et al., 2016). This suggests an evolutionary trajectory from narrowband TOA-domain timing, through wideband timing, toward profile-domain and eventually multimessenger or beyond-GR Bayesian timing engines.

Taken together, these works define Bayesian Timing Pipeline Vela as both a concrete software ecosystem and a methodological paradigm: a probabilistic timing-centered analysis framework in which pulsar rotational models, noise processes, glitch physics, gravitational-wave searches, and theory inference are integrated by Bayesian estimation and model comparison (Susobhanan, 2024, Susobhanan, 15 May 2025, Collaboration et al., 2010, Shannon et al., 2016, Montoli et al., 2020).