Pulsar glitch detection with a hidden Markov model

Abstract

Pulsar timing experiments typically generate a phase-connected timing solution from a sequence of times-of-arrival (TOAs) by absolute pulse numbering, i.e. by fitting an integer number of pulses between TOAs in order to minimize the residuals with respect to a parametrized phase model. In this observing mode, rotational glitches are discovered, when the residuals of the no-glitch phase model diverge after some epoch, and glitch parameters are refined by Bayesian followup. Here an alternative, complementary approach is presented which tracks the pulse frequency f and its time derivative ḟ with a hidden Markov model (HMM), whose dynamics include stochastic spin wandering (timing noise) and impulsive jumps in f and ḟ (glitches). The HMM tracks spin wandering explicitly, as a specific realization of a discrete-time Markov chain. It discovers glitches by comparing the Bayes factor for glitch and no-glitch models. It ingests standard TOAs for convenience and, being fully automated, allows performance bounds to be calculated quickly via Monte Carlo simulations. Practical, user-oriented plots are presented of the false alarm probability and detection threshold (e.g. minimum resolvable glitch size) versus observational scheduling parameters (e.g. TOA uncertainty, mean delay between TOAs) and glitch parameters (e.g. transient and permanent jump sizes, exponential recovery time-scale). The HMM is also applied to ∼ 1 yr of real data bracketing the 2016 December 12 glitch in PSR J0835−4510 as a proof of principle. It detects the known glitch and confirms that no other glitch exists in the same data with size 10 −7 f .

Testing for a Random Walk Structure in the Frequency Evolution of a Tone in Noise

“…The modelling and detection of such CGWs by interferometric data (such as from aLIGO) follows the work of Jaranowski [27] (and the following parts). The stochastic wander in f is incorporated into detection schemes based on hidden Markov models [32,33].…”

“…). This is a non-physical prior which will impose little to no bias on inferences, used similarly in discretised methods such as in HMMs [32,33]. Definition 2.…”

“…} k is serially correlated at lag τ. (32) The joint inference of the above hypothesis tests determines which of the four hypotheses holds as described in (9).…”

Testing for a Random Walk Structure in the Frequency Evolution of a Tone in Noise

“…The modelling and detection of such CGWs by interferometric data (such as from aLIGO) follows the work of Jaranowski [11], [12], [29]- [31]. The stochastic wander in f is incorporated into detection schemes based on hidden Markov models [37], [59]- [61].…”

“…Important targets include the persistent, sinusoidal, quasimonochromatic signals predicted to emanate from rapidly rotating neutron stars with an asymmetric mass distribution; these are termed continuouswave signals [48], [49]. In continuous-wave searches, detection and estimation algorithms often assume that the frequency wanders according to a random walk (RW) [11], [12], [29]- [31], [37], [59]- [61]. The problem also arises in a range of other contexts in structural analysis [67] and sound processing [9], [24], [63] (see §II).…”