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 follow-up. Here an alternative, complementary approach is presented which tracks the pulse frequency $f$ and its time derivative $df/dt$ with a hidden Markov model (HMM), whose dynamics include stochastic spin wandering (timing noise) and impulsive jumps in $f$ and $df/dt$ (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 $\sim 1$ 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 $> 10^{-7} f$.

Searches for persistent gravitational radiation from nonpulsating neutron stars in young supernova remnants (SNRs) are computationally challenging because of rapid stellar braking. We describe a practical, efficient, semi-coherent search based on a hidden Markov model (HMM) tracking scheme, solved by the Viterbi algorithm, combined with a maximum likelihood matched filter, the $\mathcal{F}$-statistic. The scheme is well suited to analyzing data from advanced detectors like the Advanced Laser Interferometer Gravitational Wave Observatory (Advanced LIGO). It can track rapid phase evolution from secular stellar braking and stochastic timing noise torques simultaneously without searching second- and higher-order derivatives of the signal frequency, providing an economical alternative to stack-slide-based semi-coherent algorithms. One implementation tracks the signal frequency alone. A second implementation tracks the signal frequency and its first time derivative. It improves the sensitivity by a factor of a few upon the first implementation, but the cost increases by two to three orders of magnitude.

A hidden Markov model (HMM) scheme for tracking continuous-wave gravitational radiation from neutron stars in low-mass X-ray binaries (LMXBs) with wandering spin is extended by introducing a frequency-domain matched filter, called the J-statistic, which sums the signal power in orbital sidebands coherently. The J-statistic is similar but not identical to the binary-modulated F-statistic computed by demodulation or resampling. By injecting synthetic LMXB signals into Gaussian noise characteristic of the Advanced Laser Interferometer Gravitational-wave Observatory (Advanced LIGO), it is shown that the J-statistic HMM tracker detects signals with characteristic wave strain $h_0 \geq 2 \times 10^{-26}$ in 370 d of data from two interferometers, divided into 37 coherent blocks of equal length. When applied to data from Stage I of the Scorpius X-1 Mock Data Challenge organised by the LIGO Scientific Collaboration, the tracker detects all 50 closed injections ($h_0 \geq 6.84 \times 10^{-26}$), recovering the frequency with a root-mean-square accuracy of $\leq 1.95\times10^{-5}$ Hz. Of the 50 injections, 43 (with $h_0 \geq 1.09 \times 10^{-25}$) are detected in a single, coherent 10-d block of data. The tracker employs an efficient, recursive HMM solver based on the Viterbi algorithm, which requires $\sim 10^5$ CPU-hours for a typical, broadband (0.5-kHz), LMXB search.

Gravitational wave searches for continuous-wave signals from neutron stars are especially challenging when the star's spin frequency is unknown a priori from electromagnetic observations and wanders stochastically under the action of internal (e.g. superfluid or magnetospheric) or external (e.g. accretion) torques. It is shown that frequency tracking by hidden Markov model (HMM) methods can be combined with existing maximum likelihood coherent matched filters like the F-statistic to surmount some of the challenges raised by spin wandering. Specifically it is found that, for an isolated, biaxial rotor whose spin frequency walks randomly, HMM tracking of the F-statistic output from coherent segments with duration T_drift = 10d over a total observation time of T_obs = 1yr can detect signals with wave strains h0 > 2e-26 at a noise level characteristic of the Advanced Laser Interferometer Gravitational Wave Observatory (Advanced LIGO). For a biaxial rotor with randomly walking spin in a binary orbit, whose orbital period and semi-major axis are known approximately from electromagnetic observations, HMM tracking of the Bessel-weighted F-statistic output can detect signals with h0 > 8e-26. An efficient, recursive, HMM solver based on the Viterbi algorithm is demonstrated, which requires ~10^3 CPU-hours for a typical, broadband (0.5-kHz) search for the low-mass X-ray binary Scorpius X-1, including generation of the relevant F-statistic input. In a "realistic" observational scenario, Viterbi tracking successfully detects 41 out of 50 synthetic signals without spin wandering in Stage I of the Scorpius X-1 Mock Data Challenge convened by the LIGO Scientific Collaboration down to a wave strain of h0 = 1.1e-25, recovering the frequency with a root-mean-square accuracy of <= 4.3e-3 Hz.