A timing residual is the difference between when a pulse from a pulsar actually arrived and when a fitted timing model said it would arrive. The model removes everything that is specific to one pulsar and predictable: its spin and slow-down, its position and motion, its orbit and the delay added by the interstellar plasma. A slow gravitational wave survives that fit only partly, as a wandering of the residuals that lasts years. This chapter builds the residual, separates white noise from red noise, and works out how large the nanohertz signal reported by NANOGrav is in nanoseconds.
You need the idea of a sine wave, a least-squares fit and a variance. Times are in seconds, frequencies in hertz, and one nanohertz is 10−9Hz.
From pulses to residuals
A pulsar's rotation phase grows almost linearly with time. Over years its spin frequency ν slowly decreases, so to first approximation
ϕ(t)=ϕ0+νt+21ν˙t2,
and a pulse arrives whenever ϕ passes a whole number. Observers record times of arrival, convert them to the centre of mass of the Solar System, and fit a timing model to them. In the NANOGrav 15-year analysis that model contains the spin period and its derivative, sky position, proper motion and parallax for every pulsar, five orbital elements and any needed extra terms for pulsars in binaries, and a piecewise-constant correction for the dispersion measure called DMX. NANOGrav Collaboration, 2023, Section 2
The residual is δt=tobs−tmodel. A gravitational wave makes the pulses arrive early and late in a time series that varies at the wave's frequency, with an amplitude proportional to the wave's strain. Allen, 2022, Section I
What the fit takes away
The fit cannot tell a slow signal from a slightly wrong spin frequency or spin-down. Any part of a residual that looks like a constant, a straight line or a parabola is absorbed by ϕ0, ν and ν˙. Parallax and proper motion are fitted in every pulsar for the same reason: they produce delays at one cycle per half year and one cycle per year in all pulsars, which a search could otherwise mistake for a gravitational wave.
Worked example: how much of the slowest wave survives the spin fitShow the answer
Sample a sine and a cosine with one full cycle over the span T, and remove their least-squares fit by 1, t and t2. The sine keeps 39% of its variance and the cosine keeps 7.6%. At two cycles per span the two keep 85% and 94%, and at five cycles more than 97%. The lowest frequency, where a gravitational-wave background is strongest, is exactly where the spin fit removes the most.
White noise and red noise
Each arrival time carries a measurement error. NANOGrav models it with three parameters per receiver and backend: a scaling of the quoted uncertainties, an extra term added in quadrature, and a term common to all sub-bands at the same epoch. NANOGrav Collaboration, 2023, Section 2 This noise is white: independent from one epoch to the next, with equal power at all frequencies.
Red noise has more power at low frequencies. Some comes from the pulsar itself, some from the interstellar medium, and a gravitational-wave background is red too. To describe red processes, NANOGrav uses sines and cosines at the frequencies
fi=Ti,T=16.03yr,
so f1=1.98nHz. Each sine and each cosine gets a random coefficient, and the variance of the coefficient at fi is the power in that bin. NANOGrav Collaboration, 2023, eqs. 1 to 3
White noise also has a variance per coefficient. Fit a unit sine to N evenly spaced residuals with independent errors σ, and its coefficient has variance 2σ2/N.
The power-law spectrum
A population of inspiralling supermassive black-hole binaries gives a characteristic strain hc(f)=A(f/fref)−2/3, with fref=1yr−1. The matching variance of each Fourier coefficient is
The coefficient spread is Φ1=709ns. Each further bin is smaller by i−13/6: 158ns at i=2 and 22ns at i=5.
A sine with coefficient spread Φ contributes a time-averaged residual variance Φ/2, and its cosine partner adds the same, so bin i adds Φi to the residual variance before the timing fit.
Table 2.1 Coefficient spread of the reported background and of white noise with σ=1μs on N=192 epochs.
Bin i
fi (nHz)
Background Φi (ns)
White noise (ns)
1
1.98
709
102
2
3.95
158
102
3
5.93
66
102
5
9.88
22
102
The background dominates the lowest two bins and falls below the white noise from the third. That is why a pulsar timing array needs many years: the signal lives in the few lowest frequencies, and each year of data moves f1 lower.
Lab task
Open the pulsar-timing-array Lab and click a pulsar on the sky map. The white dots are its residuals with noise; the line is the smooth common signal alone, drawn with the signals of its nearest and farthest partners. Raise the white noise from 1 to 3μs and watch the slow wander disappear into the scatter, while the correlation plot loses its shape.
Exercises
NANOGrav limits the common-signal fit to the first 14 bins. What is f14?
By what factor is Φ2 smaller than Φ1 for γ=13/3? What would the factor be for a flat spectrum, γ=0?
A pulsar has white noise σ=0.5μs on 400 epochs. What is the white-noise spread of one Fourier coefficient?
With the Table 2.1 numbers, at which bin index would the background equal the white noise, treating i as continuous?
Estimate the residual amplitude in the first bin from hc(f1)/(2πf1) and compare it with Φ1.
AnswersShow the answer
f14=14/T=27.7nHz, matching the 2 to 28 nHz range quoted at the end of Section 2 of the paper.
2−13/6=0.223. A flat spectrum gives 20=1: every bin has the same power, like white noise.
2/400×0.5μs=35ns.
709i−13/6=102 gives i=6.956/13=2.45, between the second and third bins.
hc(f1)=2.4×10−15×16.032/3=1.53×10−14, and hc/(2πf1)=1.2μs. It agrees with 709ns within a factor of two, as an order-of-magnitude estimate should.