---
title: "Deriving the Hellings–Downs curve"
description: "How one gravitational wave shifts a pulsar's ticks, why a sky average gives the Hellings–Downs curve, and why clock and ephemeris errors look different."
canonical_url: "https://sandboxphysics.com/stories/pulsar-timing-arrays/deriving-the-hellings-downs-curve/"
language: "en"
page_type: "story"
alternate_languages:
  zh-CN: "https://sandboxphysics.com/zh/stories/pulsar-timing-arrays/deriving-the-hellings-downs-curve/"
last_modified: "2026-11-14"
---

[Pulsar timing arrays and the Hellings–Downs curve](https://sandboxphysics.com/stories/pulsar-timing-arrays/)

# Deriving the Hellings–Downs curve

Sandbox Physics · 2026-11-14

The Hellings–Downs curve is the average correlation between the timing residuals of two pulsars separated by an angle $\zeta$, when an isotropic, unpolarised gravitational-wave background passes the Earth. It comes from one geometric factor per pulsar, the antenna pattern, multiplied for the two pulsars and averaged over all wave directions and both polarisations. The result depends only on $\zeta$, has no monopole and no dipole part, and so differs in shape from the correlation that a clock error or an ephemeris error produces. This chapter derives the antenna pattern, states the curve in its two common normalisations, checks its features and compares it with the other two patterns.

You need vectors, the dot product, and the idea of averaging over a sphere. The previous chapter, [what a pulsar timing residual measures](https://sandboxphysics.com/stories/pulsar-timing-arrays/what-a-timing-residual-measures/), defines the residual.

## One wave and one pulsar

Let $\hat p$ point from the Earth to a pulsar at distance $L$, and let a plane gravitational wave travel in the direction $\hat\Omega$ with metric perturbation $h_{ij}(t,\vec x)$. Integrating the perturbation along the path of the radio pulse gives the fractional change of the pulse frequency,

$$
z(t)=\frac12\,\frac{\hat p^i\hat p^j}{1+\hat\Omega\cdot\hat p}\Big[h_{ij}(t,\vec 0)-h_{ij}\big(t-L(1+\hat\Omega\cdot\hat p)/c,\;L\hat p\big)\Big].
$$

The first term is the Earth term, the second the pulsar term. The timing residual is the time integral of $z$. The factor $1/(1+\hat\Omega\cdot\hat p)$ comes from that integration along the photon path. [Jenet and Romano, 2014, eqs. 54 to 56](https://arxiv.org/abs/1412.1142v1)

Write the wave in its two polarisations, $h_{ij}=h_+e^+_{ij}+h_\times e^\times_{ij}$, with $e^+=\hat\theta\hat\theta-\hat\phi\hat\phi$ and $e^\times=\hat\theta\hat\phi+\hat\phi\hat\theta$ built from two unit vectors perpendicular to $\hat\Omega$. The antenna pattern of the pulsar for polarisation $A$ is

$$
F^A(\hat\Omega,\hat p)=\frac12\,\frac{\hat p^i\hat p^j\,e^A_{ij}(\hat\Omega)}{1+\hat\Omega\cdot\hat p}.
$$

Here $\hat\Omega$ is the propagation direction. The NANOGrav anisotropy paper writes the denominator as $1-\hat\Omega\cdot\hat p$ because its $\hat\Omega$ points toward the source. The two forms are the same function. [NANOGrav Collaboration, 2023b, eq. 2](https://arxiv.org/abs/2306.16221v1)

Worked example: the antenna pattern in the wave's own frame

Put $\hat\Omega$ along the $z$ axis and take $\hat\theta$ and $\hat\phi$ along $x$ and $y$. A pulsar at polar angle $\theta$ and azimuth $\phi$ has $\hat p\cdot e^+\cdot\hat p=\sin^2\theta\cos2\phi$ and $\hat p\cdot e^\times\cdot\hat p=\sin^2\theta\sin2\phi$. Because $\sin^2\theta/(1+\cos\theta)=1-\cos\theta$,

$$
F^+=\tfrac12(1-\cos\theta)\cos2\phi,\qquad F^\times=\tfrac12(1-\cos\theta)\sin2\phi .
$$

The response vanishes for a wave travelling toward the pulsar, $\theta=0$, and grows toward a wave arriving from behind the pulsar, $\theta\to180^\circ$. In that last direction the delay $L(1+\cos\theta)/c$ between the Earth and pulsar terms goes to zero, the two terms cancel, and the full response still goes to zero.

## From two antenna patterns to one curve

For an isotropic, unpolarised background every direction and both polarisations carry equal power, and different directions are independent. The Earth-term correlation of pulsars $a$ and $b$ is therefore proportional to

$$
\int \mathrm d^2\hat\Omega\,\Big[F^+_aF^+_b+F^\times_aF^\times_b\Big].
$$

The pulsar terms are delayed by $L(1+\hat\Omega\cdot\hat p)/c$, thousands of years for a pulsar a kiloparsec away. For two different pulsars these delays differ by many wave periods, and the pulsar-term products average to zero. For one pulsar compared with itself, the pulsar term correlates with itself and doubles the result.

Doing the integral gives the Hellings–Downs function. In the normalisation of the NANOGrav 15-year paper,

$$
\Gamma(\zeta)=\frac32\,x\ln x-\frac x4+\frac12+\frac12\,\delta_{ab},\qquad x=\frac{1-\cos\zeta}{2}.
$$

[NANOGrav Collaboration, 2023, eqs. 4 and 5](https://arxiv.org/abs/2306.16213v1) Allen averages the polarisation-summed product over the sky and writes the same shape as $\mu_u(\gamma)=\tfrac14+\tfrac{\cos\gamma}{12}+\tfrac12(1-\cos\gamma)\ln\tfrac{1-\cos\gamma}{2}$, which equals $\tfrac23\Gamma$ for two different pulsars and starts at $\tfrac13$. [Allen, 2022, eq. 2.4](https://arxiv.org/abs/2205.05637v3) When you read a paper, check which normalisation it uses before comparing numbers.

Worked example: two pulsars at a right angle

At $\zeta=90^\circ$, $x=\tfrac12$, so $\Gamma=\tfrac32\cdot\tfrac12\ln\tfrac12-\tfrac18+\tfrac12=-0.520+0.375=-0.145$. In Allen's normalisation this is $-0.0966$. Two pulsars a right angle apart are slightly anticorrelated.

## Checking the shape

The curve starts at $\tfrac12$ for two different pulsars very close together, crosses zero twice, reaches its minimum near $82.5^\circ$ and rises to $\tfrac14$ for pulsars on opposite sides of the sky. The NANOGrav paper quotes zero crossings at $49.2^\circ$ and $121.8^\circ$. Solving the formula above numerically gives $49.32^\circ$ and $121.79^\circ$; at $49.2^\circ$ the formula is still $+0.001$. The printed $49.2^\circ$ is a rounding or a typo, not a different curve.

We checked the formula independently: a direct quadrature of the polarisation-summed antenna patterns over $240\times480$ sky cells, with explicit tensors rather than the closed form. Its sky mean, multiplied by $\tfrac32$, agrees with the formula to within $3.5\times10^{-4}$ at every angle tested.

**Table 3.1** Sky quadrature of the antenna patterns against the closed-form Hellings–Downs function, both in the NANOGrav normalisation.

| Separation | Quadrature times three halves | Closed form |
| --- | --- | --- |
| $10^\circ$ | 0.4426 | 0.4425 |
| $30^\circ$ | 0.2116 | 0.2116 |
| $60^\circ$ | -0.0824 | -0.0824 |
| $82.5^\circ$ | -0.1519 | -0.1519 |
| $121.8^\circ$ | 0.0001 | 0.0001 |
| $179^\circ$ | 0.2499 | 0.2499 |

## Why it is neither a monopole nor a dipole

Any function of $\cos\zeta$ can be written as a sum of Legendre polynomials $P_l(\cos\zeta)$. A clock error is pure $l=0$, because every pair correlates with the same value $1$. An ephemeris error shifts each pulsar by $\vec e(t)\cdot\hat p/c$. For an error vector with no preferred direction, the correlation is proportional to $\hat p_a\cdot\hat p_b=\cos\zeta=P_1$, pure $l=1$. [Tiburzi et al., 2015, Section 2](https://arxiv.org/abs/1510.02363v1)

The Hellings–Downs function has none of these two terms:

$$
\Gamma(\zeta)=\sum_{l=2}^{\infty}g_l\,P_l(\cos\zeta),\qquad g_l=\frac32\,\frac{(2l+1)\,(l-2)!}{(l+2)!}.
$$

[NANOGrav Collaboration, 2023, eq. 12](https://arxiv.org/abs/2306.16213v1) The leading term is quadrupolar, $g_2=\tfrac{15}{48}=0.3125$, because a gravitational wave stretches along one axis and squeezes along the other. Projecting the closed form onto the Legendre polynomials by quadrature reproduces these coefficients to $3\times10^{-9}$ up to $l=8$, and $g_0$ and $g_1$ come out zero.

Worked example: the first three Legendre coefficients

$g_2=\tfrac32\cdot\tfrac{5\cdot0!}{4!}=0.3125$, $g_3=\tfrac32\cdot\tfrac{7\cdot1!}{5!}=0.0875$ and $g_4=\tfrac32\cdot\tfrac{9\cdot2!}{6!}=0.0375$. They fall roughly as $l^{-3}$, so a few terms already trace the curve; the sum of all $g_l$ is $\Gamma(0^+)=\tfrac12$.

## One sky is not the average sky

The curve is an average over an ensemble of skies. Allen computes how much one sky can differ. For a single point source and pulsar pairs swept around the sky at fixed $\gamma$, the correlation has mean $\mu_u$ and variance

$$
\sigma_u^2=\frac{97}{80}+\frac{c}{24}-\frac{839}{720}\,c^2+\frac{1}{12}\,\psi\,(18-10c-3\psi),\qquad c=\cos\gamma,\quad \psi=(1-c)\ln\frac{1-c}{2}.
$$

[Allen, 2022, eq. 2.6](https://arxiv.org/abs/2205.05637v3) At small separation $\sigma_u=0.30$ while $\mu_u=0.33$: one source alone barely traces the curve. Many sources at different frequencies average this out; sources at the same frequency interfere and leave a cosmic variance that more pulsars cannot remove. Our sky quadrature reproduces eq. 2.6 to within $9\times10^{-5}$.

## Lab task

Open the pulsar-timing-array Lab with the default background. Drag one pulsar slowly across the sky map and watch its pair points slide along the angle axis of the correlation plot. Then load the clock-error and ephemeris-error presets and watch the dashed reference curve change shape.

## Exercises

1. Evaluate $\Gamma(60^\circ)$ and $\Gamma(180^\circ)$.
2. What is $\Gamma$ for a pulsar compared with itself? Which term supplies the extra half?
3. Show that the average of $\Gamma$ over pulsar pairs spread uniformly on the sky is zero. What does that imply for a clock error fitted with the Hellings–Downs shape alone, if pulsars were spread uniformly?
4. Using the Legendre series and $\int_{-1}^{1}P_l^2\,\mathrm d\mu=2/(2l+1)$, show that the uniform-sky average of $\Gamma^2$ is $\sum_l g_l^2/(2l+1)$. Evaluate the first three terms.
5. Convert $\Gamma(82.5^\circ)=-0.152$ to Allen's normalisation.

Answers

1. At $60^\circ$, $x=\tfrac14$ and $\Gamma=\tfrac38\ln\tfrac14-\tfrac1{16}+\tfrac12=-0.0824$. At $180^\circ$, $x=1$, $\ln x=0$ and $\Gamma=\tfrac14$.
2. $\Gamma=\tfrac12+\tfrac12=1$. The Earth term gives $\tfrac12$ and the pulsar term, which correlates only with itself, gives the other $\tfrac12$.
3. A uniform average picks out the $P_0$ coefficient, which is $g_0=0$. For a uniform array, a monopole and the Hellings–Downs shape would be orthogonal and a clock error would not leak into the fit on average. Real arrays are not uniform, which the [optimal-statistic chapter](https://sandboxphysics.com/stories/pulsar-timing-arrays/optimal-statistic-and-null-tests/) quantifies.
4. Squaring the series and averaging over $\mu=\cos\zeta$ with weight $\tfrac12$ keeps only the diagonal terms, $\sum_l g_l^2/(2l+1)$. The first three are $0.01953$, $0.00109$ and $0.00016$; the full sum is $1/48=0.02083$.
5. Multiply by $\tfrac23$: $-0.101$.

## [Try it in the pulsar-timing-array Lab](https://sandboxphysics.com/labs/pulsar-timing-array-correlations/)

Drag one pulsar across the sky and watch its pair points slide along the curve.

## Sources

1. [Jenet and Romano, 2014, eqs. 54 to 56](https://arxiv.org/abs/1412.1142v1)
2. [NANOGrav Collaboration, 2023b, eq. 2](https://arxiv.org/abs/2306.16221v1)
3. [NANOGrav Collaboration, 2023, eqs. 4 and 5](https://arxiv.org/abs/2306.16213v1)
4. [Allen, 2022, eq. 2.4](https://arxiv.org/abs/2205.05637v3)
5. [Tiburzi et al., 2015, Section 2](https://arxiv.org/abs/1510.02363v1)

## Keep reading

Pulsar timing arrays and the Hellings–Downs curve · Reading path

### From one ticking pulsar to a Galaxy-sized detector

1. [Article **How can a pulsar timing array tell gravitational waves from a bad clock?**](https://sandboxphysics.com/stories/pulsar-timing-arrays/gravitational-waves-or-a-bad-clock/)
2. [Article **What a pulsar timing residual measures**](https://sandboxphysics.com/stories/pulsar-timing-arrays/what-a-timing-residual-measures/)
3. [You are here **Deriving the Hellings–Downs curve**](https://sandboxphysics.com/stories/pulsar-timing-arrays/deriving-the-hellings-downs-curve/)
4. [Article **The optimal statistic and its null tests**](https://sandboxphysics.com/stories/pulsar-timing-arrays/optimal-statistic-and-null-tests/)
5. [Article **Recompute a Hellings–Downs statistic from NANOGrav 15-year data**](https://sandboxphysics.com/stories/pulsar-timing-arrays/recompute-the-optimal-statistic/)
6. [Article **Reading the NANOGrav 15-year evidence for a gravitational-wave background**](https://sandboxphysics.com/stories/pulsar-timing-arrays/reading-the-nanograv-15-year-evidence/)
