Astronomy & Celestial Mechanics
the sky, computed
Concept

Prior — where it appears

The probability distribution assumed for a parameter before the data are considered. Where the data constrain a parameter well the prior is irrelevant, and along a degenerate direction the posterior is the prior — which is what a reported limit sometimes is.

Named by 6 essays across 3 fields — each of them below, with the objects they name alongside it.

One amplitude, and every point on this curve fits it. The set of distances and inclinations that produce the same measured amplitude in one detector. A binary seen face-on radiates most strongly along its spin axis, so it can be twice as far away as an edge-on binary and still arrive with the same strain — the curve is (1 + cos²ι)/2 and the factor between its ends is exactly two. Nothing in a single detector's data distinguishes the two ends. What makes this worse rather than merely awkward is that the prior pulls the other way: an isotropically oriented population has half its members beyond 60 degrees, so most binaries really are closer to edge-on, and a posterior that combines a flat likelihood along this curve with that prior returns a distance biased low with an error bar that understates the range. The whole business of standard-siren cosmology is the business of cutting across this curve.

A distance tangled with an angle

A gravitational wave carries its own distance, with no ladder underneath it and nothing to calibrate. What it also carries, inseparably, is the orientation of the orbit that made it — and one detector cannot tell a nearby binary seen edge-on from one twice as far away seen face-on.

gravitation · Gravitational waves One over a noisy parallax, at three precisions. The distribution of the distance obtained by inverting a parallax, for a star truly at 100 parsecs measured with fractional errors of 5, 10, 20 per cent. At five per cent the distribution is nearly symmetric and inverting is harmless. At twenty per cent it is strongly skewed: the mean sits at 105 parsecs rather than 100, and the tail runs to distances several times the truth, because a parallax scattered a little towards zero is a distance scattered a long way outward. The asymmetry is a Jacobian and nothing else — the parallax measurement is unbiased and symmetric throughout. Above about twenty per cent the mean of the distribution stops existing at all, because the density falls only as the inverse square of the distance and the integral of d times that diverges.

The distance is not one over the parallax

A parallax is measured with symmetric errors and a distance is one over it. Inverting a noisy positive quantity is not a change of units — it is a change of distribution, and the one that comes out is skewed, biased outward, and above about twenty per cent error has no mean at all.

starlight · Parallax The evidence for a period, one trial period at a time. The natural logarithm of the Bayes factor for a sinusoid at each trial period against no signal, for 120 radial velocities spread over 359 days with 2.0 m/s white noise per point and a planet of semi-amplitude 1.0 m/s at 11.3 days. The amplitude prior is Gaussian with scale 3.0 m/s in each component. At the planet's period the evidence is 14.4 — odds of about 1.8·10⁶ to one — and almost everywhere else it is near zero or negative, because a sinusoid that fits nothing is penalised for the amplitude range it wasted. The quantity that answers the question "is there a planet at some period" is not the peak but the average of the Bayes factor over the prior on the period, here uniform in the logarithm of the period from 1.5 to 1,000 days, and that average is ln B = 7.7: 6.7 lower than the peak, the Bayesian form of the look-elsewhere penalty, paid because the data had to find the period as well as the amplitude.

Odds for a planet that depend on a guess about its size

A false-alarm probability asks how often noise alone would produce a peak this tall. A Bayes factor asks something better — how much more probable the data are with a planet than without one — and pays for it with two numbers nobody measured. The width allowed for the period barely matters. The width allowed for the amplitude can turn strong evidence into none, from the same velocities.

starlight · Periodograms Amplitudes measured near the noise, against the amplitudes that were there. A simulated survey of 6000 stars, each with one planet whose true radial-velocity semi-amplitude K is drawn from a population rising towards small amplitudes (dN/d ln K ∝ K^−0.6, from 0.2 to 20 m/s), each measured with an amplitude uncertainty of 0.6 m/s; 1500 are drawn. Across: the true amplitude; up: the amplitude fitted at the planet's period; both logarithmic. The dashed diagonal is a perfect measurement; the horizontal line is the detection threshold, 4 times the uncertainty. Points above it are the detections (filled). Well above the threshold they sit on the diagonal; near it they sit above it — the detections within one unit of the threshold have measured amplitudes 1.33 times their true ones on average — and below it the measured amplitudes stop falling at all, because an amplitude is a length and noise alone gives it about 0.75 m/s.

A prior the survey measures through its own noise

Whether a weak periodic signal is a planet depends on how large planets are expected to be, and the best source for that expectation is the population the surveys themselves have measured. But the population's small end is built from the marginal detections whose reality depends on the prior. Breaking the circle means measuring the population and the planets together, with the noise and the threshold written into the model — and doing it that way changes the answers by factors of two where they matter most.

starlight · Periodograms A calibration sample, and the stars a precision cut keeps. Measured parallax against apparent magnitude for one star in four of a simulated sample of 8000 standard candles of absolute magnitude 0.5 and intrinsic scatter 0.15 mag, spread through a disc 15 kpc in radius and 300 pc thick and catalogued to magnitude 16. The parallax axis is stretched near zero so the negative measurements show. The solid curve is the parallax a star of exactly the calibrated luminosity would have at each magnitude, and the dashed curve is the cut ϖ/σ = 5: stars above it are the "good parallaxes" a calibration usually keeps. It keeps 3242 of the 8000, 41 per cent, and at faint magnitudes the curve it cuts through runs through the middle of the stars rather than below them, so the ones kept there are the ones whose errors pushed their parallaxes up. 45 stars have negative parallaxes; no inversion can use them, and a fit in parallax space uses them like any other.

A calibration chosen by the measurement it calibrates

A standard candle is calibrated on stars with good parallaxes, and "good" is decided by the parallax itself — so the calibrators are the stars whose errors happened to push them nearer. Correcting each star's distance afterwards does not repair it, because the correction needs a picture of where the sample's stars really lie, and that picture depends on the luminosity being calibrated. What works is to choose nothing, keep every star with the negative parallaxes included, and fit the candle by predicting the parallaxes rather than inverting them.

starlight · Parallax One likelihood, read with three priors. A Gaussian likelihood in the effective neutrino mass sum, peaking at −0.08 eV with a width of 0.05 eV — of the shape that combinations of microwave-background and baryon-acoustic data have been reported to give. Read with no prior it says the data prefer a negative sum. Read with the physical prior, sum ≥ 0 (shaded right of zero), its 95 per cent upper limit is 0.059 eV, at the normal-ordering floor. Read with the oscillation prior, sum ≥ 0.059 eV, its limit is 0.102 eV — but only 0.3 per cent of the likelihood lies above the floor at all, which is the more informative number: the data and the oscillation experiments disagree at about 2.8 standard deviations.

A bound below the floor is not a mass

Oscillation experiments prove that neutrinos weigh at least 0.059 eV between them. Cosmology's latest upper bound on their total is 0.064 eV, and a bound computed differently would sit below the floor. The usual reading is that cosmology has nearly weighed the neutrino. The better one is that the fit's likelihood peaks at a negative mass: the data see more clustering than any real neutrino allows, and the bound is a symptom of that tension rather than a measurement of anything.

cosmology · Density parameters

Named alongside it

The objects these essays reach for when they reach for this one.

Bayes factorLutz kelker biasParallaxParallax zero pointPeriodogramRadial-velocityAntenna patternBaryon acoustic oscillationsCatalogueDegeneracyDensity parameterDistance estimation

All concepts