Isaque Dutra
Projects

AGN variability as a distance probe

A simulation that teaches the idea: if how an active galactic nucleus flickers depends on how luminous it is, then measuring the flicker gives the luminosity, and comparing it with the observed brightness gives the distance. Each light curve is fitted for its amplitude σ and timescale τ. The relations between those and luminosity are toys.

Survey
Toy relationsLuminosity, and so distance, from

The σ slope is fixed at 0.090 dex/mag. bτ = 0 makes τ independent of luminosity, as measured.

The universe that generates the data

H0 is fixed at 70 km/s/Mpc, so only the Planck 2018 and DESI DR2 values of Ωm, w0 and wa are marked, for the model selected. Changing the model or these sliders moves the distances without refitting.

ResultModel fitted to the Hubble diagram

fitting…

Scatter of μ̂ − μtrue (σ and τ)
–
Mean bias of μ̂ (σ and τ)
–
Weighted mean bias (σ and τ)
–
Median survey length / τobs
–
τ unconstrained
–
Distance scatter, σ alone
–
Distance scatter, σ and τ
–
Best constant w (wa = 0)
–
Its 1σ range
–
χ² / degrees of freedom
–
Light curve of one AGN: –
Fitted σ against luminosity (brighter AGN to the left)
Fitted rest-frame τ against luminosity (faint points: τ unconstrained)
Hubble diagram (click a point to see its light curve)
Residuals from ΛCDM with the same Ωm: μ̂ − μΛCDM

What this shows

Each simulated AGN has a light curve whose brightness wanders up and down over months to years, as a damped random walk with two parameters: σ, the long-term rms of the light curve in magnitudes, and τ, the damping time over which the wander forgets where it was. Here both are tied to the absolute magnitude M by toy relations. Each light curve is fitted for σ and τ; the two together give an estimate of M, and the difference between the observed mean magnitude m and M is the distance modulus μ, which the Hubble diagram plots against redshift.

A light curve only shows the variability it had time to show. A survey a few τ long sees less than the full variance, and cannot tell a longer τ with a larger σ from a shorter τ with a smaller one. Time dilation makes this worse with distance: an AGN at redshift z is seen to vary 1 + z times more slowly, so the observed damping time τobs = (1 + z) τ is longer and the same survey covers fewer of them. Fits of τ are known to be biased and poorly constrained unless the light curve is about ten times longer than τobs (Kozłowski 2017). Shorten the survey and watch the τ chart, the count of unconstrained fits and the residuals at high redshift.

What a short light curve does measure well is the variability on short timescales, the driving amplitude σ̂ = σ √(2/τ) of Kelly et al. (2009), for which the long-term variance is σ² = σ̂² τ / 2. That is why σ alone gives poor distances here. σ and τ together do much better only because their errors are strongly correlated and the toy declares τ known to within 0.05 dex, so the fitted τ can correct σ; with a looser τ relation the advantage shrinks and the distances become biased (see below). Set the luminosity to come from σ only to see the difference in the Hubble diagram. For the default survey and sample the σ-only distances are also biased, and their quoted errors are too small and smallest where the distance is most overestimated. The weighted binned means therefore rise with redshift to about 3.5 mag too high (χ² is about twice the degrees of freedom), and the constant-w fit runs to the edge of its scan: a precise-looking result that is wrong.

The residual chart shows whether this survey could tell the generating cosmology from ΛCDM: compare the binned means, with their error bars, against the solid curve. The default generating cosmology is close to the DESI 2025 fit with CMB and DES 5-year supernovae (w0 = −0.75, wa = −0.86).

The generating universe can be ΛCDM, wCDM or CPL (w0waCDM). A slider the model fixes is greyed out, and the marks under the others are the Planck 2018 and DESI DR2 fits for that model; their values, data combinations and sources are tabulated on the dark energy explorer. H0 stays at 70 km/s/Mpc in every model, so the fits' own H0 values are not used.

How weak the indicator is

The σ relation is shallow. Its slope of 0.09 dex per magnitude means that an error of 0.05 dex in σ (12 per cent) is an error of 0.05 / 0.09 ≈ 0.56 mag in M, and so in μ. With the default intrinsic scatter of 0.05 dex in each relation and the toy τ slope, even a perfectly measured AGN gives μ only to 0.42 mag; with the default survey the simulation gives 0.52 mag per AGN on average over samples (0.56 mag for the sample the page opens with). A type Ia supernova gives about 0.15 mag, so in this toy about (0.52 / 0.15)² ≈ 12 AGN carry the distance information of one supernova. That number belongs to the toy, not to real AGN. It rests on two invented assumptions: the τ slope, and above all a τ relation with only 0.05 dex of scatter, which lets the fitted τ correct the σ it was traded against. With 0.15 dex of τ scatter the per-AGN scatter grows to 0.78 mag and the distances come out 0.2 mag too large; with 0.3 dex, to 1.3 mag and 0.6 mag. With σ alone the default survey gives 2.9 mag. With bτ = 0, when τ says nothing about luminosity, using both still narrows the scatter (0.72 against 2.2 mag), because the toy states that every AGN's τ lies within 0.05 dex of 200 days. But those distances are biased by 0.12 mag, and the 1σ range of w contains −1 in only about 7 per cent of samples. Real AGN offer no such guarantee.

How honest the error bars are

Measured from 400 simulated samples of each survey, each with a different random seed (300 AGN, 3-day cadence, distances from σ and τ, a ΛCDM generating universe, constant w fitted), with no correction applied on this page:

  • 10 years, 0.02 mag errors (the default). Averaged over the samples, the median survey covers 3.7 observed damping times and about 5 per cent of the fits leave τ unconstrained; the sample the page opens with has 4.0 and 20 of 300. The distance moduli come out too large by 0.040 mag on average, rising from 0.013 mag at z ≈ 0.5 to 0.074 mag at z ≈ 3.5: fits of short light curves are biased, and the net effect here is that the AGN look more luminous and so farther away than they are, more so at high redshift where time dilation shortens the light curve in units of τobs. The quoted errors match the actual scatter (the standard deviation of the pulls is 1.00). The 1σ range of the constant-w fit contains w = −1 in about 62 per cent of samples (247 of 400; 68 per cent would be exact), and the best fits average −1.05: the distance bias pulls w low by about half its 1σ error.
  • 3 years, 0.02 mag errors. On average the median survey covers 1.1 damping times and 20 per cent of fits leave τ unconstrained. The bias is 0.25 mag, from 0.11 mag at z ≈ 0.5 to 0.43 mag at z ≈ 3.5, and the actual scatter is 1.14 times the quoted errors. The 1σ range of w contained the true w = −1 in only 3 of 400 samples; the best fits averaged −1.42.
  • 10 years, 0.10 mag errors (the largest). The bias is 0.30 mag, from 0.15 to 0.47 mag across the redshift range, the scatter is 1.08 times the quoted errors, and the 1σ range of w contained w = −1 in 17 of 400 samples, with best fits averaging −1.32.
  • For a single light curve 50 to 100 times longer than τobs the fit recovers σ and τ with no bias above 0.01 dex and its errors match the scatter (pull standard deviations about 1). At 10 τobs the actual scatter is about 1.15 times the quoted errors; at 3 τobs, 1.4 to 1.6 times; at 1 τobs, 2.0 to 2.3 times, and σ and τ come out low on average, by roughly 0.1 and 0.2 dex (τ about twice σ in dex, along the likelihood ridge); the exact amount depends on the sampling and on whether fits with τ on its bound are counted.

What is real and what is a toy

  • Real: AGN optical light curves are well described by a damped random walk (an Ornstein–Uhlenbeck process) with a damping time of a few hundred days in the rest frame (Kelly et al. 2009; MacLeod et al. 2010). The light curves here are sampled exactly from that process, with intervals converted to the rest frame by dividing by 1 + z. σ is the long-term rms; the asymptotic structure function is SF∞ = √2 σ. More luminous AGN vary less: MacLeod et al. (2010, Table 1) fit log SF∞ against rest wavelength, absolute i-band magnitude and black hole mass, and find a luminosity coefficient of 0.131 ± 0.008 dex per magnitude at fixed black hole mass and wavelength, and 0.090 ± 0.003 when black hole mass is left out. A relation in luminosity alone corresponds to the second; since σ is SF∞ / √2, the slope is the same for σ. Measuring τ from light curves only a few τ long is hard, as above. The distances come from a flat universe with matter and w0wa dark energy (Chevallier & Polarski 2001; Linder 2003).
  • Toy: the relations log10(σ/mag) = log10 0.141 + 0.090 (M + 23) and log10(τ/day) = log10 200 + bτ (M + 23), each with its own adjustable Gaussian scatter, are taken as exact, with known zero-points; real relations also depend on wavelength and black hole mass. The τ slope is invented for this exercise so that τ carries distance information (by default bτ = −0.08: more luminous AGN flicker more slowly). MacLeod et al. (2010) find τ nearly independent of luminosity (a coefficient of 0.03 ± 0.04 at fixed black hole mass, −0.01 ± 0.02 without it), which bτ = 0 reproduces. Redshifts are uniform between 0.2 and 4; absolute magnitudes are uniform between −28 and the brighter of −22 and the flux limit m = 23, with the limit applied using distances in ΛCDM with Ωm = 0.3, so that changing the generating cosmology moves the AGN in apparent magnitude without changing them. Observations are evenly spaced with no seasonal gaps, and the photometric error is the same for every epoch and known. There are no K-corrections, extinction, host galaxy light, emission line contamination or selection effects beyond the flux limit.
  • The fit: for each light curve, the exact Gaussian likelihood of a damped random walk plus the photometric noise, with the unknown mean magnitude integrated out against a flat prior, computed by a Kalman filter. It is maximised over ln σ and ln τobs, with τobs allowed from half the cadence to ten times the survey length, and the errors come from the curvature of the likelihood at its maximum. Where the maximum lies on a bound of τ, or the curvature is not that of a peak, τ is marked unconstrained and given an error spanning the whole allowed range; such AGN are kept, not dropped, because dropping them by their measured values would bias the sample.
  • Luminosity: with d = (log10 σ̂ − aσ, log10 τ̂ − aτ), b = (bσ, bτ) and W the inverse of the measurement covariance plus the intrinsic scatter, the generalised least squares estimate is M̂ = −23 + bᵀWd / bᵀWb, with variance 1 / bᵀWb. Then μ̂ = m − M̂, with the error of the fitted mean magnitude added.
  • The cosmology fit uses the distances from whichever indicator is selected. With wCDM it varies a constant w with wa = 0, Ωm at its generating value and the zero-points of the relations assumed known, by a χ² scan with inverse-variance weights over −3 ≤ w ≤ 1; the range is where χ² is within 1 of its minimum. With ΛCDM nothing is fitted: the χ² is that of the data against ΛCDM with the generating Ωm, with one degree of freedom per AGN. There is no w0wa fit: w0 and wa are strongly degenerate for distances alone, and a fair two-parameter fit would need its own contour plot. In a real analysis the absolute calibration is degenerate with H0 and would be fitted together with Ωm, which widens the range.

For an actual measurement of a variability-luminosity relation and the Hubble diagram it gives, with real data, see Dutra, Burke, Natarajan & Yu (2025), arXiv:2512.07931; nothing on this page reproduces that paper's method, relation or data.

References

  • Kelly, B. C., Bechtold, J. & Siemiginowska, A. 2009, ApJ, 698, 895
  • MacLeod, C. L. et al. 2010, ApJ, 721, 1014
  • Kozłowski, S. 2017, A&A, 597, A128
  • Chevallier, M. & Polarski, D. 2001, Int. J. Mod. Phys. D, 10, 213
  • Linder, E. V. 2003, Phys. Rev. Lett., 90, 091301
  • DESI Collaboration 2025, arXiv:2503.14738