Reading the ruler along and across the line of sight
Sandbox Physics ·
A survey converts redshifts and angles into positions with an assumed cosmology. If the assumed distance along the line of sight is off by a factor q∥ and the distance across it by q⊥, the round acoustic shell appears as an ellipsoid. Fitting that ellipsoid gives two dilation parameters, α∥ and α⊥. Each is a ratio of distance over sound horizon, compared with the assumed cosmology: α∥ measures the expansion rate and α⊥ the transverse distance. This chapter derives both, shows how wedges of the correlation function separate them, and explains how to judge whether a fitted bump is a detection.
It uses the estimator and the jackknife from the previous chapter.
Two distances from one pair
Take two galaxies at nearly the same redshift z, separated by a small redshift difference Δz and a small angle Δθ. In a flat universe their comoving separations along and across the line of sight are
s∥=DH(z)Δz,s⊥=DM(z)Δθ,
where the Hubble distance and the transverse comoving distance are
DH(z)=H(z)c,DM(z)=∫0zH(z′)cdz′.
The survey measures Δz and Δθ. To get separations it must assume DH and DM. Call the assumed values DHfid and DMfid, for fiducial.
What a wrong assumption does to the shell
Define the true-over-assumed ratios q∥=DH/DHfid and q⊥=DM/DMfid. The separations the survey computes are
s∥fid=q∥s∥,s⊥fid=q⊥s⊥.
A shell of true radius rd becomes an ellipsoid with semi-axis rd/q∥ along the line of sight and rd/q⊥ across it. If the assumed distances are too short, q>1 and the shell looks too small.
The survey compares what it sees with a template built for a fiducial sound horizon rdfid. The template's ring sits at rdfid. The dilation that maps the template onto the data is
This is equation 1.1 of DESI 2024 III, written with DH=c/H. The survey never measures q alone. A survey that assumed the right distances but the wrong ruler sees exactly the same shift as one that assumed the wrong distances and the right ruler.
DESI often reports two combinations instead:
αiso=(α∥α⊥2)1/3,αAP=α⊥α∥.
The first scales an overall size and corresponds to the volume-averaged distance DV=(zDM2DH)1/3. The second measures only the shape, the Alcock-Paczynski distortion, and is independent of rd.
Separating the directions with wedges
Let μ=s∥/s be the cosine of the angle between a pair and the line of sight. A pair seen at fiducial separation s and angle μ has, in template units, the separation
st=sα∥2μ2+α⊥2(1−μ2).
The Lab splits pairs into two wedges: a transverse wedge with μ<1/2 and a line-of-sight wedge with μ≥1/2. Pairs are spread evenly in μ, so the average of μ2 over the transverse wedge is 1/12 and over the line-of-sight wedge 7/12. The ring then shows up in each wedge near
speak≈⟨μ2⟩α∥2+(1−⟨μ2⟩)α⊥2rdfid.
This is an approximation, because the average of a square root is not the square root of the average. A Lab test checks it against simulated catalogues and finds the wedge peaks within 5Mpc for three different pairs of ratios.
Worked example. Set q∥=1.10 and q⊥=1 with the fiducial ruler rd=rdfid=147.05Mpc. The line-of-sight wedge peaks near 147.05/(7/12)(1.21)+5/12=138.8Mpc, and the transverse wedge near 147.05/(1/12)(1.21)+11/12=145.8Mpc. The line-of-sight peak moves by 8 Mpc, the transverse one by 1 Mpc. The Lab's preset with these settings fits α∥=1.096±0.012 and α⊥=1.004±0.008.
The weights also explain why the Lab measures α⊥ better than α∥. The transverse wedge depends almost only on α⊥, with weight 11/12. The line-of-sight wedge mixes 7/12 of α∥ with 5/12 of α⊥. So α⊥ is constrained by both wedges and α∥ mainly by one. DESI 2024 III notes for the same reason that its transverse and radial measurements are correlated, at about 40 percent.
Fitting and judging a detection
In each wedge the Lab's model is the template, dilated and averaged over the wedge, times an amplitude B, plus a smooth broadband a0+a1/s+a2/s2. The amplitude and broadband terms enter linearly, so for every pair (α∥,α⊥) on a grid they are solved exactly. That leaves a surface χ2(α∥,α⊥). The best fit is its minimum, and the one-standard-deviation range of each parameter is where the minimum over the other parameter rises by Δχ2=1.
Two checks matter before reading a number.
First, does the interval close inside the grid? If χ2 is still within 1 of its minimum at α=0.8 or 1.2, the data do not bound the parameter. The Lab then labels the parameter unbounded.
Second, is there a bump at all? The Lab's naive detection is
Nσ=χnotemplate2−χmin2,
the improvement from adding the template to the broadband alone. It has no correction for trying many values of α. A template with a free position and amplitude can always find some upward wiggle in noise. In the No sound wave at all preset, where the catalogue holds no shells, the Lab still reports 2.3σ, with α∥ at the edge of the grid. Six more shell-free catalogues, with seeds 31 to 36, give 1.1 to 3.4σ, and three of the six reach about 3σ. So a naive 3 sigma from a fit like this is not yet a detection. The Lab calls a bounded peak a detection only above 4.5σ, and a real analysis would calibrate that threshold on many simulated catalogues.
For comparison, the default catalogue gives 15.7σ with χ2=19.2 for 43 degrees of freedom. That χ2 is low because the Lab uses only the diagonal of its jackknife covariance and neighbouring bins are positively correlated. DESI 2024 III quotes, for its 300,017 bright galaxies after reconstruction, αiso=0.9829±0.0185 with χ2=17.3 for 19 degrees of freedom, and a detection of 4.0σ.
What the Lab leaves out
Galaxies move. Their peculiar velocities add to their redshifts, which squashes structure along the line of sight on large scales and stretches it inside clusters. The Lab places galaxies at their true positions, so its anisotropy comes only from the distance ratios. Real analyses fit these redshift-space distortions with extra broadband freedom.
Gravity also moves galaxies over the history of the universe. Eisenstein, Seo and White estimate displacements of about 10Mpc, which broaden the peak, and a shift of its position at the level of 0.5 percent at first order. DESI uses reconstruction to undo much of this. The Lab's shells have a fixed width and no shift.
Lab tasks
Drag the box corner until the line-of-sight ratio reads 1.10 and the transverse ratio 1.00. Compare the two fitted stretches with the worked example.
Open the Every assumed distance too long preset. Predict both wedge peak positions, then check them on the chart.
Open the No sound wave at all preset and note the naive detection and whether the stretch is bounded.
Exercises
Show that αiso=DVfid/rdfidDV/rd when DV=(zDM2DH)1/3.
Predict both wedge peaks for q∥=q⊥=0.92 with the fiducial ruler.
From α∥=1.096±0.012 and α⊥=1.004±0.008, ignoring their correlation, compute αAP and its error.
A survey finds α∥=0.98. Is the true DH/rd at that redshift larger or smaller than in the fiducial cosmology? Is the expansion rate higher or lower, if rd is fixed?
Why does αAP not depend on rd?
AnswersShow the answer
α∥α⊥2=DHfid(DMfid)2/(rdfid)3DHDM2/rd3. Multiply top and bottom by z and take the cube root: the result is (DV/rd)/(DVfid/rdfid).
Both α equal 0.92, so the square root is 0.92 in each wedge and both peaks sit at 147.05/0.92=159.8Mpc. The ring grows without changing shape.
αAP=1.096/1.004=1.092. The relative errors add in quadrature: (0.012/1.096)2+(0.008/1.004)2=0.0135, so the error is 1.092×0.0135=0.015.
α∥<1 means DH/rd is 2 percent smaller than fiducial. With rd fixed, DH=c/H is smaller, so the expansion rate at that redshift is 2 percent higher than assumed.
In the ratio α∥/α⊥ the factor rdfid/rd appears in both and cancels. It depends only on DH/DM, the shape of the shell.