Milne extrapolation distances (Hopf constant in three dimensions)

edit · history · discussion · files · long url · integral physics
Numbers
$d$ 
$z_0(d)$
1:
1
comment: The two-direction rod has extrapolation distance exactly one.
2:
0.8183098861837906715377675267450287240689192914809128974953346881177935952684530701802276055325061719
comment: Flatland: $z_0(2)=1/2+1/\pi$.
3:
0.7104460895987630727325241416991536719932013339587852390927979685109726970439157681527636399809077528
comment: The Hopf constant, also called the Hopf extrapolation length.
4:
0.6366197723675813430755350534900574481378385829618257949906693762355871905369061403604552110650123438
comment: $z_0(4)=2/\pi$.
5:
0.5819457611127071632027759976263870223499328372646899182725146087148770956248604406037107166944698701
6:
0.5393348405666930242709934543316591330145456898956963405556491652996476504750085759428212248719946783
7:
0.5049086837166863670225514214326945477083736705691002373318044538976776641271757538674142774216573603
8:
0.4763395251278849844505273018787819813997885496010104533191751090692239733572839292131450164368574561
9:
0.4521333137367962250703087426932748947941514093156869785295670367311121609173774059023068368105247797
10:
0.4312808634660110920373140953489419077445879444258369838961389544236859610289717790763566354299048368
Definition
The Milne extrapolation distance $z_0(d)$ is the positive distance for which the linear asymptote of the scalar flux in a conservative isotropically scattering half-space in $d$ spatial dimensions vanishes at depth $x=-z_0(d)$, with lengths measured in mean-free-path units [1].
Parameters
$d$
—   Spatial dimension ($d\geq1$)
Formulas
(1)
$G_d(\mu)=\dfrac{2\Gamma(d/2)}{\sqrt\pi\,\Gamma((d-1)/2)}(1-\mu^2)^{(d-3)/2}$ for $0<\mu<1$ and $d\geq2$, with $\int_0^1G_d(\mu)\,d\mu=1$ [1].
(2)
$\dfrac1{H_d(\mu)}=1-\dfrac\mu2\int_0^1\dfrac{H_d(v)G_d(v)}{\mu+v}\,dv$ for $0<\mu\leq1$ and $d\geq2$ [1].
(3)
$\alpha_n(d)=\int_0^1\mu^nH_d(\mu)G_d(\mu)\,d\mu$, $\alpha_0(d)=2$, $\alpha_1(d)=2/\sqrt d$, and $z_0(d)=\alpha_2(d)/\alpha_1(d)=\sqrt d\,\alpha_2(d)/2$ for $d\geq2$ and nonnegative integer $n$ [1].
(4)
$\lambda_d(t)={}_2F_1(1,1-d/2;1/2;t^2)$, $\theta_d(t)=\operatorname{Arg}(\lambda_d(t)+i\pi tG_d(t)/2)$, and $z_0(d)=1-\pi^{-1}\int_0^1\theta_d(t)\,dt$ for $d\geq2$. This is the conservative limit of [1], equation (191).
(5)
$z_0(1)=1$, $z_0(2)=1/2+1/\pi$, and $z_0(4)=2/\pi$ [1] [2], where $\pi$ is the circle constant.
Comments
(6)
The medium occupies $x>0$. Particles have one speed, independent exponential free paths of mean one, and uniformly distributed scattering directions. The boundary has no incoming radiation or internal reflection, and the single-scattering albedo is exactly $c=1$. The case $d=1$ uses the two directions $+1$ and $-1$. Plane-parallel symmetry reduces the depth coordinate without changing the spatial dimension. The nonzero asymptotic scalar flux is normalized to have unit depth gradient, so its linear asymptote is $x+z_0(d)$.
(7)
The three-dimensional member is the Hopf constant, or Hopf extrapolation length [3]. The two-dimensional member is the Flatland extrapolation distance [2].
(8)
For $d\geq2$, $\mu$ is the direction cosine and $G_d$ is the normalized angular weight in (1). Its normalization uses $\Gamma$. The function $H_d$ is the physical positive conservative solution of (2), obtained as the limit from albedos less than one, with $H_d(0)=1$. For nonnegative integer moment order $n$, the moments $\alpha_n(d)$ in (3) are raw weighted moments; they are not divided by $\alpha_0(d)=2$.
(9)
The dispersion function $\lambda_d$ in (4) uses the Gauss hypergeometric function. The phase is the continuous argument in $[0,\pi]$ with $\theta_d(0)=0$; an uncorrected principal arctangent gives the wrong branch when $\lambda_d$ is negative.
Programs
(P1)
Python
import mpmath as m
m.mp.dps = 150
d = 3                          # any integer dimension at least 2
k = m.sqrt(m.pi)*m.gamma(m.mpf(d)/2)/m.gamma(m.mpf(d-1)/2)
def theta(t):
    if t == 0: return m.mpf(0)
    if t == 1: return m.pi/2 if d == 2 else m.pi
    re = m.hyp2f1(1, 1-m.mpf(d)/2, m.mpf(1)/2, t*t)
    im = k*t*((1-t)*(1+t))**(m.mpf(d-3)/2)
    return m.atan2(im, re)
z0 = 1-m.quad(theta, [0, m.mpf(1)/2, m.mpf(9)/10, 1])/m.pi
print(m.nstr(z0, 100))
References
[1]
E. d'Eon and N. J. McCormick, Radiative transfer in half spaces of arbitrary dimension (2019), equations (7), (50), (57), (64)-(68), (176), (179)-(181), (184), and (191). (arXiv)
[2]
E. d'Eon and M. M. R. Williams, Isotropic Scattering in a Flatland Half-Space (2018), Section 3.1, equation (57) and its conservative limit. (arXiv)
[3]
K. Kawabata, 15-Digit Accuracy Calculations of Chandrasekhar's H-function for Isotropic Scattering by Means of the Double Exponential Formula (2016), equations (11)-(13) and Appendix A. (arXiv)
Similar tables
Values of the Gauss hypergeometric function —   holds values of the function defining the dispersion kernel $\lambda_d$ in (4)
Values of the Gamma function at rational numbers —   holds the Gamma values that normalize the angular weight $G_d$
Pi —   holds the circle constant in the exact two- and four-dimensional distance formulas
Chandrasekhar H-function moments for conservative isotropic scattering —   holds the raw angular moments $\alpha_n(d)$ of the same conservative isotropic transport model; for $d\geq2$, the second moment gives $z_0(d)=\alpha_2(d)/\alpha_1(d)=\sqrt d\,\alpha_2(d)/2$, while the higher moments describe further features of the emerging angular distribution
Data properties
Entries are of type: real number
How well the digits are known: heuristic (agreement-checked)
How they were obtained:

Computed from the conservative phase integral with mpmath tanh-sinh quadrature at 150 and 200 decimal working digits and different interval subdivisions. The largest change between those computations was less than $8\times10^{-152}$. Every inexact entry is stored to 100 significant digits. The rod value is exact.

more

Independent checks use the direct distance integral in d'Eon and McCormick, equation (176), after an angular substitution and an analytic cancellation of its small-argument terms. That integral was evaluated at 170 working digits and differed from the 200-digit phase result by less than $10^{-171}$ for every entry. Closed forms in dimensions 1, 2, 4 and 6 and published numerical values provide controls. Separate angular quadratures test the zeroth and first H-function moments before the second-moment relation is tested in every dimension from 2 through 10. The finer angular quadrature gave errors less than $2\times10^{-8}$ in the moment controls and $6\times10^{-9}$ in the distance relation. These angular checks test normalization and the moment relation; they do not justify the stored 100 digits. The attached numerical evidence records the observed discrepancies. These computations have no rigorous quadrature error bounds.

Table is complete: no (it holds every integer spatial dimension $1\leq d\leq10$, covering the rod, Flatland, physical three-space, and initial even- and odd-dimensional extensions of the same model; the upper bound is a computational choice permitting independent checks of the full range. Absorption, anisotropic scattering and other free-path laws are different models and are not included)