Documente Academic
Documente Profesional
Documente Cultură
Keith Briggs
BT Exact
Adastral Park Polaris 134
Martlesham Heath, Suffolk IP5 3RE, UK
2003 December 29 10:02
Abstract
I compute some integrals involving erf and apply them to some problems related
to the maximum of Gaussian RVs.
1
1 Introduction
Let
1 2 2
f µ,σ 2 (x) = √ e−(x−µ) /2σ
2πσ
be the N(µ, σ 2 ) Gaussian density, and
1 x−µ
Fµ,σ 2 (x) = 1 + erf √
2 2σ
be its cumulative distribution. Here
Z x
2 2
erf (x) = √ e−u du.
π 0
2 Known integrals
Most of these are not useful but are given for reference. Some obvious conditions on
α, β etc. are not stated. The first integral will be used in section 3.
Z ∞ √ " #
2 π αδ − β γ
e−(αx+β ) erf (γx + δ ) dx = erf p
−∞ α α2 + γ2
Z ∞
2 x2 arctan (β /α)
e−α erf (β x) dx = √
0 α π
Z ∞ √
2 π
e−x erf (x) dx =
0 4
Z ∞ √
2 π
e−x erf 3 (x) dx =
0 8
√
π erf n+1 (x)
Z
2
e−x erf n (x) dx =
2(n + 1)
Z ∞
2 x2 β
x e−α erf (β x) dx = p
0 2α 2 α 2 + β 2
√
√ 2 −x 2 2 erf 2x
π x erf x + 2 e erf (x)
Z
erf 2 (x)dx = √ − √
π 2π
2
3 Pr[x1 < x2]
Let x1 ∼ N(µ1 , σ12 ) and x2 ∼ N(µ2 , σ22 ) be two independent Gaussian RVs. I would
like to calculate Pr[x1 < x2 ]. This is given by
Z ∞Z ∞
Pr[x1 < x2 ] = fµ 2 (x1 ) f µ 2 (x2 ) dx2 dx1
−∞ x1 1 ,σ1 2 ,σ2
Z ∞ " #
−(x1 − µ1 )2 x1 − µ2
1 1
= − √ exp erf √ dx1
2 2σ2 2π −∞ 2σ12 2σ2
1 µ − µ1
= 1 + erf q 2
2 2(σ12 + σ22 )
√ √
We have µ(2) = 1/ π, µ(3) = 3/(2 π), and it seems that cos(µ(4)π 3/2 ) = 23/27,
though I cannot prove this.
3
Using results from David Order statistics pp. 32-38:
µ(1, 1) = 0
µ(1, 2) = −π −1/2
µ(2, 2) = π −1/2
3
µ(1, 3) = − π −1/2
2
µ(2, 3) = 0
3 −1/2
µ(3, 3) = π
2
µ(1, 4) = −6 π −3/2 arctan (21/2 )
µ(2, 4) = ?
µ(3, 4) = ?
µ(4, 4) = −µ(1, 4)
µ(1, 5) = −µ(5, 5)
µ(2, 5) = −µ(4, 5)
µ(3, 5) = 0
5 −1/2
µ(4, 5) = π − 15π −3/2 arcsin (1/3)
4
5 −1/2 15 −3/2
µ(5, 5) = π + π arcsin (1/3)
4 2
Here µ(r, n) is the expectation value of the rth order statistic for a sample of n N(0,1)
RVs.