Professional Documents
Culture Documents
Distributions Derived From The Normal Distribution: Dr. Kempthorne
Distributions Derived From The Normal Distribution: Dr. Kempthorne
MIT 18.443
Dr. Kempthorne
Spring 2015
Outline
Normal Distribution
Definition. A Normal / Gaussian random variable X ∼ N(µ, σ 2 )
has density function:
1 2 2
f (x) = √ e −(x−µ) /2σ , −∞ < x < +∞.
σ 2π
with mean and variance parameters: J
+∞
µ = E [X ] = −∞ xf (x)dx
+∞
σ 2 = E [(X − µ)2 ] = −∞ (x − µ)2 f (x)dx
J
f (µ + x ∗ ) = f (µ − x ∗ ).
f (x) is a maximum at x = µ.
f "" (x) = 0 at x = µ + σ and x = µ − σ
Chi-Square Distributions
Properties:
The density function of U is:
−1/2
fU (u) = u√2π e −u/2 , 0 < u < ∞
Recall the density of a Gamma(α, λ) distribution:
λα α−1 −λx
g (x) = Γ(α) x e , x > 0,
So U is Gamma(α, λ) with α = 1/2 and λ = 1/2.
Moment generating function
MU (t) = E [e tU ] = [1 − t/λ]−α = (1 − 2t)−1/2
Chi-Square Distributions
Definition. If Z1 , Z2 , . . . , Zn are i.i.d. N(0, 1) random variables
V = Z12 + Z22 + . . . Zn2
has a Chi-Squared distribution with n degrees of freedom.
Properties (continued)
The Chi-Square r.v. V can be expressed as:
V = U1 + U2 + · · · + Un
Student’s t Distribution
Properties
The density function of T is
−(r +1)/2
Γ[(r + 1)/2] t2
f (t) = √ 1+ , −∞ < t < +∞
r πΓ(r /2) r
For what powers k does E [T k ] converge/diverge?
Does the moment generating function for T exist?
F Distribution
Definition. For independent r.v.’s U and V where
U ∼ χ2m
V ∼ χ2n
U/m
the distribution of F = is the
V /n
F distribution with m and n degrees of freedom.
(notation F ∼ Fm,n )
Properties
The density function of F is
Γ[(m + n)/2] m n/2 m/2−1 m −(m+n)/2
f (w ) = w 1+ w ,
Γ(m/2)Γ(n/2) n n
with domain w > 0.
1 n
E [F ] = E [U/m] × E [n/V ] = 1 × n × n−2 = n−2 (for n > 2).
If T ∼ tr , then T 2 ∼ F1,r .
MIT 18.443 Distributions Derived From the Normal Distribution 7
χ2 , t, and F Distributions
Distributions Derived from Normal Random Variables
Statistics from Normal Samples
Outline
Properties of X
The moment generating function of X is
t n
MX (t) = E [e tX ] = E [e n 1 Xi ]
2
n n µ(t/n)+ σ2 (t/n)2
= 1 MXi (t/n) = i=1 [e ]
σ 2 /n
µt+ 2
t 2
= e
i.e., X ∼ N(µ, σ 2 /n).
Independence of X and S 2 .
Theorem 6.3.A The random variable X and the vector of random
variables
(X1 − X , X2 − X , . . . , Xn − X ) are independent.
Proof:
Note that X and each of Xi − X are linear combinations of
X1 , . . . , Xn , i.e., Ln
X = ai Xi = aT X = U and
L1n (k) (k) T
Xk − X = 1 bi Xi = (b ) X = Vk
where X = (X1 , . . . , Xn )
a = (a1 , . . . , an ) = (1/n, . . . , 1/n)
(k) (k) (k)
b = (b1 , . . . , bn )
1 − n1 , for i = k
.
(k)
with bi =
− n1 , for i = k
U and V1 , . . . , Vn are jointly normal r.v.s
U is uncorrelated (independent) of each Vk
MIT 18.443 Distributions Derived From the Normal Distribution 10
χ2 , t, and F Distributions
Distributions Derived from Normal Random Variables
Statistics from Normal Samples
Independence of U = aT X and V = bT X
where X = (X1 , . . . , Xn )
b = (b1 , . . . , bn )
1 − n1 , for i = k
.
(k)
with bi = bi = (so V = Vk )
− n1 , for i = 6 k
Proof: Joint MGF of (U, V ) factors into MU (s) × MV (t)
T T
MU,V (s, t) = E [e sU+tVP ] = EP [e sa X+tb X ] P
= E [e s[
P ∗
i ai Xi ]+t[ i bi Xi ] ] = E [e i [(sai +tbi )Xi ] ]
[t X ] ∗
= E [e i i i ] with ti = (sai + tbi )
n
[ti∗ Xi ]
Qn n
= i=1 E [e ] = MXi (ti∗ )
i=1
2 2 Pn
ti∗ µ+ σ2 (ti∗ )2 µ( ni=1 ti∗ )+ σ2 ∗ 2
Qn P
= e = e i=1 [(ti ) ]
i=1
2
P n 2 2 2 2 +2sta b ]
= e µ(s×1+t×0)+σ /2[ i=1 [s ai +t bi P i i
2 /n)s 2 /2 2 σ 2 /2 n 2
= [e sµ+(σ ] × [e t×0+t b
i =1 i ]
Independence of X and S 2
Proof (continued):
Independence of U = L X and each Vk = Xk − X gives
X and S 2 = n1 Vi2 are independent.
Random variables/vectors are independent if their joint
moment generating function is the product of their individual
moment generating functions:
MU,V1 ,...,Vn (s, t1 , . . . , tn ) = E [e sU+t1 V1 +···tn Vn ]
= MU (s) × M(t1 , . . . , tn ).
Theorem 6.3.B The distribution of (n − 1)S 2 /σ 2 is the
Chi-Square distribution with (n − 1) degrees of freedom.
Proof:
1 Ln 2 =
Ln Xi −µ 2 Ln
σ2 1 (Xk − µ) i=1 σ = i=1 Zi2 ∼ χn2 .
1 Ln 2 1 Ln
� �2
σ2 1 (Xk − µ) = σ 2 i=1 Xi − X + X − µ
Ln �2
= σ12 i=1 (Xi − X )2 + σn2 X − µ
�
Proof (continued):
Proof:
1 Ln 2 1 Ln 2 n
� �2
σ2 1 (Xk − µ) = σ 2 i=1 (Xi − X ) + σ2
X −µ
χn2 = [distribution of (nS 2 /σ 2 )] + χ21
By independence:
1
that is
(1 − 2t)−n/2 = MnS 2 /σ2 (t) × (1 − 2t)−1/2
So
MnS 2 /σ2 (t) = (1 − 2t)−(n−1)/2 ,
=⇒ nS 2 /σ 2 ∼ χ2n−1 .
Corollary 6.3.B For a X and S 2 from a Normal sample of size n,
X −µ
T = √ ∼ tn−1
S/ n
MIT 18.443 Distributions Derived From the Normal Distribution 13
MIT OpenCourseWare
http://ocw.mit.edu
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.