TAsession note #5
Shouto Yonekura
May 18, 2016
Contents
1 chi-squre distribution 2
2 t-distribution 5
3 F-distribution 6
4 Distribution of (n−k)sσ2 2 7
5 Distribution of qβˆi−βi V [ ˆβ]ii
7
6 Distribution of ( ˆβ−β)
′(X′X)( ˆβ−β)
s2 8
1 chi-squre distribution
def 5.1
Let Xi ∼iid N(0, 1) i = 1, 2, · · · n. Then
Z:= X12+ X22+ · · · Xn2
is called chi-squre distribution with degrees of freedom n and denoted Z ∼ χ2(n). Its p.d.f is given by below: f(x) :=nΓ(n/2)21 n/2xn/2−1e−x/2 ,0 < x < ∞
where Γ(s) :=´0∞xs−1e−xdx.
prop 5.2
Let X ∼ χ2(n). Then
MX(t) = (1 − 2t)−n/2, E[X] = n, V[X] = 2n. Proof
By the definition, MX(t) can be calculated as follows:
MX(t) =´ Γ(n/2)21 n/2xn/2−1e−x/2etxdx
= constant´ xn/2−1e−x2(1−2t)dx
= constant´ 1−2tr n/2−1e−r/21−2t1 dr (r := x(1 − 2t))
= (1−2t1 )n/2´ Γ(n/2)21 n/2rn/2−1e−r/2dr
= (1 − 2t)−n/2. Next we can derive E[X] and V [X] as follws:
MX′ (t) = −n2(1 − 2t)−n/2−1(−2) = n(1 − 2t)−n/2−1. MX′ (t) |t=0= E[X] = n(1 − 0)−n/2−1= n.
MX′′(t) = −2n(−n2− 1)(1 − 2t)−n/2−2. MX′′(t) |t=0= n(n − 2).
V[X] = (MX′ (t) |t=0)2− MX′ (t) |t=0
= n2− n2+ 2n
= 2n Q.E.D.
prop 5.3
Let x ∼ Nk(µ, Σ). Then
(x − µ)′Σ−1(x − µ) ∼ χ2(k).
Proof
Let z := Σ−1/2(x − µ). Then z ∼ Nk(0 Ik). Thus ,by the definition of chi-squre distribution, z′z∼ χ(k) and we can rewrite this as follows:
z′z= (Σ−1/2(x − µ))′(Σ−1/2(x − µ))
= (x − µ)′(Σ−1/2)′Σ−1/2(x − µ)
= (x − µ)′Σ−1/2Σ−1/2(x − µ) , (Σ′ = Σ)
= (x − µ)′Σ−1(x − µ). Q.E.D
prop 5.4
Let A be a symmetric and idempotent n×n matrix and its rank is k. Let x ∼ Nn(µ, In). Then x′Ax∼ χ2(k).
Proof
Since A is a symmetric and idempotent, there exisits n×n matrix P such that (see appendix.): P′P = P P′ = In
P′AP =
1
. .. 1
0 . ..
0
.
Let y = P x. Then we get
y∼ Nn(0, P IP′)
= Nn(0, In). Furthermore, we can rewrite it as follows:
y= P′x
⇐⇒ P y = P P′x
⇐⇒ P y = x. Using these results, we can obtain
x′Ax= (P y)′A(P y)
= y′P′AP y
=Pki=1yi2. Since each yi∼iid N(0, 1), thusPki yi2∼ χ2(k) Q.E.D.
prop 5.5
Let A be a symmertric and idempotent n×n matrix and its rank is k. Let x ∼ Nn(0, σ2In). Then
1 σ2x
′Ax∼ χ2(k). Proof
Define y := 1σx. Then E[y] = 0 and
V[y] = E[(σx) xσ′]
=σ12E[xx′]
=σ12σ2In
= In.
Thus y ∼ Nn(0, In) and y′Ay∼ χ(k) (prop5.4). Finally we obtain y′Ay = (xσ)′A(xσ)
= σ12x′Ax∼ χ2(k) Q.E.D
2 t-distribution
def 5.6
Let x ∼ N(0, 1) and y ∼ χ2(n). Then if x and y are independent, z:= √x
y/n
is called t-distribution with degrees of freedom n and denoted z ∼ t(n).
prop 5.7
Let x ∼ t(n). If n → ∞, then x →dN(0, 1) where →d means convergence in probability(we will study later calass.) .
Although we provide proof of this proposision, you do not need to understand it at present. Proof
Let z ∼ N(0, 1) and y ∼ χ2(n). If vi ∼iid χ2(1) , i = 1, 2, · · · n, then Pni=1vi ∼ χ2(n).Since E[vi] = 1 (prop5.2), we can show that E[n1Pni vi] →p1 by using the law of large number. Thuspyn →p1 (Continuous mapping theorem).From these results, we can finally obtain that √z
y/n →d N(0, 1) (Slutsky’s theorem) as required. Q.E.D.
Remark
It you need to show that some r.v. is distributed from t-distribution, you have to show following 3 things: (1) a numerator is distributed N (0, 1), (2) a denominator is distributed χ2(n), and (3) they are independent.
3 F-distribution
def 5.8
Let u ∼ χ2(n) and v ∼ χ2(m). If they are independent, then z:=v/mu/n
is called F-distribution with degrees of freedom (n,m) and denoted z ∼ F (n, m).
prop 5.9
If x ∼ t(n), then x2∼ F (1, n). Proof
Let y ∼ N(0, 1) and z ∼ χ2(n). Suppose that they are independent. Then √y
z/n ∼ t(n). Thus the distribution of x2 is as same as the distribution of z/ny2 . Since y2and z are independent and y2∼ χ1(1), z ∼ χ2(n), we can get x2=z/ny2 ∼ F (1, n). Q.E.D.
Remark
If you want to show that some r.v. has F (n, m), you you have to show following 3 things: (1) a numerator is distributed from χ2(n), (2) a denominator is distributed from χ2(m), and (3) they are independent.
4 Distribution of
(n−k)sσ2 2prop 5.10
Suppose that assumptions A1 ∼ A6 hold. Let ˆβ is the OLSE of β and e := y − X ˆβ. Then
(n−k)s2 σ2 ∼ χ
2(n − k),
where s2:= n−ke′e. Proof
First we rewrite (n−k)sσ2 2 as eσ′2e. Let MX:= In− X(X′X)−1X′. Since e′e= u′MXu, we can get:
e′e σ2 =
u′MXu σ2
=uσ′MXuσ.
Since u ∼ Nn(0, σ2In), uσ ∼ Nn(0, In). Recall that MXis symmetric and idempotent matrix. Thus uσ′MXuσ ∼ χ2(rankMX) (prop5.5). As we studied last class, if a matrix A is symmetric and idempotent, then rankA = trace(A) and rankMX is could be calculated as follows:
rnakMX= traceMX
= trace(In− X(X
′X)−1X′)
= n − k. Therefore we obtain uσ′MXuσ ∼ χ2(n − k) as required. Q.E.D.
5 Distribution of √
βˆi−βiV[ ˆβ]ii prop 5.11
Suppose that assumptions A1 ∼ A6 hold. Let ˆβ is the OLSE of β. Then
βˆi−βi
q V [ ˆβ]ii =
βˆi−βi
√(s2(X′X)−1)ii ∼ t(n − k),
where s2:= n−ke′e and V [ ˆβ]ii means (i, i) component of V [ ˆβ]. Proof
As mentined above, we have to show that (1) a numerator is distributed from N (0, 1), (2) a denominator is distributed from χ2(n − k), and (3) they are independent.
(1)
First we rewrite √ βˆi−βi
(s2(X′X)−1)ii
as follows:
βˆi−βi
√(s2(X′X)−1)ii
=√ βˆi−βi
(σ2X′X)−1)ii
qσ2 s2
=√zi
s2/σ2 ,(zi:=
βˆi−βi
√(σ2X′X)−1)ii
)
= r zi
e′ e/(n−k) σ2
= qzi
l n−k
(l := eσ′2e).
Since ˆβi∼ N(βi, σ2((X′X)−ii1), ˆβi− βi ∼ N(0, σ2((X′X)−ii1).Therefore zi=√ βˆi−βi
(σ2X′X)−1)ii ∼ N(0, 1).
(2)
l∼ χ2(n − k) is already shown (prop5.10). (3)
Since both e and ˆβ are linear functions of u which has standard normal distribution, we only need to show that Cov(e, ˆβ) = 0n.
Cov(e, ˆβ) = E[MXu((X′X)−1X′u)′] , e = MXu
= E[MXuu′X(X′X)−1]
= E[(In− X(X
′X)−1X′)uu′X(X′X)−1]
= E[uu′X(X′X)−1− X(X′X)−1X′uu′X(X′X)−1]
= E[uu′]X(X′X)−1− X(X′X)−1X′E[uu′]X(X′X)−1
= σ2X(X′X)−1− X(X′X)−1X′σ2X(X′X)−1 ,(E[uu′] = σ2In)
= σ2X(X′X)−1− σ2X(X′X)−1X′X(X′X)−1
= σ2X(X′X)−1− σ2X(X′X)−1= 0n, thus ei and ˆβi are independent.Therefore qβˆi−βi
V [ ˆβ]ii ∼ t(n − k) Q.E.D.
6 Distribution of
( ˆβ−β)′(X′X)( ˆβ−β) s2
prop 5.12
Suppose that assumptions A1 ∼ A6 hold. Let ˆβ is the OLSE of β.Then
( ˆβ−β)′(X′X)( ˆβ−β)/k
s2 ∼ F (k, n − k),
where s2:= n−ke′e . Proof
First, we rewrite ( ˆβ−β)′(Xs′2X)( ˆβ−β) as follows:
( ˆβ−β)′(X′X)( ˆβ−β)/k
s2 =
( ˆβ−β)′(X′X)( ˆβ−β)/k e′e/n−k
= ( ˆβ−β)
′(X′X)( ˆβ−β)/k
σ2 × e′ e/(n−k)1
σ2
=l/n−ku/k , u:== ( ˆβ−β)
′(X′X)( ˆβ−β) σ2 , l:=
e′e σ2.
As mentioned above, we have to show that (1) u is distributed from χ2(k), (2) l is distributed from χ2(n − k), and (3) they are independent.
(1)
Since ˆβ∼ Nk(β, σ2(X′X)−1), we get u ∼ χ2(k). (prop5.3, (Xσ′2X) corresponds to Σ−1and µ = β.) (2)
l∼ χ2(n − k) is already shown. (prop5.10) (3)
Since u is a linear fuction of ˆβ and l is a linear function of e, we only need to show that Cov(e, ˆβ) = 0n
and this is already proved in proof of prop5.11.Therefore ( ˆβ−β)
′(X′X)−1( ˆβ−β)/k
s2 ∼ F (k, n − k) Q.E.D.
Appendix
lem 5.13
Let A be a n×n symmetric matrix. Then there exisits n×n orthogonal matrix P such that
P′AP =
λ1 0 · · · 0 0 λ2 0 ... ... 0 . .. 0 0 · · · 0 λn
,
where λi are eigenvalues of A. Proof
Let pi be a eigenvectors which correspond to λi and k pik= 1, ∀i. That is, Ap1= λ1p1, Ap2= λ2p2,· · · Apn= λnpn. (#)
Let P := (p1, p2,· · · pn). Since pi and pj are orthogonal for any i 6= j, P is a n×n orthogonal matrix. Let Λ be a matrix whose diagonal components are λi. Then we can rewrite (#) as AP = P Λ. Thus we obtain P′AP = Λ. Q.E.D.
prop 5.14
Let A be a n×n symmetric and idempotent matrix. Then there exisits n×n orthogonal matrix P such that
P′AP =
1
. .. 1
0 . ..
0
.
Proof
Since A is a n×n symmetric and idempotent matrix, its eigenvalues are 1 or 0 (see note #3). Thus combining this and prop5.13 yields prop5.14 Q.E.D.