Saurav De
Department of Statistics Presidency University
Likelihood Equation of MLE
MLE and Invariance Property
Let ˆθbe MLE of θ.Then for the parametric function g(θ) : Ω→Γ; MLE is g(ˆθ).
Proof. Let us define Ωγ ={θ:g(θ) =γ}.This means Ω = [
γ∈Γ
Ωγ. Again let Mx(γ) = sup
θ∈Ωγ
Lx(θ) = Likelihood function induced byg. We are to find ˆγ at which Mx(γ) is maximised.
Now Mx(ˆγ0) = sup
θ∈Ωγˆ
0
Lx(θ)≥Lx(ˆθ)
θ∈Ωγˆ
0
where Ωγˆ0 ={θ:g(θ) = ˆγ0}.As g(ˆθ) = ˆγ0 so ˆθ ∈ Ωγˆ0 Again
Mx(ˆγ0)≤sup
ˆγ∈Γ
Mx(ˆγ) = sup
ˆγ∈Ω
sup
θ∈Ωγˆ
0
Lx(θ)
= sup
θ∈Ω
Lx(θ) =Lx(ˆθ)
Likelihood Equation of MLE
Therefore
Mx(ˆγ0) =Lx(ˆθ) = sup
γ∈Γ
Mx(γ).
Hence ˆγ0 is the MLE ofγ, i.e. g(ˆθ)(= ˆγ0) is the MLE of g(θ)(=γ).
Proved
Ex.3Let X1,X2, . . . ,Xn∼Bin (1,p) ; 0≤p ≤1 Then Vp(X) =p(1−p)(=g(p)) and pˆMLE =Xn. By invariance property MLE ofVp(X) is
g( ˆpMLE) = ˆpMLE(1−pˆMLE) =Xn(1−Xn).
Application 4. Suppose that n observations are taken on a random variable X with distribution N(µ,1),but instead of recording all the observations, one notes only whether or not the observation is less than 0.
If{X <0}occursm(<n) times, find the MLE ofµ.
Let X1,X2, . . . ,Xn∼N(µ,1).
Let θ=Pµ[X1 <0] = Φ(−µ) = 1−Φ(µ).
This means µ=−Φ−1(θ), a continuous function ofθ.
Likelihood Equation of MLE
Yi = 1 ifXi <0
= 0 ifXi ≥0 Then Y1,Y2, . . . ,Yn∼Bin (1, θ) ; 0≤θ≤1, Now the MLE ofθ is Y = 1n
n
X
i=1
Yi = #{Xni<0} = mn.(See the Application 3).
Hence by the invariance property the MLE of µis −Φ−1(mn).
Application. A commuter trip consists of first riding a subway to the bus stop and then taking a bus. The bus she should like to catch arrives uniformly over the intervalθ1 toθ2.She would like to estimate bothθ1 and θ2 so that she would have some idea about the time she should be at the bus stop (θ1) and she should have too late and wait for the next bus (θ2).
Over an 8-day period she makes certain to be at the bus stop so early not to miss the bus and records the following arrival time of the bus.
5:15 PM , 5:21 PM , 5:14 PM , 5:23 PM , 5:29 PM , 5:17 PM , 5:15 PM , 5:18 PM
Estimate θ1 andθ2.Also give the MLEs for the mean and the variability of the arrival distribution.
Likelihood Equation of MLE
Solution: LetT : arrival time of the bus. Then according to the question T ∼R(θ1, θ2).
Let T1,T2, . . . ,Tn be n independent random arrival times of the bus.
Then the likelihood of θ= (θ1, θ2) is
L(θ) = 1
(θ2−θ1)n ifθ1 ≤ti ≤θ2;i = 1,2, . . . ,n
= 0 otherwise i.e. L(θ) = 1
(θ2−θ1)n ifθ1≤t(1)<t(n)≤θ2
= 0 otherwise
where t(1)= min{t1,t2, . . . ,tn}andt(n)= max{t1,t2, . . . ,tn}.
Under θ1≤t(1) <t(n)≤θ2 , t(n)−t(1) ≤θ2−θ1. Hence L(θ1, θ2) = (θ 1
2−θ1)n ≤ (t 1
(n)−t(1))n =L(t(1),t(n)).
Thus MLE of (θ1, θ2) is (T(1),T(n)), whereT(1) andT(n) are the minimum and maximum order statistic respectively.
Now E(T) = θ1+θ2 2 andV(T) = (θ2−θ121)2 are two continuous functions of (θ1, θ2).
So MLE of Mean : T(1)+T2 (n) and of variability : (T(n)√−T(1))
12 (from invariance property of MLE)
Likelihood Equation of MLE
Computation using R :
(The minute component of the time has been considered as the data) R Code and Output :
> s a m p = c (15 ,21 ,14 ,23 ,29 ,17 ,15 ,18)# g i v e n s a m p l e
> n = l e n g t h ( s a m p )# s i z e of the s a m p l e
> n [1] 8
> MLE _ t h e t a 1 = min ( s a m p )# MLE of the p a r a m e t e r s
> MLE _ t h e t a 2 = max ( s a m p )
> MLE _ t h e t a 1 [1] 14
> MLE _ t h e t a 2 [1] 29
> MLE _ M e a n =( MLE _ t h e t a 1 + MLE _ t h e t a 2 ) / 2# MLE of m e a n
> cat ( " The m e a n a r r i v a l t i m e is : " , MLE _ Mean , " m i n u t e s a f t e r 5 pm .\ n " ) The m e a n a r r i v a l t i m e is : 2 1 . 5 m i n u t e s a f t e r 5 pm .
> MLE _ Var =( MLE _ theta2 - MLE _ t h e t a 1 ) / s q r t ( 1 2 )# MLE of v a r i a b i l i t y
> MLE _ Var [1] 4 . 3 3 0 1 2 7
Likelihood Equations and Related Discussions lx(θ) = lnLx(θ) is called log-likelihood function ofθ.
Likelihood Equation: ∂l∂θx(θ) = 0.
Any MLE is a root of the likelihood equation.
Any root may be local minimaor local maxima.
Possible verification for the root ˆθto be an MLE: ∂2∂θlx(θ)2 |θ=ˆθ <0.
Ifθ = (θ1, . . . , θs)0,the likelihood equations: ∂lx(θ)
∂θi = 0 i = 1, . . . ,s.
Possible verification for the root ˆθto be an MLE of θ: The Hessian matrix ∂2lx(θ)
∂θi∂θj |θ
i=ˆθi,θj=ˆθj
is negative definite.
Likelihood Equation of MLE
Result: Let T be a sufficient statistic for the family of distributions {fθ:θ∈Ω}.If a unique MLE of θexists, it is a (nonconstant) function of T.If a MLE ofθ exists but is not unique, one can find a MLE that is a function of T.
Proof. Since T is sufficient, from Neyman-Fisher factorisability we can write
L(θ) =fθ(x) =gθ(T(x))h(x)
for all x, all θand some h andgθ, where L(θ) is the likelihood function of θ andfθ(x) is the joint pmf or pdf of sample observations x= (x1, . . . ,xn).
IfL(θ) is maximised by a unique MLE ˆθ, naturally it will also maximise the functiongθ(T(x)).
=⇒ θˆshould be a function of T.
If MLE ofθ exists but is not unique, then∃ some MLEθ that can be expressed as a function ofT.
Proved
Likelihood Equation of MLE
Result: Under regular estimation case (i.e. the situation where all the regularity conditions of Cramer-Rao Inequality hold) if an estimator ˆθ ofθ attains the Cramer-Rao Lower Bound CRLBfor the variance, the
likelihood equation has a unique solution ˆθ that maximises the likelihood function.
Proof. Let L(θ|x) denote the likelihood function of real-valued parameter θ given the sample observationsx= (x1, . . . ,xn).Iffθ(x) denotes the joint pmf or pdf of x, from the equality condition in CR inequality we get
∂
∂θlogfθ(x) =k(θ)(ˆθ(x)−θ)
that is
∂
∂θlogL(θ|x) =k(θ)(ˆθ(x)−θ) (∗) with probability 1.
=⇒ the likelihood equation ∂θ∂ logL(θ|x) = 0 has the unique solution θ= ˆθ.
Differentiating both sides of (∗) again with respect to θwe get
∂2
∂θ2 logL(θ|x) =k0(θ)(ˆθ−θ)−k(θ).
Hence
∂2
∂θ2logL(θ|x)|θ=ˆθ=−k(θ) (∗∗)
Likelihood Equation of MLE
Now, If T is an unbiased estimator of a real-valued estimable parametric functiong(θ) which is differentiable at least once, from CR regularity conditions we directly get
Z
T(x))fθ(x)dx=g(θ) Or, differentiating both sides with respect to θ we get
Z
T(x))∂
∂θlogfθ(x)dx=g0(θ) In particular choosing g(θ) =θand noting that Eθ∂
∂θlogL(θ|X)
= 0 we get
Eθ
(T(X)−θ) ∂
∂θlogL(θ|X)
= 1.
Finally substituting
T(X)−θ= [k(θ)]−1 ∂
∂θ logL(θ|X)
(looking at (∗) as T(X)is just like θ(X) by its definition) we getˆ [k(θ)]−1Eθ
∂
∂θlogL(θ|X) 2
= 1 That is
k(θ) =Eθ ∂
∂θ logL(θ|X) 2
>0 by regularity condition Thus from (∗∗) the S.O.C. for maximising L(θ) holds. Hence proved.
Likelihood Equation of MLE
Result: Suppose ∂2∂θlx(θ)2 ≤0 ∀ θ∈Ω.Then ˆθ satisfying ∂l∂θx(θ) = 0 is the global maxima.
Proof. lx(θ) =lx(ˆθ) + (θ−θ)ˆ ∂l∂θx(θ)|θ=ˆθ+ (θ−2θ)ˆ2∂2∂θlx(θ)2 |θ=θ∗ θ∗∈(ˆθ, θ).
Note that the RHS ≤lx(ˆθ) because, in RHS the 2nd factor of 2nd term vanishes and the 2nd factor of the third term ≤0.Hence proved.
Result: Suppose
(i) ∂l∂θx(θ) = 0 iff θ= ˆθ.
(ii) ∂2∂θlx(θ)2 |θ=ˆθ <0.And
(iii) θˆis an interior point of an intervalI ⊂Ω.
Then ˆθ is the global maxima.
Proof. If possible suppose ˆθ∗ is such thatlx(ˆθ∗)>lx(ˆθ).Then there must be a local minima in between the local maxima ˆθ and ˆθ∗.This means for that minima point also ∂l∂θx(θ) = 0,which is a contradiction to the
supposition (i). Hence proved.
Likelihood Equation of MLE
Ex.1Let X1,X2, . . . ,Xn∼N(µ, σ2) independently.
Then lx(θ) = constant - X
(xi−µ)2
2σ2 −n2lnσ2.
So ∂l∂µx(θ) = 0 and ∂l∂σx(θ)2 = 0 imply ˆµ=x , σˆ2= n1X
(xi −x)2=s2. Also check that ∂∂µ∂σ2lx(θ)2|(µ,σ2)=(x,s2)= 0,
∂2lx(θ)
∂µ2 |(µ,σ2)=(x,s2)=−sn2 , ∂∂(σ∂2lx(θ)2)2|(µ,σ2)=(x,s2)=−sn4. So the Hessian matrix
H =
−n/s2 0 0 −n/2s4
is negative definite. Hence (X,S2) is the global maxima point and is the MLE of (µ, σ2).
Likelihood Equation of MLE
Aliter: Ifψ(x) =x−1−lnx then ψ0(x) = 1−1/x , ψ00(x) = 1/x2>0.
Therefore ψ(x) is minimum atx = 1 and minψ(x) = 1−1−0 = 0.Based on this result we can write
lx(ˆµ,σˆ2)−lx(µ, σ2) = X
(xi−µ)2
2σ2 +n2lnσ2−n2 −n2lns2
≥ 2σns22 − n2lnσs22 − n2 = n2h
s2
σ2 −1−lnsσ22
i≥0.
TRY YOURSELF!
M4.1. LetX1,X2, . . . ,Xn∼Lognormal(µ, σ2) independently. Then MLE of (µ, σ2) is (Y,S∗2) whereY = logX and (Y,S∗2) = (sample mean , sample variance(with divisor n)) onY.
Hint. IfX ∼Lognormal(µ, σ2) thenY = logX ∼N(µ, σ2).Now proceed as in Ex. 1.
M4.2. (continuation)If inM4.1. µ= 0, find the MLE of σ2
M4.3. Consider a random sample of sizen from Exponential (mean = β).
It is given only thatk, 0<k <n, of thesen observations are≤M, where M is a known positive number. Find the MLE of β.
Likelihood Equation of MLE
TUTORIAL DISCUSSION :
Overview to the problems from MODULE 4. . .
M4.2. IfX ∼Lognormal(0, σ2) then Y = logX ∼N(0, σ2).
Given y= (y1, . . . ,yn), the loglikelihood function ofσ2 is l(σ2) = constant −n
2σ2− 1 2σ2
n
X
i=1
yi2
Now using maxima-minima principle from F.O.C. we get σ2 = 1n
n
X
i=1
yi2= ˆσ2 (say) is the only solution of the likelihood equation.
Also ∂(σ∂22)2l(σ2)|σ2=ˆσ2 =−2ˆnσ2 <0.
Moreover ˆσ2 is an interior point of the parameterspace Ω = (0,∞).
=⇒ σˆ2 is the point of global maxima of the likelihood function i.e. the unique MLE of σ2.
Likelihood Equation of MLE
M4.3. LetX1,X2, . . . ,Xn∼Exponential (mean =β).
DefineYi = 1(0) ifXi ≤M (Xi >M).
Then Y1,Y2, . . . ,Yn∼Bin (1, θ) ; 0< θ <1, where θ=Pβ[X1 ≤M] = 1−exp
h
−Mβi
.This meansβ =−log(1−θ)M , a continuous function of θ.
Now the MLE ofθ is Y = 1n
n
X
i=1
Yi = #{Xin≤M} = kn.(See the Application 4).
Hence by the invariance property the MLE of β is− M
log(1−k
n).