SlideShare a Scribd company logo
Maximum Likelihood Estimation of Beetle’s Species from Its Mass,
Length and Other Characters
LIANGKAI HU
1. Derivation of the EM steps
The likelihood of the whole data set is (suppose we know all the missing species
information):
𝐿( 𝜇, 𝜈, 𝜌, 𝛼) = ∏ 𝑃( 𝑚 𝑖, 𝑟𝑖, 𝑠𝑖| 𝑠𝑝𝑖) ∙ 𝑃(𝑠𝑝𝑖)
𝑁
𝑖=1
= ∏
1
0.08√2𝜋
∙ exp{−
( 𝑙𝑜𝑔 𝑚 𝑖 − 𝜇 𝑠)2
2 ∙ 0.082
}
𝑁
𝑖=1
∙
1
0.1 ∙ √2𝜋
∙ exp{−
(𝑙𝑜𝑔 𝑟𝑖 − 𝜈𝑠 )2
2 ∙ 0.12
} ∙ 𝜌𝑠
𝑠𝑖
(1 − 𝜌𝑠)1−𝑠𝑖 ∙ 𝛼 𝑠𝑝
𝑙( 𝜇, 𝜈, 𝜌, 𝛼) = ∑ log(
1
0.008 ∙ 2𝜋
) −
( 𝑙𝑜𝑔 𝑚 𝑖 − 𝜇 𝑠)2
2 ∙ 0.082
−
( 𝑙𝑜𝑔 𝑚 𝑖 − 𝜇 𝑠)2
2 ∙ 0.082
+ 𝑠𝑖 log 𝜌𝑠
𝑁
𝑖=1
+(1 − 𝑠𝑖)log(1 − 𝜌𝑠 ) + log 𝛼 𝑠
However, we have some observations whose species are not known, while others
whose species are not known, but genus are known. Thus we need to divide the data
into three groups U, V, W as follows:
𝑈 = { 𝑜𝑏𝑠. 𝑖: 𝑤ℎ𝑜𝑠𝑒 𝑒𝑥𝑎𝑐𝑡 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑎𝑟𝑒 𝑘𝑛𝑜𝑤𝑛}
𝑉 = { 𝑜𝑏𝑠. 𝑖: 𝑤ℎ𝑜𝑠𝑒 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑎𝑟𝑒 𝑛𝑜𝑡 𝑘𝑛𝑜𝑤𝑛 𝑏𝑢𝑡 𝑔𝑒𝑛𝑢𝑠 𝑎𝑟𝑒 𝑘𝑛𝑜𝑤𝑛}
𝑊 = { 𝑜𝑏𝑠. 𝑖: 𝑛𝑒𝑖𝑡ℎ𝑒𝑟 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑛𝑜𝑟 𝑔𝑒𝑛𝑢𝑠 𝑎𝑟𝑒 𝑘𝑛𝑜𝑤𝑛}
E step: We want to find the distribution of the missing data Z (i.e. species for some
observations) given all the known information.
We denote the probability that obs. i is actually species j as 𝑃𝑖𝑗. Note that {𝑃𝑖𝑗} is a
500𝑥10 matrix.
○1 For obs. 𝑖 in U, 𝑃𝑖𝑗 = 1 𝑓𝑜𝑟 𝑗 = 𝑡ℎ𝑒 𝑘𝑛𝑜𝑤𝑛 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑜𝑓 𝑜𝑏𝑠. 𝑖, 𝑃𝑖𝑗 = 0 𝑜. 𝑤.
○2 For obs. in V:
Note, for example, if some obs. i is of genus 1, then j can only be 1,2,3, so
𝑃𝑖1 + 𝑃𝑖2 + 𝑃𝑖3 = 1 and 𝑃𝑖𝑗 = 0 for 𝑗 = 4, …,10. Then we have
𝑃𝑖𝑗 = 𝑃{𝑍𝑖 = 𝑗|𝑚 𝑖, 𝑟𝑖, 𝑠𝑖, 𝜃( 𝑡)
} =
𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑧𝑖 = 𝑗, 𝜃( 𝑡)
) ∙ 𝑃(𝑍𝑖 = 𝑗|𝜃( 𝑡)
)
∑ 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑍𝑖 = 𝑟,𝑟∈𝐺(𝑖) 𝜃( 𝑡)
) ∙ 𝑃(𝑍𝑖 = 𝑟|𝜃( 𝑡)
)
𝐺(𝑖) means all the possible species that obs. i can be. For example, if genus of some
obs. i is known to be 1, then 𝐺( 𝑖) = {1,2,3}.
Notice that 𝑃(𝑍𝑖 = 𝑗|𝜃( 𝑡)
) = 𝛼𝑗
( 𝑡)
and
𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑧𝑖 = 𝑗, 𝜃( 𝑡)
)
=
1
0.08√2𝜋
∙ exp{−
(𝑙𝑜𝑔 𝑚 𝑖 − 𝜇𝑗
( 𝑡)
)
2
2 ∙ 0.082
} ∙
1
0.1 ∙ √2𝜋
∙ exp{−
(𝑙𝑜𝑔 𝑟𝑖 − 𝜈𝑗
( 𝑡)
)2
2 ∙ 0.12
} ∙ 𝜌𝑗
( 𝑡) 𝑠𝑖
(1 − 𝜌𝑗
( 𝑡)
)
1−𝑠𝑖
○3 For obs. in W:
Since no species information are known about these observations, so j can range from
1 to 10 for every obs.
𝑃𝑖𝑗 =
𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑍𝑖 = 𝑗, 𝜃( 𝑡)
) ∙ 𝑃(𝑍𝑖 = 𝑗|𝜃( 𝑡)
)
∑ 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑍𝑖 = 𝑟, 𝜃( 𝑡)
) ∙ 𝑃(𝑍𝑖 = 𝑟|𝜃( 𝑡)
)10
𝑟=1
Next we will find the expectation of the log likelihood w.r.t Z. Denote the log
likelihood function of obs. i as 𝑙 𝑖(𝜇 𝑠, 𝜈𝑠 , 𝜌𝑠 , 𝛼 𝑠)where s is the possible species of i.
𝐸 𝑍(𝑙|𝜃( 𝑡)
) = ∑ ∑ 𝑃𝑖𝑗 ∙ 𝑙 𝑖(𝜇 𝑗, 𝜈𝑗, 𝜌𝑗 , 𝛼𝑗)
10
𝑗=1
500
𝑖=1
M Step:
In M step, we want to find 𝜇, 𝜈, 𝜌, 𝛼 such that 𝐸 𝑍𝑖
(𝑙(𝜇, 𝜈, 𝜌, 𝛼)|𝜃( 𝑡)
) is maximized.
To achieve this, we take partial derivatives of 𝐸 𝑍𝑖
(𝑙(𝜇, 𝜈, 𝜌, 𝛼)|𝜃( 𝑡)
) w.r.t. each
parameters and set them to zero. We get:
1) For 𝜇 𝑠:
𝜕𝐸( 𝑙)
𝜕 𝜇 𝑠
= ∑ 𝑃𝑖𝑠
log 𝑚𝑖−𝜇 𝑠
0.082
500
𝑖=1
2) For 𝜈𝑠 : (Similar as above)
3) For 𝜌𝑠 :
𝜕𝐸( 𝑙)
𝜕 𝜇 𝑠
=
1
𝜌 𝑠
(∑ 𝑃𝑖𝑠 ∙ 𝑠𝑖
500
𝑖=1 ) +
1
1−𝜌 𝑠
(∑ 𝑃𝑖𝑠 ∙500
𝑖=1 (1 − 𝑠𝑖))
By solving these equations, we get the following result:
𝜇 𝑠 =
∑ 𝑃𝑖𝑠 ∙ log 𝑚𝑖𝑖
∑ 𝑃𝑖𝑠𝑖
𝜈𝑠 =
∑ 𝑃𝑖𝑠 ∙ log 𝑟𝑖𝑖
∑ 𝑃𝑖𝑠𝑖
𝜌𝑠 =
∑ 𝑃𝑖𝑠 ∙ 𝑠𝑖𝑖
∑ 𝑃𝑖𝑠𝑖
To find 𝛼 𝑠, we need to solve the following problem:
max ∑ ∑Pis ∙ log 𝛼 𝑠
500
𝑖=1
10
𝑠=1
𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 ∑ 𝛼 𝑟
10
𝑟=1
= 1
Lagrange multiplier was used to solve this optimization problem.
The Lagrange equation is:
𝐿( 𝛼, 𝜆) = ∑ ∑ Pis ∙ log 𝛼 𝑠
500
𝑖=1
10
𝑠 =1
+ 𝜆 ∙ (∑ 𝛼 𝑟
10
𝑟=1
− 1)
Set its first order derivative to zero, and we get:
𝜕𝐿
𝜕𝛼 𝑠
= ∑
𝑃𝑖𝑠
𝛼 𝑠
500
𝑖=1 = −𝜆 𝑓𝑜𝑟 𝑠 = 1,… ,10
𝜕𝐿
𝜕𝜆
= ∑ 𝛼 𝑠
10
𝑖=1
= 1
Solving the equations above, we get the final result:
𝛼 𝑠 =
∑ 𝑃𝑖𝑠
500
𝑖=1
𝑁
Finally, the log likelihood function of the observed data is:
𝑙( 𝜇, 𝜈, 𝜌, 𝛼) = ∑ log (∑ 𝐿 𝑖(𝜇𝑗, 𝜈𝑗, 𝜌𝑗 , 𝛼𝑗)
𝑗
)
500
𝑖=1
Here for each 𝑖, 𝑗 belongs to the set of all the possible species of obs. 𝑖.
2. Results and discussion
After running 120 iterations, the result become stable within a small interval.
Species mu nu rho alpha
1 0.89429230531416382 1.03426438306816149 0.265733858674903622 0.067298939652255960
2 1.38996762809172703 0.96230142865978008 0.099347396713654113 0.194401480749592737
3 1.71685888287966826 1.05462477433474833 0.155178243458189369 0.076713456308002470
4 0.53124009522079751 1.26205166839992011 0.817691807359486988 0.073477794818745140
5 1.50946001386076034 1.40867770953983351 0.918358415865703770 0.055110000451895985
6 1.72333687565397997 0.99940488257136595 0.146446373184280970 0.105822215585868865
7 1.32082385152815229 0.93800398966601317 0.429516035555412845 0.043668802805622346
8 1.05193497565973648 1.25530539642827943 0.440694262394571323 0.157577787120826457
9 1.02392291475561303 1.23667032047351566 0.460623206492330128 0.192135243325587707
10 0.90466366880025695 1.73042299432002511 0.532830415964590243 0.033794279181602327
The log likelihood is -122.12229914568172. (Please see the output page for the log
likelihood after each iteration.)
The plot of the data and the MLE of these parameters is shown in the graph below:
In this plot, ratios are plotted against masses, with the variable swamp represented by two
different signs. The vertical and horizontal lines represent the MLE estimators of 𝜇 and 𝜈,
respectively, of various species. We may tell from the graph that the MLE we have found fit the
data well, as the points are much denser around the lines.
It took 120 steps for the MLE to have 14 digits’ precision. We also want to look at the rate of
convergence of the log likelihood in order to make a thorough conclusion.
For the log likelihood at Step 110,111,112, we have:
Final log likelihood: -122.12229914568185
At Step 100: -122.12229914572499
At Step 101: -122.12229914571668
At Step 102: -122.1222991457101
(P110- final)/(P111-final)= 1.228583029
(P111- final)/(P112-final)= 1.206761391
Thus we conclude the rate of convergence of the log likelihood is approximately 1.
3. Derivation of Gibbs Sampling steps.
Step 1: Update unknown species indicators
As in Assignment 2, denote the probability that obs. i is actually species j as 𝑃𝑖𝑗. Note
that {𝑃𝑖𝑗} is a 500𝑥10 matrix.
Then the species indicator of an observation 𝑖 is a multinomial distribution, with
probabilities indicated by 𝑃𝑖𝑗, 𝑗 = 1 …10. We can use the sample function in R to
generate simulation of species indicator.
Step 2: Obtain a sample of parameters based on their posterior distribution
The posterior probability of parameters is:
𝑃(𝜇, 𝜈, 𝜌|𝑋, 𝑌, 𝑍, 𝜃( 𝑡)
) ∝ 𝑃( 𝜇) ∙ ∏ 𝑃( 𝑋𝑖| 𝜇) ∙ 𝑃( 𝜈) ∙ ∏ 𝑃( 𝑌𝑖| 𝜈) ∙ 𝑃( 𝜌) ∙ ∏ 𝑃(𝑍𝑖|𝜌𝑠 )
𝑁
𝑖=1
𝑁
𝑖=1
𝑁
𝑖=1
𝑤ℎ𝑒𝑟𝑒
𝜇 = ( 𝜇1,… , 𝜇10) 𝜈 = ( 𝜈1, … , 𝜈10) 𝜌 = (𝜌1, … , 𝜌10)
𝑋𝑖 = log 𝑚𝑎𝑠𝑠 𝑜𝑓 𝑜𝑏𝑠. 𝑖 𝑌𝑖 = log 𝑟𝑎𝑡𝑖𝑜 𝑜𝑓 𝑜𝑏𝑠. 𝑖 𝑍𝑖 = 𝑠𝑤𝑎𝑚𝑝 𝑖𝑛𝑑𝑖𝑐𝑎𝑡𝑜𝑟 𝑜𝑓 𝑜𝑏𝑠. 𝑖
Then for some 𝑠, we have
𝑃( 𝜇 𝑠|𝑋) ∝ exp(−
1
2
∙
( 𝜇 𝑠 − 1)2
22
) ∙ ∏ exp(−
1
2
∙
( 𝑋𝑖 − 𝜇 𝑠)2
0.082
)
{ 𝑖:𝑠𝑝𝑒𝑐𝑖𝑒𝑠( 𝑖)=𝑠}
We want that 𝑃( 𝜇 𝑠|𝑋) to have the form exp(−
1
2
( 𝜇 𝑠−𝑚)2
𝑠2 ). After expansion and comparison, we
get
𝑚 =
1
22 ∙ 1 +
𝑛 𝑠
0.082 ∙ 𝑋𝑠
̅̅̅
1
22 +
𝑛 𝑠
0.082
𝑠2
= (
1
22
+
𝑛 𝑠
0.082
)
−1
Thus 𝜇 𝑠|𝑋~𝑁𝑜𝑟𝑚𝑎𝑙 (𝑚, 𝑠2
).
Similarly,
𝑃( 𝜈𝑠|𝑋) ∝ exp(−
1
2
∙
( 𝜈𝑠 − 1)2
22
) ∙ ∏ exp(−
1
2
∙
( 𝑋𝑖 − 𝜈𝑠 )2
0.102
)
{ 𝑖:𝑠𝑝𝑒𝑐𝑖𝑒𝑠 ( 𝑖)=𝑠}
𝜈𝑠 |𝑋~𝑁𝑜𝑟𝑚𝑎𝑙 (𝑚, 𝑠2
)
𝑤ℎ𝑒𝑟𝑒
𝑚 =
1
22 ∙ 1 +
𝑛 𝑠
0.102 ∙ 𝑌𝑠
̅
1
22 +
𝑛 𝑠
0.102
𝑠2
= (
1
22
+
𝑛 𝑠
0.102
)
−1
Finally, we have
𝑃( 𝜌𝑠 | 𝑍) ∝
1
1 − 0
∙ ∏ 𝜌𝑠
𝑍𝑖 ∙ (1 − 𝜌𝑠)1−𝑍𝑖
{ 𝑖:𝑠𝑝𝑒𝑐𝑖𝑒𝑠( 𝑖)=𝑠}
∝ 𝜌𝑠
(∑ 𝑍𝑖 +1)−1
∙ (1 − 𝜌𝑠)(𝑛 𝑠−∑ 𝑍𝑖 +1)−1
Thus 𝜌𝑠 |𝑍~𝐵𝑒𝑡𝑎(∑ 𝑍𝑖 + 1, 𝑛 𝑠 + 1 − ∑ 𝑍𝑖)
4. Results and discussions
Initial points are chosen in the same fashion as we did in Assignment 2, i.e. we set them
to their corresponding means of data, regardless of species. The result is as below:
After looking at the graphs, we decide to set the first 20 iterations as burn-in. We get the mean and
standard deviation of the parameters:
𝜇
Gibbs sampling of all parameters
𝜈
𝜌
The followings are two plots of the simulated species indicators for observation No. 3 (genus
known as 4) and No. 8 (genus unknown).
From the plots of 𝜇 and 𝜈, we can determine that it takes only about 10 steps for Gibbs sampling
to converge to the true distribution. If we set initial points farther away from the true value, we
get results like below (in this case all 𝜇′𝑠 are set to 2, 𝜈′𝑠 to 2 and 𝜌′𝑠 to 0.5):
Simulated species indicators for obs. 3 and 8
Gibbs Sampling for parameters (initialpoints set far awayfrom true values)
We may notice that the results do not change very much. Thus we conclude that Gibbs Sampling
converges very fast.
We can judge the correlation between subsequent points by looking at the autocorrelation
function of the series of 𝜇1:
From this graph, we can tell that the autocorrelation at lag 1 and 2 are significant. This is a little
different from what we would expect from generating a series using Markov chain, as the current
state is only dependent on the last state.

More Related Content

What's hot

Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...
Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...
Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...
IJMER
 
Btech_II_ engineering mathematics_unit5
Btech_II_ engineering mathematics_unit5Btech_II_ engineering mathematics_unit5
Btech_II_ engineering mathematics_unit5
Rai University
 
BSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICSBSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICS
Rai University
 
Runge Kutta Method
Runge Kutta MethodRunge Kutta Method
Runge Kutta Method
ch macharaverriyya naidu
 
Backpropagation
BackpropagationBackpropagation
Backpropagation
강민국 강민국
 
International Refereed Journal of Engineering and Science (IRJES)
International Refereed Journal of Engineering and Science (IRJES)International Refereed Journal of Engineering and Science (IRJES)
International Refereed Journal of Engineering and Science (IRJES)irjes
 
Fismat chapter 4
Fismat chapter 4Fismat chapter 4
Fismat chapter 4
MAY NURHAYATI
 
BSC_Computer Science_Discrete Mathematics_Unit-I
BSC_Computer Science_Discrete Mathematics_Unit-IBSC_Computer Science_Discrete Mathematics_Unit-I
BSC_Computer Science_Discrete Mathematics_Unit-I
Rai University
 
Runge kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...
Runge  kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...Runge  kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...
Runge kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...
Prashant Goad
 
Btech_II_ engineering mathematics_unit4
Btech_II_ engineering mathematics_unit4Btech_II_ engineering mathematics_unit4
Btech_II_ engineering mathematics_unit4
Rai University
 
B.tech ii unit-3 material multiple integration
B.tech ii unit-3 material multiple integrationB.tech ii unit-3 material multiple integration
B.tech ii unit-3 material multiple integration
Rai University
 
Exercices calculs de_primitives
Exercices calculs de_primitivesExercices calculs de_primitives
Exercices calculs de_primitives
ZaakXO
 
Vector calculus
Vector calculusVector calculus
Vector calculus
sujathavvv
 
Z transforms
Z transformsZ transforms
Z transforms
sujathavvv
 
Study Material Numerical Solution of Odinary Differential Equations
Study Material Numerical Solution of Odinary Differential EquationsStudy Material Numerical Solution of Odinary Differential Equations
Study Material Numerical Solution of Odinary Differential Equations
Meenakshisundaram N
 
Some properties of two-fuzzy Nor med spaces
Some properties of two-fuzzy Nor med spacesSome properties of two-fuzzy Nor med spaces
Some properties of two-fuzzy Nor med spaces
IOSR Journals
 
B.tech ii unit-5 material vector integration
B.tech ii unit-5 material vector integrationB.tech ii unit-5 material vector integration
B.tech ii unit-5 material vector integration
Rai University
 
Dyadics
DyadicsDyadics
Dyadics
Solo Hermelin
 
Tenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear Equations
Tenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear EquationsTenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear Equations
Tenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear Equations
QUESTJOURNAL
 
New Information on the Generalized Euler-Tricomi Equation
New Information on the Generalized Euler-Tricomi Equation New Information on the Generalized Euler-Tricomi Equation
New Information on the Generalized Euler-Tricomi Equation
Lossian Barbosa Bacelar Miranda
 

What's hot (20)

Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...
Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...
Further Results On The Basis Of Cauchy’s Proper Bound for the Zeros of Entire...
 
Btech_II_ engineering mathematics_unit5
Btech_II_ engineering mathematics_unit5Btech_II_ engineering mathematics_unit5
Btech_II_ engineering mathematics_unit5
 
BSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICSBSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-2_DISCRETE MATHEMATICS
 
Runge Kutta Method
Runge Kutta MethodRunge Kutta Method
Runge Kutta Method
 
Backpropagation
BackpropagationBackpropagation
Backpropagation
 
International Refereed Journal of Engineering and Science (IRJES)
International Refereed Journal of Engineering and Science (IRJES)International Refereed Journal of Engineering and Science (IRJES)
International Refereed Journal of Engineering and Science (IRJES)
 
Fismat chapter 4
Fismat chapter 4Fismat chapter 4
Fismat chapter 4
 
BSC_Computer Science_Discrete Mathematics_Unit-I
BSC_Computer Science_Discrete Mathematics_Unit-IBSC_Computer Science_Discrete Mathematics_Unit-I
BSC_Computer Science_Discrete Mathematics_Unit-I
 
Runge kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...
Runge  kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...Runge  kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...
Runge kutta method -by Prof.Prashant Goad(R.C.Patel Institute of Technology,...
 
Btech_II_ engineering mathematics_unit4
Btech_II_ engineering mathematics_unit4Btech_II_ engineering mathematics_unit4
Btech_II_ engineering mathematics_unit4
 
B.tech ii unit-3 material multiple integration
B.tech ii unit-3 material multiple integrationB.tech ii unit-3 material multiple integration
B.tech ii unit-3 material multiple integration
 
Exercices calculs de_primitives
Exercices calculs de_primitivesExercices calculs de_primitives
Exercices calculs de_primitives
 
Vector calculus
Vector calculusVector calculus
Vector calculus
 
Z transforms
Z transformsZ transforms
Z transforms
 
Study Material Numerical Solution of Odinary Differential Equations
Study Material Numerical Solution of Odinary Differential EquationsStudy Material Numerical Solution of Odinary Differential Equations
Study Material Numerical Solution of Odinary Differential Equations
 
Some properties of two-fuzzy Nor med spaces
Some properties of two-fuzzy Nor med spacesSome properties of two-fuzzy Nor med spaces
Some properties of two-fuzzy Nor med spaces
 
B.tech ii unit-5 material vector integration
B.tech ii unit-5 material vector integrationB.tech ii unit-5 material vector integration
B.tech ii unit-5 material vector integration
 
Dyadics
DyadicsDyadics
Dyadics
 
Tenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear Equations
Tenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear EquationsTenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear Equations
Tenth-Order Iterative Methods withoutDerivatives forSolving Nonlinear Equations
 
New Information on the Generalized Euler-Tricomi Equation
New Information on the Generalized Euler-Tricomi Equation New Information on the Generalized Euler-Tricomi Equation
New Information on the Generalized Euler-Tricomi Equation
 

Similar to Maximum Likelihood Estimation of Beetle

Fourier series
Fourier series Fourier series
Fourier series
Santhanam Krishnan
 
Periodic Solutions for Non-Linear Systems of Integral Equations
Periodic Solutions for Non-Linear Systems of Integral EquationsPeriodic Solutions for Non-Linear Systems of Integral Equations
Periodic Solutions for Non-Linear Systems of Integral Equations
International Journal of Engineering Inventions www.ijeijournal.com
 
Functions of severable variables
Functions of severable variablesFunctions of severable variables
Functions of severable variables
Santhanam Krishnan
 
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
IJMER
 
Semana 24 funciones iv álgebra uni ccesa007
Semana 24 funciones iv álgebra uni ccesa007Semana 24 funciones iv álgebra uni ccesa007
Semana 24 funciones iv álgebra uni ccesa007
Demetrio Ccesa Rayme
 
Dual Spaces of Generalized Cesaro Sequence Space and Related Matrix Mapping
Dual Spaces of Generalized Cesaro Sequence Space and Related Matrix MappingDual Spaces of Generalized Cesaro Sequence Space and Related Matrix Mapping
Dual Spaces of Generalized Cesaro Sequence Space and Related Matrix Mapping
inventionjournals
 
A Non Local Boundary Value Problem with Integral Boundary Condition
A Non Local Boundary Value Problem with Integral Boundary ConditionA Non Local Boundary Value Problem with Integral Boundary Condition
A Non Local Boundary Value Problem with Integral Boundary Condition
IJMERJOURNAL
 
publisher in research
publisher in researchpublisher in research
publisher in research
rikaseorika
 
International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)
IJERD Editor
 
Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...
Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...
Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...
International Journal of Engineering Inventions www.ijeijournal.com
 
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
IJMER
 
Product Rules & Amp Laplacian 1
Product Rules & Amp Laplacian 1Product Rules & Amp Laplacian 1
Product Rules & Amp Laplacian 1
NumanUsama
 
Generalized Laplace - Mellin Integral Transformation
Generalized Laplace - Mellin Integral TransformationGeneralized Laplace - Mellin Integral Transformation
Generalized Laplace - Mellin Integral Transformation
IJERA Editor
 
Lecture 11 state observer-2020-typed
Lecture 11 state observer-2020-typedLecture 11 state observer-2020-typed
Lecture 11 state observer-2020-typed
cairo university
 
E0561719
E0561719E0561719
E0561719
IOSR Journals
 
Matlab lab manual
Matlab lab manualMatlab lab manual
Matlab lab manual
nmahi96
 
Matrix Transformations on Some Difference Sequence Spaces
Matrix Transformations on Some Difference Sequence SpacesMatrix Transformations on Some Difference Sequence Spaces
Matrix Transformations on Some Difference Sequence Spaces
IOSR Journals
 
BSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICSBSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICS
Rai University
 

Similar to Maximum Likelihood Estimation of Beetle (20)

Fourier series
Fourier series Fourier series
Fourier series
 
Periodic Solutions for Non-Linear Systems of Integral Equations
Periodic Solutions for Non-Linear Systems of Integral EquationsPeriodic Solutions for Non-Linear Systems of Integral Equations
Periodic Solutions for Non-Linear Systems of Integral Equations
 
Functions of severable variables
Functions of severable variablesFunctions of severable variables
Functions of severable variables
 
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
 
Semana 24 funciones iv álgebra uni ccesa007
Semana 24 funciones iv álgebra uni ccesa007Semana 24 funciones iv álgebra uni ccesa007
Semana 24 funciones iv álgebra uni ccesa007
 
Dual Spaces of Generalized Cesaro Sequence Space and Related Matrix Mapping
Dual Spaces of Generalized Cesaro Sequence Space and Related Matrix MappingDual Spaces of Generalized Cesaro Sequence Space and Related Matrix Mapping
Dual Spaces of Generalized Cesaro Sequence Space and Related Matrix Mapping
 
A Non Local Boundary Value Problem with Integral Boundary Condition
A Non Local Boundary Value Problem with Integral Boundary ConditionA Non Local Boundary Value Problem with Integral Boundary Condition
A Non Local Boundary Value Problem with Integral Boundary Condition
 
publisher in research
publisher in researchpublisher in research
publisher in research
 
International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)
 
Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...
Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...
Periodic Solutions for Nonlinear Systems of Integro-Differential Equations of...
 
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
On ranges and null spaces of a special type of operator named 𝝀 − 𝒋𝒆𝒄𝒕𝒊𝒐𝒏. – ...
 
Product Rules & Amp Laplacian 1
Product Rules & Amp Laplacian 1Product Rules & Amp Laplacian 1
Product Rules & Amp Laplacian 1
 
AJMS_480_23.pdf
AJMS_480_23.pdfAJMS_480_23.pdf
AJMS_480_23.pdf
 
lec32.ppt
lec32.pptlec32.ppt
lec32.ppt
 
Generalized Laplace - Mellin Integral Transformation
Generalized Laplace - Mellin Integral TransformationGeneralized Laplace - Mellin Integral Transformation
Generalized Laplace - Mellin Integral Transformation
 
Lecture 11 state observer-2020-typed
Lecture 11 state observer-2020-typedLecture 11 state observer-2020-typed
Lecture 11 state observer-2020-typed
 
E0561719
E0561719E0561719
E0561719
 
Matlab lab manual
Matlab lab manualMatlab lab manual
Matlab lab manual
 
Matrix Transformations on Some Difference Sequence Spaces
Matrix Transformations on Some Difference Sequence SpacesMatrix Transformations on Some Difference Sequence Spaces
Matrix Transformations on Some Difference Sequence Spaces
 
BSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICSBSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICS
BSC_COMPUTER _SCIENCE_UNIT-1_DISCRETE MATHEMATICS
 

Maximum Likelihood Estimation of Beetle

  • 1. Maximum Likelihood Estimation of Beetle’s Species from Its Mass, Length and Other Characters LIANGKAI HU 1. Derivation of the EM steps The likelihood of the whole data set is (suppose we know all the missing species information): 𝐿( 𝜇, 𝜈, 𝜌, 𝛼) = ∏ 𝑃( 𝑚 𝑖, 𝑟𝑖, 𝑠𝑖| 𝑠𝑝𝑖) ∙ 𝑃(𝑠𝑝𝑖) 𝑁 𝑖=1 = ∏ 1 0.08√2𝜋 ∙ exp{− ( 𝑙𝑜𝑔 𝑚 𝑖 − 𝜇 𝑠)2 2 ∙ 0.082 } 𝑁 𝑖=1 ∙ 1 0.1 ∙ √2𝜋 ∙ exp{− (𝑙𝑜𝑔 𝑟𝑖 − 𝜈𝑠 )2 2 ∙ 0.12 } ∙ 𝜌𝑠 𝑠𝑖 (1 − 𝜌𝑠)1−𝑠𝑖 ∙ 𝛼 𝑠𝑝 𝑙( 𝜇, 𝜈, 𝜌, 𝛼) = ∑ log( 1 0.008 ∙ 2𝜋 ) − ( 𝑙𝑜𝑔 𝑚 𝑖 − 𝜇 𝑠)2 2 ∙ 0.082 − ( 𝑙𝑜𝑔 𝑚 𝑖 − 𝜇 𝑠)2 2 ∙ 0.082 + 𝑠𝑖 log 𝜌𝑠 𝑁 𝑖=1 +(1 − 𝑠𝑖)log(1 − 𝜌𝑠 ) + log 𝛼 𝑠 However, we have some observations whose species are not known, while others whose species are not known, but genus are known. Thus we need to divide the data into three groups U, V, W as follows: 𝑈 = { 𝑜𝑏𝑠. 𝑖: 𝑤ℎ𝑜𝑠𝑒 𝑒𝑥𝑎𝑐𝑡 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑎𝑟𝑒 𝑘𝑛𝑜𝑤𝑛} 𝑉 = { 𝑜𝑏𝑠. 𝑖: 𝑤ℎ𝑜𝑠𝑒 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑎𝑟𝑒 𝑛𝑜𝑡 𝑘𝑛𝑜𝑤𝑛 𝑏𝑢𝑡 𝑔𝑒𝑛𝑢𝑠 𝑎𝑟𝑒 𝑘𝑛𝑜𝑤𝑛} 𝑊 = { 𝑜𝑏𝑠. 𝑖: 𝑛𝑒𝑖𝑡ℎ𝑒𝑟 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑛𝑜𝑟 𝑔𝑒𝑛𝑢𝑠 𝑎𝑟𝑒 𝑘𝑛𝑜𝑤𝑛} E step: We want to find the distribution of the missing data Z (i.e. species for some observations) given all the known information. We denote the probability that obs. i is actually species j as 𝑃𝑖𝑗. Note that {𝑃𝑖𝑗} is a 500𝑥10 matrix. ○1 For obs. 𝑖 in U, 𝑃𝑖𝑗 = 1 𝑓𝑜𝑟 𝑗 = 𝑡ℎ𝑒 𝑘𝑛𝑜𝑤𝑛 𝑠𝑝𝑒𝑐𝑖𝑒𝑠 𝑜𝑓 𝑜𝑏𝑠. 𝑖, 𝑃𝑖𝑗 = 0 𝑜. 𝑤. ○2 For obs. in V: Note, for example, if some obs. i is of genus 1, then j can only be 1,2,3, so 𝑃𝑖1 + 𝑃𝑖2 + 𝑃𝑖3 = 1 and 𝑃𝑖𝑗 = 0 for 𝑗 = 4, …,10. Then we have 𝑃𝑖𝑗 = 𝑃{𝑍𝑖 = 𝑗|𝑚 𝑖, 𝑟𝑖, 𝑠𝑖, 𝜃( 𝑡) } = 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑧𝑖 = 𝑗, 𝜃( 𝑡) ) ∙ 𝑃(𝑍𝑖 = 𝑗|𝜃( 𝑡) ) ∑ 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑍𝑖 = 𝑟,𝑟∈𝐺(𝑖) 𝜃( 𝑡) ) ∙ 𝑃(𝑍𝑖 = 𝑟|𝜃( 𝑡) )
  • 2. 𝐺(𝑖) means all the possible species that obs. i can be. For example, if genus of some obs. i is known to be 1, then 𝐺( 𝑖) = {1,2,3}. Notice that 𝑃(𝑍𝑖 = 𝑗|𝜃( 𝑡) ) = 𝛼𝑗 ( 𝑡) and 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑧𝑖 = 𝑗, 𝜃( 𝑡) ) = 1 0.08√2𝜋 ∙ exp{− (𝑙𝑜𝑔 𝑚 𝑖 − 𝜇𝑗 ( 𝑡) ) 2 2 ∙ 0.082 } ∙ 1 0.1 ∙ √2𝜋 ∙ exp{− (𝑙𝑜𝑔 𝑟𝑖 − 𝜈𝑗 ( 𝑡) )2 2 ∙ 0.12 } ∙ 𝜌𝑗 ( 𝑡) 𝑠𝑖 (1 − 𝜌𝑗 ( 𝑡) ) 1−𝑠𝑖 ○3 For obs. in W: Since no species information are known about these observations, so j can range from 1 to 10 for every obs. 𝑃𝑖𝑗 = 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑍𝑖 = 𝑗, 𝜃( 𝑡) ) ∙ 𝑃(𝑍𝑖 = 𝑗|𝜃( 𝑡) ) ∑ 𝑃(𝑚 𝑖, 𝑟𝑖, 𝑠𝑖|𝑍𝑖 = 𝑟, 𝜃( 𝑡) ) ∙ 𝑃(𝑍𝑖 = 𝑟|𝜃( 𝑡) )10 𝑟=1 Next we will find the expectation of the log likelihood w.r.t Z. Denote the log likelihood function of obs. i as 𝑙 𝑖(𝜇 𝑠, 𝜈𝑠 , 𝜌𝑠 , 𝛼 𝑠)where s is the possible species of i. 𝐸 𝑍(𝑙|𝜃( 𝑡) ) = ∑ ∑ 𝑃𝑖𝑗 ∙ 𝑙 𝑖(𝜇 𝑗, 𝜈𝑗, 𝜌𝑗 , 𝛼𝑗) 10 𝑗=1 500 𝑖=1 M Step: In M step, we want to find 𝜇, 𝜈, 𝜌, 𝛼 such that 𝐸 𝑍𝑖 (𝑙(𝜇, 𝜈, 𝜌, 𝛼)|𝜃( 𝑡) ) is maximized. To achieve this, we take partial derivatives of 𝐸 𝑍𝑖 (𝑙(𝜇, 𝜈, 𝜌, 𝛼)|𝜃( 𝑡) ) w.r.t. each parameters and set them to zero. We get: 1) For 𝜇 𝑠: 𝜕𝐸( 𝑙) 𝜕 𝜇 𝑠 = ∑ 𝑃𝑖𝑠 log 𝑚𝑖−𝜇 𝑠 0.082 500 𝑖=1 2) For 𝜈𝑠 : (Similar as above) 3) For 𝜌𝑠 : 𝜕𝐸( 𝑙) 𝜕 𝜇 𝑠 = 1 𝜌 𝑠 (∑ 𝑃𝑖𝑠 ∙ 𝑠𝑖 500 𝑖=1 ) + 1 1−𝜌 𝑠 (∑ 𝑃𝑖𝑠 ∙500 𝑖=1 (1 − 𝑠𝑖)) By solving these equations, we get the following result: 𝜇 𝑠 = ∑ 𝑃𝑖𝑠 ∙ log 𝑚𝑖𝑖 ∑ 𝑃𝑖𝑠𝑖
  • 3. 𝜈𝑠 = ∑ 𝑃𝑖𝑠 ∙ log 𝑟𝑖𝑖 ∑ 𝑃𝑖𝑠𝑖 𝜌𝑠 = ∑ 𝑃𝑖𝑠 ∙ 𝑠𝑖𝑖 ∑ 𝑃𝑖𝑠𝑖 To find 𝛼 𝑠, we need to solve the following problem: max ∑ ∑Pis ∙ log 𝛼 𝑠 500 𝑖=1 10 𝑠=1 𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 ∑ 𝛼 𝑟 10 𝑟=1 = 1 Lagrange multiplier was used to solve this optimization problem. The Lagrange equation is: 𝐿( 𝛼, 𝜆) = ∑ ∑ Pis ∙ log 𝛼 𝑠 500 𝑖=1 10 𝑠 =1 + 𝜆 ∙ (∑ 𝛼 𝑟 10 𝑟=1 − 1) Set its first order derivative to zero, and we get: 𝜕𝐿 𝜕𝛼 𝑠 = ∑ 𝑃𝑖𝑠 𝛼 𝑠 500 𝑖=1 = −𝜆 𝑓𝑜𝑟 𝑠 = 1,… ,10 𝜕𝐿 𝜕𝜆 = ∑ 𝛼 𝑠 10 𝑖=1 = 1 Solving the equations above, we get the final result: 𝛼 𝑠 = ∑ 𝑃𝑖𝑠 500 𝑖=1 𝑁 Finally, the log likelihood function of the observed data is: 𝑙( 𝜇, 𝜈, 𝜌, 𝛼) = ∑ log (∑ 𝐿 𝑖(𝜇𝑗, 𝜈𝑗, 𝜌𝑗 , 𝛼𝑗) 𝑗 ) 500 𝑖=1 Here for each 𝑖, 𝑗 belongs to the set of all the possible species of obs. 𝑖.
  • 4. 2. Results and discussion After running 120 iterations, the result become stable within a small interval. Species mu nu rho alpha 1 0.89429230531416382 1.03426438306816149 0.265733858674903622 0.067298939652255960 2 1.38996762809172703 0.96230142865978008 0.099347396713654113 0.194401480749592737 3 1.71685888287966826 1.05462477433474833 0.155178243458189369 0.076713456308002470 4 0.53124009522079751 1.26205166839992011 0.817691807359486988 0.073477794818745140 5 1.50946001386076034 1.40867770953983351 0.918358415865703770 0.055110000451895985 6 1.72333687565397997 0.99940488257136595 0.146446373184280970 0.105822215585868865 7 1.32082385152815229 0.93800398966601317 0.429516035555412845 0.043668802805622346 8 1.05193497565973648 1.25530539642827943 0.440694262394571323 0.157577787120826457 9 1.02392291475561303 1.23667032047351566 0.460623206492330128 0.192135243325587707 10 0.90466366880025695 1.73042299432002511 0.532830415964590243 0.033794279181602327 The log likelihood is -122.12229914568172. (Please see the output page for the log likelihood after each iteration.) The plot of the data and the MLE of these parameters is shown in the graph below:
  • 5. In this plot, ratios are plotted against masses, with the variable swamp represented by two different signs. The vertical and horizontal lines represent the MLE estimators of 𝜇 and 𝜈, respectively, of various species. We may tell from the graph that the MLE we have found fit the data well, as the points are much denser around the lines. It took 120 steps for the MLE to have 14 digits’ precision. We also want to look at the rate of convergence of the log likelihood in order to make a thorough conclusion. For the log likelihood at Step 110,111,112, we have: Final log likelihood: -122.12229914568185 At Step 100: -122.12229914572499 At Step 101: -122.12229914571668 At Step 102: -122.1222991457101 (P110- final)/(P111-final)= 1.228583029 (P111- final)/(P112-final)= 1.206761391 Thus we conclude the rate of convergence of the log likelihood is approximately 1. 3. Derivation of Gibbs Sampling steps. Step 1: Update unknown species indicators As in Assignment 2, denote the probability that obs. i is actually species j as 𝑃𝑖𝑗. Note that {𝑃𝑖𝑗} is a 500𝑥10 matrix. Then the species indicator of an observation 𝑖 is a multinomial distribution, with probabilities indicated by 𝑃𝑖𝑗, 𝑗 = 1 …10. We can use the sample function in R to generate simulation of species indicator. Step 2: Obtain a sample of parameters based on their posterior distribution The posterior probability of parameters is: 𝑃(𝜇, 𝜈, 𝜌|𝑋, 𝑌, 𝑍, 𝜃( 𝑡) ) ∝ 𝑃( 𝜇) ∙ ∏ 𝑃( 𝑋𝑖| 𝜇) ∙ 𝑃( 𝜈) ∙ ∏ 𝑃( 𝑌𝑖| 𝜈) ∙ 𝑃( 𝜌) ∙ ∏ 𝑃(𝑍𝑖|𝜌𝑠 ) 𝑁 𝑖=1 𝑁 𝑖=1 𝑁 𝑖=1 𝑤ℎ𝑒𝑟𝑒 𝜇 = ( 𝜇1,… , 𝜇10) 𝜈 = ( 𝜈1, … , 𝜈10) 𝜌 = (𝜌1, … , 𝜌10) 𝑋𝑖 = log 𝑚𝑎𝑠𝑠 𝑜𝑓 𝑜𝑏𝑠. 𝑖 𝑌𝑖 = log 𝑟𝑎𝑡𝑖𝑜 𝑜𝑓 𝑜𝑏𝑠. 𝑖 𝑍𝑖 = 𝑠𝑤𝑎𝑚𝑝 𝑖𝑛𝑑𝑖𝑐𝑎𝑡𝑜𝑟 𝑜𝑓 𝑜𝑏𝑠. 𝑖 Then for some 𝑠, we have
  • 6. 𝑃( 𝜇 𝑠|𝑋) ∝ exp(− 1 2 ∙ ( 𝜇 𝑠 − 1)2 22 ) ∙ ∏ exp(− 1 2 ∙ ( 𝑋𝑖 − 𝜇 𝑠)2 0.082 ) { 𝑖:𝑠𝑝𝑒𝑐𝑖𝑒𝑠( 𝑖)=𝑠} We want that 𝑃( 𝜇 𝑠|𝑋) to have the form exp(− 1 2 ( 𝜇 𝑠−𝑚)2 𝑠2 ). After expansion and comparison, we get 𝑚 = 1 22 ∙ 1 + 𝑛 𝑠 0.082 ∙ 𝑋𝑠 ̅̅̅ 1 22 + 𝑛 𝑠 0.082 𝑠2 = ( 1 22 + 𝑛 𝑠 0.082 ) −1 Thus 𝜇 𝑠|𝑋~𝑁𝑜𝑟𝑚𝑎𝑙 (𝑚, 𝑠2 ). Similarly, 𝑃( 𝜈𝑠|𝑋) ∝ exp(− 1 2 ∙ ( 𝜈𝑠 − 1)2 22 ) ∙ ∏ exp(− 1 2 ∙ ( 𝑋𝑖 − 𝜈𝑠 )2 0.102 ) { 𝑖:𝑠𝑝𝑒𝑐𝑖𝑒𝑠 ( 𝑖)=𝑠} 𝜈𝑠 |𝑋~𝑁𝑜𝑟𝑚𝑎𝑙 (𝑚, 𝑠2 ) 𝑤ℎ𝑒𝑟𝑒 𝑚 = 1 22 ∙ 1 + 𝑛 𝑠 0.102 ∙ 𝑌𝑠 ̅ 1 22 + 𝑛 𝑠 0.102 𝑠2 = ( 1 22 + 𝑛 𝑠 0.102 ) −1 Finally, we have 𝑃( 𝜌𝑠 | 𝑍) ∝ 1 1 − 0 ∙ ∏ 𝜌𝑠 𝑍𝑖 ∙ (1 − 𝜌𝑠)1−𝑍𝑖 { 𝑖:𝑠𝑝𝑒𝑐𝑖𝑒𝑠( 𝑖)=𝑠} ∝ 𝜌𝑠 (∑ 𝑍𝑖 +1)−1 ∙ (1 − 𝜌𝑠)(𝑛 𝑠−∑ 𝑍𝑖 +1)−1 Thus 𝜌𝑠 |𝑍~𝐵𝑒𝑡𝑎(∑ 𝑍𝑖 + 1, 𝑛 𝑠 + 1 − ∑ 𝑍𝑖) 4. Results and discussions Initial points are chosen in the same fashion as we did in Assignment 2, i.e. we set them to their corresponding means of data, regardless of species. The result is as below:
  • 7. After looking at the graphs, we decide to set the first 20 iterations as burn-in. We get the mean and standard deviation of the parameters: 𝜇 Gibbs sampling of all parameters
  • 8. 𝜈 𝜌 The followings are two plots of the simulated species indicators for observation No. 3 (genus known as 4) and No. 8 (genus unknown). From the plots of 𝜇 and 𝜈, we can determine that it takes only about 10 steps for Gibbs sampling to converge to the true distribution. If we set initial points farther away from the true value, we get results like below (in this case all 𝜇′𝑠 are set to 2, 𝜈′𝑠 to 2 and 𝜌′𝑠 to 0.5): Simulated species indicators for obs. 3 and 8 Gibbs Sampling for parameters (initialpoints set far awayfrom true values)
  • 9. We may notice that the results do not change very much. Thus we conclude that Gibbs Sampling converges very fast. We can judge the correlation between subsequent points by looking at the autocorrelation function of the series of 𝜇1:
  • 10. From this graph, we can tell that the autocorrelation at lag 1 and 2 are significant. This is a little different from what we would expect from generating a series using Markov chain, as the current state is only dependent on the last state.