Seminar IV - GMM

Krenar Avdulaj
October 27th, 2014
GMM estimation was formalized by Hansen (1982), and since has become one of the most widely used
methods of estimation for models in economics and finance.
1. Unlike MLE, GMM does not require complete knowledge of the distribution of the data. Only specified
moments derived from an underlying model are needed for GMM estimation.
2. In some cases in which the distribution of the data is known, MLE can be computationally very
burdensome whereas GMM can be computationally very easy. (e.g. log-normal stochastic volatility
model.)
3. In models for which there are more moment conditions than model parameters, GMM estimation
provides a straightforward way to test the specification of the proposed model.

Single Equation Linear GMM

Short review
Consider the linear regression

yt = zt

t = 1,

,n

where zt is a Lx1 vector of explanatory variables, 0 a vector of unknown coefficients and t is a random
error term. Some of zt elements are possibly correlated with t (possibly being endogenous variable). In
addition assume xt is a vector of instrumental variables of size Kx1. Let wt represent the vector of unique
and non-constant elements of {yt , zt , xt } .

Basic idea
GMM estimator of in yt

= zt

E[gt (wt ,

exploits the orthogonality condition

)] = E[xt

] = E[x (y z

)] = 0

The idea is to create a set of equations for by making sample moments match the population moments.
Sample moments:

gn ( ) =

g(wt ,

) =

t=1

1
n

n
t=1

x1t (y

1
n

Population moment E[xt t ]

xt (y

z ))

t=1

n
t=1

xK t (y

z )

K set of equations with L unknowns

z )

= 0

Sample moment=population moment

xt yt

n
t=1

xt zt

=0

t=1

or
Sxy

Necessary condition for identification of is K

S =0

xz

L.

If is just identified i.e. (K=L) and Sxz is invertible the GMM estimator of is

= S

xz

Sxy

This case is also known as indirect least squares.

If K

> L

there may not be a solution for (). Thus we need to find that makes () as close as possible

to 0 . Denote
W K xK symmetric and positive definite weight matrix such that
W
n

Wsym.p.d.

as

Then GMM estimator of ,

(
W ) is defined as

( W ) = argmin

J ( , W)

where

J( ,
W ) = ngn ( )
W gn ( )

= n(Sxy

S )
xz

W (Sxy

S )
xz

Solving for we get

1

( W ) = (Sxz
W Sxz )
Sxz W Sxy

J-statistics
Value of the GMM objective function evaluated using an efficient GMM estimator.
J = J(

where
(S

(S

), S

) = ngn (

(S

)) S

gn (

(S

))

represents any efficient GMM estimator of and

S a consistent estimate of S .
K = L
K > L

J = 0
J > 0

(K

L)

as

If the model is mis-specified or some of the moment conditions do not hold e.g.
E[xit

] = E[x
t

it

(yt

)]

for some i, the J-statistics will be large relative to 2 random variable with K-L d.o.f.
A large J-statistics indicates a mis-specification. It does not, however, indicate about the source of the
mis-specification.

Examples

1. Linear regression model by GMM

Let us take the classical case

yt = xt

where xt = (x1t ,
, xmt ) is a vector of explanatory variables (all exogenous), 0 is a m -vector of unkown
coefficients and t a random error term.

The population orthogonality condition is E[gt (wt , 0 )]

where gt (wt , )

= xt

= E[xt

] = E[x (y x

)] = 0

. The sample analogue moment conditions would be

t

[xi (yi

x )]

= 0

i=1

xi yi =

i=1

[n

xi x

i=1

[n

xi x

i=1

= (X X )

GMM

xi yi

i=1

X Y

OLS

As an example of a simple linear regression model, consider the Capital Asset Pricing Model (CAPM)
Rt

ft

for

+ (R

Mt

t = 1,

ft

) +

,n

R Excercise CAPM
Note: The code below is for exercise purposes only! In case you need to do some research on CAPM it is
advised to get a more precise risk free rate e.g. for the US from Kenneth R. French
(http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html) website, FRED
(http://research.stlouisfed.org/fred2/categories/116) or some other trusted source. You should also
consider the time span of you dataset according to your research objectives.
Below, the S&P500 returns serve as market return proxy while Chicago Board Options Exchange (CBOE)
10y interest rate T-note as risk free rate (it is easy to obtain the data by using R command line). We
estimate using the CAPM model for Intel Corporation. You need to have internet connection to be able to
run this example! However, you can connect only once and download/save the data and then call them
locally anytime.
Load the tseries, zoo, lmtest and gmm package
rm(list=ls()) # clear the memory
library(gmm)

## Loading required package: sandwich

library(tseries)
library(zoo)

##
## Attaching package: 'zoo'
##
## The following objects are masked from 'package:base':
##
##

as.Date, as.Date.numeric

library(lmtest)
# get prices
SP500 = get.hist.quote(instrument = "^gspc", start = "1992-01-31",end="2005-12-31", quot
e="AdjClose",compression = "m")
INTC = get.hist.quote(instrument = "intc", start = "1992-01-31",end="2005-12-31", quote="A
djClose",compression = "m")
p <- cbind(SP500,INTC)
colnames(p) <- c("SP500","INTC") # rename column names
ret=diff(log(p)) # estimate continuous returns

Let us plot the data and see how the time series look like. In adition we also create the excess returns for
SP500 and INTC.
par(mfcol=c(2, 2)) # create a a subplot 2x2
plot(p$SP500,main="Price of SP500",ylab="price",xlab="")
plot(ret$SP500,main="Returns of SP500",ylab="return",xlab="")
plot(p$INTC,main="Price of Intel Corporation",ylab="price",xlab="")
plot(ret$INTC,main="Returns of Intel Corporation",ylab="return",xlab="")

# Chicago Board Options Exchange (CBOE) 10y interest rate T-note

rf <- get.hist.quote(instrument = "^tnx", start = "1992-02-01",end="2005-12-31", quote="Ad
jClose",compression = "m")

## time series starts 1992-02-03

## time series ends

2005-12-01

rf <- (1+rf/100)^(1/12)-1 # transform to monthly returns

par(mfcol=c(1, 1)) # reset graph for only 1 plot
plot(rf,main="CBOE 10y interest rate T-Note",ylab="risk free rate",xlab="")

ret <- merge(ret,rf,(ret$SP500-rf),(ret$INTC-rf)) # add excess return to data

ret <- na.omit(ret)
colnames(ret)[3:5] <- c("rf","exRetSP500","exRetINTC")

The purpose of this example is to estimate the CAPM model in three different ways (OLS, MLE and GMM)
and show that the results are the same i.e. (indeed) OLS and MLE are special cases of the GMM.

a. OLS estimation
This is straight forward using built in function lm (I am not going to code the OLS because you have
already done it in previous seminars.)
ols.model <- lm(ret$exRetINTC~ret$exRetSP500,data=ret)
summary(ols.model)

##
## Call:
## lm(formula = ret$exRetINTC ~ ret$exRetSP500, data = ret)
##
## Residuals:
##

Min

1Q

Median

3Q

Max

## -0.49270 -0.06578 -0.00390 0.07606 0.27573

##
## Coefficients:
##

Estimate Std. Error t value Pr(>|t|)

## (Intercept)

0.007363

0.008150

0.903

## ret$exRetSP500 1.810853

0.202574

8.939 7.59e-16 ***

0.368

## --## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.1052 on 165 degrees of freedom
## Multiple R-squared: 0.3263, Adjusted R-squared: 0.3222
## F-statistic: 79.91 on 1 and 165 DF, p-value: 7.586e-16

coeftest(ols.model, df=Inf,vcov = NeweyWest(ols.model,lag=4,prewhite=FALSE))

##
## z test of coefficients:
##
##
## (Intercept)

Estimate Std. Error z value Pr(>|z|)

0.0073635 0.0076591 0.9614

0.3363

## ret$exRetSP500 1.8108526 0.2296546 7.8851 3.143e-15 ***

## --## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

b. MLE estimation
# extract only the data from returns series
exRetSP500 <- coredata(ret$exRetSP500)
exRetINTC <- coredata(ret$exRetINTC)
data=cbind(exRetINTC,exRetSP500)

We create the objective function which should be in the form of -LL and use the R optimizer optim.
Assuming the error term

N (0,

we incorporate its log-likelihood

+ log )

(log 2
2

i=1

2
i

in the model (do not forget to take the negative of LL when you write the R function because we will use the
general optimizer optim and not the maxLik function). One of the ways how to do it is:

# MLE estimation
LL <- function(param,data=data){
y=data[,1]
x=cbind(1,data[,-1]) # # add the intercept and remove y (1st data col)
beta <- param[-1] # exclude the first
sigma2 <- param[1]
if(sigma2<=0) return(NA)
epsilon=y-x%*%beta # calculate residuals
# log-likelihood of errors
logLik=-0.5*(log(2*pi)+log(sigma2)+(epsilon)^2/sigma2)
-sum(logLik)
}
# The maxLik version would be (uncomment to try that you get the same result):
# library(maxLik)
# LL1 <- function(param,data=data){
#

y=data[,1]

x=cbind(1,data[,-1])

beta <- param[-1]

sigma2 <- param[1]

if(sigma2<=0) return(NA)

epsilon=y-x%*%beta # calculate residuals

# log-likelihood of errors

logLik=-0.5*(log(2*pi)+log(sigma2)+(epsilon)^2/sigma2)

#}

Let us estimate the CAPM using MLE

theta.start = c(0.007,0, 1)
MLE <- optim(theta.start,LL,gr=NULL,data,method="L-BFGS-B",hessian=TRUE)
mle.param <- as.matrix(MLE$par)
fish <- MLE$hessian
stdErr <- sqrt(diag(solve(fish)))
tStat <- mle.param/stdErr
mle.model <- cbind(mle.param,stdErr,tStat)
rownames(mle.model) <- c("sigma2","alpha","beta")
colnames(mle.model) <- c("Estimates","Std. errors","t.Stat")
mle.model

##

Estimates Std. errors

t.Stat

## sigma2 0.010987984 0.001178909 9.3204710

## alpha 0.007373047 0.008123249 0.9076476
## beta

1.810627878 0.201905727 8.9676896

# The maxLik version would be (uncomment the lines below to try maxLik)
# MLE <- maxLik(LL1,start=theta.start,data=data,method="BFGS")
# coef(MLE)

Note: if your initial guess for the parameters is too far off then things can go seriously wrong! This applies
especially when objective function is (almost) flat or in boundary solutions.

c. GMM estimation

c. GMM estimation
Moment conditions for linear regression model (introduced in section 1) can be written as follows.
ols.moments = function(param,data=NULL) {
data = as.matrix(data)
y=data[,1]
x=cbind(1,data[,-1]) # add the intercept and remove y (1st data col)
x*as.vector(y - x%*%param)
}

Let us estimate the model using gmm.

start.vals=c(0,1)
names(start.vals) <- c("alpha","beta")
gmm.model=gmm(ols.moments,data,t0=start.vals,vcov="HAC")
summary(gmm.model)

##
## Call:
## gmm(g = ols.moments, x = data, t0 = start.vals, vcov = "HAC")
##
##
## Method: twoStep
##
## Kernel: Quadratic Spectral
##
## Coefficients:
##

Estimate

Std. Error t value

Pr(>|t|)

## alpha 7.3635e-03 8.1270e-03 9.0605e-01 3.6491e-01

## beta

1.8111e+00 2.2639e-01 7.9998e+00 1.2459e-15

##
## J-Test: degrees of freedom is 0
##

J-test

P-value

## Test E(g)=0:

4.87416130913001e-11 *******

##
## #############
## Information related to the numerical optimization
## Convergence code = 0
## Function eval. = 77
## Gradian eval. = NA

print(specTest(gmm.model))

##
## ## J-Test: degrees of freedom is 0 ##
##
##

J-test

P-value

## Test E(g)=0:

4.87416130913001e-11 *******

Let us graphically check whether the estimates from different models are the same.
plot(exRetSP500,exRetINTC,main="Comparison of OLS, MLE and GMM")
abline(ols.model,col="blue")
abline(a=mle.param[2],b=mle.param[3],col="green")
abline(gmm.model,col="red")
legend('topleft',c("OLS","MLE","GMM"),lty=c(1,1,1),lwd=c(2.5,2.5,2.5), col=c("blue","gree
n","red"))

Indeed, as expected, the fitted lines overlap (we see only the last one, the red colour of the GMM).

Non-linear GMM estimation

1. MA(1) model
Yt =

+ + ,
iid(0, ),
(, , )
t

t = 1,

,n

| | < 1

Some of MA(1) population moments conditions we can use are:

] =
] =
] =

E[Yt ] =
2

E[Yt

E[Yt Yt1
E[Yt Yt2

(1 +
+ =
+
+

) =
+

where k is the autocovariance of lag k (when k=0 we get the variance. Autocorrelation of lag k is obtained

. What is maximum autocorrelation you can get for MA(1) process?). Notice that we have 4
k

moment conditions and 3 unknowns (K>L), thus our model is over-identified.

The parameters we will estimate 0

, ,

= (

. Let wt

= (yt , yt , yt yt1 , yt yt2 )

The moment vector then is

g(wt ,

which should satisfy E[g(wt , 0 )]

= 0

y y
y y

yt

) =

yt
2

(1 +

at the solution 0 .

The sample moment conditions on the other hand are

gn ( ) =

1
n

g(wt ,

) =

t=3

yt

t=3

yt

yt yt1

yt yt2

t=3
n
t=3
n
t=3

(1 +

Note: our sample now has size n-2 due to the 4th moment condition (time index t-2).
Since the moment conditions K=4 are greater than the number of model parameters L=3 0 is
overidentified and the GMM objective function has the form

J ( ) = (n

where
S is a consistent estimate of S

2) g ()
n

( ))
= avar(g
n

gn ( )

Let us write a function for population moment conditions.

# function to compute four moments from MA(1) model
# y(t) = mu + e(t) + psi*e(t-1)
# e(t) ~ iid (0,sig2)
ma1.moments <- function(parm,data=NULL) {
# parm = (mu,psi,sig2)'
# data = (y(t),y(t)^2,y(t)*y(t-1),y(t)*y(t-2)) is assumed to be a matrix
m1 = parm[1]
m2 = parm[1]^2 + parm[3]*(1 + parm[2]^2)
m3 = parm[1]^2 + parm[3]*parm[2]
m4 = parm[1]^2
t(t(data) - c(m1,m2,m3,m4))
}

Simulate data from MA(1) model.

# simulate from MA(1) using arima.sim

set.seed(345)
ma1.sim = arima.sim(model=list(ma=0.6),n=500)
par(mfrow=c(3,1))
plot(ma1.sim,main="Simulated MA(1) Data")
abline(h=0)
tmp = acf(ma1.sim,plot=F)
tmp2 = acf(ma1.sim,type="partial",plot=F)
plot(tmp,main="SACF")
plot(tmp2,main="SPACF")

par(mfrow=c(1,1))
summary(ma1.sim)

##

Min. 1st Qu.

Median

Mean 3rd Qu.

Max.

## -3.14300 -0.87890 -0.07527 -0.05805 0.79450 3.04200

# check moment function

# data = (y(t),y(t)^2,y(t)*y(t-1),y(t)*y(t-2)) is assumed to be a matrix
nobs = length(ma1.sim)
ma1.data = cbind(ma1.sim[3:nobs],ma1.sim[3:nobs]^2,
ma1.sim[3:nobs]*ma1.sim[2:(nobs-1)],ma1.sim[3:nobs]*ma1.sim[1:(nobs-2)])
start.vals = c(0,0.6,1)
names(start.vals) = c("mu","psi","sig2")
ma1.mom = ma1.moments(parm=start.vals,data=ma1.data)
head(ma1.mom)

##

[,1]

[,2]

[,3]

[,4]

## [1,] -0.3874713 -1.20986599 -0.4724574 0.29078143

## [2,] -0.2418895 -1.30148946 -0.5062747 0.07962193
## [3,] -0.6740394 -0.90567094 -0.4369569 0.26117091
## [4,] -1.3078362 0.35043556 0.2815331 0.31635188
## [5,] 1.1541366 -0.02796864 -2.1094217 -0.77793351
## [6,] 2.6812313 5.82900105 2.4945072 -3.50661133

Let us check the sample moment conditions mean. It should be close to population moments, i.e.0.
colMeans(ma1.mom)

## [1] -0.056110194 0.008832286 -0.050796758 -0.134745903

# check scaling, correlation and autocorrelations of moments at true parameters

var(ma1.mom)

##

[,1]

[,2]

[,3]

[,4]

## [1,] 1.36843179 -0.2253970 -0.1119853 0.06846235

## [2,] -0.22539703 3.0823141 1.4099666 -0.37816954
## [3,] -0.11198534 1.4099666 2.0915223 0.72412086
## [4,] 0.06846235 -0.3781695 0.7241209 2.02030788

cor(ma1.mom)

##

[,1]

[,2]

[,3]

[,4]

## [1,] 1.00000000 -0.1097484 -0.06619396 0.04117479

## [2,] -0.10974839 1.0000000 0.55531467 -0.15154420
## [3,] -0.06619396 0.5553147 1.00000000 0.35226625
## [4,] 0.04117479 -0.1515442 0.35226625 1.00000000

tmp = acf(ma1.mom)

Estimate the simulated data using GMM. We should use a HAC (heteroskedasticity and autocorrelation
consistent) estimator because the MA(1) process is autocorrelated (1

(1+

=
)

1+

start.vals = c(0,0.5,1)
names(start.vals) = c("mu","psi","sigma2")
# estimate using truncated kernel with bandwith = 1
ma1.gmm = gmm(ma1.moments,ma1.data,t0=start.vals,vcov="HAC",kernel="Truncated" )
summary(ma1.gmm)

##
## Call:
## gmm(g = ma1.moments, x = ma1.data, t0 = start.vals, vcov = "HAC",
##

kernel = "Truncated")

##
##
## Method: twoStep
##
## Kernel: Truncated(with bw = 2.44423 )
##
## Coefficients:
##

Estimate

## mu

-4.8682e-02

Std. Error

t value

Pr(>|t|)

6.6346e-02 -7.3376e-01

4.6310e-01

## psi

6.1822e-01

5.7027e-02

1.0841e+01

2.2051e-27

## sigma2

9.7198e-01

5.1216e-02

1.8978e+01

2.6013e-80

##
## J-Test: degrees of freedom is 1
##

J-test

P-value

## Test E(g)=0:

3.968984 0.046346

##
## Initial values of the coefficients
##

mu

psi

sigma2

## -0.04405648 0.50078659 1.09281877

##
## #############
## Information related to the numerical optimization
## Convergence code = 0
## Function eval. = 100
## Gradian eval. = NA

print(specTest(ma1.gmm))

##
## ## J-Test: degrees of freedom is 1 ##
##
##

J-test

P-value

## Test E(g)=0:

3.968984 0.046346

The GMM estimates are close to the parameters of simulated data. The low J statistics indicates the model
is correctly specified.

2. Normal distribution GMM estimation (bonus example

This example is from gmm vignette, which you can access from here (http://cran.rproject.org/web/packages/gmm/vignettes/gmm_with_R.pdf).
The ML estimators of the mean and the variance of a normal distribution are more efficient because the
likelihood carries more information than few moment conditions. For two parameters of a normal
distribution (, 2 ) the vector of moments condition is

E[g( , xi )]

E (x )

x ( + 3
2

= 0

# vector of moment conditions

g1 <- function(tet,x)
{
m1 <- (tet[1]-x)
m2 <- (tet[2]^2 - (x - tet[1])^2)
m3 <- x^3-tet[1]*(tet[1]^2+3*tet[2]^2)
f <- cbind(m1,m2,m3)
return(f)
}

If we provide the gradient of moment conditions to the gmm function it will be used for computing the
covariance matrix of

. Derivative of moment conditions wrt to vector of parameters theta is

g ()
G

2(x

3(

Dg <- function(tet,x)
{
G <- matrix(c( 1,
2*(-tet[1]+mean(x)),
-3*tet[1]^2-3*tet[2]^2,0,
2*tet[2],-6*tet[1]*tet[2]),
nrow=3,ncol=2)
return(G)
}

Generate normal distributed random numbers

set.seed(123)
n<-200
x1<-rnorm(n,mean=4,sd=2)

estimate distributiom parameters using GMM package

print(res<-gmm(g1,x1,c(mu=0,sig=0),grad=Dg))

6
2

## Method
## twoStep
##
## Objective function value: 0.01287054
##
##

mu

sig

## 3.8762 1.7887
##
## Convergence code = 0

# print summary of results

print(summary(res))

##
## Call:
## gmm(g = g1, x = x1, t0 = c(mu = 0, sig = 0), gradv = Dg)
##
##
## Method: twoStep
##
## Kernel: Quadratic Spectral(with bw = 1.62663 )
##
## Coefficients:
##

Estimate

Std. Error

t value

Pr(>|t|)

## mu

3.8762e+00

1.2143e-01

3.1922e+01 1.3309e-223

## sig

1.7887e+00

8.3299e-02

2.1474e+01 2.7440e-102

##
## J-Test: degrees of freedom is 1
##

J-test

P-value

## Test E(g)=0:

2.57411 0.10863

##
## Initial values of the coefficients
##

mu

sig

## 4.022499 1.881766
##
## #############
## Information related to the numerical optimization
## Convergence code = 0
## Function eval. = 55
## Gradian eval. = NA

# The J-test of over-identifying restrictions

print(specTest(res))

##
## ## J-Test: degrees of freedom is 1 ##
##
##

J-test

P-value

## Test E(g)=0:

2.57411 0.10863

If we compare ML and GMM by using simulations we notice that ML produces estimators with smaller
mean squared errors than GMM based on the above moment conditions. However, it is not GMM but the
moment conditions that are not efficient, because ML is GMM with the likelihood derivatives as moment
conditions.
sim_ex <- function(n,iter)
{
tet1 <- matrix(0,iter,2) # preallocate space for theta 1
tet2 <- tet1
for(i in 1:iter)
{
x1 <- rnorm(n, mean = 4, sd = 2) # generate from normal distribution
tet1[i,1] <- mean(x1)
tet1[i,2] <- sqrt(var(x1)*(n-1)/n)
tet2[i,] <- gmm(g1,x1,c(0,0),grad=Dg)$coefficients
}
par(mfcol=c(2, 2),oma=c(0,0,2,0)) # create a a subplot 2x2
hist(tet1[,1],main="ML mean",xlab="est. mean")
hist(tet2[,1],main="GMM mean",xlab="est. mean")
hist(tet1[,2],main="ML sd",xlab="est. sd")
hist(tet2[,2],main="GMM sd",xlab="est. sd")
title(paste("ML and GMM estimated parameters comparison (sample size=",n,sep=" ",", sim
s=",iter, ")"), outer=TRUE)
bias <- cbind(rowMeans(t(tet1)-c(4,2)),rowMeans(t(tet2)-c(4,2)))
dimnames(bias)<-list(c("mu","sigma"),c("ML","GMM"))
Var <- cbind(diag(var(tet1)),diag(var(tet2)))
dimnames(Var)<-list(c("mu","sigma"),c("ML","GMM"))
MSE <- cbind(rowMeans((t(tet1)-c(4,2))^2),rowMeans((t(tet2)-c(4,2))^2))
dimnames(MSE)<-list(c("mu","sigma"),c("ML","GMM"))
return(list(bias=bias,Variance=Var,MSE=MSE))
}
set.seed(345)
sim_ex(100,200)# 100 sims of sample size 200

## $bias
##
## mu

ML

GMM

-0.01406445 -0.01619955

## sigma -0.03723073 -0.07366727

##
## $Variance
##
## mu

ML

GMM

0.04530069 0.05574631

## sigma 0.02069728 0.02171330

##
## $MSE
##
## mu

ML

GMM

0.04527199 0.0557300

## sigma 0.02197992 0.0270316

If we increase the sample size (sample mean approaches the population mean) to 2,000, we notice that the
GMM estimates improve, however the ML is still better.
set.seed(345)
sim_ex(100,2000) #100 sims of sample size 2000

## $bias
##
## mu

ML

GMM

-7.842598e-05 -0.01057923

## sigma -1.590007e-02 -0.05185603

##
## $Variance
##
## mu

ML

GMM

0.04078809 0.04754989

## sigma 0.02060341 0.02293479

##
## $MSE
##
## mu

ML

GMM

0.04076770 0.04763803

## sigma 0.02084592 0.02561237

Nice treatment of GMM (with examples) can be found in Chapter 21 of book Modeling Financial Time
Series with S-PLUS(r) (http://www.amazon.co.uk/dp/0387279652/ref=rdr_ext_tmb) . There are parts from
examples above which follow this book.