You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

220 lines
6.6 KiB
Plaintext

#LyX 1.6.2 created this file. For more info see http://www.lyx.org/
\lyxformat 345
\begin_document
\begin_header
\textclass article
\use_default_options true
\language english
\inputencoding auto
\font_roman default
\font_sans default
\font_typewriter default
\font_default_family default
\font_sc false
\font_osf false
\font_sf_scale 100
\font_tt_scale 100
\graphics default
\paperfontsize default
\use_hyperref false
\papersize default
\use_geometry false
\use_amsmath 1
\use_esint 1
\cite_engine basic
\use_bibtopic false
\paperorientation portrait
\secnumdepth 3
\tocdepth 3
\paragraph_separation indent
\defskip medskip
\quotes_language english
\papercolumns 1
\papersides 1
\paperpagestyle default
\tracking_changes false
\output_changes false
\author ""
\author ""
\end_header
\begin_body
\begin_layout Standard
Notes on mlogit.
\end_layout
\begin_layout Standard
Assume that
\begin_inset Formula $J=3$
\end_inset
, so that there are
\begin_inset Formula $2$
\end_inset
vectors of parameters for
\begin_inset Formula $J-1$
\end_inset
.
For now the parameters are passed around as
\begin_inset Formula \[
\left[\beta_{1}^{\prime}\beta_{2}^{\prime}\right]\]
\end_inset
\end_layout
\begin_layout Standard
So if
\begin_inset Formula $K=3$
\end_inset
(including the constant), then the matrix of parameters is
\begin_inset Formula \[
\left[\begin{array}{cc}
b_{10} & b_{20}\\
b_{11} & b_{21}\\
b_{12} & b_{22}\end{array}\right]^{\prime}\]
\end_inset
\end_layout
\begin_layout Standard
(changed to rows and added prime above, so this all changes and the score
is also just transposed and flattend along the zero axis.) This is flattened
along the zero axis for the sake of the solvers.
So that it is passed internally as
\begin_inset Formula \[
\left[\begin{array}{cccccc}
b_{10} & b_{20} & b_{11} & b_{21} & b_{12} & b_{22}\end{array}\right]\]
\end_inset
Now the matrix of score vectors is
\begin_inset Formula \[
\left[\begin{array}{cc}
\frac{\partial\ln L}{\partial b_{10}} & \frac{\partial\ln L}{\partial b_{20}}\\
\frac{\partial\ln L}{\partial b_{11}} & \frac{\partial\ln L}{\partial b_{21}}\\
\frac{\partial\ln L}{\partial b_{12}} & \frac{\partial\ln L}{\partial b_{22}}\end{array}\right]\]
\end_inset
\end_layout
\begin_layout Standard
In Dhrymes notation, this would be column vectors
\begin_inset Formula $\left(\partial\ln L/\partial\beta_{j}\right)^{\prime}\text{ for }j=1,2$
\end_inset
in our example.
So, our Jacobian is actually transposed vis-a-vis the more traditional
notation.
So that the solvers can handle this, though, it gets flattened but the
score gets flattened along the first axis to make things easier, which
is going to make things tricky.
Now, in traditional notation, the Hessian would be
\begin_inset Formula \[
\frac{\partial^{2}\ln L}{\partial\beta_{j}\partial\beta_{j}}=\frac{\partial}{\partial\beta_{j}}\vec{\left[\left(\frac{\partial\ln L}{\partial\beta_{j}}\right)\right]}\]
\end_inset
\end_layout
\begin_layout Standard
where
\begin_inset Formula $\vec{}$
\end_inset
denotes a vectorized matrix, i.e., for a
\begin_inset Formula $n\times m$
\end_inset
matrix
\begin_inset Formula $X$
\end_inset
,
\begin_inset Formula $\vec{X}=\left(x_{\cdot1}^{\prime},x_{\cdot2}^{\prime},...,x_{\cdot m}^{\prime}\right)^{\prime}$
\end_inset
such that
\begin_inset Formula $x_{\cdot1}$
\end_inset
is the first
\begin_inset Formula $n$
\end_inset
elements of column 1 of
\begin_inset Formula $X$
\end_inset
.
This matrix is
\begin_inset Formula $mn\times n$
\end_inset
.
In our case
\begin_inset Formula $\ln L$
\end_inset
is a scalar so
\begin_inset Formula $m=1$
\end_inset
, so each second derivative is
\begin_inset Formula $n\times n$
\end_inset
and
\begin_inset Formula $n=K=3$
\end_inset
in our example.
Given our score
\begin_inset Quotes eld
\end_inset
matrix,
\begin_inset Quotes erd
\end_inset
our Hessian will look like
\begin_inset Formula \[
H=\left[\begin{array}{cc}
\frac{\partial\ln L}{\partial\beta_{1}\partial\beta_{1}} & \frac{\partial\ln L}{\partial\beta_{1}\partial\beta_{2}}\\
\frac{\partial\ln L}{\partial\beta_{2}\partial\beta_{1}} & \frac{\partial\ln L}{\partial\beta_{2}\partial\beta_{2}}\end{array}\right]\]
\end_inset
\begin_inset Formula \[
H=\left[\begin{array}{cccccc}
\frac{\partial^{2}\ln L}{\partial b_{10}\partial b_{10}} & \frac{\partial^{2}\ln L}{\partial b_{10}\partial b_{11}} & \frac{\partial^{2}\ln L}{\partial b_{10}\partial b_{12}} & \frac{\partial^{2}\ln L}{\partial b_{10}\partial b_{20}} & \frac{\partial^{2}\ln L}{\partial b_{10}\partial b_{21}} & \frac{\partial^{2}\ln L}{\partial b_{10}\partial b_{22}}\\
\frac{\partial\ln L}{\partial b_{11}\partial b_{10}} & \frac{\partial\ln L}{\partial b_{11}\partial b_{11}} & \frac{\partial\ln L}{\partial b_{11}\partial b_{12}} & \frac{\partial\ln L}{\partial b_{11}\partial b_{20}} & \frac{\partial\ln L}{\partial b_{11}\partial b_{21}} & \frac{\partial\ln L}{\partial b_{11}\partial b_{22}}\\
\frac{\partial\ln L}{\partial b_{12}\partial b_{10}} & \frac{\partial\ln L}{\partial b_{12}\partial b_{11}} & \frac{\partial\ln L}{\partial b_{12}\partial b_{12}} & \frac{\partial\ln L}{\partial b_{12}\partial b_{20}} & \frac{\partial\ln L}{\partial b_{12}\partial b_{21}} & \frac{\partial\ln L}{\partial b_{12}\partial b_{22}}\\
\frac{\partial^{2}\ln L}{\partial b_{20}\partial b_{10}} & \frac{\partial^{2}\ln L}{\partial b_{20}\partial b_{11}} & \frac{\partial^{2}\ln L}{\partial b_{20}\partial b_{12}} & \frac{\partial^{2}\ln L}{\partial b_{20}\partial b_{20}} & \frac{\partial^{2}\ln L}{\partial b_{20}\partial b_{21}} & \frac{\partial^{2}\ln L}{\partial b_{20}\partial b_{22}}\\
\frac{\partial\ln L}{\partial b_{21}\partial b_{10}} & \frac{\partial\ln L}{\partial b_{21}\partial b_{11}} & \frac{\partial\ln L}{\partial b_{21}\partial b_{12}} & \frac{\partial\ln L}{\partial b_{21}\partial b_{20}} & \frac{\partial\ln L}{\partial b_{21}\partial b_{21}} & \frac{\partial\ln L}{\partial b_{21}\partial b_{22}}\\
\frac{\partial\ln L}{\partial b_{22}\partial b_{10}} & \frac{\partial\ln L}{\partial b_{22}\partial b_{11}} & \frac{\partial\ln L}{\partial b_{22}\partial b_{12}} & \frac{\partial\ln L}{\partial b_{22}\partial b_{20}} & \frac{\partial\ln L}{\partial b_{22}\partial b_{21}} & \frac{\partial\ln L}{\partial b_{22}\partial b_{22}}\end{array}\right]\]
\end_inset
\end_layout
\begin_layout Standard
But since our Jacobian is a row vector that alternate
\end_layout
\end_body
\end_document