enbid-ananke 0.3.0__cp310-cp310-manylinux_2_39_x86_64.whl

enbid_ananke/Enbid-2.0/Documentation/users-guide.tex

@@ -0,0 +1,679 @@
+\documentclass{article}
+\title{User's Guide for EnBiD}
+\author{Sanjib Sharma \thanks{E-mail:sanjib.sharma@gmail.com}}
+\begin{document}
+\maketitle
+
+\tableofcontents
+\section{Introduction} \label{sec:intro}
+EnBiD ({\bf En}tropy Based {\bf Bi}nary {\bf D}ecomposition) is a code
+for calculation of densities from a discrete set of data points 
+distributed in a multi-dimensional space.  The code is completely 
+metric free. The most important feature of EnBiD  is its ability 
+to determine the appropriate metric at  each point in space. 
+It is based upon the
+idea of binary space partioning tree (kd-tree). The algorithm EnBiD 
+was inspired by an earlier algorithm dubbed FiEstAS 
+developed by Ascasibar \& Binney 2005 ,MNRAS,356,872. 
+The main feature of EnBiD is its unique node splitting
+criterion, which is  based on {\it Shannon Entropy} and helps
+to extract maximum information from the data. 
+After the spatial decomposition has been done the code 
+offers a variety of methods for obtaining smoothed estimates of density.
+The code determines the appropriate metric at each point in space
+from  the shape of the leaf nodes which are generated by the spatial decomposition.
+The main method of density estimation is the kernel based density 
+estimation.  Spherical kernels as well as kernels of the product form  can be 
+used. In addition a variety of different kernel functions can also be used.
+FiEstAS style smoothing as originally proposed by Ascasibar \& Binney 2005 
+can also be done. The algorithm and its test results are described in the 
+paper  astroph/0507550. The code is distributed under the GNU general 
+public license. 
+
+This guide provides the technical details which 
+will help others to use the code. 
+Some routines in this  code are derived from the code GADGET 
+which is a code for cosmological N-Body/SPH simulations,
+developed by Springel,Yoshida \& White 2001, New Astronomy,6,51,
+The algorithm for nearest neighbor search is based 
+on the algorithm of code SMOOTH 
+\newline
+http://www-hpcc.astro.washington.edu/tools/smooth.html
+
+\section{Compilation}
+The code is written in C++. Unpacking the code
+using \newline{\bf tar -zxvf Enbid-2.0.tar.gz} \newline generates a 
+directory \verb$Enbid-2.0$. The \verb$src$ directory 
+contains \verb$.cpp$ and \verb$.h$ files
+along with a \verb$Makefile$. Typing {\bf make} in the \verb$src$
+directory should compile the code on most Unix and linux based systems.
+Below are listed some options from the \verb$Makefile$
+which need to be specified during compilation time.
+\begin{verbatim}
+OPT1    =  -DWARN     # to put print statements for debugging
+OPT2    =  -DDIM3     # For doing calculations in 3 dimensions
+OPT2    =  -DDIM6     # For doing calculations in 6 dimensions
+OPT2    =  -DDIMO     # For other spaces 
+OPT3   =   -DMEDIAN   #enables optimizations for faster kernel smoothing
+OPT4   =   -DPERIODIC #enables periodic boundary conditions
+\end{verbatim}
+Option \verb$WARN$ should normally be kept enabled. In case of a problem
+it makes the code print a warning and diverts it into an interactive mode 
+asking the user for an input.  To run it in an un-interrupted fashion 
+this option should be disabled.
+The number of dimensions need to be specified during the compilation. $3$ and
+$6$ dimensional spaces  can be invoked by selecting the appropriate
+\verb$OPT2$. For other spaces one
+has to use option \verb$DIMO$ and then alter the variable
+\verb$#define ND$ in file \verb$allvars.h$.  When \verb$OPT3$ is 
+enabled special optimizations are enabled which  work with the median
+splitting criterion, this speeds up the neighbor search by about $10\%$ .
+Periodic boundary conditions are enabled by option \verb$OPT4$.
+Any change in the options in Makefile should generally be followed by
+a {\bf make clean} procedure before compiling it again.
+
+Compiling the code generates the executable file \verb$enbid$
+in the directory \verb$Enbid-2.0$ which can be run from command line along with a parameterfile. \newline
+{\bf ./enbid parameterfile}\newline
+
+\section{Parameters}
+The main method for calculating densities is the 
+tree based scheme followed by Kernel  smoothing. 
+Other than this FiEstAS based smoothing option 
+is also available. The parameter file 
+helps the user to choose the available options and also
+provides various parameters to fine tune the efficiency 
+and accuracy of the code. 
+A short description of these parameters is provided below.
+\begin{verbatim}
+InitCondFile            Examples/hernquist1_small/snapshot_ici
+\end{verbatim}
+This is the name of the input file which provides
+the data for density calculation.
+
+
+\begin{verbatim}
+SnapshotFileBase        _ph
+\end{verbatim}
+The output file name is derived from the input file name
+by adding the above base name and a suffix \verb$.est$. For the above
+choice of parameters the file generated is
+{\bf snapshot\_ici\_ph.est}. 
+
+\begin{verbatim}
+ICFormat                   1
+\end{verbatim}
+For a value of 1 the code reads files in 
+GADGET format. For a value of 0 it can read
+data written in ASCII format.
+Files of other formats can be made to read in 
+by writing ones own reading routine. A template 
+function is provided for this. This can be invoked
+by setting the above parameter to 2.
+
+
+\begin{verbatim}
+SpatialScale             1
+\end{verbatim}
+This parameter is used to define global scaling
+relations between various co-ordinates.
+This choice is useful when one wants to use 
+normal SPH (isotropic kernels and non-adaptive metric) 
+otherwise the parameter should be set to $1$.
+For \verb$SpatialScale$ $=0$ the scaling is done
+based on the global co-variance of the data.
+If $\sigma_i=\sqrt{<x_i^2>}$ is the standard deviation of the 
+$i$th co-ordinate then the code performs the following transformation
+$x_i \rightarrow x_i/\sigma_i$. 
+For phase space density calculations (i.e. number of dimensions $d=6$),
+if \verb$SpatialScale$ $> 0$, the co-ordinates of real space 
+(dimensions $1,2$ and $3$) are scaled globally with 
+respect to velocity space (dimensions $4,5$ and $6$) by the following
+transformation,  for $i=1$ to $3$ $x_i \rightarrow x_i/$SpatialScale.
+
+\begin{verbatim}
+PartBoundary              7
+\end{verbatim}
+During tree generation if the number of particles in a node is
+greater than or equal to the above parameter 
+a boundary correction is applied to its surfaces. 
+Optimum choice of this paramater is $d+1$, $d$ being the dimensionality 
+of space under consideration.  
+If  the densities are found to be underestimated then lower it or vice versa.  
+If the value is too small this can give rise to Poisson errors, so 
+irrespective of the number of dimensions  a minimum  value of $7$ 
+should be used.
+When the system is known to have 
+well defined rectangular boundaries e.g. 
+systems with periodic boundary conditions, 
+the boundary correction can be switched
+off by setting this parameter to zero.  Read option  
+\verb$PeriodicBoundaryOn$ for more details.
+If this parameter has value $1$ boundary correction is 
+applied to the leaf nodes only (as done originally in FiEstAS).
+
+\begin{verbatim}
+NodeSplittingCriterion     1
+\end{verbatim}
+This parameter is used to alter the node splitting 
+criteria used to generate the tree. 
+For a value of $1$ \emph{Shannon Entropy} is evaluated
+in each dimension and the dimension with lowest entropy 
+is chosen to be split. Setting this parameter to $0$ 
+disables entropy based node splitting criterion and 
+splits each dimension alternately. 
+
+\begin{verbatim}
+CubicCells                 0
+\end{verbatim}
+When this parameter has a value
+of 1 the \newline
+\verb$NodeSplittingCriteria$ is used to 
+select the subspace (real or velocity) to be split rather than the 
+dimension.
+Then from this subspace the axis having maximum  
+variance ($<x_i^2>-<x_i>^2$) is selected to be split. 
+For systems whose subspaces are known to be Euclidean this
+gives better results. It results in cubic cells in 
+each sub-space. This option only works for 3 and 6 dimensional spaces.
+
+
+\begin{verbatim}
+MedianSplittingOn          0
+\end{verbatim}
+Ideally the position of splitting should be close to
+the median of the data points so that there are equal 
+number of data points on both sides of the split.
+For building the tree it is computationally  more efficient
+to instead choose the splitting position close 
+to the mean of the data points.
+When the code is compiled with the \verb$-DMEDIAN$ flag 
+setting this parameter to 1 results in faster kernel density estimates.
+
+\begin{verbatim}
+TypeOfSmoothing             3
+\end{verbatim}
+This parameter helps to select various different smoothing options.
+For a value of zero no smoothing is done
+this is useful to get fast estimates of 
+density but at the price of a large dispersion.
+For a value of $1$ FiEstAS style smoothing is done,
+for $2$ normal spherical Kernel  based scheme is used (having 
+isotropic metric same scale in all dimensions)  and
+for $3$ spherical Kernel based scheme with an appropriate 
+adaptive metric (estimated by the size and the shape of the 
+leaf node containing the data point) is used.
+For a value of $4$ and $5$ the kernel used is of product form with 
+the metric being isotropic for the former and adaptive for the later. 
+For  FiEstAS style smoothing particles of a given type should not 
+have multiple masses. 
+
+\begin{verbatim}
+DesNumNgb                  40
+\end{verbatim}
+This gives the number of smoothing neighbors 
+that are used for density calculation.
+For FiEstAS style smoothing optimum range of this
+parameter is $2$ to $10$. For SPH smoothing the optimum 
+choice depends upon the number of dimensions. 
+For a Hernquist sphere in 6 dimensions 
+\verb$DesNumNgb=40$ was found to give smoothing 
+equivalent to FiEstAS smoothing with \verb$DesNumNgb=2$. 
+
+
+\begin{verbatim}
+VolCorr                    1
+\end{verbatim}
+During smoothing if the smoothing
+box extends outside the boundary of the system
+density might be underestimated.  Setting this parameter 
+to 1 enables this correction for both Kernel  and FiEstAS 
+smoothing. 
+
+
+
+\begin{verbatim}
+TypeOfKernel               3
+\end{verbatim}
+For Kernel based smoothing 
+three different types of kernels
+can be used. For a value of $0$ 
+the B-Spline kernel is used,
+for $1$ top hat kernel is used, for $2$ Bi-weight kernel 
+is used and for $3$ Epanechnikov kernel is used. 
+For $4$ cloud in cell (CIC) type of linear kernel is 
+used and for $5$ triangular shaped cloud (TSC) type 
+of kernel is used.
+In our tests Epanechnikov kernel was found to give 
+best results. 
+
+
+\begin{verbatim}
+KernelBiasCorrection          1
+\end{verbatim}
+Normal Kernel based schemes have significant bias in
+their estimated densities for irregularly
+distributed data. This bias arises when the densities are calculated
+at the location of the data points and can be eliminated by suitably  
+displacing the central data point within the smoothing volume.
+This is enabled by setting the parameter to 1. For regularly spaced 
+data (e.g.  glass or lattice like systems)  this option  should be 
+disabled by setting the value  to 0.
+
+
+\begin{verbatim}
+AnisotropicKernel           0
+\end{verbatim}
+This enables the use of Anisotropic kernels which 
+can have both shear and rotation.
+The kernel instead of being spherical is 
+now an rotated ellipsoid in general.
+The local co-variance matrix of the data is used to 
+determine the orientation and the smoothing
+lengths of the kernel. For this option the \verb$TypeOfSmoothing$
+should be either 2 or 3 (i.e spherical kernel smoothing). 
+
+\begin{verbatim}
+Anisotropy                  0
+\end{verbatim}
+Depending upon the anisotropy of the 
+problem, density estimation with 
+anisotropic kernels can be 
+computationally very intensive. The 
+parameter \verb$Anisotropy$ can be used to
+set the minimum allowable minor to major axis ratio of the
+kernel smoothing lengths. Maximum 
+accuracy is achieved by setting the
+value to zero while a value of $1$
+corresponds to an isotropic kernel.
+
+\begin{verbatim}
+DesNumNgbA                  100
+\end{verbatim}
+This is the number of neighbors used for 
+estimation of the co-varaince matrix, which is 
+in turn used for calculating the anisotropic kernel.
+Its value should be between $100-200$.  Setting this  value 
+to less than 100, results  in poor estimation of the 
+co-variance matrix.    
+
+
+
+
+\begin{verbatim}
+TypeListOn               0
+\end{verbatim}
+It is possible to specify multiple species (types) of particles in the same initial 
+condition file. Each particle type is treated independently for the purpose 
+of density estimation and separate tree is constructed for each of them.
+For Gadget format initial condition  files the number of particle types  is limited 
+to $6$.  To overcome this limitation and to offer more flexibility  the parameter 
+\verb$TypeListOn$ can be enabled by setting it to $1$. This overrides 
+the type of particles specified in the initial condition file.
+If this parameter is set to 1 a typelist file in ascii text, with the name  \verb$InitCondFile$ + suffix  
+\verb$_typelist$, should be provided. 
+If a total of $N$ particles are divided into $l$ independent groups of different types 
+having particles $n_1$, $n_2$ ...  $n_l$ respectively such that $\sum n_i=N$, then 
+they can be read in by specifying this list in an ascii format  in the typelist file as follows
+\newline
+\newline
+$n_1 n_2  ...  n_l $ 
+\newline
+
+\begin{verbatim}
+PeriodicBoundaryOn               0
+\end{verbatim}
+Periodic boundary conditions can be enabled for Kernel based smoothing 
+by setting this parameter to $1$ and compiling the code with option \verb$-DPERIODIC$.
+Periodicity of each dimension can be set independently. 
+A file \verb$periodic_lenghts.txt$ should be provided in the current 
+working directory, which contains the periodic lengths $l_i$, for each of the dimensions. 
+Boundary correction to the tree nodes as specified by parameter \verb$PartBound$ 
+are not applied in the dimensions that  are periodic.
+If $l_i$  is set to zero the code automatically determines the appropriate 
+periodic length from the range of the particles in that dimension  ($l \sim (x_{max}-x_{min})$). 
+If it is set to less than zero then that dimension is not considered periodic. In the code 
+this is  achieved by setting $l \sim10 (x_{max}-x_{min})$)  and enabling 
+ \verb$PartBound$ for that dimension.   Periodic support is as such not available
+ for FiEstAS smoothing, enabling periodic boundaries in this case 
+ only effects  the tesselation process (i.e boundary correction to the tree nodes).
+ 
+
+
+
+\section{Input}
+The data can be read from either  an ASCII file, GADGET format file
+or a user defined file format these can be invoked by setting the
+parameter \verb$ICFormat$ to 0 , 1 and 2 respectively.
+For GADGET format files the code can 
+read both little and big endian format binary files.
+The endianness of the output file is the same as that of  input  file.
+
+\subsection{ASCII input file format}
+In ASCII format the co-ordinates of the data points are arranged in rows
+as follows.
+\newline
+$x_1 y_1 z_1 $ \newline
+$x_2 y_2 z_2 $ \newline
+........so on \newline
+
+\subsection{User defined file format}
+The user can read in files of his own format by writing his own
+reading routine. A template function \verb$ void read_ic2(char *fname)$ 
+in file \verb$read_ic.cpp$ is provided for this.
+The user should read his data and assign the co-ordinates to
+a predefined structure. 
+Make the following changes in the function \verb$read_ic2$
+\begin{verbatim}
+  /* SPECIFY TOTAL NUMBER OF PARTICLES*/
+  NumPart= specify total number of particles;
+
+/* leave this unchanged */
+//-------------------------------------
+  All.MaxPart = NumPart; 
+  for(i=0;i<6;i++)
+    {
+      header1.npart[i]=0;
+      header1.npart[i]=0;
+      header1.mass[i]=0;
+    }
+  header1.npart[1]=NumPart;
+  header1.npartTotal[1]=NumPart;
+  allocate_memory();
+//-------------------------------------
+
+  /* READ THE DATA HERE AND ASSIGN IT TO P[i].Pos */
+  for(i=1; i<=NumPart; i++) 
+   for(k=0;k<ND;k++)
+     P[i].Pos[k]=assign positions;
+  for(i=1; i<=NumPart; i++) 
+    P[i].Mass=assign mass or set it to 1;
+\end{verbatim}
+
+
+
+\subsection{GADGET input file format}
+In a GADGET format file the data is written in binary mode and consists 
+of a header followed by the main data containing position 
+,velocities and id's of particles. 
+The data fields are
+separated  by block-size fields so that it can be read in with 
+unformatted fortran format. A description of these fields is given below.
+Read statements in fortran would be as follows.
+\begin{verbatim}
+read (1) npart,massarr,a,redshift,flag1,flag2,nall,unused
+read (1) pos
+read (1) vel
+read (1) id
+read (1) masses
+\end{verbatim}
+The code can handle multiple species of particles and calculates the density 
+of each specie separately, but for a given specie all the particles must have 
+same mass or else the results might be erroneous.
+For 6 dimensional phase space density estimation,
+each particle has $6$ co-ordinates, the position of the particle 
+constitutes the first three co-ordinates and the three velocity 
+components are assigned co-ordinate numbers $4,5$ and $6$.
+\newpage
+\begin{table}[here]
+\caption{Header}
+\begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+npart(6) & int & $4\times6$ & Array specifying number of particles of various types\\
+massarr(6) & double & $8\times6$ & mass of different types of particles\\
+a & double & 4 & Time or expansion factor\\
+redshift & double & 4 & Redshift\\
+flag1 & int & 4 & unused here \\
+flag2 & int & 4 & unused here \\
+nall(6) & int & $4\times6$ & Number of particles of each type, same as npart(6) here\\
+flag3 & int & 4 & unused here \\
+numfiles & int & 4 & number of files holding the data \\
+boxsize & double & 4 & size of periodic box\\
+omega-0 & double & 4 & Cosmological parameter specifying matter density \\
+omega-l & double & 4 &  Cosmological parameter specifying vacuum energy density\\
+hubble-param & double & 4 & Hubble parameter \\
+unused &  &  & used to fill the header to a total size of 256 bytes \\
+\hline
+\end{tabular}
+%\end{table}
+%\begin{table}[here]
+  \caption{Particle Positions}
+  \begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+Pos(3,N) & float & $4 \times 3 \times N$ & position of particles\\
+\hline
+\end{tabular}
+%\end{table}
+%\begin{table}[here]
+  \caption{Particle Velocities}
+  \begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+Vel(3,N) & float & $4 \times 3 \times N$ & velocities of particles\\
+\hline
+\end{tabular}
+\end{table}
+\newpage
+\begin{table}[here]
+  \caption{Particle ID's}
+  \begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+id(N) & int & $4 \times N$ & id of particles used to uniquely identify
+them in  case the order is not same among different files. \\
+\hline
+\end{tabular}
+%\end{table}
+%\begin{table}[h]
+\caption{Particle Masses}
+\begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+mass(N) & int & $4 \times N$ & mass of those particles types which
+have zero entries in massarr.\\
+\hline
+\end{tabular}
+\end{table}
+
+
+\section{Output}
+The output file contains a list of densities at the location 
+of the data points (particles). The order is same as that 
+specified in the input file. For \verb$ICFormat=1$ ,
+if the mass of the particles is specified the output 
+density is mass density.
+For \verb$ICFormat=0$ mass of each particle is assumed
+to be unity so the density evaluated by the code 
+is the number density instead of mass density.
+The output in this case is written in ASCII  format.
+One can also specify ones own output format a template function 
+\verb$void savepositions_ioformat2(int )$ is provided for this in file
+\verb$io.cpp$.
+
+\setcounter{table}{0} 
+\subsection{EnBiD output file format}
+For \verb$ICFormat=1$ the output is written in 
+a binary file with a GADGET format header and a list of densities. 
+Some extra flags are added in the header which might be useful for
+reading the output densities. The data fields are
+separated  by block-size fields. A description of these fields 
+is given below. Read statements in fortran 
+would be as follows.
+\begin{verbatim}
+read (1) npart,massarr,a,redshift,flag1,flag2,nall,unused1
+         ,flag-dim,flag-density,unused
+read (1) density
+\end{verbatim}
+\newpage
+\begin{table}[here]
+  \caption{Header}
+  \begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+npart(6) & int & $4\times6$ & Array specifying number of particles of various types\\
+massarr(6) & double & $8\times6$ & mass of different types of particles\\
+a & double & 4 & Time or expansion factor\\
+redshift & double & 4 & Redshift\\
+flag1 & int & 4 & unused here \\
+flag2 & int & 4 & unused here \\
+nall(6) & int & $4\times6$ & Number of particles of each type, same as npart(6) here\\
+flag3 & int & 4 & unused here \\
+numfiles & int & 4 & number of files holding the data \\
+boxsize & double & 4 & size of periodic box\\
+omega-0 & double & 4 & Cosmological parameter specifying matter density \\
+omega-l & double & 4 &  Cosmological parameter specifying vacuum energy density\\
+hubble-param & double & 4 & Hubble parameter \\
+\hline
+\textsf{flag-id} & int & 4 & unused here\\
+\textsf{flag-dim} & int & 4 & number of dimensions used for density calculations \\
+\textsf{flag-density} & int & 4 &  indicates that it is a \verb$.est$
+file containing densities and not a GADGET format file\\
+\hline
+unused &  &  & used to fill the header to a total size of 256 bytes \\
+\hline
+\end{tabular}
+%\end{table}
+%\begin{table}[h]
+  \caption{Particle Density}
+  \begin{tabular}{|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{6cm}|}
+\hline
+\hline
+  Variable     &  Type   &   Bytes & Description\\
+\hline
+Density(N) & float & $4 \times N$ & Calculated Density of particles \\
+\hline
+\end{tabular}
+\end{table}
+
+
+\section{Examples}
+Two example data files are provided in the 
+directory \verb$Examples$. The file \newline 
+\verb$snapshot_ici$ in 
+directory \verb$uniform_6d_box_4$ is a GADGET format data
+of $10^4$ particles distributed uniformly in a cubic box of 
+length $100$ units in both real and velocity space. 
+The file \verb$snapshot_ici.ascii$ is the corresponding ASCII
+format file. The directory \verb$hernquist1_small$ contains 
+the data for a Hernquist sphere modelled with $10^4$ particles.
+
+\section{General usage tips}
+The main method of density estimation is the kernel based 
+method along with adaptive metric which is shown as option $3$ 
+below, but the code also offers various other options for calculating densities. 
+More accurate estimates require more CPU time.
+Below we list four main options in increasing order of accuracy
+and CPU time. Parameterfiles for these can be found in the directory 
+parameterfile.
+\begin{enumerate}
+\item {\bf Unsmoothed estimate:}  \newline
+For fast crude estimates of any general data use
+\begin{verbatim}
+NodeSplittingCriterion=1
+CubicCells= either 1 or 0 % 1 for symmetric systems 0 for general
+TypeOfSmoothing=0 
+\end{verbatim}
+\item {\bf FiEstAS smoothing:}  \newline
+For smooth densities with FiEstAS  
+style smoothing 
+\begin{verbatim}
+TypeOfSmoothing=1
+DesNumNgb=2-10 
+All.VolCorr=1
+\end{verbatim}
+\item {\bf Kernel smoothing with adaptive metric:} \newline 
+For kernel smoothed estimates  set
+\begin{verbatim}
+TypeOfSmoothing=3
+TypeOfKernel=3
+KernelBiasCorrection=1 
+DesNumNgb=10-50
+All.VolCorr=1
+\end{verbatim}
+\item {\bf Anisotropic kernel smoothing:}  \newline
+If one has more time one can also 
+use Anisotropic kernels with 
+\begin{verbatim}
+TypeOfSmoothing= 3 or 2 % more faster with 2 but less accurate
+AnisotropicKernel=1
+Anisotropy=0
+DesNumNgbA=100-200
+\end{verbatim}
+\end{enumerate}
+\newpage
+The table below gives the CPU time in seconds required  on an AMD  1666.7 MHz
+processor for calculating the density of a data 
+of 1 million particles modelled as a Hernquist sphere.
+\begin{table}[here]
+\begin{center}
+\caption{Cpu Time}
+\begin{tabular}{|p{6.0cm}|p{2.5cm}|p{2.5cm}|}
+\hline
+Type of Estimate                  & 6 Dimensions \\
+\hline
+\hline
+Unsmoothed                        &   10-24 s          \\
+Fiestas smoothing                  &   356 s         \\
+Kernel smoothing with adaptive metric   &   863 s \\
+\hline 
+\end{tabular}
+\end{center}
+\end{table}
+
+
+\begin{itemize}
+\item For a hernquist sphere options 3 and 4 
+were found to give nearly same results.
+This is because the anisotropy has already been corrected by means
+of an adaptive metric. 
+\item For real and phase space density calculation of 
+systems with special symmetries e.g. spherical
+systems use \verb$CubicCells=1$. Hernquist sphere and galactic halos
+fall into this category.
+\item For kernel smoothing on regularly spaced
+data e.g. a lattice like system disable the parameter
+\verb$KernelBiasCorrection$ by setting it to $0$.
+\item
+For FiEstAS smoothing all the particles of a given type must have same mass.
+For other types of density estimation multiple mass support is available 
+but the density estimates are most accurate if the particles have same mass 
+or if at least  their distribution in space is smooth. 
+\item The code has mostly been tested in 3 and 6 dimensions 
+and on systems having spherical symmetry in real and velocity sub-spaces. 
+For other situations it will be useful to first test the accuracy and
+efficiency of the code by experimenting with the   
+parameters and options provided here.
+\item Although the code can support arbitrary number of dimensions
+but for kernel based smoothing, for dimensions greater than 20, 
+one needs to specify appropriate normalization constants of the kernel
+in routine \newline\verb$set_sph_kernel$ (file \verb$begrun.cpp$).
+For other schemes there is no restriction.
+\item  For FiEstAS smoothing of systems with periodic boundary conditions or well defined
+rectangular boundaries use \verb$VolCorr=1$ to avoid under-estimation of densities 
+in boundary regions. Periodic support is as such not available
+ for FiEstAS smoothing, enabling periodic boundaries in this case 
+only effects  the tesselation process (i.e boundary correction to the tree nodes).
+\item 
+The mean splitting criterion gives better results as compared to median criterion, 
+for systems having substructures.  But if  one does not want to use adaptive metric  
+and instead wants to use isotropic kernels  (i.e \verb$TypeOfSmoothing=2$ or $4$)  
+the code can be compiled with \verb$-DMEDIAN$  
+and used with \verb$MedianSplittingOn=1$, this results in faster density estimates.
+\end{itemize}
+
+
+\end{document}