Mca
Mca
\begin{document}
\def\emline#1#2#3#4#5#6{%
\put(#1,#2){\special{em:moveto}}%
\put(#4,#5){\special{em:lineto}}}
But we know only one world, we live in, only one physics, one
chemistry, one biology. There should be only one mathematics,
too.
Algebra is very abstract science, except for its one branch, linear
algebra. There are studied operations with vectors and matrices. And
on these notions the heart of the book is based.
If this book were written one hundred years ago, it could save one
human life, if it were published fifty years ago, it could prevent
erroneous interpretation of the information
theory.
\item ``{\bf explanations of matrices did not make any sense}''\\ and
vectors in the string are written using the formalism
Sums and differences of two naive matrices are studied in the second
part of the book. They are known as graphs. The block designs could
form the next step. They are exploited in advanced combinatorics.
Blocks have matrix form and the numbers of distinguishable blocks are
searched for. Hall \cite{8} began his book by chapters surveying
classical combinatorics before he treated block designs, but no
attempt was made to use an unified matrix technique to traditional
combinatorics and explain combinatorial problems as counting naive
blocks.
The difference is that the system starts from the first power of 1,
which is undistinguishable from its zero power
addition
subtraction
and division
We will use this system implicitly without mentioning it. There will
be some problems with notation. Not enough letters are available to
use a special one for each function. We will use some letters for
different functions without warning. Figures, tables, and equations
are indexed separately in~each chapter.
\tableofcontents
\listoffigures
\listoftables
\pagenumbering{arabic}
\newpage
Our notion of the space is based on our book form: a point has three
coordinates corresponding to the page, line, and column numbers,
respectively\footnote{There exist polar coordinates giving positions
as on reels, too, but they are outside our study.}.
Three dimensions of a book are formed by the given convention from
a~string of symbols. Too long strings are cut into lines, too long
strings of lines are cut into pages and eventually too long sequences
of pages are cut into volumes forming the fourth dimension since we
must determine at first positions of symbols in lines. There exist
different forms of books, as for examples scrolls. Strings of symbols
can be wound on reels, or rolled up into balls, and they remain
essentially unchanged. Similarly the points of the space can be
indexed in different ways.
Books exist without any movement but when we read them, we need time
to transfer their symbols into our brain, to remember essential facts
and thoughts, to transcribe the book into our brain. The world is a
word, a~very long one, in a foreign language. We must learn, how to
understand it.
There exist one essential difference between a book and our world.
The world is moving. As if a book itself were constantly transcribed.
Some parts seem to us to be constant, but somewhere invisible
corrections are made constantly. The world is in instant a the book A,
in the next instant b the book B. All possible states of the world
form a library.
But we will analyze at first the simpler case, the unmoving text.
The three dimensions of the space are not equivalent. To move forward
is easy, backward clumsy, left or right movements, as crabs do, are
not normal, up and down we can move only in short jumps (long falls
down end dangerously). In books eyes must jump on the next line, a
page turned, a new volume opened. Increasing efforts are needed in
each step.
From ancient times the ortogonality means that between two straight
lines the right angle R exists. Actually there must be always 4 R if
two lines cross $$\begin{array}{ccc} R &\vline & R \\ & \vline & \\
\hline & \vline & \\ R & \vline & R \end{array}$$
The third straight line in the plane must be either parallel to one
them, and then it crosses the other one, or it crosses both of them,
and then they form a triangle, except lines going through the cross
of the first two lines.
The most important property of right triangles is, that the squares
of their hypotenuses are equal to the sums of squares of both other
sides as on Fig. \ref{Pythagorean}
\begin{figure}
\caption{Pythagorean theorem. $a^2\ +\ b^2\ = \ c^2$}
\label{Pythagorean}
\linethickness{0.6 pt}
\begin{picture}(150.00, 130.00)
\put(50.00,80.00){\line(0,1){30.00}}
\put(50.00,110.00){\line(1,0){30.00}}
\put(80.00,110.00){\line(0,-1){30.00}}
\put(80.00,80.00){\line(-1,0){70.00}}
\put(10.00,80.00){\line(0,-1){40.00}}
\put(10.00,40.00){\line(1,0){40.00}}
\put(50.00,40.00){\line(0,1)
{40.00}} \put(50.00,80.00){\line(1,0){30.00}}
\put(80.00,80.00){\line(0,1){0.00}}
\put(80.00,80.00){\line(-1,0){30.00}}
\put(50.00, 40.00){\line(3,4){30.00}}
\put(80.00,80.00){\line(4,-3){40.00}}
\put(120.00,50.00){\line(-3,-4){30.00}}
\put(90.00,10.00){\line(-4,3){40.00}}
\put(63.00,85.00){\makebox(0,0)[cc]{a}}
\put(71.00,56.00){\makebox(0,0)[cc]{c}}
\put(43.00,60.00){\makebox(0,0)[cc]{b}}
\end{picture}
\end{figure}
The smallest right triangle which sides are whole numbers has sides 3,
4, 5 and their squares are 9 + 16 = 25. The relation between the
sides of right triangles is known as the {\em Pythagorean theorem}.
The knowledge of right triangles was one from the first mathematical
achievements of mankind. The pyramids have square bases. Their
triangulation was very accurate due to exploitation of this
knowledge.
But similarly as we are not able to find the fourth dimension, we are
not able to decide if a set of numbers does not correspond to a set
of orthogonal straight lines which lengths are corresponding to the
given numbers. We form consecutively right triangles as on
Fig.\ref{Consecutive}
\begin{figure}
\caption{Consecutive Pythagorean addition. New vectors
are added as orthogonal to the sum of previous ones}
\label{Consecutive}
\linethickness{0.6 pt}
\begin{picture}(90.00,
90.00) %\vector(20.00,80.00)(70.00,80.00)
\put(70.00,80.00){\vector(1, 0){0.2}}
\put(20.00,80.00){\line(1,0){50.00}} %\end %\vector(70.00, 80.00)(70.00,50.00)
\put(70.00,50.00){\vector(0,-1){0.2}}
\put(70.00,80.00){\line(0,-1){30.00}} %\end %\vector(70.00,50.00)(55.00,25.00)
\put(55.00,25.00){\vector(-2,-3){0.2}}
\multiput(70.00,50.00)(-0.12,
-0.20){126}{\line(0,-1){0.20}} %\end %\vector(55.00,25.00)(30.67,8.33)
\put(30.67,8.33){\vector(-3,-2){0.2}}
\multiput(55.00,25.00)(-0.18, -0.12){139}{\line(-1,0){0.18}} %\end
\put(43.67,84.00){\makebox(0,0)
[cc]{a}} \put(75.00,64.00){\makebox(0,0)[cc]{b}}
\put(68.33,34.33){\makebox(0,0)[cc]{c}}
\put(48.00,12.67){\makebox(0,0)[cc]{d}}
\put(63.67,73.00){\makebox(0,0)[cc]{R}}
\put(60.67,48.67){\makebox(0,0)[cc]{R}}
\put(45.67,25.67){\makebox(0,0)[cc]{R}}
\put(20.00,80.00){\circle*{0.00}}
\put(20.00,80.00){\circle*{0.00}}
\put(20.00,80.00){\circle*{2.75}} %\vector(20.00,80.00)(70.00,50.00)
\put(70.00,50.00){\vector(3,-2){0.2}}
\multiput(20.00,80.00)(0.20,-0.12){251}{\line(1,0){0.20}} %\end
%\vector(20.00,80.00)(55.00,25.00)
\put(55.00,25.00){\vector(2,-3){0.2}}
\multiput(20.00,80.00)(0.12,-0.19){292}{\line(0,-1){0.19}} %\end
%\vector(20.00,80.00)(30.67,8.33) \put(30.67,8.33)
{\vector(1,-4){0.2}}
\multiput(20.00,80.00)(0.12,-0.81){89}{\line(0,
-1){0.81}} %\end
\end{picture}
\end{figure}
\begin{equation}
L^2\ = \ \sum m^2_j \end{equation}
\label{sum}
where $m^2_j$ stands for $n$ different abscissa and $L^2$ is the
square of the length of all $n$~abscissa. We can rotate consecutively
each vector of the plane in such a way that it forms a right triangle
with the sum of all other $(n~-~1)$ vectors, but we must not consider
simultaneously more lines or we find that they are not orthogonal as
we clearly see on Fig.\ref{Consecutive} where a series of right
triangles was drawn.
The sum of $n$ squares with side values (lengths) $m^2_j$ can be in
its turn decomposed into a Pythagorean triangle which squared sides a,
b, and c are
\begin{equation}
a^2 = n\overline{m}
\end{equation}
\begin{equation}
b^2 = \sum~m^2~-~n\overline{m}
\end{equation}
\begin{equation}
c^2 = \sum m^2_j,
\end{equation}
\label{mean}
\begin{equation}
\overline{m} = \sum m_j/n.
\end{equation}
\label{bar}
The straight length of the side is its square root. Here the square
root of $n$ appears somewhat surprisingly, but it is the length of
the diagonal of $n$ dimensional cube. Similarly the third side of the
triangle (1.3) can be normalized by dividing it with $n$. Then we get
the value $\sigma^2$ known as the {\em dispersion}
\begin{equation}
\sigma^2 = 1/n\sum (m^2_j - n\overline{m}^2).
\end{equation}
\label{standard}
Its square root, comparable with the mean, is the {\em standard
deviation} $\sigma$. For example, take the values 1,~2,~3,~4. Their
mean is 2.5, the sum of squares $30 = 1 + 4 + 9 + 16$. The dispersion
is $1/4(30 - 4 \times 6.25) = 1.25$.
Calculating the mean and the standard deviation we need not to know
the directions of both legs, since they are determined automatically
by their lengths, as when a triangle is constructed from the known
lengths of its three sides. We draw two circles with diameters $a$
and $b$ on both ends of the side $c$. Where the circles cross, the
third vertex lies. The direction of all sides in the multidimensional space is
abstract for us.
\section{Euclidean space}
\label{Euclidean space}
The space of right triangles and never crossing parallel lines is known as the
{\em Euclidean space}. Its generalization to infinite many dimensions
$n~\rightarrow~\infty$ for the sum \ref{mean} is known as the {\em Hilbert space}.
An Euclidean space with $n$~dimensions forms a subspace in it.
We take a square ABCD. All its right angles are according to the 4. postulate
right, and all its sides are straight lines.
We add to this square ABCD a new square CDEF and align the sides AE and BF
according to the 2. postulate.
To the obtained rectangle ABEF, we and a new square EFGH and align again the sides
AG and BH according to the 2. postulate. In such a way we continue with adding of
squares infinitely, eventually on the other shorter side of the rectangle.
There are two possibilities that the long sides of the infinite rectangle meet or
diverge. Either these long sides are not straight lines meeting the demands of the
2. postulate, or the right angles of consecutive squares did not meet the demands
of the 4. postulate.
The problem of ortogonality is loosing its importance in the Hilbert space. If you
have a store of infinite many vectors, you can pick any two as the first ones. You
can be sure, that you find the third one which is orthogonal to the first two. So
you continue. You will be exhausted before you will be able to empty the store. Or
you can be lazy, and to use alternatively vectors with angles greater and smaller
than orthogonal. The errors will compensate.
Euclides introduced axioms into mathematics. Space and its elements are defined by
a set of propositions. A disadvantage of this approach is that we don't know a
priori which elements form the space. We will use another approach and generate
elements consecutively.
We encounter spaces of many dimensions by recognizing that we are not alone in the
space. There are other people living and many things exist in~this particular
space. Each entity has its own position in the space. Listing these positions, we
need for each object its specific line with its coordinates. In each line must be
as many coordinates as the space, entities are embedded in, has dimensions. If
there were $m$~entities in $n$~dimensional space it would be necessary to know
$mn$ coordinates, in the 3~dimensional space we need for $m$~entities
$3m$~coordinates and in 1~dimensional space still $m$~coordinates to determine
positions of all objects.
Spaces with $m$~objects are known in physics as phase spaces\footnote {The number
of objects is there given as $n$. To avoid confusion with $n$ dimensions, we use
the symbol $m$.}. They have curious properties and we can sense directly some of
them, as for example temperature and the wind velocity of a system of molecules of
air are, which correspond to mathematical notions. Each molecule has at ambient
temperature mean velocity several hundred meters per second. Impacts of tiny
molecules on walls of the container produce the pressure. Chaotic collisions of
the molecules moving in different directions lead to a stable distribution of
particle velocities. These velocities decompose into two components. One component
is formed by a part of movement which all particles in the given volume have
common. This component, the mathematical arithmetical mean, usually great only few
meters per second, as compared to the above mentioned hundred meters per second,
we feel, when we are inside it as its part, as the wind, the physical property of
the system of molecules. The other component is the dispersion from the mean
vector velocity. It is known as the thermal motion of molecules, the temperature.
We will show that all phase spaces are {\em isomorphic}. Some of their properties
do not depend on the dimensionality $n$ of the space the system of $m$~entities is
embedded, but they are given only by the number of the entities.
Next we introduce a special place in our $n$~dimensional space from which we will
measure all translations. This point we will call the {\em center of
the~coordinate system}. Then we define $n$~points on a sphere (circle) with its
center in the~center of the coordinate system. We accept the radius of the~sphere
as the unit length. We can imagine that the points on the sphere are the
translated center of coordinate system and we will call each vector connecting the
center of the coordinate system with the defined $n$~points on a sphere the {\em
unit vector} ${\bf e}_j$. The notation of the unit vector ${\bf e}_j$ is a row
in~round brackets with $n$~elements. In physics symbols with arrows are used as
$\vec \j$. $(n~-~1)$ elements of the vector ${\bf e}_j$ are zeroes and there is
only one unit element on the j-th place
\begin{equation}
{\bf e}_j = (0_1, 0_2,\dots, 1_j, \dots, 0_n).
\end{equation}
\label{unit}
Equal length of all unit vectors ${\bf e}_j$ in (\ref{unit}) is not an essential
condition of the existence of a vector space. We could define unit vectors ${\bf
e}_j$ in (\ref{unit}) as having different lengths, make all operations as with
vectors having equal lengths, and only then modify results according the defined
lengths. A cube which sides are not equal is a rectangular parallelepiped. Its
volume, for example, can be calculated as
\begin{itemize}
\item {side $a = 4.4$~cm}
\item {side $b = 3.9$~cm}
\item {side $c = 0.4$~cm}
\begin{itemize}
\item {volume = $4.4 \times 3.9 \times 0.4 = 6.864$.}
\end{itemize}
Vectors which begin in other places of the space are compared with these specimen
vectors ${\bf e}_j$ beginning in the center. They are considered to be {\em
identical} with unit vectors ${\bf e}_j$, if they are {\em collinear}. Of course,
vectors can be shorter or longer, can have opposite directions, but such
differences will be remedied by algebraic means later.
Sometimes vectors do not coincide with unit vectors, but with their linear
combinations. We suppose, that the unit vectors ${\bf e}_j$ are orthogonal
by~definition.
How can be two vectors added? Suppose that the center ${\bf 0}$ was at first
translated into a point $a$ by the vector ${\bf e}_a$ and then to the point with
coordinates $ab$ by the translation ${\bf e}_b$. There are another possibilities
how to reach the same point. We can at first make translation ${\bf e}_b$ and then
${\bf e}_a$. In textbooks of algebra you can find that the summation is a {\em
commutative} operation. This word means that the result of the operation does not
depend on the ordering of terms in the operation. It is true: The final position
in space does not contain information about the way, how it was reached. But there
is still another possibility how vectors can be added: Both vectors can act
simultaneously and the point is shifted directly in direction between both of them
as pulling a car by two ropes as on Fig.\ref{Vector action}
\begin{figure}
\caption{Vector action. Consecutive actions A and B and the simultaneous action S
of two vectors ${\bf a}$ and ${\bf b}$ lead to the same final position R}
\label{Vector action}
\linethickness{0.6 pt}
\begin{picture}(150.00,100.00) %\vector(9.00,
55.00)(9.00,85.00)
\put(9.00,85.00){\vector(0,1){0.2}}
\put(9.00,55.00){\line(0,1){30.00}} %\end %\vector(9.00,85.00)(39.00,85.00)
\put(39.00,85.00){\vector(1,0){0.2}}
\put(9.00,85.00){\line(1,0){30.00}} %\end %\vector(44.00,55.00)(74.00,55.00)
\put(74.00,55.00){\vector(1,0){0.2}}
\put(44.00,55.00){\line(1,0){30.00}} %\end %\vector(74.00,55.00)(74.00,85.00)
\put(74.00,85.00){\vector(0,1){0.2}}
\put(74.00,55.00){\line(0,1){30.00}} %\end %\vector(89.00,55.00)(89.00,85.00)
\put(89.00,85.00){\vector(0,1){0.2}}
\put(89.00,55.00){\line(0,1){30.00}} %\end %\vector(89.00,55.00)(119.00,55.00)
\put(119.00,55.00){\vector(1,0){0.2}}
\put(89.00,55.00){\line(1,0){30.00}} %\end
\put(9.00,55.00){\circle*{4.00}}
\put(44.00,55.00){\circle*{4.00}}
\put(89.00,55.00){\circle*{4.00}}
\put(44.00,10.00){\circle*{4.00}} %\vector(44.00,10.00)(74.00,40.00)
\put(74.00,40.00){\vector(1,1){0.2}} \multiput(44.00,10.00)(0.12,0.12){251}
{\line(0,1){0.12}} %\end \put(24.00,70.00){\makebox(0,0)[cc]{A}}
\put(54.00,70.00){\makebox(0,0)[cc]{B}}
\put(104.00,70.00){\makebox(0,0)[cc]{S}}
\put(55.00,35.00){\makebox(0,0)[cc]{R}}
\end{picture}
\end{figure}
\section{Matrices}
%\addcontentsline{toc}{section}{Matrices}
\label{Matrices}
Thus we need three possibilities how to write a sum of two vectors. We must have
the opportunity to write them as consecutively acting vectors or as simultaneously
acting vectors. Simultaneously acting unit vectors can be written easily as a sum
of two unit vectors in a single row. The rule is simple, elements are added in
their places:
\begin{center}
(1, 0, 0) + (0, 1, 0) = (1, 1, 0).
\end{center}
In this notation we have already $n$ simultaneously acting vectors in a row. Thus
we must write consecutive vectors in such a sum as a column of row vectors. We get
two different columns for our examples
$$\begin{array}{ccc}
\left( \begin{array}{ccc} (1, & 0, & 0) \\
(0, & 1, & 0)
\end{array}
\right) & \qquad & \left(
\begin{array}{ccc} (0, & 1, & 0) \\ (1, & 0, & 0)
\end{array} \right)\;.
\end{array}$$
Such columns of $m$~vector-rows having in each row $n$~elements are known as {\em
matrices}. The row brackets and commas are wiped out of matrices. Notice that in a
matrix its elements are arranged in columns similarly as in rows. It is thus
possible to use the convention that a matrix is formed from $n$ consecutive
$m$~dimensional vector-columns. Since we have introduced for individual columns
the lower index $j$ going from 1~to~$n$, we can use for rows of matrices the index
$i$ going from 1~to~$m$. Remember, the index $i$ is present in texts implicitly,
as the natural order of consecutive symbols. It need not to be given explicitly.
Sometimes it is convenient to let both indices start from zero. Then they go to $
(m~-1)$ or to $(n~-1)$, respectively. It can be found one matrix index written
over the other one. But it is better to reserve the upper index for powers. When
two same symbols follow, for example ${\bf aa}$, is written shortly ${\bf a}^2$.
Doing so, we treat consecutive vectors as if they were multiplied, and the
multiplication is a {\em noncommutative} operation. The result depends on the
ordering of terms in the operation. We will not use any symbol for multiplication
of~vectors and or matrices.
We have in our examples the small round brackets in all rows of matrices within
the larger brackets used for matrices. Matrices are also bordered by double
vertical lines or they are written into a frame. We will write them sometimes
without any trimmings, but when they touch, we separate them by simple lines.
$$\begin{array}{ccccc}
\left(
\begin{array}{cc} 0 & 0 \\
1 & 1
\end{array}
\right) & \qquad &
\left( \begin{array}{cc} 1 & 1 \\
0 & 0
\end{array}
\right) & \qquad & \left(
\begin{array}{cc} 1 & 0 \\
1 & 0
\end{array}
\right)
\end{array}\;.$$
Matrices with empty rows are, maybe, superfluous, since no action corresponds to
the given row, but notice that the third matrix can be obtained from the second
one by rotating the elements around to the main diagonal, or changing the row $i$
and column $j$ indices. Matrices ${\bf M}$ are transposed into matrices ${\bf
M}^{\rm T}$. A matrix with two identical unit vectors in consecutive rows can be
interpreted as two consecutive translations going in the same direction. The
resulting position in space can be obviously described by the vector $(2,~0)$. But
if we try to interpret this vector with another numbers than 0 and 1, keeping in
mind our convention that vectors in a row are simultaneous, we have some
difficulties with the interpretation of these elements. We can imagine that the
translation requires a greater force to be performed and that it has double
intensity as in music a forte. To be consistent, we can not interpret other matrix
elements than 0 and 1 simply as the length of a~vector, unless we introduce such
vectors by some algebraic operation which will make such multivectors allowed
elements in our space.
Only after we get acquainted with all naive matrices by counting them,
we will study their sums and differences, that means the properties
of matrices having in each row a sum or a difference of two unit
vectors. Before we move to matrix arithmetic we are going to learn,
how to operate with matrices. At first we introduce matrix
products.
Having two vectors ${\bf a}$, ${\bf b}$, we can find mutual
projections of both vectors as on Fig. \ref{Scalar
Products}.
$$\begin{array}{rrr|r} & & & 3 \\ & & & 1 \\ & & & 0 \\ \hline & & &
\\ 1 & 1 & 1 & 4 \end{array}$$
The result was obtained as $1 \times 3 + 1 \times 1 + 1 \times 0 =
4$. Otherwise multiplying matching elements vertically
$$\begin{tabular}{cr} (1, 1, 1) & $\times$ \\ (3, 1, 0) & \\ \hline &
\\ (3, 1, 0) & = 4 \end{tabular}$$
Changing the row and column position, we get the same result
$$\begin{array}{rrr|r} & & & 1 \\ & & & 1 \\ & & & 1 \\ \hline & & &
\\ 3 & 1 & 0 & 4 \end{array}$$
$$\begin{array}{ccc} \begin{array}{rrr|r} & & & 1 \\ & & & 1 \\ & & &
1 \\ \hline & & & \\ 1 & 1 & 1 & 3 \end{array} & {\rm and} &
\begin{array}{rrr|r} & & & 3 \\ & & & 1 \\ & & & 0 \\ \hline & & & \\
3 & 1 & 0 & 10 \end{array} \end{array}$$
The scalar products can be made from matrix vectors, too. Scalar
products of matrix vectors multiplied by vector-rows from the left
are just vector-rows and matrix vectors multiplied by vector-columns
from the right are just vector-columns, respectively. The
multiplication is decreasing the dimensionality of the matrix
vector:
The quadratic form ${\bf J}^{\rm T}{\bf J}$ counts the elements of
the unit vector $\bf {J}$. It is simultaneously an operator
we get the sum of elements of the matrix ${\bf M}$. ${\bf J}^{\rm T}
{\bf M}$ is a $n$ dimensional vector row and ${\bf MJ}$ is $m$
dimensional vector column. The next multiplication by ${\bf J}$
(or by ${\bf J}^{\rm T}$) sums the elements of these vectors,
respectively.
We supposed that the original matrix ${\bf M}$ had $m$ rows and $n$
columns, both $m$ and $n$ being different. Therefore the transposed
matrix ${\bf M}^{\rm T}$ had $n$ rows and $m$ columns and was
different from the original matrix ${\bf M}$. We say that such
matrices are {\em asymmetrical}. Both quadratic forms are symmetrical
matrices. ${\bf M}^{\rm T}{\bf M}$ has $n$ rows and $n$ columns,
${\bf M}{\bf M}^{\rm T}$ has $m$ rows and $m$ columns. On~the traces
of both product matrices there are the sums of squared elements
$m_{ij}^2$ of the matrix ${\bf M}$. This is the Hilbert length of the
matrix vector and both traces which have the same length lie on a
sphere with the diameter of~the matrix vector. Off diagonal elements
of both quadratic forms form with their traces the right triangles
having both unit projections ${\bf J}^{\rm T}{\bf M}$ and ${\bf
MJ}^{\rm T}$ as hypotenuses (Fig. \ref{Matrix vector system}).
(0) --- (1) --- (2) --- (3) --- (4) --- (5)
(0)$\longrightarrow$(1)$\longrightarrow$(2)$\longrightarrow$(3)
$\longrightarrow$(4)$\longrightarrow$(5)
we see that the only difference is, that the vector scale is oriented
and the number scale is not.
We have introduced the unit vectors ${\bf e}_j$ in Sect. 1.2 as basic
units of our space. At first, we will allow only positive
translations corresponding to natural numbers. This means that a
matrix vector can go only forwards from the~center of coordinates and
never back. A string of consecutive vectors forms a {\em path}. All
possible paths in this space form a~{\em lattice}. We already know,
that the distinction must be made between a path and its final point,
the {\em position vector}. This distinction is the same as between
reading and mere counting words. Suppose that we have two vector
strings for example $${\bf aababac}$$
where we need to distinguish the meaning of terms $2a$ and $a^2$. The
multiplier gives the number of~the~strings, the power determines the
length of~the~vector. Now it is convenient, that the base of the unit
vectors is 1. The upper indexes, having meaning as powers of the unit
vectors do not change them. When we accept that $x^0~=~$1, the zero
power vector is just a multiplier one. Thus it is not necessary to
write this 1, because it does not change the product as $a\times$ 1
$=$ 1$\times a = a$.
$$\begin{array}{cccc} 1 & a & a^2 &\ldots \\ b & ab & a^2b & \ldots
\\ b^2 & ab^2 & a^2b^2 & \ldots \\ \vdots & \vdots & \vdots & \ddots
\end{array}$$
This square can be multiplied later by the third axis and the 3
dimensional cube is obtained, then the fourth axis can be applied and
so higher dimensional cubes, sometimes called hypercubes, are
obtained. We could speak about hyperplanes, hyperedges and so on, but
we will not use this prefix because it would hyperinflate our
text.
The individual products in the sum are vector strings ending on the
lines orthogonal to the diagonal vector ${\bf I}$. The square with
the side $0-2$ is obtained from these points by truncating points
$a^3$ and $b^3$ from the incomplete layer $3$ and adding 6 strings
$a^2b^2$ from the fourth level product $(a + b)^4$:
$$\begin{array}{ccc} 1 & a & a^2 \\ b & 2ab & 3a^2b\\ b^2 & 3ab^2 &
6a^2b^2\;. \end{array}$$
The sum of n~unit vectors ${\bf e}_j$ multiplied m~times is the {\em
generator} of the~vector space. When the three dimensional generator
is applied, the vector strings go to the triangle planes (Fig.
\ref{Three dimensional}).
Our vectors were limited only to natural numbers and therefore planes
are generated by the operator \begin{equation} [\,\sum^n_{j=1}\,{\bf
e}_j]^m \end{equation}
are the elements of the {\em natural space}. It includes its limit,
points with the coordinates $a^0b^0$, $ab^0$, $a^0b$ and so on. The
elements of the natural space are countable and are formed by the
vector strings going to points with nonnegative coordinates. We will
call the individual layers the {\em plane simplexes}. If you have
heard something about simplexes than you know that a simplex in
$n$~dimensional space should be determined by $(n~+~1)$~points and we
have just $n$ points. But remember that we speak about planes. A
plane in~$n$~dimensional space is a body\footnote{My son suggested
here to add the adjective `solid'. But a solid body is a solid body,
whereas the plain term `body' includes a code, a system, thus it is a
more abstract notion.} in~$(n~-~1)$~dimensional space and the missing
point is restored.
The planes mentioned above are orthogonal to the {\em diagonal unit
vector} ${\bf I}$. It is necessary to explain, why there are three
unit vectors: ${\bf I}$, ${\bf J}$ and ${\bf J}^{\rm T}$. We have
shown that the unit vector row ${\bf J}^{\rm T}$ and the unit vector
column ${\bf J}$ have different effects on the naive matrices ${\bf
N}$, which are basic elements of~our space, or generally on any
matrices ${\bf M}$. They transform them into vector rows or columns,
respectively. Therefore we need a new unit vector invariant to the
matrices. This vector is the unit diagonal vector ${\bf I}$. It is
the square matrix having the unit elements on the diagonal, where
both indices are equal, $i~=~j$.
When the unit diagonal matrix ${\bf I}$ multiplies any matrix either
from the left or from the right, it leaves the matrix unchanged:
\begin{equation} {\bf IM} = {\bf MI} = {\bf M}.
\end{equation}
The unit diagonal matrix ${\bf I}$ is known as the {\em identity
matrix} and was already mentioned its sophisticated formulation as
the Kronecker symbol $\delta_{ij}$, where $\delta_{ij} = 1$, if $i=j$,
and $\delta_{ij} = 0$ otherwise.
and you will see that we just inverted the ordering of formal
operations. The multiplication in (\ref{multiplication}) is done by
upper index i. But we obtained another kind of space. Our space of
vector strings is {\em noncommutative}, whereas the space formed by a
lattice of points is {\em commutative}. The transition between both
spaces is made by finding of scalar products. This formal operation
corresponds to logical abstraction as it was shown in the previous
Chapter.
The higher simplexes differ from lower ones not only by an addition
of a~new edge but also by increased number of strings leading to all
points except vertices.
$$\begin{array}{ccccl} a & b & c & d & 0 \\* a & b & c & 0 & e \\* a
& b & 0 & d & e \\* a & 0 & c & d & e \\* 0 & b & c & d & e\;.
\end{array}$$
Until now we used only natural numbers and vectors. But we will need
fractional numbers and vectors, too. Now we are able to introduce
them because we have enough space for necessary constructive
operations.
The position vector (1,1) goes through the plane simplex $(a + b)^1$
in a~point which has until now no name in our world. We introduce it
by finding its coordinates on both axes. This is done using parallel
lines with both axes. The new numbers are defined as the ratio of the
coordinate $a$ of the~position vector and the power of its simplex,
or as the ratio of the coordinate $b$ of the position vector and the
power of its simplex, respectively. In the example the~ratio is
1/2.
When this operation is done with all simplexes going to infinity (or
equivalently with the infinite simplex), we obtain infinite many
points in the interval $<0,~1>$. All these points are countable by
indices i of the infinite plane. They are known as {\em rational
numbers}. The rational numbers outside the interval $<0,~1>$ are
obtained by adding the rational number and the natural number (or by
multiplying).
New position vectors appear on it. They cross the unit simplex in
points, which all lie before the first rational number. They divide
the angle between the first countable rational vector from the
primary infinite division of~the~unit interval and therefore form a
new set of infinite many points on~the~number scale. The operation
can be repeated ad infinitum. The first set of the irrational
numbers is sufficient for representation of {\em continuum}. Its
elements are not be countable because the infinite store of numbers
is exhausted by counting the first crop of the rational numbers. The
uncountable numbers of the second crop are {\em irrational
numbers}.
We have shown that a matrix vector ${\bf M}$ can be projected onto
the unit vector row ${\bf J}^{\rm T}$ or column ${\bf J}$, and that
the quadratic forms ${\bf M}^{\rm T}{\bf M}$ and ${\bf MM}^{\rm T}$
can be separated into right triangles. This is true for matrix
vectors in which all elements are either positive or negative. If a
matrix vector contains numbers of both signs, its projections are
shorter than the matrix vector itself. Then the hypotenuse of the
right triangle (Fig. 1.2), represented by the trace of~the~quadratic
form, is longer than the outer product, where the off-diagonal
elements form the legs. The off-diagonal elements can be either
positive or negative. For example:
You have probably heard about imaginary numbers i, square roots from
the negative number $\sqrt{-1}$. When they appeared as possible
solutions of quadratic equations, mathematicians feared them as
ghosts. Only later Euler showed, how they can be exorcised by mapping
them onto a complex plane (Fig. \ref{Complex numbers}).
Because strings ${\bf x}_a{\bf x}_b$ and ${\bf x}_b{\bf x}_a$ are
undistinguishable in the commutative process, it was considered as
impossible to formulate a generating function which would exhibit the
order of symbols in products (permutations). Nevertheless, the
enumerators are easy to find in the form
The formula
or
The von Neumann set model generates numbers by counting sets. The
empty set \{0\} has one element, it generates 1. The set containing
\{0, 1\} has two elements, it generates 2, and so
on.
Number notations had different forms: In the primitive form, one cut
on a stick corresponded to each counted object. Egyptians introduced
specific signs for powers of 10 to $10^7$, but numerals one to nine
expressed primitively by the corresponding number of signs.
Phoenicians introduced letters for 1 -- 9, 10 -- 90 and 100 -- 900. It
shortened the notation considerably. This system has been taken over
by Hebrews and Greeks. Romans used their own system. Specific symbols
were reduced on I, V, X, L, C, D, and M and the number of necessary
symbols in one number by using a position system IV = one hand
without one, IX = two hands without one. Finally, we have
Indian-Arabic decimal position system.
The empty set \{0\} is for obvious reasons unsuitable as the modular
set \{m\}.
The set \{1\} as the modular set $\{m\}$ generates the natural number
0, only, since $$\{n\}\ {\rm mod}\ \{1\} \equiv 0\ .$$
Although the set \{1\} seems to be a natural base for a number system,
and the objects in sets already exist in such a form, a series of
modular comparisons with 1 gives only a series of zeroes. A division
by 1 does not decrease the digit size of a number and it does not
compress the notation. Therefore, such a~number system is
impractical. The binary system is the first
applicable.
\begin{center} \begin{tabular}{lrrrr} ***** & mod & **: & ** & \ \\*
\ & \ & \ & ** & \ \\* Rest & \ & \ & * & = 1. \end{tabular}
\end{center}
One column of the complete rows is transposed into the row and
the~operation is repeated
\begin{center} \begin{tabular}{lrrl} ** & mod & **: & ** \\* Rest & &
& \ 0 = 0. \end{tabular} \end{center}
The result is the binary notation ***** = 101. The third modular
operation was in fact the division by the second power of 2, the
third rest gives the~number of fours in the original set. In the
binary notation, they are determined by their third position from the
last digit giving the number of $1 = 2^0$. A number of a smaller
modulo is simultaneously a number of a greater modulo. The binary
number four looks like the decadic number hundred (4 = 100)
.
Two natural numbers are equal if they are obtained from the same set
$\{n\}$, and comparable if they are determined using the same modular
set $\{m\}$.
Compared with The von Neumann set model, where joined sets \{\{0\},
\{1\}\} produce the number 2, here the generating set \{2\} covers
the numbers 0 and 1.
\section{Introduction} \label{Introduction 3}
For example
The unit vectors ${\bf e}_j$ are {\em primary} vectors, their sums or
differences ${\rm s}_{ij}$ are {\em secondary} vectors. Their space
is bent in angle $45^0$ to the primary space. To each sum $(i + j)$
belong two differences, $(i - j)$ and $(j - i)$.
The sum of two unit vectors $({\bf e}_j + {\bf e}_i)$ is orthogonal
to the difference $({\bf e}_j - {\bf e}_i)$ and the corresponding
matrices ${\bf G} = {\bf N}_1 + {\bf N}_2$ differ from matrices ${\bf
S}$ only by positive signs relating both unit vector strings. Since
each consecutive element in the string is orthogonal, the ${\bf G}$
represent vectors orthogonal to~the~operators ${\bf S}$. Matrices
${\bf G}$ are linear vectors orthogonal to the surface
of~the~operator ${\bf S}$. They form a secondary vector space, which
is not complete, as we will see in the second part of this
book.
$$\begin{tabular}{rrrrcrrrr} \ & \ & \ & \ & \vline\ & 0 & 1 & 0 & 0
\\ \ & \ & \ & \ & \vline\ & 0 & 0 & 1 & 0 \\ \ & \ & \ & \ & \vline\
& 0 & 0 & 0 & 1 \\ \ & \ & \ & \ & \vline\ & 1 & 0 & 0 & 0 \\ \hline
& & & & \vline\ & & & & \\ 1 \ & 0 & 0 & 0 & \vline\ & 0 & 1 & 0 & 0
\\ 1 \ & 0 & 0 & 0 & \vline\ & 0 & 1 & 0 & 0 \\ 0 \ & 1 & 0 & 0 &
\vline\ & 0 & 0 & 1 & 0 \\ 0 & 0 & 1 & 0 & \vline\ & 0\ & 0\ & 0\ & 1
\end{tabular}$$
The first column appears as the second one in the product since
the~matrix ${\bf P}$ has 1 in the second column in the first row. The
last (zero) column is similarly shifted on the first place by the
last unit element in the first column.
The multiplication from the left changes the ordering of rows of the
multiplied matrix. For example
$$\begin{tabular}{rrrrcrrrr} \ & \ & \ & \ & \vline\ & 1 & 0 & 0 & 0
\\ \ & \ & \ & \ & \vline\ & 1 & 0 & 0 & 0 \\ \ & \ & \ & \ & \vline\
& 0 & 1 & 0 & 0 \\ \ & \ & \ & \ & \vline\ & 0 & 0 & 1 & 0 \\ \hline
& & & & \vline\ & & & & \\ 0 \ & 1 & 0 & 0 & \vline\ & 1 & 0 & 0 &
0\\ 0 \ & 0 & 1 & 0 & \vline\ & 0 & 1 & 0 & 0 \\ 0 \ & 0 & 0 & 1 &
\vline\ & 0 & 0 & 1 & 0 \\ 1 & 0 & 0 & 0 & \vline\ & 1\ & 0\ & 0\ & 0
\end{tabular}$$
When we have a number, say 5, we can define its inverse element again
by two modes, additive and multiplicative. Similar elements can be
defined for vectors.
$$\begin{array}{ccc} \begin{tabular}{rrrcr} & & & \vline & 3 \\ & & &
\vline & 1/2 \\ & & & \vline & 1 \\ \hline & & & \vline & \\ 1/6 & 1
& 0 & \vline & 1 \end{tabular} & \qquad & \begin{tabular}{rrrcr} & &
& \vline & 1/6 \\ & & & \vline & 1 \\ & & & \vline & 0 \\ \hline & &
& \vline & \\ 3 & 1/2 & 1 & \vline & 1 \end{tabular} \end{array}$$
But such inverses have one essential disadvantage: They are not
unique. There exist infinite many such inverses which balance each
vector-column (or each vector-row), therefore they are {\em
undetermined}, for example another suitable solution is:
$$\begin{array}{ccc} \begin{tabular}{rrrcr} & & & \vline & 3 \\ & & &
\vline & 1/2 \\ & & & \vline & 1 \\ \hline & & & \vline & \\ 1/9 &
2/3 & 1/3 & \vline & 1 \end{tabular} & \qquad &
\begin{tabular}{rrrcr} & & & \vline & 1/9 \\ & & & \vline & 2/3 \\ &
& & \vline & 1/3 \\ \hline & & & \vline & \\ 3 & 1/2 & 1 & \vline & 1
\end{tabular} \end{array}$$
If we try to find a left (right) inverse for a matrix, its rows must
be left (right) inverses for corresponding columns (rows), but
simultaneously zero vectors for other columns (rows). In the given
case the zero
is again undetermined:
$$\begin{array}{ccc} \begin{tabular}{lrrcr} & & & \vline & 3 \\ & & &
\vline & 1/2 \\ & & & \vline & 1 \\ \hline & & & \vline & \\ 1 & 0 &
-3 & \vline & 0 \\ -4/3 & 2 & 3 & \vline & 0 \end{tabular} & \qquad &
\begin{tabular}{rrrcrr} & & & \vline & 1 & -4/3 \\ & & & \vline & 0 &
2 \\ & & & \vline & -3 & 3 \\ \hline & & & \vline & \\ 3 & 1/2 & 1 &
\vline & 0 & 0 \\ & & & & & \end{tabular} \end{array}$$
Another difficulty with inverse elements of vectors is, that one can
not find a right inverse to a vector-column (left inverse to a
vector-row):
There were 1/3 as the first inverse element ${\rm m}^{-1}_{ij}$, but
it can not be nullified in following rows of the first column. For
its nullification we needed some nonzero elements in the second and
third columns of the left matrix. For matrix vectors we can, at least
sometimes, find matrices which transform all their vector columns
into a diagonal matrix. One vector column does not have any inverse
from the right, but their system has. Which properties a matrix must
have for being inversifiable will be shown later. If a matrix has
inverses both from the left and from the right, then both inverses
are identical and there exist only one inverse which action is equal
from both sides. This is the true inverse of the given
matrix.
The first column is the zero eigenvector, all values in its column
product are zero, and the second eigenvector eigenvalue is 3, the
eigenvalue of the last eigenvector is 1. There is yet another
condition on eigenvectors, see the next Section.
$$\begin{tabular}{ccccccc} & & & \vline & 1 & & \\ & & & \vline & -2
& 1 & \\ & & & \vline & 1 & -2 & 1 \\ \hline & & & \vline & & & \\ 1
& & & \vline & 1 & & \\ 2 & 1 & & \vline & & 1 & \\ 3 & 2 & 1 &
\vline & & & 1 \\ \end{tabular}$$
Here the second row balances are $1 \times 2 - 2 \times 1 = 0$, and
$1 \times 1 = 1$. In detail, the problem of inverse matrices will be
treated in Chapt. 16.
or equivalently when their action does not change the unit diagonal
matrix in their frame:
Then, if moreover
The vector set used for finding of eigenvalues gives the diagonal
matrix, but not the unit matrix ${\bf I}$:
$$\begin{tabular}{ccccccc} & & & \vline & 1 & 1 & 1 \\ & & & \vline &
1 & -2 & 0 \\ & & & \vline & 1 & 1 & -1 \\ \hline & & & \vline & & &
\\ 1 & 1 & 1 & \vline & 3 & 0 & 0 \\ 1 &-2 & 1 & \vline & 0 & 6 & 0
\\ 1 & 0 &-1 & \vline & 0 & 0 & 2 \\ \end{tabular}$$
The row indices go in each matrix from 1 till $m$, the column indices
go in each matrix from 1 till $n$. This is internal counting.
Similarly as Jewish, Christian and Islamic epochs, sets of indices in
compared matrices can be unequal, or one set be the same, or both
sets can match. Thus the rule of matrix arithmetic for addition and
subtraction of matrices is simply addition and subtraction of
individual matrix elements according the rule:
The unit diagonal matrix forms root to itself, since we can not
distinguish forms
\chapter{Partitions} \label{Partitions}
A partition splits a number $m$ into $n$ parts which sum is equal to
the~number $m$, say 7:\ \ 3, 2, 1, 1. A partition is an ordered set.
Its objects, {\em parts} are written in a row in decreasing order: $$
m_{j-1} \geq m_j \geq m_{j+1}\;.$$
Ferrers graphs are used in the theory of partitions for many proofs
based simply on their properties. Ferrers graphs are tables (see Fig.
\ref{Ferrers Graphs}) containing $m$ objects, each object in its own
box. The square boxes are arranged into columns in nonincreasing
order $m_j \geq m_{j+1}$ with the sum
We write here first the number of rows $m$, then the number of parts
$n$, and last the sum of unit elements (the number of filled boxes)
M.
All partitions into exactly $N$ parts are divided into two sets. In
one set are partitions having in the last column 1, their number is
counted by the~term $p(*,N-1,M-1)$ which is the number of partitions
of the number $(M-1)$ into exactly $(N-1)$ parts to which 1 was added
on the n~th place and in other set are partitions which have in the
last column 2 and more. They were obtained by adding the unit row
${\bf J}^{\rm T}$) with $n$ unit elements to the~partitions of $(M -
N)$ into $N$ parts.
The term p(*, *=N-1, M) are partitions of $M$ into $(N - 1)$ parts
transformed in partitions into $N$ parts by adding zero in the n-th
column, the~term p(*, *=N, M-N) are partitions of $(M - 1)$ into $N$
parts to which the~unit row was added.
Table 4.2 is obtained from the Table 4.1 as partial sums of its rows,
it means, by multiplying with the unit triangular matrix ${\bf T}^T$
from the right. The elements of the matrix ${\bf T}^{\rm T}$
are
On the other hand, the Table 4.1 is obtained from the Table 4.2 by
multiplying with a matrix ${\bf T}^{\rm -T}$ from the right. The
inverse elements are
Notice, that the elements of the Table 4.2 right of the diagonal
remain constant. They are equal to the row sums of the Table 4.1.
Increasing the number of zeroes does not change the number of
partitions.
When we multiply Table 4.1 by the matrix ${\bf T}^{\rm T}$ again, we
obtain partitions having as the smallest allowed part the number 2.
The effect of these operators can be visualized on the 2~dimensional
complex, the operators shift the border of counted orbits (Fig.
\ref{Limiting}). The operator ${\bf T}^{\rm T}$ differentiates
n~dimensional complexes, shifting their border to positive numbers
and cutting lover numbers. Zero forms the natural base
border.
If the number of equal parts $n_k$ is written as the vector row under
the vector formed by the~number scale, the number of partitions is
independent on shifts of the number scale, see Table 4.3. Partitions
are always derived by shifting two vectors, one 1 position up, the
other 1 position down. Each partition corresponds to a vector. If we
write them as columns then their scalar product with the number scale,
forming the vector row ${\bf m}^{\rm T}$, gives constant sum:
The parameter $r$ shifts the table of partitions, its front rotates
around the~zero point. If $r$ were $-\infty$, then $p(-\infty, 1) =
1$ but $p(-\infty, 2)$ were undetermined, because a sum of a finite
number with an infinite number is again infinite. The parameter $r$
will be written to a partition as its upper index to show that
different bases of partitions are differentiating plane
simplexes.
Notice how the scarce matrix of odd partitions is made from Table
4.1. Its elements, except the first one in each column, are shifted
down on cross diagonals. An odd number must be partitioned into an
odd number of odd parts and an even number into even number of odd
parts. Therefore the~matrix can be filled only in half. The
recurrence is given by two possibilities how to increase the number
$m$. Either we add odd 1 to odd partitions of~$(m - 1)$ with exactly
$(j - 1)$ parts or we add $2j$ to odd numbers of partitions of $(m -
2j)$ with exactly $j$ parts. The relation is expressed as
$$\begin{tabular}{rrrr} 10 & & & \\ & 9,1 & & \\ & 8,2 & & \\ & 7,3 &
7,2,1 & \\ & 6,3 & 6,3,1 & \\ & & 5,4,1 & \\ & & 5,3,2 & \\ & & & 4,3,
2,1 \end{tabular}$$
The differences are due to Franklin blocks with growing minimal parts
and growing number of parts (their transposed notation is used),
which are minimal in that sense that their parts differ by one, the
shape of corresponding Ferrers graphs is trapeze:
we find rather easily its inverse matrix which is given in the second
part of~the same Table.
The nonzero elements in the first column of the Euler inversion (and
similarly in the next columns which are only shifted down one row)
appear at indices, which can be expressed by the Euler identity
concerning the coefficients of expansion of
For example the last row of the partition Table \ref{Partitions and
their Euler inversion} is eliminated by multiplying it with the Euler
inversion as: $$(7 \times 1) + (5 \times -1) +( 3 \times -1) + (2
\times 0) + (1 \times 0) + (1 \times 1) = 0.$$
The Euler function has all parts $t^i$ different. We have constructed
such partitions in Table 4.5. If the coefficient at $t^i$ is obtained
as the product of~the~even number of $(1 - t^i)$ terms then the sign
is positive, and if it is the result of~the~uneven number of~terms
then the sign is negative. The coefficients are determined by the
difference of the number of partitions with odd and even number
of~unequal parts. This difference can be further explained according
to Franklin using Ferrers graphs.
Whereas the columns of the Table 4.8 are irregular and elements of
each column must be found separately, columns of the Table 4.9 repeat
as they are only shifted in each column one row down, similarly as
the elements of~their parent matrix are. They can be easily found by
multiplying the matrix of~the~Euler function (Table 4.7) by the
matrix ${\bf T}$ from the left.
The Equation 4.3 can be applied for cubes. It shows their important
property, they are symmetrical along the main diagonal, going from
the center of the coordinates, the simplex $n^0$ to the most distant
vertex of the cube in~which all $n$ coordinates are $(m-1)$. The
diagonal of~the~cube is represented on Table 4.10 by $k$ indices.
Moreover, a cube is convex, therefore
and if
and so on till
When the multiplication's for all parts are made and terms on
consecutive plane simplexes counted, we get:
& 1 \\ \hline $\Sigma$ & 1 & 6 & 14 & 18 & 18 & 14 & 11 & 7 & 5 & 3&
2 & 1 & 1 \\ \hline \end{tabular} \end{table}
We have shown in Sect 4.4 that one orbit can be obtained from another
by shifting just two vectors, one up and other down on the number
scale. We can imagine that both vectors collide and exchange their
values as two particles of the ideal gas exchange their energy. If we
limit the result of such an exchange to 1 unit, we can consider such
two orbits to be the nearest neighbor orbits. The distance inside
this pair is $\sqrt{2}$. We connect them in~the scheme by a~line.
Some orbits are thus connected with many neighbor orbits, other have
just one neighbor, compare with Fig. 5.1. Orbits (3,3,0) and (4,1,1)
are not nearest neighbors, because they must be transformed in~two
steps: $$(3,3,0) \leftrightarrow ((3,2,1) \leftrightarrow (4,1,1)$$
or
or
The number of right hand neighbors is the sum of two terms. The right
hand diagonal neighbors exist for all $p(m,n-1)$. We add 1 to all
these partitions and decrease the largest part. Undetermined remain
right hand neighbors in rows. Their number is equal to the number of
partitions $p(m-2)$. To each partition $p(m-2, n-1)$ are added two
units, one in the n~th column, the second in the (n-1)~the
column.
$$\begin{array}{cccc} x & x & x & x \\ x & & & \\ x & & & \\ x & & &
\end{array}$$
The values in brackets are the numbers for partitions which lie
inside the L frame having $(2k-1)$ units. At higher diagonal layers
appear these possibilities to add new elements later. Partitions 4, 4,
4 and 3, 3, 3, 3, for $n=12$, are counted in the seventh layer. For
$n=13$, the layer counts seven partitions:
$$\begin{array}{ccccc} 5,5,3; & & \\ 5,4,4; & & \\ & 4,4,4,1; & \\ &
4,4,3,2; & \\ & 4,3,3,3; & \\ & & 3,3,3,3,1; \\ & & 3,3,3,2,1.
\end{array}$$
$$\begin{array}{ccccc} air & & {\bf humid} & & water \\ \\ & {\bf
warm} & 0 & {\bf cold} & \\ \\ fire & & {\bf dry} & & earth\;.
\end{array}$$
The elements have always only two properties. The properties adjacent
vertically and horizontally exclude themselves. Something can not be
simultaneously warm and cold, or humid and dry\footnote { More
precisely, it is necessary to draw a borderline (point zero) between
these properties. Depending on its saturation, water vapor can be dry
as well as wet.}.
The divisors form a regular pattern, they are in all rows $i\ \equiv
0\ {\rm mod}\ j$. The prime numbers become scarcer, as the matrix
grows, but it is always possible to find another prime number $p(n)$
as the product of all n previous prime numbers \begin{equation} p(n)
= \prod^n_{j=1}\; p_j +1\;. \end{equation}
This equation does not generate all prime numbers. Between $p(2)=3$
and $p(3)=7$ is $p=5$.
The row sums of the Erasthothenes sieve $({\bf EJ})$ are the numbers
of divisors. They appear on the diagonal of the quadratic form ${\bf
EE}^{\rm T}$ of the matrix ${\bf E}$. They are known as the {\em
Euler function} $\sigma^0\;(n)$. This function is related with
logarithms of divisors. If we use as the base of logarithms the
number $n$ itself, we get (except n = 1)
The divisors appear in pairs, $d_id_j = n$, except the lone divisor
which~is the~square root of $n$. The sum of logarithms with the base
$n$ is thus only a~half of the number of divisors of~the~number $n$.
The sum of divisors values $\sigma^1 (n)$ sometimes gives twice the
number itself as $2\times6 = 6+3+2+1$ or $2\times28 = 28+14+7+4+2+1$.
Such numbers are known as the {\em perfect numbers}.
In Table 6.1 the Moebius function was shown as the inverse matrix
${\bf E}^{-1}$. The~elements of its first column are
where [n/j] means the whole part of the given ratio. Therefore the
sum $\sum[\sigma^0(n)]$ has as a limit the product $n\sum_{j=1}^n\;
n/j$. For example $$\sum[\sigma^0(3)] = 5 < 3(1 + 1/2 + 1/3) = 11/2\;
.$$
The row elements of the previous matrix ${\bf M}$ are ${\bf J}^{\rm
T}{\bf E}$, thus the Moebius function is ${\bf M}^{-1}{\bf J}$.
A still more important function is the sum of the divisor values. It
can be expressed as the matrix product having in the frame ${\bf E}(*)
{\bf E}^{\rm T}$ the diagonal matrix of indices $\Delta(j)$. ${\bf
E}\Delta(j)$ is the matrix of divisor values. The sums of~divisor
values $\sigma^1 (n)$ are the diagonal elements of the matrix ${\bf
E}\Delta(j){\bf E}^{\rm T}$:
The number of divisors $j$ which also gives ratios $n/d$ is obtained
as another matrix product:
The rows of ${\bf E}$ are multiplied by the corresponding index $i$
and the columns are divided by the corresponding index $j$. The
elements of the matrix product are $e_{ij} = i/j$, if $i\ \equiv\ 0\
{\rm mod}\ j$, and $e_{ij} = 0$ otherwise.
$$\left( \begin{array}{ccccc} 1 & & & & \\ 2 & 1 & & & \\ 3 & 0 & 1 &
& \\ 4 & 2 & 0 & 1 & \\ 5 & 0 & 0 & 0 & 1 \end{array} \right)\;.$$
For example 1 in the sixth row divides 1,5; 2 divides 2,4; 3 and 6
divide themselves and 4 and 5 are not divisors.
This inverse function has again the table form (see Table
\ref{Inverse function of numbers of numbers}).
where $p$ are prime numbers that are divisors of $n$. The ratio $n/p$
is split from the number $n$ by each inverse of the prime number
$1/p$. The sum counts all subtracted parts from the total $n$. The
function $\varphi(n)$ of the product of~two numbers is simply the
product of the values for each number
$$\begin{array}{ccc} \left( \begin{array}{ccccc} 5 & & & & \\ & 3 & &
& \\ & & 2 & & \\ & & & 1 & \\ & & & & 1 \end{array} \right) & \left(
\begin{array}{ccccc} 1 & & & & \\ 1 & 1 & & & \\ 1 & 0 & 1 & & \\ 1 &
1 & 0 & 1 & \\ 1 & 0 & 0 & 0 & 1 \end{array} \right) & \left(
\begin{array}{ccccc} 1 & & & & \\ & 2 & & & \\ & & 3 & & \\ & & & 4 &
\\ & & & & 5 \end{array} \right) \end{array}$$
$$\begin{tabular}{rrrrr|rrrrr} & & & & & 1 & 1 & 1 & 1 & 1 \\ & & & &
& & 2 & 2 & 2 & 2 \\ & & & & & & & 3 & 3 & 3 \\ & & & & & & & & 4 & 4
\\ & & & & & & & & & 5 \\ \hline & & & & & & & & & \\ 1 & & & & & 1 &
1 & 1 & 1 & 1 \\ 2 & 1 & & & & 2 & 4 & 4 & 4 & 4 \\ 4 & 1 & 1 & & & 4
& 6 & 9 & 9 & 9 \\ 7 & 3 & 1 & 1 & & 7 & 13 & 16 & 20 & 20 \\ 12 & 4
& 2 & 1 & 1 & 12 & 20 & 26 & 30 & 35 \end{tabular}$$
The left matrix counts the numbers $n_k$ in the partitions, the right
matrix weights them as $m_k$.
The numbers $mp(m)$ appear again on the diagonal of the product. This
elementary relations can be expressed in a formal abstract form. We
write the generating function of unrestricted partitions
Matrix elements of the Table 6.2, except its first column, are
obtained as the partial result of the matrix product used for finding
the sum of values of~parts in partitions with equation 5.3. They are
elements of products of two matrices $\Delta[p(m-i)] {\bf E}$. The
first column is in its turn again a matrix product of the matrix of
partitions into exactly $n$ parts (Table 4.2) and the matrix
of~positive $(j -i)$ elements, and the unit vector column ${\bf J }$,
which sums the~row values of the intermediate product. It is easy to
explain this relation: In each partition, when $m$ splits into
exactly $n$ parts, there are $(m - n)$ zeroes. For example for $m =
4: 8 = 3\times1 +2\times2 +1\times1 +0\times1$. The number of~zeroes
is balanced by other numbers. This leads to the simple form of some
elements of inverse matrix $$m^{-1}_{i0} = (1-i)\;.$$
The first three objects are permuted in a {\em cycle of length 3}, the
first object appeared after the third one, next two objects form a
cycle of length 2, they exchanged their places, and the last object
remained in its place. By~repeating the procedure we obtain
permutations 2 to 6: \item Step 2: \qquad 3 \qquad 1 \qquad 2 \qquad
4 \qquad 5 \qquad 6
The string returns in the 6~th step into the initial order and a new
cycle starts in the 7~th step.
The index labeling objects is the column index $j$. The position
in~the~permutation is labeled by the row index $i$ at the element
$1_{ij}$. Thus permutations are isomorphic with matrices. The
starting position, corresponding to the diagonal unit matrix ${\bf
I}$, can be considered as the zero order. The last element remained
in all steps in its original position and the first three elements
returned to their positions twice and two elements made three turns.
The~length of the total cycle is the product of individual
cycles:$3\times2\times1 = 6$. The elements belonging to the same
cycles are usually written in brackets: $(2,3,1)(5,4)(6)$.
The number of elements $n$ splits into $k$ cycles, $k$ going from 1
to $n$. The~cycle structure is described by {\em partition
orbits}.
We could map the cycle changes by the additive operators ${\bf S}$
having $-1_{ij}$ for the leaving object j, $1_{ij}$ for the becoming
object j, zero rows for unmoved objects (+1 and -1 appear on the same
place). This operator was introduced in Chapt. 3 and in more detail
will be studied in Chapt. 12. Now we will study the {\em
multiplication operators} ${\bf P}$. Their matrices, {\em unit
permutation matrices}, are naive, they have in each row only one unit
element and moreover they have only one unit element in each column.
The matrices ${\bf P}$ are simultaneously notations of permutations,
since their row unit elements $p_{ij}$ correspond to indexes (or
equivalently to alphabetical symbols) $j$.
The last but one power of any permutation matrix is its inverse (Fig.
\ref{Cycle of permutation matrices}). It is rather easy to find this
matrix, because it is identical with the transposed matrix ${\bf
P}^{\rm T}$:
The set of all permutation matrices ${\bf P }$, with $n$ rows and $n$
columns, represents all possible permutations. A special class of
permutation matrices are the symmetrical ones, which
have
The third element 1 jumps in the first column and shifts 3 down, then
2 shifts 4 down. Or:
$$\begin{array}{ccc|ccc} & & & \ 1& 0 & 0 \\ & & & 0 & 0 & 1 \\ & & &
0 & 1 & 0 \\ \hline & & & & & \\ 0 & 1& 0 & 0 & 0 & 1 \\ 1 & 0& 0 & 1
& 0 & 0 \\ 0 & 0& 1 & 0 & 1& 0 \end{array}$$ $$(a,c,b)\times(b,a,c) =
(c,a,b)$$
We will explain this function later, when we will look for other
formulas determining the number of permutations. Before then we will
study convolutions. Here an example is given how equation
(\ref{factorial}) works:
The inverse table has the same elements, only the signs of elements
which indices i differ from indices j by the value $(4k + 2)$ are
negative. Their recurrence is
where the terms in the first column of the Table 7.1 $y_{k0}$ are
considered as powers of $y^k$ when the sum $(1 + y)$ is multiplied
with itself. For example $$(1 + y)^6 = 1\times1 + 6\times0 +
15\times1 + 20\times0 + 15\times3 + 6\times0 +1\times15 = 76\;
.$$
From it, other $n/2$ values of the gamma function are calculated
easily which fits excellently in holes between factorials to plot one
smooth function (Fig. \ref{Plot of the function Gamma}).
The row sums of consecutive schemes give the Table 7.2. Its elements
are known as the {\em Stirling numbers of the first kind}. Their name
suggests that there are more kinds of Stirling numbers. They are
related by many ways as we will see later.
The recurrence is
Notice that here the parameter $k$ does not end at the number $n$ but
continues to~the~value $n(n-1)/2$. It is as if we counted these
values on~the~diagonal of a square. The maximal moment $k$ is just
the sum of values $(i-1)$ where $i$ is going from 1 to
$n$
Till now we have worked with permutations which were read from one
side, only. Most of our symbols determine from which side they must
be read (with some exceptions as W, A, T, 8, and other symmetric
symbols are). Imagine now, that a permutation is represented by a
string of colored beads as
Before we study all naive matrices ${\bf N}$, we will deal at first
with the naive matrices in the {\em lower triangular form} which form
a subgroup of naive matrices. The remaining naive matrices can be
obtained from them by permuting columns with the unit permutation
matrices ${\bf P}$ from right. Recall that naive matrices ${\bf N}$
have one unit element in each row. If a matrix is in the lower
triangular form then all its nonzero elements must be on or below the
main diagonal. Similarly, if a matrix is in the upper triangular form,
then all its nonzero elements must be on or over the main diagonal.
From all permutation matrices ${\bf P}$ only the {\em identity
matrix} ${\bf I }$ has the triangular form. But it exists
simultaneously in both triangular forms as all diagonal matrices
do.
There is only one place in the first row of the lower triangular form
for the unit element, two places are in the second row and always one
place more in each consecutive row for the unit element. This
situation is just opposite to the construction of permutation
matrices. There the possibilities, of placement of the unit element
decreased in every row. Nevertheless both approaches give the same
result. Therefore there are n! naive matrices in the lower triangular
form (or in the case of transposed naive matrices ${\bf N }^{\rm T}$
in the upper triangular form). The transposed naive matrices can be
mapped onto points with natural coordinates in m~dimensional
cubes.
For example two naive matrices $n=3$ are excluded by this rule:
${\bf A}$ is excluded since $b^2 > a$, ${\bf B}$ is excluded since $c
> b^0$.
There are $(n - 1)$ places under the diagonal in the n~th row which
can be added to each naive matrix with $k$ elements on the main
diagonal without changing $k$. This multiplies the first term.
This is not all what can be said about the Stirling numbers of the
first kind. If we multiply the Stirling numbers of the first kind
directly with powers $2^{j-1}$ we get a table which row sums are
equal to the half of the higher factorial $(i+1)!/2$ as
in
$$\begin{tabular}{rrrr|rrrrr|rrrrr} & & & & \ & 1 & 1 & 1 & 1 & \ & 1
& -1 & & \\ & & & & & & 2 & 2 & 2 & & & 1 & -1 & \\ & & & & & & & 4 &
4 & & & & 1 & \\ & & & & & & & & 8 & & & & & 1 \\ \hline 1 & & & & &
1 & 1 & 1 & 1 & &1 & & & \\ 1 & 1 & & & & 1 & 3 & 3 & 3 & &1 & 2 & &
\\ 1 & 4 & 1 & & & 1 & 9 & 13 & 13 & &1 & 8 & 4 & \\ 1 & 11 & 11 & 1
& & 1 & 23 & 67 & 75 & &1 & 22 & 44 & 8 \\ \end{tabular}$$
The row sums of the Table 7.1 are interesting. They are generated
directly by the formal equation
from it E(1) = 1 and so on. These numbers will appear later many
times, as differences of complete plane simplexes, but here they
appear as an~extension of the factorial simplexes or as the product
of the matrix of the Euler numbers with the diagonal matrix of powers
$2^{j-1}$.
\section{Mac Mahon Numbers} \label{Mac Mahon Numbers}
When we look at the sequence $aac$, we see that in its matrix the
column $b$ is missing. We excluded such strings as faulty Young
tables or convolutions, but the strings as $abb$ were also not
allowed, where $b$ appeared twice, and $a$ only once. There is a
difference in these two cases: If not all columns are occupied
successively, we jump in the space over some positions. Therefore we
will now count all naive matrices in the lower triangular form with
successively occupied columns. Their recurrence is
The row sums of Stirling numbers of the second kind, which count
naive matrices in the lower triangular form with successively
occupied columns are obtained as selfgenerating function
$$\begin{tabular}{rrrr|rrrr|rrrr} \hline & & & & \ 1 & 1 & 1 & 1 & \
1 & & & \\ & & & & & 1 & 2 & 3 & & 2 & & \\ & & & & & & 1 & 3 & & &3
& \\ & & & & & & & 1 & & & & 4 \\ \hline & & & & & & & & & & & \\ 1 &
& & & 1 & 1 & 1 & 1 &1 & 2 & 3 & 4 \\ 1 & 1 & & & 1 & 2 & 3 & 4 &1 &
4 & 9 &16 \\ 1 & 1 & 2 & & 1 & 2 &5 &10 &1 & 4 &15 &40 \\ 1 & 1 & 2 &
5 & 1 & 2 & 5 &15 &1 & 4 &15 &60 \\ \end{tabular}$$
Notice that the matrix of the Stirling sums begins with two 1 and
on~the~diagonal of the product is only one 1 and then immediately the
higher sums follow. In the final product the sums are multiplied with
corresponding powers. If we label the matrix of binomial coefficients
as ${\bf B}^{\rm T}$, then we can consider its product with the
diagonal matrix $\Delta(i)$ as the logarithmic difference $d(log{\bf
S})$, similarly as it was derived in Sect. 6.4. The inverse matrix of
sums of Stirling numbers of the second kind has elements:
where we put again $S^k = S_k$. For example: $S_6 - S_5 = 1\times1
+4\times2 + 6\times5 + 4\times15 + 1\times52 = 151$.
The Stirling numbers of the second kind are defined by the formal
relation
\begin{equation} \Delta_n^1(m)^n = m^{n-1}[(1 +1/m)^{n-1} + (1 + 1/m)
^{n-2} \dots +(1 - 1/m)^0]\;. \end{equation}
\section{Substirlings} \label{Substir}
Another possibility, how the Stirling numbers of the second kind are
obtained, is the direct count of the corresponding matrices arranged
according to the powers of a. For example:
$$\begin{array}{ccc} a & & \\ ab & aa & \\ abb,\ abc;& aab,\ aba; &
aaa \\ \end{array}$$
We have compared three statistics, but a fourth one appeared here and
that is on the diagonal of the Stirling and Euler statistics. The
Euler numbers divide naive matrices in the lower triangular form
according to the number of occupied columns. The Stirling numbers of
the second kind count matrices with rows occupied consecutively and,
these matrices appear on the crossection of both statistics. The
Euler statistics splits 6 naive matrices in the lower triangular form
with one unit element on the diagonal in three groups the Stirling
numbers evaluate differently naive matrices in the lower triangular
form with two unit elements on the diagonal.
I do not know what you think about these coincidences. The Euclidean
space is full of surprises. It seems to be alive and if we try to
analyze it, new layers appear just on elementary levels. Euclides was
wrong when he told to the king Ptolemaos that there was no other way
to his space then his axioms. Different combinatorial functions lead
through this maze as the Ariadne's thread.
The first term 4 counts strings ${\bf aaab}$, ${\bf aaba}$, $ {\bf
abaa},$ and ${\bf baaa}$, the third term 4 counts strings ${\bf
abbb}$, ${\bf babb}$, ${\bf bbab}$, and ${\bf bbba}$. The binomial
coefficient is written as the number m stacked over the number k in
the brackets
There are $n!$ permutation matrices, but not as many permuted vector
columns, when some elements of the vector row are not
distinguishable. Vectors with equal length $m_k$ point to the sphere
and if rotated, their permutations are undistinguishable. If all
elements of the vector are equal, then no permutations have any
effect on the partition vector.
We divide vector elements into two groups, one with all zero elements,
that is $n_0$ elements, and the second group with all remaining
$(n-n_0)$ elements. The number of possible permutations will be
reduced from the factorial $n!$ to the binomial coefficient ${ n
\choose n_0 }$, or $n!/n_0!(n-n_0)!$.
In the next step we single out from the second group the vectors with
the length 1, their number is $n_1$. All other vectors will be
counted by the~third term $(n- n_0 -n_1)$, and corresponding
permutations by the binomial coefficient $(n-n_0)!/n_1!(n- n_0 -n_1)
!$. In this way we proceed till all possible values of $m_k$ are
exhausted. If some $n_k = 0$, then conveniently $0! = 1$ and
the~corresponding term is ineffective. At the end we obtain a product
of binomial coefficients:
We limited the index $k$ by the lower limit 0. The coefficient could
be used actually also for vectors with negative elements. The numbers
$n_k$ of equal vectors are always positive even if the vectors
themselves are negative. We count by the polynomial coefficient
(\ref{polynomial coefficient}) points on the partition orbits of
the~positive cone of the $n$~dimensional space.
Please note the importance of this step. We know the vector ${\bf m}$
exactly, but we replace it by the corresponding partition. All points
on the given orbit are considered to be {\em equivalent}. Replacing
the vector ${\bf m }$ by the partition is a~logical abstraction. We
can proceed further, the partition is compared with an analytical
function and the orbit is described by a density
distribution.
\begin{eqnarray} \left(\frac{(m+n-2)!}{m!(n-2)!}\right) +
\left(\frac{(m+n-2)!}{(m-1)!(n-1)!}\right) = \nonumber \\
\left(\frac{(m+n-2)![(n-1)+m]}{m!(n-1)!}\right) = { m+n-1 \choose m
}\;. \end{eqnarray} \label{sum of 2}
In the first column vertices of the plane simplex are counted, in the
second column points on 2~dimensional edges, in the third column
points of its 3~dimensional sides. Only the last point lies inside of
the 6~dimensional plane, all other 461 points are on its borders.
This is a rather surprising property of~the high dimensional spaces
that the envelope of their normal plane simplexes is such big. But we
can not forget, that usually $m\gg n$ and then there are more points
inside than on the border.
The numbers $b_{i0}$ are formed by vectors without any unit elements.
They can be called {\em subplane numbers}, because they generate the
number of points of the normal plane simplex by multiplying with
binomial coefficients:
Its values $b(i,j)$ for small $m$ are: \begin{itemize} \item b(0,n) =
1; \item b(1,n) = 0; \item b(2,n) = ${ n \choose 1 }$; \item b(3,n) =
${ n \choose 1 }$; \item b(4,n) = ${ n \choose 1 }+{ n \choose 2 } =
{ n+1 \choose 2 }$; \item b(5,n) = ${ n \choose 1 } + 2{ n \choose 2
} = n^2$; \item b(6,n) = ${ n \choose 1 } + 3{ n \choose 2 } + 3{ n
\choose 3 }= (n^3-n)/2$. \end{itemize}
Till now zero elements were permuted with the other elements. We
exclude the zero element and count only existing (nonzero) vectors
and not virtual vectors. It means, that we count consecutively all
k~dimensional vectors $(k = 1$ to $n)$ with constant sums $m$. If we
draw the tetrahedron (Fig. \ref{Difference of}), then the counted set
of points is formed by one vertex, one edge without the~second vertex,
the inside of one side and by the four dimensional core.
In~combinatorics these vectors are known as {\em compositions}. They
can be arranged onto partition schemes. For $m=5$ we
get:
The inverse matrix ${\bf B}^{-1}$ to the matrix ${\bf B}$ is obtained
from the formal binomial
The vectors counted for $m=7$ are: 7; 52, 25, 43, 34; 322, 232, 223.
Notice, that the elements of the Table \ref{Fibbonacci numbers} are
binomial coefficients shifted in each column for 2 rows. Fibbonacci
numbers $F_m$ have the recurrence
In each row all elements of both preceding rows are repeated which
gives the recurrence of the Fibbonacci numbers.
For example the last row counts the compositions: $51,\ 15,\ 33;\
4\times(3111);\ 111111$.
\section{Fibbonacci Spirals} \label{Fibbonacci Spirals}
On the diagonal the values $F_{n+1}$ and $F_{n-1}$ are. This fact
gives a possibility to prolongate the Fibbonacci numbers to negative
indices. This series must be: $1, -1, 2, -3, 5, -8, \dots$. We obtain
these numbers again as the sums of the two consecutive Fibbonacci
numbers, row sums of the elements of ${\bf B }^{-1}$ or as
the~elements of their generating matrix
Both actions are independent and therefore the final result is just
the~product of both coefficients
The difference against the cycle index (Equation 7.15) is the second
factorial $m!$ and factorials of $m_k$ instead their first powers.
When we use (10.2) for the partition $1^m$ we obtain $(n!/n_1!)
(m!/1!^{n_1}) =m!$. The cycle index splits the $S_m$ group according
to the cycle structure.
\section{Differences in Power Series} \label{Differences in Power
Series}
In Table \ref{Power series}, only the first column and the row sums
are clearly connected with $m$ and $n^m$. Moreover there appear
factorials but other elements grow too fast to be analyzed directly.
But all elements are divisible by $m$. In this way the Table
\ref{Power series} is decomposed into the direct product of two
matrices. One from them is the matrix of binomial coefficients ${ m
\choose k }$. This is the matrix ${\bf B}^{\rm T}$. The other one is
the matrix of differences $\Delta^n{\bf 0}^m$:
$$\begin{tabular}{|l|c|ccc|ccc|c|} \hline n & 1 & & 2 & & & 3 & & 4
\\ \hline Basic string & aaaa &aaab& aabb &abbb &aabc& abbc &abcc
&abcd \\ Permutations & 1 & 4 & 6 & 4 &12 & 12 & 12 & 24 \\ \hline
Counts & 1 & & 14 & & & 36 & & 24 \\ \hline
\end{tabular}$$
$$\begin{tabular}{rrrr|rrrrr} & & & & 1 & 1 & 1 &1 & 1 \\ & & & & & 1
& 2 & 3 & 4 \\ & & & & & & 1 & 3 & 6 \\ & & & & & & & 1 & 4 \\ \hline
1 & & & & 1 &1 &1 &1 & 1 \\ & 1 & & & & 1 & 2 & 3 & 4 \\ & 1 & 2 & &
& 1 & 4 & 9 &16 \\ & 1 & 6 & 6 & & 1 & 8 &27 &64 \\
\end{tabular}$$
In the zero column are counted strings of the simplex which are not
in~its difference. The elements in other columns are consecutive
differences. For example the elements in $d_{30}=14$ are: $b^3$,
$c^3$, $b^3$, 3$b^2c$, 3$bc^2$, 3$a^2c$, 3$ac^2$. The column indices
correspond to the powers of the first index, for example $d_{41}=13=
ab^3 + 3ab^2c + 3abc^2 +6abcd,\ d_{42}=3 = a^2b^2 + 2a^2bc$. When we
multiply this matrix with the transposed matrix of binomial
coefficients ${\bf B}^{\rm T}$, we get on the diagonal of the product
corresponding powers $n^n$. The binomial coefficient permutes the
first vector with other already permuted vectors.
We used the operator notation many times. Now we shall explain its
notation. There exist the {\em identity function} $E$ and the {\em
difference function} $\Delta$. Moreover there are formal powers
$0^n$. These functions are defined reciprocally
as
The row sums of the Table 10.2 taken with alternating signs (the
difference of even and odd columns) gives $(-1)^i$. Let suppose that
this is true for some row. The elements of the next row are just
multiplied sums of the preceding row:
When we make the difference $d_1 -2(d_1 +d_2) +3(d_2 +d_3) - \dots =
-d_1 + d_2 -d_3 \dots$, we get the elements of the preceding row
with the other signs which sum was +/-1.
For the sums of the first rows the following identities are found
easily
It should be noted that the i-th row of the Table 10.2 is obtained
consecutively by multiplying this matrix by the ${\bf Q }$ from the
right from the (i-1)-th row. ${\bf Q }$ is the diagonal matrix of
indices which repeat once again just under the main diagonal as in
the following example
$$\begin{array}{cccc|cccc} & & & & 1 & & & \\ & & & & 1 & 2 & & \\ &
& & & & 2 & 3 & \\ & & & & & &3 &4\\ \hline 1 & & & & 1 & & & \\ & 1
& & & 1 & 2 & & \\ & 1 & 2 & & 1 & 6 & 6 & \\ & 1 & 6 &6 & 1 &14 &36
&24\;.\\ \end{array}$$
We will not analyze these recurrences, but show another one. If the
strings in plane simplexes are classified according to the number of
unit vectors $n_1$, we obtain the difference Table 10.5.
The first column elements of the Table 10.5 can be named {\em
subpowers}, because they generate the other elements in rows which
sums give the powers $n^n$. The recurrence is
The row and column sums of two vector schemes give the one vector
classification
This identity is known as the {\em rising factorial} and the notation
$(n)^m$ is used. Both rising and falling factorials are related as
We have already counted quadratic forms ${\bf N}^{\rm T}{\bf N}$. Now
we shall study the~other quadratic forms ${\bf NN}^{\rm T}$. In them
blocks ${\bf JJ}_k^{\rm T}$ obtained as outer products of~the~unit
vector columns ${\bf J}_k$ appear.
When the Bell polynomials are compared with the~cycle index (7.15),
we see that here instead of the simple $m$~terms their factorials
appear. The~elements in columns do not form cycles but
undistinguishable subsets. The Stirling numbers generate differences,
if multiplied with the matrix of factorials, and with the matrix of
powers, if multiplied with falling factorials:
$$\begin{array}{cccc|rrrrr|rrrrr} & & & & \quad & 1 & 1 & 1 & 1
&\quad & 1 & 1 & 1 & 1 \\ & & & & & & 2 & 2 & 2 & & & 1 & 2 & 3 \\ &
& & & & & & 6 & 6 & & & & 2 & 6 \\ & & & & & & & & 24 & & & & & 6 \\
\hline 1 & & & & & 1 & 1 & 1 & 1 & & 1 & 1 & 1 & 1 \\ 1 & 1 & & & & 1
& 3 & 3 & 3 & & 1 & 2 & 3 & 4 \\ 1 & 3 & 1 & & & 1 & 7 &13 &13 & & 1
& 4 & 9 &16 \\ 1 & 7 & 6 & 1 & & 1& 15 &51 &75 & & 1 & 8& 27& 64 \\
\end{array}$$
When the lowest allowable value $m_j = 2$, the polynomials give
associated Stirling numbers of the second kind which recurrence
is
Actually, the balloting numbers are all positive. The negative signs
appear by multiplication with ${\bf I}*$ from both sides. They count
binary strings in~which one side has always an advantage given by the
sieve rule $m_{ai} \geq m_{bi}$. The counted strings lead only in a
half of the two dimensional cone (Fig. 10.1). The inverse Fibbonacci
matrix counts strings which elements are ${\bf b}$ and two
consecutive ${\bf aa} = {\bf a}^2$. For example $f_{75} = 5$ counts
strings ${\bf b}^5{\bf a}^2$, ${\bf b}^4{\bf a}^2{\bf b}$, ${\bf
b}^3{\bf a}^2{\bf b}^2$, ${\bf b}^2{\bf a}^2{\bf b}^3$, ${\bf b}{\bf
a}^2{\bf b}^4$.
$$\begin{tabular}{|r|rrrrrrrrr|} \hline & 1 & 2 & 3 & 4 & 5 & 6 & 7 &
8 & 9 \\ \hline n=0 & 1 & & & & & & & & \\ 1 & & 1 & & & & & & & \\ 2
& & & 1 & & & & & & \\ 3 & 1 & & & 1 & & & & & \\ 4 & & 2 & & & 1 & &
& & \\ 5 & & & 3 & & & 1 & & & \\ 6 & 3 & & &4 & & & 1 & & \\ 7 & & 7
& & & 5 & & & 1& \\ 8 & & & 12 & & & 6 & & & 1 \\ \hline
\end{tabular}$$
Actually, the original Lah numbers ${\bf L}$ have odd rows negative
signs and then
\begin{equation} {\bf L}^2 = {\bf I};\ {\rm or}\ {\bf L}^{-1} = (-1)
^{i+j}{\bf L}\;. \end{equation}
The first three simplexes are complete, but the last two are
truncated. Moreover, not all strings are produced. Now we will treat
cubes systematically. Especially we will show how the vector strings
are transformed into points of cubes and points of plane simplexes
into orbits. This transformation is possible by interpretation of the
transposed naive matrices ${\bf N}^{\rm T}$ as faces (Fig. ), the
vectors determining the coordinates of the points in m~dimensional
space. Each vector string corresponds to one point and all strings
of~the~plane simplex n are mapped onto points of m~dimensional cube
which side is $(n-1)$. This transformation is not a simple task. It
can be demonstrated on mapping a 3~dimensional plane onto
4~dimensional cube with the sides 0-2.
The strings from different orbits are counted together because they
have equal {\em moments}. New orbits go from 0 to $m(n-1)$. Some of
the known functions receive a new interpretation, but it still will
be necessary to introduce some new functions.
The unit cubes are most instructive to start with. They have n~sides
and on each side there are just two points, 0 and 1. They are
generated by~the~function
There are $(m+1)$ partition orbits in the unit cubes, from each plane
simplex there is just one orbit. The number of points on each orbit
is determined by the corresponding binomial coefficient. What remains
to be determined is the number of strings in the unit cubes, but we
have studied even this function, and this number is given by the
falling factorial $(i)_{(i-j)}$. We will look on it again. We write
it in the inverse order against the Table 10.6
For example if $n=5, m=3$, and we choose $k=2$, the result is $$(5)_3
= 60 = { 3 \choose 0 }(2)_3(3)_0 + { 3 \choose 1 }(2)_2(3)_1 + { 3
\choose 2 }(2)_1(3)_2 + { 3 \choose 3 }(2)_0(3)_3 =$$ \\
The partition orbits in the m~dimensional cubes which sides are 0-2
are easily found. The results are given in the Table \ref{Partition
orbits in cubes 0-2}. It was shown in Sect. 11.1, how its row m=4 is
obtained from points of the plane simplex.
When we observe row differences in Table 11.3, we see that they are
always 1 on the last $(m+1)$ occupied places. These numbers are just
the~numbers of partition orbits in m~dimensional unit cubes. In
3~dimensional space, it can be drawn (Fig. \ref{Formation of three
dimensional cube}). To a square with the sides 0-2 the unit three
dimensional cube is added, forming the back of the cube with sides
0-2. The~orbit 111 is formed from the orbit 11, which was not in the
square, 211 or 221 is obtained from 21, 22 generates 221 and 222. This
suggests the recurrence of the partition orbits. It can be formulated
graphically:
We know the total number of the points with the natural coordinates
(otherwise the volume $m^n$) in the cubes, and now we want to
determine their distribution according to their moments in the plane
simplexes. If starting simplexes are not truncated, these numbers
must be the binomial coefficients ${ m+k-1 \choose k }$. Similar
numbers appear on the tails of distributions. From the first cubes
with $c=2$, the recurrence can be easily deduced
A new vector with all its allowed values is added to each partition
on~a~suitable place.
$$\begin{tabular}{rrrr|rrrrr} & & & &\quad& 1 & 1 & 1 & 1 \\ & & & &&
& 1 & 2 & 3 \\ & & & && & & 1 & 3 \\ & & & && & & & 1 \\ \hline & & &
&& & & & \\ 1 & & & && 1 & 1 & 1 & 1 \\ 3 & 1 & & && 3 & 4 & 5 & 6 \\
9 & 3 & 1 & && 9 &12 &16 &21 \\ 27& 9 & 3 & 1&&27 &36& 48 &64 \\
\end{tabular}$$
In Sect. 11.2, we have shown that in the unit cubes the strings are
counted by~the~falling factorials. For other cubes the numbers of
strings are not determined as easily, but it is not as that difficult,
if we make it consecutively. For example for $ c=2$ we obtain the
Table \ref{Vector strings in cubes with c=2}.
The recurrence is
It is also possible to shift a cube in its space, when its point with
the~lowest moment is not incident with the beginning of the
coordinate system. The number of orbits and points is not changed by
this operation, but the~number of strings is.
We have shown that the unit cubes are generated by the formula 1.3.
The~term 1 in $(1+{\bf e}_j)$ was interpreted as ${\bf e}_j^0$. The
volume of a cube depends on its base $m$ and on the its
dimensionality $n$. Now we will study, what volume $e$ a cube has, if
its side nears to one and its dimensionality to infinity. We try to
find what value has the limit
The base of $e$ cube lies between cubes with whole numbers $1 <
(1+1/z) < 2$. When $z =1$, the result is $1.5$ instead $2^1$. When $z
=2$, the result is $1.5^2=2.25$ instead $2^2$. Evaluating the
binomial development of (11.7), we obtain
inequalities
The generating function of the $e$ cube has some important properties,
which make from it a useful tool. When a substitution $x = az$ is
applied, the~limit of the expression
is the $a$-th power of the number $e$. This property of the number
$e$ is exploited using $e$ as the base of natural logarithms.
We started our study with permutation matrices having in each row and
column exactly one unit symbol. Then we added the naive matrices,
having this restriction only for rows and the transposed naive
matrices, where it was applied for columns. The next step is to allow
units to be inserted to any available place of a matrix. We already
know, that the number of~these matrices will be determined by a
binomial coefficient. For matrices with $m$ columns and $n$ rows,
with k unit elements in the matrix, the number of~the~possible
configurations will be determined by the binomial coefficient ${ mn
\choose k }$. These configurations can be counted using tables having
two partition orbits in rows as well as in columns. For example for
$m=n=k=4$ we obtain Table 12.1.
The Table 12.1 gives some perspective. In the space, new vector
strings appeared. They lead to the same points as the naive matrices,
but their orbits are not simple partition orbits but the {\em pattern
orbits }which are the products of two partitions, one for rows and
the other one for columns.
The total number of unit vectors with constant sums is given by the
row or column sums elements of tables similar to Table 12.1. Due to
the difficulties with the notation, we will give the formula only for
column sums, where we can use the symbol $n_i$ for the number of
identical binomial coefficients
The sum is made over all possible partitions. The product of the
binomials is not restricted by any conditions on column sums, and
therefore units in each row can be distributed independently, then
the rows obtained by such a~way are permuted ($n=m$) but $n!$
overestimates permutations of rows with equal sums, therefore the
result must be divided by the partial factorials.
gives for two different points (or objects) the same address. This is
possible, if the address $(1, 0)$ is for example a house or a box.
Thus it is necessary to study the possibility that matrices define
positions of $m$ points in space, that they are lists of coordinates
in orthogonal axes. Such a list forms the {\em coordinate matrix}
${\bf C}$ which elements $c_{ij}$ are coordinates of $m$ points
(vertices, objects) $i$ on~$n$~axes.
The matrix column $$(0, 1, 2, 3, 4)^{\rm T}$$
The matrix
Here the four points are placed on the vertices of the three
dimensional cube. Another configuration is
Here all four coordinates in the first column are zeroes. They can be
thus neglected. The first point lies in the center of the coordinate
system, the second one on the end of the second unit vector, the
third one on the end of the third unit vector. The points are related
as in the three dimensional plane complex. The distances between them
are not equal. The first point is in the unit distance to the other
three points, the distances between these three points are
doubled.
The quadratical forms are composed from two parts: The diagonal
matrix ${\bf V }$ formed by the sum of quadratic forms of the two
naive matrices ${\bf N}_a$ and ${\bf N}_b$. The diagonal elements
$v_j$ are known as {\em degrees} of the corresponding
vertices.
The other quadratic forms ${\bf GG}^{\rm T}$ and ${\bf SS}^{\rm T}$
have on the diagonal 2, the number of unit vectors in the rows of the
incidence matrices. This is in accord with the fact that each line is
registered twice in matrix ${\bf V }$ as well as in matrix ${\bf
A}$. Off diagonal elements are $\pm 1$, if two lines are adjacent
having a common vertex. The off-diagonal elements form in such a way
the adjacency matrices of line graphs. But at oriented graphs this
explanation is complicated by signs which signs can be positive and
negative. This sign pattern depends on mutual orientation of arcs. It
is unpredictable and must be determined separately.
The unit matrices ${\bf J}$ (${\bf J}^{\rm T}$) are operators which
sum row (or column) elements of the matrix they are acting on, or
transfer them into the resulting vector-row (or vector-column). In
canonical form of incidence matrices of complete graphs $K_n$ the
unit matrices ${\bf J}$ are combined with the unit matrices ${\bf I}$
with the negative signs. The incidence matrices of complete graphs
$K_n$ are frame operators\footnote{It is curious that such elementary
things can be discovered at the end of the twenties century. Maybe
they were just forgotten.}. The framing operation is applied to
quadratic forms of coordinate matrices twice. At first ${\bf CC}^{\rm
T}$ is framed
or
or
The unit diagonal matrix ${\bf I}$ gives ${\bf S}({\bf I}){\bf
S}^{\rm T}$. This is matrix ${\bf SS}^{\rm T}$ of the complete graph
$K_4$. Four diagonal elements of ${\bf I}$ exploded into six diagonal
elements of the product. The diagonal elements (2) are differences of
the coordinates (or squared distances, since ${\bf I} = {\bf I}^2$)
of the four vertices of the regular tetrahedron. The diagonal
elements are rearranged back into four dimensions as in \ref{c} or
\ref{d}.
The equation \ref{e} shows that each unit vector ${\bf e}_j$ must
appear in the scheme l-times and each pair of elements r-times. The
numbers m, n, k, l ,r are limited by following conditions
It means that all rows and columns of the Hadamard matrices are
orthogonal. The examples of two lowest Hadamard matrices are:
$$\left( \begin{array}{rr} {\bf H}_n & {\bf H }_n\\ {\bf H }_n &
-{\bf H }_n \\ \end{array} \right)\;.$$
\chapter{Graphs} \label{Graphs}
Euler formulated the basic idea of the graph theory when solving
the~puzzle of~the~seven bridges in K\"onigsberg (Fig. \ref{Seven
bridges}). Is it possible to take a~walk over all the bridges, and
returning back to the starting place, crossing each bridge only once?
Euler has shown that the demanded path exists only if in all the
crossing points of the roads even number of the roads meet. Three
roads intersected in some crossing points of the roads in the Euler's
graph. Thus in K\"onigsberg a simple path was
impossible.
The graph theory has two basic notions. The first one is the {\em
vertex} which is usually depicted as a point, but a vertex can be
identified with anything, even with a surface comprising many
vertices, if the graph theory is applied to practical problems. The
second notion is the {\em line} representing a relation between two
vertices. Lines can be {\em oriented}, as vectors are, going from a
vertex into another, then they are called {\em arcs}, and/or {\em
unoriented}, just connecting two vertices without any preference of
the direction. Then they are called {\em edges} (Fig. 3.2).
We can construct a {\em line graph} \ref{Graph and its line graph},
changing lines into vertices, and introducing new incidences defined
now by the common vertices of two original lines. If the parent graph
had m edges, the sum of its vertex degrees $v_j$ was $2m$. Its line
graph has m vertices and the sum of its vertex degrees $v_i$
is
Graphs are {\em connected}, if there exists at least one path or walk
between all pairs of~vertices. It is uninterrupted string of lines
connecting given pair of vertices. Mutually unconnected parts of a
graph are known as its {\em components}. At least $(n-1)$ lines are
needed to connect all n vertices of a graph and n~lines to form a
cycle. Connected graphs with $(n-1)$ lines are known as {\em trees}
(\ref{Examples of unoriented graphs}, A and they are acyclic. A graph
formed from more trees is the {\em forest}.
The {\em linear chains} $L_n$ are a special class of trees which all
vertices except two endings have the degree $v_j = 2$. The vertex
degree counts lines incident to the vertex. Linear chains have the
longest distance between their extremal vertices and the greatest
diameters from all graphs. Another extremal trees are the {\em stars}
$S_n$. All $(n-1)$ of their vertices are connected to the central
vertex directly. The diameter of stars is always 2. The {\em decisive
trees} are trees with one vertex of the degree 2 and all other
vertices with degrees 3 or 1. If the vertex of the degree 2 is chosen
as the root (Fig. \ref{Decision tree}) then on a walk it is necessary
to make a binary decision on each step which side to go. The vertices
with degrees 1 are known as the {\em leaves}. They are connected
by~the~{\em branches} to the {\em stem} of the tree. We already know
the decision trees as strings in the unit cubes. In a tree, they are
joined into bifurcating branches. The~indexing of the leaves is known
as the binary coding.
The {\em complete graph} $K_n$ has n(n-1)/2 lines which connect
mutually all its vertices. Its diameter is 1 and it has no center.
The {\em complement} $\overline{G}$ of a graph $G$ is defined as the
set of lines of the graph $G$ missing in the complete graph $K_n$ on
the same vertices, or by the sum
The canonical forms ${\bf P}e$ and ${\bf S}$ of $K_4$ are
$$\begin{array}{cc} \begin{array}{c} ${\bf P}e$ \\ \\ \left(
\begin{array}{ccc} 1 & 0 & 0 \\ 1 & 1 & 0 \\ 0 & 1 & 0 \\ 1 & 1 & 1
\\ 0 & 1 & 1 \\ 0 & 0 & 1 \end{array} \right) \\ \end{array}&
\begin{array}{c} ${\bf S}$ \\ \\ \left( \begin{array}{cccc} -1 & 1 &
0 & 0 \\ -1 & 0 & 1 & 0 \\ 0 & -1 & 1 & 0 \\ -1 & 0 & 0 & 1 \\ 0 &
-1& 0 & 1 \\ 0 & 0 & -1& 1 \end{array} \right)\;. \end{array}
\end{array}$$
From the consecutive units in a row of the Petrie matrix only the
first and the last are mapped in the product, all intermediate pairs
are annihilated by consecutive pairs of unit symbols with opposite
signs from the incidence matrix of the linear chain, which vertices
are indexed consecutively: $1-2-3-\dots-n$. For example
$$\begin{array}{cc} \begin{array}{ccc|cccc} & & & \ -1 & 1 & 0 & 0 \\
& & & 0 &-1 & 1 & 0 \\ & & & 0 & 0 &-1 & 1 \\ \hline & & & & & & \\
1& 0& 0& -1 & 1& 0& 0 \\ 0& 1 &0& 0&-1& 1 &0 \\ 0 &0 &1& 0& 0&-1 &1
\\ \end{array} & \begin{array}{ccc|cccc} & & &\ -1& 1& 0 & 0 \\ & & &
0&-1& 1 &0 \\ & & & 0& 0&-1 &1 \\ \hline & & & & & & \\ 1& 0 &0 & -1
&1& 0& 0 \\ 1& 1 &0 & -1& 0& 1 &0 \\ 1 &1 &1 & -1 &0& 0 &1 \\
\end{array} \end{array}$$
2. Only the Petrie matrices of trees are nonsingular. The trees have
$(n-1)$ arcs. Therefore their Petrie matrices are square matrices and
because trees are connected graphs, their Petrie matrices are without
empty columns. The~importance of this property will be clear in the
Chapt. 15.
The Petrie matrices define the trees in space of arcs. The another
possibility of~coding trees is in the space of their vertices. There
exist the {\em descendant code matrices} and their inverses, showing
the relation of vertices as the relation of children to parents. In
the descendant code both ends of arcs are used, but the vertices on
the path only once. Moreover, the root itself is induced as
the~element ${\bf e}_{11}$ in the first row. The convention is, that
the arcs are always going from the root. The resulting
code\footnote{The code matrix ${\bf C}$ is simultaneously the
coordinate matrix.} has the matrix ${\bf C}$ the lower triangular
form and on the diagonal is the unit matrix ${\bf I}$. At trees, the
first column is at trees the unit matrix ${\bf J}$, but the code
allows forests, too.
The element ${\bf e}_{11}$ is the vector going from the origin of the
coordinate system to the vertex 1, or using the graph convention, the
arc going from the vertex 0 to the vertex 1. In this case, the zero
column containing one $-1$ element is deleted.
The permutation of the columns is (3,1,4,2). The row with the unit
element is inserted into the incidence matrix as the second one and
all arcs are going from the vertex 2.
We already applied the code matrix of the linear chain $L_n$ and its
inverse as the operators ${\bf T}^{\rm T}$ and ${\bf C}^{-1}$ in
Sect. \ref{Partition Matrices}. Recall that
$$\begin{array}{cccc|cccc} & & & & \ 1& 0 & 0& 0 \\ & & & & \ 1 & 1 &
0 & 0 \\ & & & & \ 1 & 1& 1 & 0 \\ & & & & \ 1 & 1 & 1 & 1 \\ \hline
& & & & \ & & & \\ 1& 0& 0& 0 & \ 1 & 0& 0 & 0 \\ -1& 1& 0& 0 & \ 0 &
1& 0 & 0 \\ 0& -1& 1 & 0 & \ 0 & 0 & 1 & 0 \\ 0 & 0& -1& 1 & \ 0 & 0
& 0 & 1. \\ \end{array}$$
It seems that the distinction between the Petrie matrices ${\bf P}e$
and the~code matrices ${\bf C}$ is due to the unit column ${\bf J}$
which transforms $(n-1)$ square matrix to $n$ dimensional square
matrix. But both sets are different.
The incidence matrices of trees ${\bf G}*$ rooted by the unit column
${\bf J}$ are nonsingular and they have inverses ${\bf G}^{-1}$ which
in their turn are code matrices ${\bf C}$ of unoriented trees. These
code matrices ${\bf C}$ must contain negative elements.
For example, for the star we get using the principle of inclusion and
exclusion
$$\begin{array}{cccc|cccc} & & & & \ 1 & 0 & 0 & 0 \\ & & & & \ -1 &
1 & 0 & 0 \\ & & & & \ -1 & 0& 1 & 0 \\ & & & & \ -1 & 0& 0 & 1 \\
\hline & & & & & & & \\ 1& 0& 0& 0 & \ 1 & 0 & 0 & 0 \\ 1& 1 & 0 & 0
& \ 0 & 1 & 0 & 0 \\ 1 & 0 & 1 & 0 & \ 0& 0 & 1& 0 \\ 1& 0 & 0 & 1 &\
0 & 0 & 0 & 1\;. \\ \end{array}$$
The incidence matrices of the unoriented stars ${\bf S}*$ and the
oriented stars ${\bf G}*$ are selfinverse.
Acyclic connected graphs, known as the trees, form the base of the
graph space. We explain later why, now we only show some
complications of enumeration of graphs on them, as compared to the
naive matrices.
Every tree which vertices are labeled can be connected with a string
of~symbols using the Pr\"{u}fer algorithm: We choose the pending
vertex with the lowest label, mark its neighbor and prune it from the
tree (its branch is cut and discarded). This pruning is repeated till
from the original tree only $K_2 = L_2$ remains. In such a way we
obtain a string of $(n-2)$ symbols. If all n vertices of the original
tree had a specific label, then there is obviously $n^{n-2}$ strings
corresponding to all possible labeling of trees. For example: $L_5$
1-5-4-3-2 gives 5,3,4, $L_5$ 2-1-4-3-5 gives 1,4,3. The sequence 4,4,
4 is obtained by pruning the star $S_5$ rooted in 4.
The sum is made over all partitions of $(n-2)$ into n parts and $v_k$
replaces $m_k$.
The equation 14.1 counts trees successfully, but there appears one
inconvenience: Different types of trees are counted together when
they have the~same partition structure. The partition orbits split in
the graphs into the~suborbits. The smallest pair of trees split in
such a way, two different trees on the orbit 322111 are on Fig.
14.1.
The partition orbits are split into the graph orbits with different
structures. Similar vertices of graphs are known as the {\em orbits
of a graph}. This use of one notion on different levels is somewhat
confusing\footnote{The Sun has its planets and the planets have in
their turn their trabants all with their own
orbits.}.
The tree A on Fig. 14.1 has 5 different orbits and B only 4. The
number of different edges, connecting vertices on different orbits,
is lesser than the~number of vertex orbits, except for the symmetric
edges connecting vertices on the same orbit, as the central edge is
in C on Fig. 14.1.
We must explain why the partition orbits are important and find
techniques how to count the number of unlabelled trees. But before
that we mention another two problems connected with the labeling of
trees.
where $m$ is the number of edges in a tree with $(m + 1)$ vertices.
The powers of $x$ can be interpreted as the number of edges connected
to the added vertex forming the root and the terms of the polynomial
at $x^k$ can be interpreted as the number of trees rooted in the n-th
vertex having the corresponding vertex degree k. For example for
$m=4$ we get: $$64x^1 + 48x^2 + 12x^3 + 1x^4 = 125\;
.$$
16 trees with 4 vertices are attached to the fifth vertex at 4
different places This gives the first coefficient. The second
coefficient is obtained by rooting $(L_3 + K_1)=3\times12$ and
$2K_2=3\times4$. The last term corresponds to the star rooted in the
fifth vertex.
Cycles with even length transform the added unit cycle into two
cycles, one having the same length as the original cycle and the
other with half length. For this case, we have in our example the
cycles of the length 2: $$[s_1^1 + (s_1^1s^1_2)] = (s_1^1 +
s_1^1s^1_2) = s_1^2s^2_2\;.$$
When there are two cycles of different length, which have not a
common divisor, they induce as many cycles as their common divisor is
of the length which is equal to their lowest multiple. For example at
$n=5: 2\times3=6$, and there remain 4 elements to be permuted by
smaller cycles. This is possible as $s_1^1s^1_3$. The cycle $s_1$ is
induced by the cycle $s_2$ which permutes two vectors of~only one
edge and leaves the identity. The cycle $s_3$ permutes columns of
three edges only and reproduces itself. Some examples of induced
graph cycles
$$\begin{array}{cccccc} S_6 & s_6^1\;; & S_7 & s_1^1s^1_6\;; & S_8 &
s^1_2s^1_6\;; \\ & & & & & \\ G_6 & s_3s^2_6\;; & G_7 & s_3s^3_6\;; &
G & s_1^1s_3^1s^4_6\;. \\ \end{array}$$
The graph groups $G_n$ can be used for determining the number of all
simple graphs with $n$ vertices, similarly as the cycle indices were
used. An~edge can be present in a graph or not. In a simple graph
multiple edges are not allowed and we can imagine that the graphs are
placed on vertices of~$n(n-1)/2$ dimensional unit cubes which sides
are formed by diagonals as on~Fig. 12.1, where two diagonal strings
in the 3 dimensional cubes were shown.
It gives
The count of the number $b$ of the permutations which only permute
the~rows of the incidence matrix ${\bf G}$ of a graph determines the
symmetry of the graph. When we divide the number of all permutations
$n!$ by the {\em symmetry number} $b$, we obtain the number of the
different labeled graphs of the given type. $b$ of a single edge on
4 vertices is 4, and there are indeed $24/4 = 6$ different edges on
the set of 4 vertices. The symmetry number of this graph $K_4$ is 24
and therefore there is only one distinguishable labeling of this
graph. The relation of the number $b$ to distinguishable labeling
is known as the~{\em Burnside lemma}.
At graphs with two edges, 15, 27, and 6 permutations belong to two
different graphs, either $L_3$ and one isolated vertex or two $L_2$.
When we try to divide the permutations into the graph orbits, we can
use the fact that both $b$ and the number of different labeling of a
graph must be divisors of $n!$. 15 can then be split only as
$12+3$. Then 27 can be divided as $12+12+3$. We can use also another
criterion, to decide which from both possibilities is right. We
exploit possible partitions of vertex degrees. Graphs with two edges
have the sum of vertex degrees 4 and for 4 vertices two partitions:
2110 and 1111. There are 12 distinguishable permutations of the first
partition and only 1 permutation of the second one. This partition is
stable at all permutations, including the cycle of the length 4,
therefore the group structure is $s^1s^1_4$. Both criterions leave as
the only possible splitting is 12+12+3. There are 12 linear chains
$L_4$ with $b=2$ and the group structure $(s^4_1 + s^2_2)$, and 3
graphs $2K_2$ with $b=8$. Their group structure is $s^4_1 + 2s^2s^1_2
+ 3s^2_2 + 2s_4^1$. the graphs with five and six edges are
complementary the graphs with none and one edge.
The complete tournament with 2k vertices has $(4k^2 - 2k)$ arcs, the
complete oriented graph with 4k vertices has $(8k^2 -2k)$ arcs. It is
necessary to complete a graph corresponding to a selfcomplementary
tournament with 2k vertices, and to generate from each arc two arcs.
It can be done as follows: We generate 2k new vertices indexed by
dashed indices of the tournament and we connect all odd dashed and
undashed vertices having equal index by k arcs. If in the tournament
the arc i-j exists, we induce arcs i-j and i-j' in~the~complementary
graph, if there is the arc j-i, we introduce arcs i'-j and i'-j'. The
arcs missing in the induced graph are present in the
selfcomplementary graph, they correspond to the arcs in the
complementary tournament or connect even dashed and undashed
vertices. The difference is formed by $4k^2$ arcs and $2k$
vertices.
When compared with the naive matrices, one property is clear: The
diagonal matrix must have the same length as the matrix vector ${\bf
M }$ itself. From this property follows that at the diagonalization,
the matrix vector ${\bf M }$ is rotated to decrease the importance of
the off-diagonal elements. Alternatively, the vector position is
stable and we move the coordinate system, exactly as if we were going
around the matrix till a place is found, from where it is possible to
see through. Such a point of view has its own set of coordinates.
The going around the matrix is similar with the function of the
polarizing filters rotating the light (the set of the eigenvalues is
known as the {\em spectrum} of the matrix) has a pair of matrices
known as matrices of {\em eigenvectors}. The~matrix ${\bf M }$ is put
between a pair of eigenvector matrices ${\bf Z}^{\rm T}$ and ${\bf Z
}$ and the~resulting product is the equivalent diagonal matrix
$\Delta({\bf M})$:
$$\begin{array}{rr|rrr} & & \ & 1/\sqrt{ 2} & 1/\sqrt{ 2} \\ & & &
1/\sqrt{ 2} &-1/\sqrt{ 2} \\ \hline 0 & 1 & & 1/\sqrt{ 2} & -1/\sqrt{
2} \\ 1 & 0 & & 1/\sqrt{ 2} & 1/\sqrt{ 2} \\ \hline 1/\sqrt{ 2} &
1/\sqrt{ 2} & & 1 & 0 \\ 1/\sqrt{ 2} & -1/\sqrt{ 2} & & 0 & -1 \\
\end{array}$$ $$\begin{array}{rr|rrr} & & \ & 1/\sqrt{ 2} & 1/\sqrt{
2} \\ & & & 1/\sqrt{ 2} &-1/\sqrt{ 2} \\ \hline 2 & 1 & \ & 3/\sqrt{
2} & 1/\sqrt{ 2} \\ 1 & 2 & & 3/\sqrt{ 2} & -1/\sqrt{ 2} \\ \hline
1/\sqrt{ 2} & 1/\sqrt{ 2} & & 3 & 0 \\ 1/\sqrt{ 2} & -1/\sqrt{ 2} & &
0 & 1 \\ \end{array}$$
For example
$$\begin{array}{rr|rrr} & & \ & 2^{-1/2} & 2^{-1/2} \\ & & & 2^{-1/2}
& -2^{-1/2} \\ \hline 2^{-1/2} & 2^{-1/2} & & 1 & 0 \\ 2^{-1/2} &
2^{-1/2} & & 0 & 1 \\ \end{array}$$
All the above equations were written for the quadratic matrices ${\bf
M}$, representing quadratic forms. For rectangular matrices we can
fill their missing row or column elements by zeroes and for any
vector taken as the eigenvector, we obtain a zero eigenvalue. We will
not be interested in the eigenvalues of the rectangular matrices, but
in the eigenvalues of their quadratic forms, which are known as the
{\em singular values} of the rectangular matrices and
of~the~asymmetric square matrices.
We repeat once again the important fact that on the diagonal of both
quadratic forms as well as squared symmetric matrices appear the
squared elements $m_{ij}$. If a matrix is symmetric, both quadratic
form coincide with the square of the matrix ${\bf M}^{\rm T}{\bf M} =
{\bf M}^2$, therefore the singular values of~the~symmetric matrices
coincide with their squared eigenvalues.
The term $a_1$ is just the sum of all the eigenvalues and it is
identical with the trace of the matrix, the last term is the product
of all eigenvalues and determines if a system of eigenvalues has a
solution. Therefore it is called the {\em determinant}. If a matrix
has at least one zero eigenvalue, then the solution of the matrix
equations is undetermined and the matrix is {\em singular}.
Until now the {\em permanents} were not defined and without them we
had difficulties describing how the polynomials are obtained from the
matrix elements. Let suppose that we have square matrices which
elements are either symbols or numbers, for
example
$$p({\bf A}) = aei + afh + bdi + bfg + cdh + ceg$$ $$p({\bf B}) = 110
+ 131 + 100 + 131 + 201 + 211 = 8\;.$$
We use the whole set of the permutation matrices ${\bf P}$ as the
templates and write the elements incised by them from the matrix as
the products.
Before we start with them, we show at least one result from the rich
theory of permanents, namely the permanents of matrices $({\bf
JJ}_n^{\rm T}+k{\bf I}_n)$: \begin{itemize} \item If $k=0$, we have a
square unit matrix. All $n!$ strings of the permanent are equal 1 and
their sum gives factorial $n!$. \item If $k= -1$, then zeroes are on
the main diagonal and all strings containing at least one diagonal
element are zero. We count the~elements of the permanent as the
permutation matrices ${\bf P }$ without elements on the main
diagonal. You might remember (if not, see Chapt. 7) that they are
counted by subfactorials $z_{i0}$, Table 7.3. It gives for the matrix
$({\bf JJ}^{\rm T}-{\bf I})$ the result $({\bf JJ}^{\rm T}_n-{\bf
I}_n) = (r_n -1)^n$. \item If k=1, we have on the main diagonal 2 and
elements of~the~permanent containing the diagonal elements are powers
of 2. Inserting this value into the generalized polynomial $({\bf
JJ}_n+{\bf I}_n) = (r_n -1)^n$. This is the Apple polynomial. \item
Similarly the permanents for any k are found.
\end{itemize}
$$\begin{array}{c|ccc|c} &\ a & b & c &\ \\ (-) & d & e & f &\ (+)\\
ceg & g & h & i & aei\\ \hline fah & a & b & c & dhc\\ ibd & d & e &
f & gbf\\ \end{array}$$
This is true for any matrix and this fact gives another definition of
the determinant as the volume of a rectangle formed by its
eigenvalues. If an~eigenvalue is zero, the rectangle does not form a
body in n-dimensional space and its volume is zero.
\item 2. The first column was subtracted from the last one.
\item 3. The first column was subtracted from the second one.
\end{itemize}
The adjacency matrices ${\bf A}$ of simple graphs without loops have
all off-diagonal elements either 1 or 0, and all diagonal elements
are zero and they are symmetrical $a_{ij}= a_{ji}$. If we try to find
their polynomial by the above described method, we find for 3
vertices
For $L_6$, we have 5 edges. Six two-tuples and one three-tuple are
shown on Fig. \ref{Six two-tuples and one three-tuple}.
The elements of the Table 15.1 (compare with Table 10.7) are
the~binomial coefficients and the row sums of the absolute values of
coefficients are the Fibbonacci numbers. The coefficients of the
polynomials of the linear chains are the greatest ones which are
obtainable for the trees. It is clear that not too many combinations
of these coefficients are possible. Since the number of trees is a
fast growing function, and the coefficients are limited, their
combinations, compared to the number of trees, are scarcer and as the
result, trees must be {\em isospectral}. This means that different
types of trees must have identical spectra. On Fig. \ref{A pair of
the smallest isospectral trees} is a pair of the smallest isospectral
trees which polynomial is $x^8 - 7x^6 + 9x^4$.
The other quadratic form of the incidence matrix with its transpose
${\bf S}\ {\bf S}^T$ has off--diagonal elements corresponding to the
adjacency matrix ${\bf A}$ of the line graph. For trees, this matrix
has dimension $(n-1)$ and has the true inverse which is the quadratic
form of the walk and path matrices ${\bf W}$ defined on arcs (edges,
respectively).
The walk (path) matrices ${\bf P}$ are defined on vertices for trees,
too. The elements of ${\bf P_p}$ (path) are for oriented trees
$p_{ij} = 1$, if the vertex j is incident with the path i, $p_{ij}
=0$, otherwise. The elements of ${\bf P_w}$ (walk) are for unoriented
trees $p_{ij} = 1$, if the vertex j is on the end of the path i,
$p_{ij} = -1$, if the vertex j is an inner vertex in the path i,
$p_{ij} =0$, otherwise.
The sum
The Cluj matrices of trees are the scalar products of the transposed
walk matrix ${\bf P_p}^T$ with the incidence matrix ${\bf G}_K$ (this
convention can be transposed)
The diagonal elements of the scalar product count $(n-1)$ walks going
from the vertex $j=i$ to the other vertices. The off--diagonal
elements of the scalar product count walks incident with both
vertices $i$ and $j$. The off--diagonal matrix is the Cluj matrix
${\bf C}_e$
which leaves only adjacent elements of the Cluj matrix ${\bf C}_e$
(or equivalently Cluj weighted adjacency matrix ${\bf A}_C$, for
example for the linear chain $L_4$ (n-butane) above
1) The sum of their elements is n(n - 1). Each from the (n - 1) edges
has n vertices on its ends.
since on the trace of ${\bf A}_C^2$ appear twice the products of the
number of vertices $N_{i,(i,j)} \ N_{j,(i,j)}$ on both sides of all
edges.
The pair of the largest eigenvalues $\pm (n-1)$ of the stars are
their only nonzero eigenvalues. This is consistent with their Wiener
number $S_n$: $W_S = (n-1)^2$.
The eigenvalues of the linear chains $L_n$ with odd n (from the
inspection of the first chains) have values $(0, \pm [2, 4, \dots,
(n-1)])$, the eigenvalues of the linear chains $L_n$ with even n have
values $(\pm [1,\ 3,\ \dots,\ (n-1)])$.
These values are compatible with the combinatorial identities for the
sequences of the binomial coefficients:
for odd n:
for even n:
In the first case two loose vertices $K_1$ correspond to the term
$x^2$, in~the~second case the graph $K_2$ corresponds to the term
$(x^2-1)$.
$$\begin{tabular}{|cc|cccc|ccccc|cc|} \hline
\multicolumn{2}{|c|}{Loop} & \multicolumn{4}{c}{1-tuples} &
\multicolumn{5}{|c|}{2-tuples} & \multicolumn{2}{|c|}{3-tuples} \\
\hline *& * & * 0& 0*& & & *0& *0& 0*& 0 * & &*0 & 0* \\ *& & & &* &
& * & & *& & * &* & * \\ *& & & & & *& & * & & * & * & * & * \\
\hline \multicolumn{2}{|c|}{$\Sigma$} & \multicolumn{4}{|c|}{4} &
\multicolumn{5}{|c|}{5} & \multicolumn{2}{|c|}{2} \\ \hline
\end{tabular}$$
The loop polynomial is $P(V) = x^3 - 4x^2 + 5x^1 - 2$. This makes
possible to find the polynomials of the quadratic forms ${\bf G}^{\rm
T}{\bf G}$ or ${\bf S}^{\rm T}{\bf S}$ (${\bf V}\pm{\bf A})
$.
The loop figures are combined with the edge or the arc figures. All
pairs of~loops are counted together with of one edge figures. The
loops figures formed from a loop and an edge are counted together
with the 3-tuples of the loops. Therefore the polynomials of the
quadratic forms of the incidence matrices of the oriented and
unoriented graphs contain all terms of the polynomial, and not only
the every second term as the acyclic polynomial does. The final loop
polynomial of $L_4$ has 3 components
The effect of the diagonal elements is simple, when all the diagonal
elements are equal r, as at the {\em regular graphs}. The unknown $x$
can be replaced by substitution $y = (x + r)$ and the matrix treated
as being without diagonal elements. This can be exploited in some
cases for the calculation of the~determinants, as we will see
later.
The set of n subgraphs of a graph $G$, obtained from the parent graph
by deleting each vertex with all its incident arcs or edges, is known
as the {\em Ulam subgraphs}. Ulam conjectured that the parent graph
can be reconstructed from this set. This appears trivial but it is
difficult to prove it for the unlabelled graphs, where there is no
simple way, how to mach the unlabelled vertices of~two graphs. There
exist another relation, the polynomials of the Ulam subgraphs are the
differences of the polynomial of the parent graph. It means that the
vertex erased subgraph $\delta_j G$ is the partial difference of the
parent graph according to the erased vertex $\delta_j P(G)$ or the
difference of the corresponding matrix, obtained by eliminating the
corresponding row and column. The rules of differentiation and
integration are the same as in the differential and the integral
calculus
For example the graph on the Fig. \ref{The graph A and its vertex
erased subgraphs $A_1$ -- $A_5$} the matching polynomials of its Ulam
subgraphs are
$$\begin{tabular}{l|rrrr} ${\bf A}_1$ & $x^4$ & $-4x^2$ & $-2x$ &
$+1$ \\ ${\bf A}_2$ & $ x^4$ & $-2x^2$ & & \\ ${\bf A}_3$ & $ x^4$ &
$-5x^2$ & $-4x$ & \\ ${\bf A}_4$ & $ x^4$ & $-3x^2$ & & \\ ${\bf
A}_5$ & $x^4$ & $-4x^2$ & $-2x$ & $+1$ \\ \hline $\sum$ & $5x^4$ &
$-18x^2$ & $-8x$ & $+2 $ \\ \hline ${\bf A}$ & $x^5$ & $-6x^3$ &
$-4x^2$ & $2x $ \\ \end{tabular}$$
In edge (or arc) erased graphs, only the edge (arc) itself is
eliminated without eradicating incident vertices, which corresponds
to the elimination of the corresponding row and column in the
quadratic form ${\bf GG}^{\rm T}$ or ${\bf SS }^{\rm T}$,
respectively. The set of the edge erased subgraphs has $m$ subgraphs,
$m$ being the number of edges of the graph. In trees, each subgraph
has always two components. Here also the sum of the polynomials of
the edge erased subgraphs of trees is a difference of the polynomial
of the parent tree, but the rules of differentiation are different.
The coefficients at $(n-2k)$ powers of $x$ are not multiplied by the
power of $x$ and the power of $x$ is not decreased, but they are
divided by $(m-k)$ and the power of $x$ is left
unchanged.
An edge erased tree is a forest with $n$ vertices and the first term
of its polynomial is $x^n$. There are $m$ subgraphs and therefore the
sum of all subgraph polynomials is divisible by $m$. All subgraphs
contain $(m-1)$ edges and therefore the coefficient of the second
term of the sum, when divided by this number gives $m$. The following
coefficients can be deduced using the full induction. If the relation
of polynomials is true for the parent tree, it must be true also for
its subgraphs (forests), containing one edge less, and their
polynomials. Corresponding coefficients of all subgraphs must be $0\
{\rm mod}\ (m-k)$. This is true also for the term $a_{n-k}$ if
$n=(2k+1)$. Among the subgraphs of~the~linear chain there exist $k$
subgraphs containing the term corresponding to $(k+1)$ tuple. For
example the graph on the Fig. \ref{The tree B and its edge erased
subgraphs $B_1$ -- $B_5$} the matching polynomials of its edge erased
subgraphs are
$$K_4 : x^4 -6x^2 + 3\\ \Sigma_i \delta (P) = 6(x^4 -5x^2 + 2) = (6/6)
x^4 -(30/5)x^2 + (12/4)\;.$$
Using the same argumentation, the quadratic form ${\bf SS }^{\rm T}$
of the bipartite cycles (n even), which spectra are equivalent to the
Laplace-Kirchhoff matrices ${\bf S}^{\rm T}{\bf S}$, have all
off-diagonal elements either negative or one off-diagonal element in
each row can be negative and the other positive. If we combine ${\bf
K}(C_{2k})$ with ${\bf K}(\overline{C_{2k}})$ the result is identical
with the difference ${\bf K}(kK_2) - {\bf K}(k\overline{K_2})$.
Therefore the Seidel adjacency matrices of $k$ complete graphs $K_2$
and the cycles $C_{2k}$ are isospectral. For example:
The adjacency matrix ${\bf A}^2[S(G)]$ has two blocks ${\bf G}^{\rm
T}{\bf G }$ and ${\bf GG}^{\rm T}$. Both blocks have the identical
spectra. Their square roots with both signs form the~spectrum of the
the adjacency matrix of the subdivision graph. The~difference between
the number of vertices and edges are zero eigenvalues.
This can be exploited for calculations. For example, the cycle $C_3$
has the~adjacency matrix ${\bf A}$ equivalent with its incidence
matrix ${\bf G }$. The subdivision graph of the cycle $C_3$ is the
cycle $C_6$. Its adjacency matrix ${\bf A}$ is
$$\left( \begin{array}{cccccc} 0 & 0 & 0 & 1 & 1 & 0 \\ 0 & 0 & 0 & 1
& 0 & 1 \\ 0 & 0 & 0 & 0 & 1 & 1 \\ 1 & 1 & 0 & 0 & 0 & 0 \\ 1 & 0 &
1 & 0 & 0 & 0 \\ 0 & 1 & 1 & 0 & 0 & 0 \end{array} \right)\;.$$
The subdivision graph of the star graph $S_4$ has the adjacency
matrix ${\bf A}$
$$\left( \begin{array}{ccccccc} 0 & 0 & 0 & 1 & 1 & 0 & 0 \\ 0 & 0 &
0 & 1 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 & 1 \\ 1 & 1 & 1 & 0 & 0 &
0 & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 & 0 & 0 \\ 0
& 0 & 1 & 0 & 0 & 0 & 0 \end{array} \right)\;.$$
All subdivision graphs of stars $S_n$ have spectra derived from the
spectra of their line graphs ${\bf GG}^{\rm T}= {\bf I }+ {\bf
JJ}^{\rm T}$. The corresponding spectra are $n,1^{n-1}$ and it is
easy to find their square roots. The signs are determined by the zero
trace of the adjacency matrix ${\bf A}$.
The quadratic form ${\bf GG }^{\rm T}$ of the incidence matrix ${\bf
G }$ defines the line graph $L(G)$ of the parent graph $G$. A line
graph is obtained from its parent graph if its edges are transformed
into vertices which are incident if they have in~the~parent graph a
common vertex. The relation between the quadratic form ${\bf GG
}^{\rm T}$ and the adjacency matrix ${\bf A}[L(G)]$ of the line graph
for parent graphs with simple the edges is
The line graph of the linear chain $L_n$ is the linear chain
$L_{n-1}$. The subdivision graph of the linear chain $L_n$ is the
linear chain $L_{2n-1}$.
These relations lead to the formula for the eigenvalues the adjacency
matrix ${\bf A}$
The linear chain $L_n$ behaves as a rod fixed in its center. This is
opposite to a string which is fixed on its ends. Its vibrations are
described by the sinus function.
of the oriented linear chains $L_n$ and the bipartite (n even) simple
cycles $C_n$ have identical spectra, and that the adjacency matrices
of their line graphs must have eigenvalues having the form
$\pm(\lambda_j \pm 2)^{1/2}$. Simple cycles, which are subdivision
graphs of cycles with uneven number of vertices, have eigenvalues in
the form $\pm(\lambda_j^2 - 2)^{1/2}$. The eigenvalues of the
subdivision graphs of the bipartite graphs have eigenvalues
$\pm(\lambda_j^2 + 2^{1/2}$), where $\lambda_j$ are eigenvalues of
the corresponding line graphs.
For the regular oriented graphs the relation (\ref{Pe}) holds for all
orientations of the subdivision graphs. For other graphs it is
necessary to solve the effects of the different orientations of arcs
in an oriented graph on the spectra of~the~corresponding subdivision
graphs individually.
Notice that ${\bf B}_3{\bf M}$ is the diagonal matrix with equal
values. It means that ${\bf B}_3$ is a multiple of the inverse of
${\bf M}^{-1}$.
$$\left( \begin{array}{ccc} 1 &-1 & 0\\ -1& 2& -1\\ 0 &-1 & 1
\end{array} \right)$$
The corresponding sums are ${\bf Q }= 3{\bf B}+ 1{\bf C}$ and ${\bf I
}= {\bf A}+ {\bf B}+ {\bf C}$.
The inverse matrices were mentioned more times, but now they shall be
explained more systematically.
At first recall Sect 3.4. There two types of inverse elements were
described, additive and multiplicative. The additive inverse is
defined by the identity $a + b = 0$, from it $b = -a$. The negative
element has the same value and opposite sign of its parent element.
The multiplicative inverse element is defined as the product $ab =
1$. From it $b = 1/a$ and $a = 1/a$. The distinction between an
element and its inverse is determined by convention. We have already
shown that the multiplicative inverses are additive on the
logarithmic scale (Fig. 3.5).
The complementary graph together with the parent graph gives the
complete graph $K_n$. The matrices ${\bf S}^{\rm T}{\bf S}$ and
$\overline{{\bf S}^{\rm T}{\bf S}}$ can be considered as the {\em
generalized additive inverses} as we see later.
The right side matrix in brackets is the left hand side inverse ${\bf
M}$ of the matrix in the lower triangular form ${\bf M}$.
We have shown how the inverse matrix elements are related to minors
of the matrix elements. But in some cases these inverses can be
deduced directly from the structure of graphs without no apparent
connection to the minors and determinants.
This is the case of matrices ${\bf SS}^{\rm T}$ or ${\bf GG}^{\rm T}$
of trees. They have $(n-1)$ rows and columns and are nonsingular
because the corresponding quadratic forms ${\bf S}^{\rm T}{\bf S}$
and ${\bf G}^{\rm T}{\bf G}$ have just one zero eigenvalue. In a tree
there are no cycles and therefore there exist only one walk between
each pair of vertices (in the case of unoriented graphs we speak
about paths). Matrices\footnote{Only one symbol is used for both
matrices for economy.} ${\bf W}$ with rows corresponding to all walks
or paths in a tree, and with columns representing the arcs or edges,
can defined. The elements $w_{ij}$ of these matrices are $\pm 1$ if
the arc or edge $j$ is a part of the path or walk $i$ and 0
otherwise. The definition is complicated, especially for unoriented
trees, by the signs necessary to eliminate unwanted elements, when
the walk matrices are multiplied with the matrices ${\bf GG}^{\rm T}$
which all elements are positive. The oriented trees can have the
configuration of all arcs head to tail, since trees are bipartite
graphs. Then all off-diagonal elements of ${\bf GG}^{\rm T}$ are
negative and all elements of ${\bf W }$ positive. Otherwise $w_{ij}$
has the positive sign, if the edge $j$ is in the even distance from
the last edge in the walk (path) or the arc $j$ has the same
orientation as the last arc, and it has the negative sign, if the
corresponding edge is in the odd distance from the last edge, or the
corresponding arc has the opposite orientation as the last
one.
The path matrices of the oriented linear chains looks like the Petrie
matrices of complete graphs (see Sect. 13.3), only the elements of
both matrices have different interpretations.
The true inverses of quadratic forms ${\bf GG}^{\rm T}$ and ${\bf
SS}^{\rm T}$ are 1/n multiples of the corresponding quadratic forms
${\bf W}^{\rm T}{\bf W}$, and matrices ${\bf G}^{\rm T}{\bf W}^{\rm
T}{\bf W}$ and ${\bf S}^{\rm }{\bf W}^{\rm T}{\bf W}$ are the right
inverses of ${\bf G }$ or ${\bf S}$, respectively, similarly as ${\bf
W}^{\rm T}{\bf WG }$ and ${\bf W}^{\rm T}{\bf WS }$ are the left
inverses of ${\bf G }$ or ${\bf S}$, respectively. The diagonal
elements of both quadratic forms count how many times the
corresponding arc or edge was used in all walks or paths, the
off-diagonal elements count common exploitations of the given pair of
lines. These simply obtained numbers are simultaneously minors of the
corresponding quadratic forms of the incidence matrices. The trace of
${\bf W}^{\rm T}{\bf W}$ is the sum of distances between the vertices
in the tree. It is known as the {\em Wiener number}, see the next
Chapt..
The walk and path matrices of trees include all walks or paths of the
given tree, whereas the code matrices of trees include only the walks
(or paths) from the root. For the oriented trees both kinds of
matrices are related as
For example
$$\begin{tabular}{rrrr|rrrrrr} & & & &\ -1 & -1 & 0 & -1 & 0 & 0 \\ &
& & & 1 & 0 & -1& 0 & -1& 0 \\ & & & & 0 & 1 & 1 & 0 & 0 & -1\\ & & &
& 0 & 0 & 0 & 1 & 1 & 1 \\ \hline & & & & & & & & &\\ 1 & & & & -1 &
-1 & 0 & -1 & 0 & 0 \\ 1 & 1 & & & 0 & -1 & -1& -1 & -1& 0 \\ 1 & 1 &
1 & & 0 & 0 & 0 & -1 & -1& -1\\ 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0
\\ \end{tabular}$$
We begin the search for the inverse for the quadratic form of a cycle
matrix ${\bf C}^{\rm T}{\bf C}$. It is easy to find it for small
cycles, for example for 7 member cycle this symmetrical matrix $({\bf
G}^{\rm T}{\bf G})$ starts as
$$\begin{array}{cc} {\bf G}^{\rm T}{\bf G} & \left(
\begin{array}{rrrrrrr} 2 & 1 & 0 & 0 & 0 & 0 & 1 \\ 1 & 2 & 1 & 0 & 0
& 0 & 0 \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots &
\vdots \end{array} \right)\;. \end{array}$$
This matrix is the quadratic form of the basic matrix ${\bf C}$ of
uneven cycles which elements are $c_{ij} = (-1)^{d(ij)}$. The upper
d(ij) indices are the distances of the vertices j from the diagonal
vertex i. There are k positive elements and (k+ 1) negative elements
in each row and column, for example
Graph B
Graph C
For example
For example
The weight of arcs is decreased for cycles $C_n$. The inverse of the
difference $(\delta_1{\bf S}^{\rm T}{\bf S})^{-1}$ is always the
matrix ${\bf SS}^{\rm T}$ of the linear chain $L_n$ which inverse is
the quadratic form ${\bf W}^{\rm T}{\bf W}$ of the path matrix. The
chain forms the spanning tree. Its square matrix must be decomposed
into the triangular form and added to the matrix ${\bf JJ}^{\rm T}$.
Since ${\bf W}^{\rm T}{\bf W}$, as defined, gives $n{\bf I }$ as the
product with ${\bf SS }^{\rm T}$, it is necessary to divide by n. For
example the triangular decomposition
The rooting technique at trees gives as the same result as the code
matrices. The multiplicity k of arcs can be expressed as repeating of
rows or by weighting the arcs. These weights in the incidence
matrices must be square roots of the multiplicity k of the
arc.
This means that an {\em eigenvalue orbit} can be defined which volume
is determined by the multiplicities of the eigenvalues.
thus
The proof used properties of graph matrices with simple arcs but the
relation between eigenvalues holds also for multigraphs and their
complementary graphs as calculated from the relation
For example
Or
The unit vector-column ${\bf J}$ or the unit vector-row ${\bf J}^{\rm
T}$ are the zero eigenvectors of the Laplace-Kirchhoff matrices of
all graphs and the Laplace-Kirchhoff matrices of all subgraphs of the
complete graph $K_n$ are not orthonormal eigenvectors to its
Laplace-Kirchhoff matrix.
We find the inverse matrix and multiplying it with the vector {\bf b}
we should obtain the unknowns.
The last column of this matrix with $m = (n+1)$ rows and columns is a
linear combination of the first n columns. The weights are given by
the elements of the vector ${\bf x}$. This is true also for the m-th
row. The determinant of the block matrix is zero and therefore when
we develop it according to the last row we get
If each row has its own weight vector, or if the vector ${\bf b }$ is
combined with an error vector, then the vector ${\bf x}$ can be far
from all vectors ${\bf x}_j$. For example a matrix
$$\left( \begin{array}{ccccc} 12 & 4 & 3 & 2 & 1\\ 14 & 5 & 5 & 3 &
2\\ 14 & 5 & 5 & 4 & 1\\ 16 & 6 & 6 & 6 & 3\\ 16 & 6 & 8 & 4 & 3
\end{array} \right)$$
has a well defined inverse and it gives to the vector ${\bf b } = (32,
46, 45, 64, 62)^{\rm T}$ as the solution the vector ${\bf x }= (1, 1,
2, 3, 4)^{\rm T}$. Inducing an error vector ${\bf r} = (2, 0, 0, 0,
0) ^{\rm T}$ which gives the vector ${\bf b } = (34, 46, 45, 64, 62)
^{\rm T}$ the vector ${\bf b }$ changes to $(8.5, -24, 4, 5, 6)^{\rm
T}$. It means that a slight error induced the error of the input
vector $(7.5, -25, 2, 2, 2)^{\rm T}$, which completely distorted the
true vector, or a slight change of the specific vector ${\bf x}$
distorted the result for the whole bundle of identical vectors. This
property of vector systems is very unfortunate, because we can not be
sure, if we do not know the exact values, using approximate values
only, our reconstruction corresponds to original values.
Distances were mentioned before but now they and their matrices will
be studied systematically, using all our knowledge.
We can move between two points i and j on different paths. The length
of the path depends on circumstances, as on accessible ways, or means
of transportation. The length of the path between the points i and j
is the {\em distance} $d_{ij}$.
The topological distance matrix ${\bf D}$ of a graph has the same
unit elements as its adjacency matrix ${\bf A}$. Both matrices are
obtained by the same operation described in Sect. 12.8 from the
coordinate matrices.
where ${\bf Q}$ is the diagonal matrix of the row or column sums of
the distances elements of the vertex i to all other vertices. The
negative off-diagonal elements show distances between corresponding
pairs of vertices:
The cycle $C_4$ bent on the regular tetrahedron with the distance
matrix corresponding to the distance matrix of the complete graph
$K_4$ gives another matrix angles. The neighboring arcs form
60-degree angles and each arc is orthogonal to its opposite arc. They
form a pair which has no common vertex.
There exist three embeddings of the cycle $C_6$ onto vertices of the
3 dimensional cube. The first one is identical with the usual
topological distance matrix and leads to three collinear pairs of
orthogonal arcs
Two another forms of $C_6$ have some distances shorter and lead to
the another arrangement of collinear arcs.
The planar conformation of $C_6$ has the following matrix and the
resulting matrix of angles between bonds
where the angles are $120^0$, $60^0$, $180^0$, $300^0$, and $240^0$,
respectively.
The uneven cycles have each arc orthogonal to its neighbors on both
sides but the pair of its opposites forms $60^0$ angles to it. This
conformation is obtained by a rotation of two consecutive right
angles for $60^0$ through the given arc. The result appears for the
arc closing the cycle.
The outer product of the incidence matrix of a graph with simple arcs
has on the diagonal 2. The off-diagonal elements are either 0, if the
arcs do not have any common vertex, or 1, if two arcs meet in a
vertex. The cosine of $60^0$ is 0.5. Therefore the equilateral
structures appear in complete graphs. $K_3$ is the equilateral
triangle, $K_4$ is the equilateral tetrahedron. Six arcs of the
equilateral tetrahedron form three pairs of orthogonal arcs.
$$\begin{tabular}{rr|rr} & & \ ${\bf S}^{\rm T}$ & $-{\bf I}$\\ & &
${\bf 0}$ & ${\bf J}^{\rm T}$\\ \hline & & & \\ ${\bf S}$ & ${\bf 0}$
& ${\bf SS}^{\rm T}$ & $-{\bf S}$\\ -${\bf I}$ & ${\bf J}$ & ${\bf
-S}^{\rm T}$ & ${\bf I + JJ}$ \end{tabular}$$
Increasing the dimension of the complete graph there will appear $(n
- 3)$ orthogonal arcs to each parent arc.
Inserting the distance matrix of the star rooted in the n-th vertex
into ${\bf SS }^{\rm T}$ of the complete graph, then we get for the
star graph the product
The arcs of the star are orthogonal. The arcs connecting its loose
vertices have the double length (on the diagonal fours appear). These
arcs are the diagonals of the corresponding squares. This can be
checked by calculation of cosines. $2/8^{1/2}$ is cosine of $45^0$.
The direct verification is possible only for $K_5$ with three
orthogonal axes.
where x goes from 0 to $(n - 1)$. If the chain increments are two
vertices then the change between the consecutive counts gives a
possibility to use full induction
$$\begin{array}{ccccc} a & 1 & 1 & 1 & $\ldots$ \\ 0 & 1 & -1/(n-2) &
-1/(n-2) & $\ldots$ \\ 1 & -a/(n-1)& -a/(n-1) & -a/(n-1) & $\ldots $
\end{array}$$
The maximal eigenvalue of the even planar cycles on the circle with
unit radius is $2n$ and its eigenvector is the unit vector (this
corresponds to $2n/4$ for topological distance matrices). The even
distances on the circle form the right triangles over the diameter as
the hypotenuse and their pairs sum to 4.
The matrix $({\bf JJ}^{\rm T}- {\bf I}$ (otherwise the distance
matrix of the complete graph) is thus the distance matrix ${\bf D}(0)
$. The changes of eigenvalues and eigenvectors between the adjacency
matrices ${\bf A }$ and the distance matrices ${\bf D}$ are then
continuous transformations produced by powers of given distances, or
in some cases, by changes of the geometrical conformations. We will
study some special examples.
As an the first example we use linear chains, which exist in the form
of stiff rods. It was found that to express this geometrical property,
it is necessary and sufficient to write the distances $d_{ij}$ as
squares of linear distances. The topological distance matrices are
then just second power geometrical distance matrices of linear chains
bent on vertices of n dimensional unit cube. Their apparently linear
distances are squares of the corresponding $d_{ij}$ as diagonals in
the n dimensional cube. In Table 1 eigenvalues of different power
distance matrices $L_5$ are tabulated. This chain is long enough to
show main properties of such a system, where the second power
geometrical distance matrices always have only 3 nonzero
eigenvalues.
Another exceptional case is the cycle $C_4$, which can be bent from
the regular tetrahedron shape to the plane square by increasing two
distances or to a rod by decreasing them evenly. Its topological
distance matrix is thus undistinguishable from the second power
geometrical distance matrix of the square and the matrix $[{\bf JJ }-
{\bf I}]$ is one of the possible geometrical conformations (similarly
for the chain $L_4$, but there the adjacency matrix is different)
.
The second case is the extreme, all vertices lie on a straight line.
The third case represents two double bonds bending to $60^0$, or the
adjacency matrix of the graph on Fig. 13.2 b or a distance matrix of
one of its conformations. The eigenvectors are also deformed, going
to lover values and to higher ones again (in the third case it is
0.7808) and having zero values which are possible for other
conformations or moments, too:
Here the zero distance appears as the permuted adjacency matrix and
the changes are:
Here we will study the formation of the cube from two cycles $C_4$.
The adjacency matrix of two cycles $C_4$ can be written similarly as
for two chains $L_2$ in the block form as
The ancient Greeks were very good geometricians and had some
knowledge of algebra, but were not able to imagine a trajectory of a
moving object as a geometrical problem. Everybody knows the Zenon
aporea.
It was a cultural shock, when Zenon came out with his discoveries.
Imagine, Achilles can never catch a turtle, if it has an handicap.
Achilles running it, the turtle changes its position, and remains
ahead. Achilles running the second handicap, the turtle changes its
position, again, and so in infinite many intervals. Ancients
mathematicians did not find that a sum of an infinite number of ever
decreasing fractions is finite. But curiously enough, they were not
able to imagine the situation as a figure, as Fig. \ref{Zenon plot of
the Achilles and turtle aporea}.
It was Descartes, who with his analytical geometry found that a simple
plot of two straight lines solves the Zenon aporea about Achilles and
turtle.
The straight line of 6 points in the axis x was copied and projected
into the axis y The resulting positions of the original points in the
axis b are described either as $ y = 1x$ or as $ y = 0.5x\;.$
But this equation is true not only for the set of six points with
natural coordinates, but for all points between them on the straight
line. The unit vectors are not needed. The equation of the straight
lines in two dimensions is
The term a is the value of y when $x =0$. In the given example $a=0$.
The term a is the slope of the line determined as the ratio $y/x$, it
is tangents of the angle $\alpha. If we know y, we can find x solving
the Equation (\ref{y}) as $x = (y - a)/b$.
The two dimensional plane simplexes are straight lines having the
form $y+x=m$, their slopes are negative, and they are defined only in
the positive cone.
In the plain many straight lines can be defined. They can be parallel
or they can cross. Crossing takes place, when both coordinates x and
y of both straight lines are equal, as for example
$$ y = 2 + 3x $$ $$ y = 3 + 2x.$$
and this gives, when multiplied with the vector ${bf b} = (2, 3)^{\rm
T}$ the solution $(1, 5)^{\rm T}$.
Now again let the baseline x to represent the position in time t, the
vertical axis y the position in time $t+1$, 1 representing an
interval of time $\Delta t$. The coordinates of the exponential curve
are
$$\begin{tabular}{|l|ccccccccc|} \hline Interval& 1 & 2 & 3 & 4 & 5 &
6 & 7 & 8 & 9\\ \hline x & 0 & 128 & 192 & 224 & 240 & 248 & 252 &
254 & 255 \\ y & 128 & 64 & 32 & 16 & 8 & 4 & 2 & 1 & ? \\ \hline
\end{tabular}$$
The x values are growing, the y values are decreasing. Both changes
are not linear. Nevertheless, if the values x are plotted against the
corresponding values y, the plot is linear, see Fig.
\ref{Linearization}.
The Zenon aporea is now transformed into the question, when the last
radioactive atom will decay, and when their starting number is $x =
256$.
The graph of the process is the same as in the case of the runners,
if both axes, time and position, are replaced by positions
(concentrations) in consecutive time intervals, $t$ and $(t +1)$ as
if both positions were on two different orthogonal scales. By doing
so, these positions are considered to be orthogonal, the exponential
movement is transformed into linear, as if we used the logarithmic
scale\footnote{The linear movement is the limit of the exponential
movement when the constant $k=0$}.
Markov was a Russian mathematician who got the somewhat childish idea
to study the order in which the consonants follow the vowels in a
Pushkin's poem. After a consonant another consonant or a vowel can
follow with some statistical probability which is determined by the
structure of the language and its use by the
author.
\begin{center} A vv A vc M cv A vc R cc K cv O vc V \end{center}
Probabilities vv, vc, cc and cv are obtained from the direct counts
by dividing them with all possibilities of the transitions (here 7
transitions of 8 letters). When arranged among into matrices, they
form the {\em stochastic matrices} ${\bf M}$ which row sums are 1.
The theory of processes connected with these matrices forms a part of
the theory of {\em stochastic processes}.
$$\begin{tabular}{|r|rrrrrrrrr|} \hline & ? & A & A & M & A & R & K &
O & V \\ & A & A & M & A & R & K & O & V& ? \\ \hline & & & & & & & &
& \\ c & * & 0 & 1 & -1 & 1 & 0 &-1 & 1 & * \\ v & * & 0 & -1 & 1 &
-1 & 0 & 1 & -1 & * \\ \hline \end{tabular}$$
A transition matrix ${\bf P}$ is formed from two parts, the Markov
matrix ${\bf M}$ and the identity matrix ${\bf I}$
The adjacency matrices ${\bf A}$ which we used till now were the
symmetrical. They were obtained as the off-diagonal elements of
quadratic forms of incidence matrices of either an oriented graph
${\bf S}$, or an unoriented graph ${\bf G}$ (see Sect. 12.7)
.
therefore the sums $\Sigma k_{ij}$ on the diagonal are the column
sums.
Let ${\bf S}$ and ${\bf G}$ be the incidence matrices of the same
oriented multigraph, where $ {\bf S}$ and ${\bf G}$ are identical
matrices except for the signs. An unoriented edge corresponds to each
arc. The rows of ${\bf S }$ and ${\bf G }$ are the mutually
orthogonal vectors.
The corresponding scalar products ${\bf S}^{\rm T}{\bf G}$ and ${\bf
G}^{\rm T}{\bf S}$ are the asymmetric matrices showing differences in
the orientation of arcs. As an example we use the multigraph defined
by the transposed incidence matrix ${\bf S}^{\rm T}$ (see Fig.
\ref{Reaction multigraph })
$$\left( \begin{array}{cccc} -3 & 1 & 1 & 1\\ -1 & 1 & 1 & -1\\ -1 &
-1 & 2 & 0\\ -1 & 1 & 0 & 0 \end{array} \right)$$
$$\left( \begin{array}{cccc} -3 &-1 & -1 & -1\\ 1 & 1 & -1 & 1\\ 1 &
1 & 2 & 0\\ 1 &-1 & 0 & 0 \end{array} \right)$$
The same operation with ${\bf S}^{\rm T}{\bf G }$ gives the pattern
$${\bf S}^{\rm T}{\bf G}- {\bf S}^{\rm T}{\bf S }= 2({\bf A}_r - {\bf
V}_{out})$$
$${\bf S}^{\rm T}{\bf G}+ {\bf G}^{\rm T}{\bf G }= 2({\bf V}_{in}+
{\bf A}_r)$$
The double entry accounting of the arcs using the orthogonal vector
strings, their sums and differences, quadratic forms, scalar products
and transposes, gives a web of related matrices describing the graphs
and to them isomorphic objects and their transformations.
The matrices $({\bf I }+ {\bf M})$ have one eigenvalue exactly 1, the
other eigenvalues are in the circle $0 < \lambda_j < 1$. The matrix
${\bf M}$ has exactly one eigenvalue equal to zero and the remaining
$(n-1)$ eigenvalues in the range limited by the circle given by the
rate sums $\Sigma -k_{ij}$. Because a transformation of any species
can not be greater that its concentration, the sum of the rate
constants must be lesser than 1. If the regular unit matrix ${\bf I
}$ is added to ${\bf M}$, all eigenvalues are increased evenly by 1.
This has an important consequence which remained unnoticed: The
equilibrium state of the operator ${\bf P}$ has one eigenvalue
exactly 1, all other eigenvalues are 0. The product of any
concentration vector ${\bf c }$ with the equilibrium operator $({\bf
I}+ {\bf M})^{\infty}$ must give the equilibrium concentration vector
${\bf c}^*$. Therefore $(1/n){\bf I}({\bf I}+ {\bf M})^{\infty}$ has
the form of n identical columns of the equilibrium concentration
vectors ${\bf c}^{\rm T}$. Because the sum of concentrations is
always $\Sigma_{j=1}^n = 1$ this result conforms with the condition
${\bf c}({\bf I}+ {\bf M})^{\infty} = {\bf c}^{*{\rm T}}$.
Because any equilibrium state of the operator ${\bf P}$ has exactly
one eigenvalue 1 and other $(n - 1)$ eigenvalues are 0, it is easy to
find the corresponding eigenvectors. The unit eigenvector is the unit
row ${\bf J}^{\rm T}$ or the unit column ${\bf J}$, respectively. The
zero eigenvectors can be chosen as any $(n -1)$ rows or columns of
the Markov matrix. Any Markov matrix is therefore a system of
eigenvectors of its equilibrium state.
A Markov matrix describes its own equilibrium state and all the paths
to the equilibrium from any point of the n dimensional concentration
simplex. This simplex is a plane orthogonal to the unit vector ${\bf
I}$, for example for 3 substances it is an equilateral triangle. Each
point of the simplex can be the equilibrium point of the system and
to each equilibrium point there go infinitely many paths. Therefore
it is necessary to classify the Markov matrices according to the
character of paths the matrix produces. If we exclude matrices going
to concentrations outside the simplex, there are three possibilities.
Easily they can be found for the two dimensional case:
\item ${\bf C}$. The steepest approach. The reaction path should be a
straight line going from any concentration point to the equilibrium.
This requires that the reaction constants of each substance were
proportional to the equilibrium concentrations of the target
substances. For example for 3 substances: $c_1k_{12} = ac^*_2$ and
$c_1k_{13} = ac^*_3$. From the microscopic reversibility conditions
$c^*_2k_{23} = c^*_3k_{32}$ we obtain the relation of reaction
constants $k_{23}/k_{13}=k_{23}/k_{12}$. For other two substances we
obtain similarly for $c_2$: $k_{21}/k_{31}=k_{23}/k_{12}$ and for
$c_3$: $k_{31}/k_{21}=k_{32}/k_{12}$. \end{itemize}
We have shown the exact methods for solving the equation ${\bf Mx }=
{\bf b}$ in Chapt. 16, based on the inverting of the matrix ${\bf M
}$ or finding its eigenvalues. In case we are not able to do such
sophisticated mathematical operations, we can try to guess the right
answer. We have counted the matrices and we know, that if we limit
ourselves to natural numbers, their number is not infinite. Therefore,
using computers, it is possible to find the solution by the trial
and error methods, especially, if the results are compared with the
target values and impossible combinations excluded. This technique of
fluctuation can be compared with the process by which a system seeks
its equilibrium.
Let us start with the guess vector {\bf y}. After multiplication with
the matrix {\bf M }we get the guess vector {\bf g}. Comparing it with
the target vector ${\bf b }$ we obtain the difference $d_{{\bf
g}-{\bf b}}$. If it is zero, our guess coincides with the searched
vector and we can end our search. Similarly if the difference
$d_{{\bf g}-{\bf b}}$ is negligible we can stop our search. Otherwise
we must correct the original guess vector using $d_{{\bf g}-{\bf
b}}$. But we cannot apply the whole difference, because the next
guess could be as a pendulum on the other side of the true values. We
must lessen the fluctuations. The correction must be smaller than the
difference, which is achieved by using a constant c: 0$<$c$<$1. If we
choose the constant too low, we need too many steps to find an
acceptable value of ${\bf g}$, if c is too close to 1, the results
could fluctuate, similarly as was shown for the Markov matrices.
The least necessary number of digits for each object from m objects
is close to $\log_2 m$. These digits count edges of a binary decision
graph on which leaves the counted objects are placed (Fig.
\label{Binary decision tree is isomorphic with indexing of m objects
by binary digits})\footnote{Please arrange the leaves onto the
vertices of the cube and draw the decision tree yourselves. I tried
it but my figure was too ugly. The cube as well as the decision tree
must be deformed}.
For all m objects we need at least $m\log_2 m$ digits (in the example
24 digits). This limit is obtainable only if $m$ is a power of $2$.
Nevertheless it can be used for elementary calculations of logarithms
with a satisfactory precision. The number of digits $m_j$ is the
distance of the leave j from the root in the decision tree. Therefore
the logarithms are related to the distances.
Till now 15 branches each with 16 leaves from 16 stems of the fourth
degree were used fully for indexing 240 leaves (objects) by 1920
digits. The shorter tree budding from the last stem is used for the
last three leaves \item \ \ 2 * 6 = \ \ 12 \item \ \ 1 * 5 = \ \ \ 5
\end{itemize}
The sum of the distances of the leaves from the root is 1937. Thus $
1937: 243 = 7.971$. The result of the division is the mean distance
which equals to $\log_2 3^4$. The estimation of the binary logarithm
of 3 is $7.971 : 5 = 1.597$. Since $\lg_2 3 = 1.585$, the precision
for such simple calculation is good and could be improved using
higher powers of the searched number close to the power of the base
number.
For example the string $aaaabbcd$ and its permutations need only 10
digits:
We have shown that the surfaces of constant energy in the phase space
are planes orthogonal to the unit vector ${\bf I}$. The system of the
ideal gas moves on this plane and for most of the time it remains on
each orbit proportionally to its volume. Therefore the system exists
in the largest orbit or orbits nearest to it for most of time. We
already know the formula for the evaluation of volumes of individual
orbits. This is the polynomial coefficient for n permutations $n!/Pi\
n_k!$. The logarithm of this coefficient was proposed by Boltzmann as
a mathematical equivalent of entropy, the H function. If n and $n_k$
are large numbers, and in the case of the ideal gas they certainly
are (the Avogadro number, determining the number of molecules in 1
mole of gas, is of order $10^{23}$), the Stirling approximation of n!
can be used and the result is
have the arithmetical mean for all sizes somewhat lesser than 1.
Starting $m_k$ values from the lowest value r, the arithmetical mean
will be always $r+1$, since we add to the basic distribution $r\times
2^{k+1}-1 $ units. The exponential slopes can be flattened by
combining several such distributions:
$$\begin{tabular}{|l|rrrrrrrr|l|} \hline $n_k$ & 8& 8& 4 & 4 & 2 & 2
& 1 &1 & $\sum 30= 2\times(2^4-1)$ \\ $m_k$ & 0 & 1 & 2 & 3 & 4 & 5 &
6 & 7 & \\ \hline $n_k \times m_k$ & 0 & 8 & 8 & 12 & 8 & 10 & 6 & 7
& $\sum 59$ \\ \hline \end{tabular}$$
The arithmetical mean grows slowly and the slopes can be flattened by
equilibrating neighbor values.
One can speculate, who was Jack with a Lantern, who changed the great
enigma connected with entropy into a greater error. Its consequences
are spread from mathematics, over physics, biology, social sciences
to philosophy.
The coefficient
chchchchchchchchchchchchchchchchchchchch
cccccccccccccccccccchhhhhhhhhhhhhhhhhhhh
2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,
and
1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,20, respectively.