Visualization of general defined space data

International Journal of Computer Graphics & Animation (IJCGA) Vol.3, No.4, October 2013

Visualization of General Defined Space Data
John R Rankin
La Trobe University, Australia

Abstract
A new algorithm is presented which determines the dimensionality and signature of a measured space. The
algorithm generalizes the Map Maker’s algorithm from 2D to n dimensions and works the same for 2D
measured spaces as the Map Maker’s algorithm but with better efficiency. The difficulty of generalizing the
geometric approach of the Map Maker’s algorithm from 2D to 3D and then to higher dimensions is
avoided by using this new approach. The new algorithm preserves all distances of the distance matrix and
also leads to a method for building the curved space as a subset of the N-1 dimensional embedding space.
This algorithm has direct application to Scientific Visualization for data viewing and searching based on
Computational Geometry.
Categories and Subject Descriptors: I.3.3 Viewing algorithms, I.3.5 Computational Geometry and Object
Modelling

1. Introduction
The Map Maker’s problem is to convert a measured space into a two-dimensional coordinatized
space. This is well achieved by the Map Maker’s algorithm [Ran13]. However the assumption in
that process is that the distance matrix of the measured space corresponds at least roughly to a flat
two-dimensional space. When this is true the Map Maker’s algorithm is distance preserving. In
general however this would not be the case and then the Map Maker’s algorithm does not
preserve all distances. A measured space is defined by a set of sites i = 1 to N and their measured
separations dij where dij is the shortest distance between site i and site j. The set of NxN distances
dij is called the Distance Matrix. This matrix is symmetric and trace-free with zeros down the
principal diagonal. Thus a set of known sites and their distance matrix constitutes what is called a
defined (or measured) space.
Research based on distance matrices has had wide applications such as in signal processing
[PC92], pattern recognition [ZWZ08] and medical imaging [PE*09]. One of the foundational
research efforts [LSB77] resulted in Lee’s algorithm for projecting n-dimensional distance matrix
data to two dimensions for visualization. Lee’s algorithm does not determine the n-dimensional
coordinates of the points involved but goes straight to a projection of those points in 2D. In the
process of this projection most of the distance information is not preserved in the visualization.
Other approaches such as LLE preserve distances just in the local neighbourhood of a chosen
point [RS00, TSL00]. In the research presented here the coordinatization and projection processes
are separated so that one is free to apply any desired projection transformation to the
coordinatized data. Surprisingly, from the distance matrix dij one can determine the
dimensionality and signature of the space that it defines as well as the coordinates for the sites
and the algorithm presented in this paper demonstrates this.

DOI : 10.5121/ijcga.2013.3402

15


The algorithm presented here, called the Coordinatizator algorithm, does not require any special
ordering of the sites before processing. The given ordering of the sites however determines the
directions and orientations of all axes in the building of the n-dimensional space. Re-orderings of
the sites therefore may be done for other reasons such as desired axial directions or reduction in
numerical errors. The Coordinatizator algorithm coordinatizes by considering the sequence of
simplices Sk = P1P2…Pk with vertices Pi for i = 1 to k and setting up orthogonal axes for these
simplices.

2. Coordinatizator Process
The Coordinatizator algorithm uses the given sites to construct axes and produce coordinates for
each of the sites. As in the Map Maker’s algorithm, site 1 is taken as the origin so it is
coordinatized as the point:
P1 = (0, 0, 0, 0...)
The direction from site 1 to site 2 is taken as the x-axis or the first axis so therefore site 2 is
coordinatized as the point:
P2 = (x21, 0, 0, 0...)
and obviously x21 = d12.
The triangle of site 1 to site 2 to site 3 defines the x-y plane and thus the 2-axis (i.e. the y-axis
which is perpendicular to the 1-axis which is the x-axis). Therefore P3 is coordinatized as the
point:
P3 = (x31, x32, 0, 0, 0...)
The tetrahedron formed by the first 4 sites defines the third axial direction perpendicular to the
first two. Thus site 4 is coordinatized as the point:
P4 = (x41, x42, x43, 0, 0, 0...)
and so forth so that in general:
Pi = (xi1, xi2, xi3, …, xii-1,0, 0, 0, 0...)
for i = 1 to N where N is the number of sites and the size of the Distance Matrix.
This process of converting the Distance Matrix into a set of N points in n-dimensional space
generally creates a non-Euclidean space of n = N-1 dimensions. We add the points progressively
by building up a simplex from 0 dimensions to n = N-1 dimensions. However it is possible that at
some step the added point is degenerate and the dimension does not increase for that step so that
in general the dimensionality of the measured space is bounded above by N-1, n  N-1. It is
interesting that every time another point say the k+1th point is added, we only have to consider k
more edges since dij = dji.
The Coordinatizator algorithm solves the following mathematical equations for the xij given the
dij. Investigation into these equations found that the algebraic solution was possible for the
general n-dimensional case.
16


Given:
P1 = (0, 0, 0, 0...)
P2 = (x21, 0, 0, 0...)
P3 = (x31, x32, 0, 0, 0...)
P4 = (x41, x42, x43, 0, 0, 0...)
P5 = (x51, x52, x53, x54, 0, 0, 0...)
:::::
Pi = (xi1, xi2, xi3, …, xii-1,0, 0, 0, 0...)
The unknowns xij in the above equations can be solved for k = 1 to N by the following procedure.
Every coordinatization is extended to higher dimensions by appending the additional required
number of zeros in the n-tuple.
For k = 1 we consider only point P1 and coordinatize it by assign it to the 0 dimensional origin.
For k = 2 we bring in point P2 and get:
d122 = (P2 - P1)(P2 - P1) = x212
and without loss of generality we choose:
x21 = d12
and thereby coordinatize P2 as the 1 dimensional point (x21).
For k = 3 we bring in point P3 and coordinatize it by getting:
d132 = (P3 - P1)(P3 - P1) = x312 + x322
d232 = (P3 - P2)(P3 - P2) = (x31-x21)2 + x322
Subtracting gives:
d132 - d232 = 2x21x31 - x212
so that
x31 = (d132 - d232 + x212)/(2x21)
and therefore
√
Now if the term in the square root is negative then we have non-Euclidean flat space. We will
consider this type of situation later on. Here we will take the positive root without loss of
generality since this will be the first point with a y component.
For k = 4 we are adding in P4 and we have these three equations to consider:
d142 = (P4 - P1)(P4 - P1) = x412 + x422 + x432
d242 = (P4 - P2)(P4 - P2) = (x41-x21)2 + x422 + x432
d342 = (P4 - P3)(P4 - P3) = (x41-x31)2 + (x42-x32)2 + x432
Therefore:
d142 - d242 = 2x21x41 - x212
so that
x41 = (d142 - d242 + x212)/(2x21)
Secondly
d242 - d342 = (x41-x21)2 - (x41-x31)2 + 2x32 x42 - x322
17


so that
x42 = (d242 - d342 - (x41-x21)2 + (x41-x31)2 + x322)/(2x32)
Finally we get:
√
generality since this will be the first point with a z component (i.e. 3rd or kth component).
For k = 5 we are adding in P5 and we have these four equations to consider:
d152 = (P5 - P1)(P5 - P1) = x512 + x522 + x532 + x542
d252 = (P5 - P2)(P5 - P2) = (x51-x21)2 + x522 + x532 + x542
d352 = (P5 - P3)(P5 - P3) = (x51-x31)2 + (x52-x32)2 + x532 + x542
d452 = (P5 - P4)(P5 - P4) = (x51-x41)2 + (x52-x42)2 + (x53-x43)2 + x542
Therefore:
d152 - d252 = 2x21x51 - x212
so that
x51 = (d152 - d252 + x212)/(2x21)
Secondly
d252 - d352 = (x51-x21)2 - (x51-x31)2 + 2x32 x52 - x322
so that
x52 = (d252 - d352 - (x51-x21)2 + (x51-x31)2 + x322)/(2x32)
Thirdly
d352 - d452 = (x51-x31)2 - (x51-x41)2 + (x52-x32)2 - (x52-x42)2 + 2x43 x53 - x432
so that
x53 = (d352 - d452 - (x51-x31)2 + (x51-x41)2- (x52-x32)2 + (x52-x42)2 + x432)/(2x43)
Finally we get:
√
generality since this will be the first point with a kth component.
This forms a clear pattern for any value of k >= 1 and hence the algorithm can be worked for any
value of N.

3. Imaginery Components
The quantities under the square root signs in the equations above could be positive, zero or
negative giving rise to three important cases to consider. If the quantity is positive we have an
extra Euclidean dimension for housing the points. If the quantity is zero then we have sufficient
dimensions already and an extra dimension is not (yet) required for embedding the points
considered so far. If the quantity is negative then the Euclidean space becomes pseudo-Euclidean
[Wik13a] like Minkowski space in Special Relativity where time is an imaginary extra axial
dimension for three-dimensional space. As in Relativity, the invariant quadratic form then is no
18


longer positive definite and is properly termed a pseudo-metric though it is common to continue
to call it a metric anyway).
Sylvester’s Theorem [Wik13b] states that the matrix for any quadratic form can be diagonalized
by a suitable change of coordinates and with rescalings the diagonal elements are only {-1,0,1}
and furthermore the number of +1s (called the positive index of inertia n+) and the number of 0s
(n0)and the number of -1s (called the negative index of inertia n-) is invariant no matter what
recoordinatization is chosen. Clearly
n+ + n0 + n- = N-1
The dimensionality of the measured space is given by
n = n + + nThe signature of the measured space is given by
s = n+ - nKnowing the dimensionality, the signature and N is equivalent to knowing n+, n0 and n- which the
Coordinatization algorithm provides. The Corrdinatizator algorithm automatically determines
these quantities by noting the sign of the quantities under the square root. The presentation of the
algorithm below explicitly indicates these quantities.

4. Coding the Algorithm
We have an outer loop for k = 1 to N. In each loop we determine k constants which are the nonzero components of Pk. Computing the next component depends on the values of all previously
computed components. Essentially we are constructing a new output NxN matrix called x from a
given input NxN matrix called d. The input matrix cannot have negative values, must be
symmetric and its principal diagonal should consist only of zeros. The output matrix is lowertriangular so that it also has zeros on its principal diagonal. In both cases the matrix indices run
from 1 to N. So the algorithm looks like this:
Inputs: The NxN distance matrix d[i,j]
Outputs: The NxN matrix of point components x[i,j].
Initialize all x[i,j] = 0 -:
for i = 1 to N
for j = 1 to N
x[i,j] = 0
PositiveIndex = 0
NegativeIndex = 0
for k = 1 to N
l = k-1
for i = 1 to l-1
sum = Sq(d[i,k]) - Sq(d[i+1,k])
for j = 1 to i-1
sum += - Sq(x[k,j] - x[i,j])
sum += + Sq(x[k,j] - x[i+1,j])
sum += Sq(x[i+1,i])
x[k,i] = sum/2/x[i+1,i]
19


sum = Sq(d[l,k])
for i = 1 to l-1
sum = sum - Sq(x[l,i]-x[k,i])
if sum > 0
PositiveIndex++
x[l,k] = SqRoot(sum)
if sum < 0
NegativeIndex++
x[l,k] = SqRoot(-sum)

5. Curvature, Dimensionality and Degeneracy
We will start by assuming that the N sites define a space of dimensionality n = N-1. If for any k
in the algorithm we find xk+1k is zero or close to it then we will reduce the dimensionality n of the
data space by 1. The Coordinatizator algorithm does not search for curvature in the space nor the
dimensionality of the embedding space, so the defined space be either a flat n-dimensional
Euclidean space or a flat n-dimensional pseudo-Euclidean space. When there is degeneracy one
has to check whether the negative is the right solution i.e. it gives the right distances to the other
points before it according to the distance matrix as described in the Map Maker’s algorithm.

6. Conclusions
The Coordinatizator algorithm presented in this paper maps sites to an n-dimensional pseudoEuclidean space preserving all distances of the given distance matrix. The algorithm returns the
positive index of inertia and the negative index of inertia of the pseudo-metric of the pseudoEuclidean space. If the negative index is zero then the space is pure Euclidean and the triangle
inequality [Wik13c] is valid for all triplets of points within it. If the negative index is non-zero
then the triangle inequality is reversed but only for the subspace of points spanned by the
imaginary axial directions. Once this algorithm is applied to defined space data projection
mappings can be applied to view the n-dimensional data on a flat screen. The Coordinatizator
algorithm effectively maps one NxN lower triangular matrix (dij) to another NxN lower triangular
matrix (xij) and this would allow us to reduce memory needs by storing data in a single square
NxN real matrix.

References
[1]. [Fer57] Ferrar W, “Algebra – A Textbook of Determinants, Matrices and Algebraic Forms”, Oxford
University Press, 2nd edition, 1957, p 154.
[2]. [LSB77] Lee R, Slagle J & Blum H, “A Triangulation Method for the Sequential Mapping of Points
from N-Space to Two-Space”, 1977, IEEE Transaction on Computers, pp 288-292.
[3]. [PC92] Potter L & Chiang D, “Distance Matrices and Modified Cyclic Projections for Molecular
Conformation”, Acoustics, Speech and Signal Processing 1992, Vol 4, pp 173-176.
[4]. [PE*09] Pham T, Eisenblatter U, Golledge J, Baune B & Berger K, “Segmentation of Medical Images
Using Geo-theoretic Distance Matrix in Fuzzy Clustering”, Image Processing (ICIP) 2009, pp 33693372.
20

[5]. [Ran13] Rankin J, “Map Making From Tables”, International Journal of Computer Graphics and
Animation, vol 3, no 2, 2013, pp 11-19
[6]. [RS00] Rowels S & Saul L, “Nonlinear Dimensionality Reduction by Locally Linear Embedding”,
Science, Vol 290, December 2000, pp. 232-233.
[7]. [Sle10] Sledge I, “A Projection Transform for non-Euclidean Relational Clustering”, Neural
Networks (IJCNN) 2010, pp 1-8.
[8]. [TSL00] Tenembaum J, de Silva V & Langford J, “A Global Geometric Framework for Nonlinear
Diemsionality Reduction”, Science, Vol 290, December 2000, pp. 232-233.
[9]. [TT*12] Tzagkarakis G, Tsagkatakis G, Starck J & Tsakalides P, “Compressive Video Classification
in a Low-dimensional Manifold with Learned Distance Matrix”, 20 th European Signal Processing
Conference [EUSIPCO 2012], 2012, pp 155-159.
[10]. [Wik13a] “PseudoEuclidean spaces” Wikipedia, 2011
[11]. http://en.wikipedia.org/wiki/Pseudo-Euclidean_space
[12]. [Wik13b] “Sylvester’s Law of Inertia, Wikipedia, 2013
[13]. http://en.wikipedia.org/wiki/Sylvester's_law_of_inertia
[14]. [Wik13c] “Triangle Inequality”, Wikepedia
[15]. http://en.wikipedia.org/wiki/Triangle_inequality
[16]. [YR13] Youldash M & Rankin J, “Evaluation of a Geometric Approach to the Coordinatization of
Measured Spaces”, 8th International Conference on Information Technology and Applications (ICITA
2013), Sydney, Australia July 2013.
[17]. [ZWZ08] Zuo W, Wang K & Zhang D, “Assembled Matrix Distance Metric for 2DPCA-based Face
and Palmprint Recognition”, Machine Learning and Cybernetics 2008, Vol 8, pp4870-4875.

21

Visualization of general defined space data

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Visualization of general defined space data (20)

Recently uploaded (20)

Visualization of general defined space data