USE OF BARYCENTRIC COORDINATES TO SOLVE CERTAIN COMPUTATIONAL PROBLEMS* S. 6.
LAVROV
(Moscow) (Received
IT is
shown in this
formulae tional
problems
of
functions,
of
functions. 1.
position,
that,
for
functions
solution
of
November
using
algorithms
Let us recall 121). Let the
[ll , the
paper
and useful
12
barycentric
can be derived of
numerous
nonlinear
variables,
equation
be selected
arbitrarily
in the
1
viz.
systems
simple
several
computa-
interpolation
and finding
extrema
n-th
Euclidean
space,
so that
. ..I
=10*11
=
.
.
.
. * * Xl,, .
Zno=nl*
not
point
equal
is
t0
If
M,).
mass
equal
zero
(Xlj,
.
a mass ~j to
is
.
.
.
centre
Cartesian
of
.
.
(1.1)
.
. . r,,,
are
Xnj
placed
the
at each
Cartesian
M, point
the
of total
gravity,
-.*
M. of
this
+I&=
(1.2)
1,
system
of
masses
will
have
the
coordinates zi =
The quantities
POxi
$
I1lTil+
p,-,, uI,
** *
+
i =
l',~ri,,*
. . . , p, are termed
l’ych.
.Math.
Vol.
4.
No.
5,
905-910. 157
1,
2, . . ., n.
barycentric
Zh.
coordinates
so that
1 PO+!%+
.
solve
determinant
D
the
coordinates,
to
the determination of barycentric coordinates (see basic points, Mc, MI, . . . . M, which are in a common
1
iS
1963)
1964.
coordinates
(1.3) of
the point M. For an arbitrary point 4! with coordinates xl, . . . , x,, the barycentric coordinates ~~~ pl, . . . . )I,, are unambiguously determined from the system of equations (1.21 and (X.3) because the determinant of this system is equal to Ct and. by hypothesis, is not zero. Some of the ctj masses for an arbitrary point M will be negative. x,) be a polynomial of the n-th degree in the vari2. Let P(xl, .+., We obtain an algebraic polynomial of the same m-th degree. It ables xi* can be transformed into a homogeneous polynomial of m-th degree in the variables ~j by multiplying each term of degree s < tn by the polynomial co+!A+
a..
06% * . -1 p,)
exactly
+&J"-**
can be written
Q (/lo,
pl,
. . .,
equal
to I. The transformed
polynomial
as
P,) =
(2.1)
where the summation is carried out over all possible combinations of the number II as the sua of a + 1 integral non-negative items, among which with a varying order of some might be equal to zero. The combinations It is easy to prove that two homoitems are considered to be different. geneous polynomials Q~@o, . . ., p,) and Q&O, . . ., p,) of the same degree condition and of identical admissible values, i.e. values which satisfy coefficients. Thus, the prob(1.2) of the variables clj, have identical lem of constructing a polynomial P(zr, . . ., zn) of the m-th degree approximating (in some sense) to a certain function F(xl, . . ., 2,). is equivalent to the probiem of approximating this function by a homogeneous i_to, . . . , gn. polynomial Q&LO, . . ., p,) of the same degree in the variables We consider the problem of constructing for the function F with a special selection polatfon.
an interpolation Polynomial of the basic points of inter-
We will split the unit mass into m equal parts and distribute them between the points MO, hf~, . . . , hi,. The number of ways in which this can be done is also equal to the number of representations of A) in the form of the ordered sum of a + 1 of integral non-negative items and, consequently, is the same as the number of coefficients of polynomial (2.1). where The point with barycentric coordinates (kw’m, kdm, . . ., k,,lm), (j = 0, I,..., nl, are integers. ko -P kl + ... +kn=mrkj>O ko, kl, . . ., k, corresponds to this distribution of masses. this point are denoted in terms of YlX3...u 11.,.X -k. rime* 8, timer
...
Values of nn...n -’
k,,eimes
the function
F at
(2.“)
Use of
We determine
having
. . ., RI)
We restrict For
the
tions
the
tarvcentric
coefficients values
our
hi.,i,, .,,,i,,
epual
to
consideration
to
the
bio9i,,
of
the
case
it
type
of
is
notations
., i,, =
m = I,
the
polynomial
of
the
values
interpolation,
of
polynomial
m = 2.
to
such
small
values
introduce
(2.2),
points. of
second
m.
nota-
namely
.,. n,,...,, -c '
ilLtime*
takes
QlPl
Q (pO,pl,
F at all
actual
×
Q (po, . . ., p,,)
the
-i- ***
+
(2 and function
form (2.3)
~A*
F at the
basic
Point
we obtain aj =
If
of
Il...1 ---
-
Q (PO,~1, . . ., p,) = UOPO-IAdjusting
function
of
“00 . 0
159
the polynomial
convenient
i. time8
If
of
those
to
unknown coefficients
similar
coordinates
the
coefficients
0, 1, . . ., n.
i =
Yj7
of
the
Q Co, . . ., pn)
(2.4)
polynomial
= i
have
cljjfl”j2j
f
j-0
j
to
be determined
&‘ikp?k’
(2.5)
‘j
By the
same method
we obtain ajj = 2yjj,
j = 0,i , . . ., n,
(2.6)
and then
We note
that
formula
(2.6).
If
m = 3.
this
formula
polynomial
remains
Q
correct
@e,. . ., p,,)
also
contains
if
j = k.
terms
of
simplifying
three
to
different
types
Q Co, . . .
...)=& j=O
For
its
coefficients
=
-
5~;,~
we successively
+
18Y.. -gYjh Jfi
+
ajk#p#r.
(2.8)
j. k. r=o j
h$
ajjj = 6~;~~ a,,k
n
I,
x -&ujj&lr + x
Ujjjpj +
obtain
/ = 0, I, . . ., n, 2Ykkk,
j.
k =
0, i, . . . , n,
(2.9) i # k,
(2.10)
160
.s. Lavrov
A.
=jkr
2
=
Y,-).rP) -
Y&f f
(y,tjj +
$ (yjjk + Yjjr f
iv k, r = 0, 1, . . ., n,
Here we are formula
restricted
(2.10)
Formulae nomial
of
verified
to
and if
can be derived any degree.
that
all
for
to
such
formulae
k f
which,
if
formula
coefficients
For polynomials
r9
i*
(2.11)
changes
Y!&r $- Yjrr f
f
y&r) f
27Yjkrl (2.11)
i # k,
one formula
k = j,
yjkk
of
a particular
by all
possible
r = j.
becomes
(2.9). the
up to the
are
r.
interpolation
5th degree case
of
poly-
it the
has
been
general
formula
where
summation
I into
is
r positive
from
1 to
notes
the
carried
III), while
items
ql>
sum in the
out
subdivisions
of
the
. . . . 9r (r assumes all integral S Ql... Qr 0'1, . . .( qs >, --- >, q,.. gxpression
whole
91,
number values
i,)
form (2.13)
'Yip ,... jp,... jp ...jp.,
-L 91time* extended
to
all
de-
combinations
of
qr times
indices
pI,
. . . . pr,
satisfying
the
limitations 1
It but
can be assumed
it
is
hardly a = 1.
its
(2.12)
Let
if
q,=qp,
zz. is
application
correct
for
any value
of
m.
to large values of I can the particular cases where and will
be used
equation
system
in the
sections. a solution
be required
fr h where
formula
that
r,
be justified in practice. However, 2 are, no doubt, of practical interest
following 3.
that
so cumbersome
PaZPpp
Pa + Pp, a, p=l,2 ,...,
fi are sufficiently
*-
-9
4
for
= 09
regular
a linear i=
1, 2, . . .) n,
continuous
functions.
(3.1) Under favourable
conditions this can be achieved by a method which is a generalization of the secant method for functions of one variable. We assume that, by a given method, the (a + 11th point of Mj has been selected with the
Cse
oj
barycentric
161
coordlnaies
coordinates I ’ = 0, 1, . . ., n,
21jV”*1xnj9 where fi
these
at the
these
points basic
are
in a general
points
of
Mj are
we approximate
values,
position.
denoted
Values
by
of
yij = fi (zij,
fi by linear
functions
the
functions
. . ., z,.,~). Using
polynomials
i = 1, 2, . . ., n. We find
a point
purpose
the
M at
Y#o
together
with
coordinates quent ly, the
which
following +
Y&l
equation
of
the
values
yi
existence
of
all
system +
the
’
-
If
the fi
Ed are
one of
the
following
should
change
can
the
M,; point
given
of
be replaced
be effected that
of
M, at which
any other
with points the
norm of
should
determinant
also
M and the order,
$j +
should
(3.1).
one of
on replacing I...
1
the
1,
. . . , n) can also
‘**
+
.
. x k=o
= /
be replaced
that
the
new by
(3.4)
has the that
highest the
Since
point
Mj by point
. . .
1
value.
new system the
.
.
.
.
.
.
pk=lk .
.
M, is
1
* . . %I .
.
.
.
=
.
coordinate
~j
should
not
However, of
new value
P,D,
k=o
sufficient
The
M as the
YXj
position.
1
2-10
is
only
basic
repeated.
point
n
it
M and
If
the
calculations
taking
be taken
be in a general
D’
Subse-
at point
system
satisfied,
(yij, . . ., Y,j)
care
(1.1))
Cartesian
(1.3).
(3.3)
in the
not
= 0.
+
vector
any substitution
the
formulae
value
I/Y:j
or
is
by point
the
from
verified
errors
(3.3)
Mj (j
this
i = i, 2, . . ., n,
in a cyclic
point
For
(3.2)
has a solution, be calculated
inequality
permissible
inequalities
points point
the
system
should
zero.
solution
i = 1, 2, . . ., n,
O*
M can be determined
functions
become
requires
Yin!& =
+
-1Yi I < ei9 where
polynomials
equations
*
(1.2).
point of
these of
be zero.
equal
basic D’ of to
162
.s. s.
If
the
smallness hazard
of
determinant,
its
change
same time not
system of
points
the
of
These
a solution
approximation
to
(1.2)
and (3.2)
is
not
solved
owing
This
Mj,
where
points,
system
the
value
of
(3.4)
is
low,
should
apparently,
(3.1)
can be obtained
method
is
not
series
of
for
the
are close to the point which using them, a more satisfactory
and,
for
ge also systems
the
functions
to be solved, it from the solution
tion
of
the
next
cannot
fi
solution
of
computational
note
that
if
to
each
other
close
have left tions
numerical
new* but the
and flexible.
fi
in the
systems
system
not
one system (in
the
is advisable to retain of one system as the
system.
vicinity
of
for
the
nonlinear here
(3.1)
value
the
functions
values
at certain
near
Bittner
the
of
points ate.
of
obtained However,
by the method view, it
ally
in the
course
at each
step,
as recommended 4.
made a detailed study of the that method. He confirms
We will
of
(3.4)
the process
the
have
to
find
In the
the
function
direction
of
the
f allows There
to
Gauss.
fi at such of
these
and rate of basic
and Interrupt increasing
this
the
tendency
volume
of
than
to do
calculations,
is
the
local
extremum
(the
definite
maximum) of
which we will consider to be sufficiently reguin addition to the existence of a maximum* moving function
maximum point
increase,
sought
a satisfactory no need to
nomial. mate formulations. It is lem into two stages: the
As indicated
of
new system.
by Blttner.
function f(xl, . .., lar. Be assume that. of
are
solution
has a tendency to degenerhere described, is more satisfactory to Intervene occasion-
considerably
x,)
vicinity
func-
points,
conditions the system
a related
in our
this
the
solution.
141 recently
convergence
fi)
points the solu-
basic
a slight deviation in the values of the functions Therefore, points should not markedly affect the quality of approximation functions
simple
but a whole
of
the
new system
equa-
is
the set of basic initial basis for
in many cases
Moreover,
be rechecked
of
described
function e. g, at those where the values of the estimation these points are far from the sufficiently high. In fact. the previous system and probably far from the solution of
*
the
unknown point.
tions
of
to
an attempt should be made to make a hapcoordinates of one or several basic points. At the
in the
those
be affected.
gives
equations
Lolirov
define
for
it
is
approximation
with
more accurately
advisable to subdivide result in the vicinity
by A.M. Ostrovskii
[31,
possible
and that.
this
in this
to
a second these
reach
the
vicinity, degree
somewhat
polyapproxi-
the solution of the probof the maximum and final
method can be traced
back
lise
of
barycentrtc
conrdlnntes
determination of the point of maximum. In the first stage, search is conducted in a region where the function f allows a linear approximation to be made, in the second stage, such an approximation has already ceased to reflect the behaviour of the function (see [51). Let us again select the basic points Mo, Ml, . . . , M,, the coordinates of which satisfy the condition D f 0. We put
Using these polynomial
values
we approximate
of the function,
‘I -;I YOPO + YlPl i
function
f by a linear
(4.1)
* * * + ?f#,,*
In the space of variables pe, vi, , . . , &i,, the gradient of the function o has components ye, ~1, . . . , _vn. We exclude from the gradient the component which does not conform to the relationship (1.2) and term the vector with the components as affine gradient
gj = Yj -Lafr n-+-l
y,,
i=
0, 1,. . ., n.
ai0
The direction of this vector depends only on the selection of the “j and the values of the function f at these points and does not with any affine transformation of coordinates in the xl, . . . . X, Subsequently, we will be interested only in the direction of the gradient; we therefore normalize it as follows:
l,b 73
rj
=
gjlg,
g=
a gs t
j=
points vary space. affine
0, 1, . . ., n.
s=1
We assume that the points values of the function f:
i.e.
,Vj are numbered in the order
of
We proceed from the point M, in the direction of the affine we take the point M with barycentric coordinates pj = rljl Pn =
1 + tr*
j = 0, 1, . . ., n -
increasing
gradient,
1,
(4.21
for t = 1. If the value y of the function f at the point M appears to be higher than yn we exclude from the sequence of points Me, Ml, . . . , , M, the first, to which coordinate Llj of point M corresponds which is not
equal
to
points the If it
zero.
We will
adopt
and assume point
calculations
new numbers
M to
starting
from the
a certain
previously
the
set
second
of
given
to
points
case
it
by a linear
the
positive of
value
one of
can be considered
polynomial
(4.1)
that
does
not
Then,
of
the
t in formula t does
not
of
t,
will
(4.2)
until than
case
point
of
a satisfactory
in
In the
function result
f because
direction of insolving the prob-
be used.
We approximate
f by a second
function
in addition
purpose,
half
same method.
approximation
give
to
all
gradient.
smaller
first
the gradient of this polynomial does not determine the crease of function f. In this case the second stage of lem should
affine
be incorporated
them by the the
subsequent we repeat
become
77. In the
number,
determined
Mj instead
by 1 for M,.
determination
we will reduce Y d Yn* subsequently, either does not show that Y > Y,,, or
M, corresponding
reduced
be the new point,
to
the
values
of
degree
function
we will use the ~111 now denote by Yjj, middle of all segments which join pairs
=
Yjk
f
( =lj
The approximating
+
Ilk
2
)
.
.
z”j~ ‘Ilk)
.(
polynomial
polynomial.
f at points
values
of
of
basic
)
can be written
the
this
For this “j
which
function points
k = 0, 1, . . . , n.
j,
we
in the
(4.3)
as (4.4)
where
(see
formulae
(2.7)
“jk =
We will tionary are
find
point).
related
and equate
the
4Yjk -
Yjj
-
partial
purpose,
we derive
j, k = 0, 1, . . ., n.
Ykk*
extremum of
For this
by (1.2)
its
and (2. g))
the polynomial q (or at least its stabearing in mind that the variables pi
the
derivatives
(4.5)
function
l.~a,pi,. . .,p,,
to zero (4.61
These
equations
Using
the
should relations
be solved (4.5)
together
and (1.2)
this
with
equation
system
(1.2).
can be changed
to the
(‘se
of
barycentric
coordrnatcs
165
form
(4.7)
where
we have
If the
this
system
point
place
put
found
one
of
has is
the
Vj basic
points
If,
at
unique this
distance
of
In
practice
comparing mating
can
value
polynomial
quality (1.2)
this the
of
of
the the
or
its
presence
often
be
the
9 at
the
approximation.
overcome
stationary It
the
an
extremum
from
if re-
value set
of
a
of
the
are
If
other
function
process f.
at
points
a of
derivatives.
of
the
f
function
problem
sufficient
in
9.
thus
equations
of
the
evaluating (4.4).
(4.6)
at
itself.
reliability
value
and
have
random. by
course the
8 difficult the
not point,
at
there
partial
point
follows
does basic
simply
of
first
with
the new
elucidated
points,
with
f
function
or be in
9 is
that the
(4.7) a new
point
of
polynomial
of
find
should
basic
ascertnin we again
again.
difficulty the
to not,
For
system to
extremum
to
function of
point
it
If
ensuring
gradient
cause
any
f.
replaced. (4.7)
made
affine
the
examined
point
be
process, be
its
from
this
verification
a stationary
the
near
order
of
this
that
when.
one
discontinuity The
of
new
system
can
recurs, shows
arises
the
be
function
Mj to
solve of
should
the
by
attempt
direction
Experience
described
and
feature
for
point
stage an
the
special
means.
the
we derive a certain
in
points
to
solution,
moving
maximum
basic
= 0 corresponds
it
a solution,
the
by approxithe and
that
h
(Jo=-As the
a control, point
it with
is
recommended
barycentric
.
= -; that
the
quality
of
the
approximation
at
coordinates
p. =
PI
=
.
.
.
=
fin =
L n +I
should
also
In the ordinates
be
verified.
problem are
investigated
particularly
the clear.
advantages Uith
movement
of
using along
barycentric the
affine
cogradient
166
each
,‘. .>.
vaiue
of
definition
new
of
the
mation
coefficients
the
the
function
direction
calculated
of
of
Lavro1,
this
the
facilitates
approximating
polynomial
simple to calculate (and on changing fron equation (4.7) the need to calculate these coefficients is for only
each
succeeding the
(n f l),
calculated
which
gradient
and more accurate new value is
for
(i.e.
the
the
function what is
are
of
this
investigated,
approximation)
approxi-
extremely
(4.6) to equation generally avoided)
determination
minimum of
a linear
of
the more precise
With a mean square
gradient.
required
has to
to
and
polynomial,
calculate
by a conventional
be the
difference
method. Translated
by 2.
Semere
REFERENCES 1.
MGbius,
A. F. , Dcr
analytischen 2.
Byushgens. Part
3.
4.
S. S.,
I,
Ostrovskii,
A.M.,
nie
uxavnenii
Bittner,
L.,
Box.
Calcul,
dcr
Analyticat
Solution
i sistem
G. E. and Wilson.
conditions,
of
J,
Roy.
Sot.,
and
zur
132’7.
geometriyaf.
Equation
Moscow.
Izd-vo
zur Auflijsung
Mech.,
K. B. , On the Stat,
Hilfsmittel
Barth,
~Analiticheskaya
Equations
uravnenii),
Math.
neues
GONTI, 1939.
Mebrpunktverfahren Z. Angew.
ein
Leipzig,
Ceometrie.
Geometry
Moscow - Leningrad,
systemen. 5.
barycentrische
Eehandlung
43,
No.
No.
lit.,
1963.
von Gleichungs-
3,
experimental 3 13,
(Reshe-
Systems
in.
1,
111-126,
1963.
attainment l-45,
1951.
of
optimum