Documente Academic
Documente Profesional
Documente Cultură
A. S. SOLODOVNIKOV
SYSTEM S
OF
LINEAR
INEQUALITIES
Mir
Moscow
A. C.
CJ4CTEMM
JTI4HEhHMX HEPABEHCTB
<<HAYKA>
MOCKSA
A. S. Solodovnikov
SYSTEM S
OF
LINEAR
INE QUALITI E S
Translated from the Russian
by
Vladimir Shokurov
MIR PUBLISHERS
MOSCOW
Ha
I43LtaTeJThCTBO
((HayKa>), 1977
CONTENTS
7
17
22
25
31
44
52
57
9. Incqmpatible Systems
64
69
81
84
91
101
107
Preface
First-degree or, to use the generally accepted term, linear inc qua!-
ities. In this connection one can mention the names of such mathe7
skian geometry"), G. F. Voronoi (one of the fathers of the "Petersburg school of number theory"), A. Haar (a Hungarian mathematician who won recognition for his works on "group integration"),
H.Weyl (one of the most outstanding mathematicians of the first half
of this century; one can read about his life and work in the pamph-
(x1,
Thus
similar coordinates.
Fig.
OM' = k x OM; for k < 0 the point M' lies on the extension of
the ray OM beyond the point 0, with OM' = IkI x OM (Fig. 2).
7k<O)
Fig. 2
the point M1 + M2 is the end ofjhe vector 0M1 + 0M2 and the point kM
is the end of the vector k x OM (on condition that the point 0 is the
where s1, s2 are any two nonnegative numbers the sum of which
equals 1.
Fig. 3
Fig. 4
and 0M1 we obtain the point N1 on the segment 0M1 and the
point N2 on the segment 0M2 (Fig. 3). Let
M2M
M1M
'
M2M
ON2
0M1 =
M2M1
0M2
M1M
M1M2
1.
M1=sM+(1s)M2 (O<s<1)
0
Fig. 5
Fig. 6
Hence
+(1t)M2
where t = 1/s. Let the case where M2 lies on the segment MM1
be considered by the reader.
(3) When a parameter s increases from 0 to 00, the point sB
runs along the ray OB * and the point A + sB is the' ray emerging
from A in the direction of Oil. When s decreases from 0 to 00,
the points sil and A + sB run along the rays that are supplementary
to those indicated above. To establish this, it is sufficient to look
at Figs. 5 and 6.
It follows from proposition (3) that, as s changes from
co
The
dinates 0.
by definition,
z1 + z2)
M1 + M2 = (x1 + X2, Yi +
kM = (kx, ky, kz)
All the propositions proved above will obviously be true for
space as well.
Fig 8
Fig. 7
(the
is a
+2'
Keir and Le2'.
So
one
then all sets obtained in this way must be united into one. It
the latter that will be .K + 2'.
is
2. ir
and
parallel to each other (Fig. 9). Then the set Jr +2' is a parallelogram with sides equal and parallel to ir and 2' (respectively).
What will result if the segments Jr and 2' are parallel?
0
Fig. 9
4.
Fig. 10
0,
Fig. 11
second by all points lying "below" or on the line (Fig. 12). If,
however, b = 0, then the inequality is reduced to one of the
following forms
or
the first of them being satisfied by all points lying to the "right"
of or on the straight line x = h and the second by all points to
the "left" of or on the line (Fig. 13).
Thus equation (1) gives a straight line in the coordinate plane
and inequality (2) gives one of the two half-planes into which
this line divides the whole plane (the line itself is considered to
belong to either of these two half-planes).
We now want to solve similar questions 'with regard to the
equation
ax+by+cz+d=0
(3)
of course, here x, y, z
are
Fig. 12
z=kx+ly+p
(5)
Denote by 2' the set of all points M (x, y, z) which satisfy (5).
Our aim is to show that 2-' is a plane.
Find what points in 2' belong to the yOz coordinate plane.
To do this, set x 0 in (5) to obtain
z=ly+p
(6)
Thus the intersection of 2' with the yOz plane is the straight line u
given in the plane by equation (6) (Fig. 14).
Similarly, we shall find that the intersection of 2-' with the xOz
(x,
s, kx + p + si)
'V
\\\\\ /////
x<h
Fig. 14
Fig. 13
It is
satisfy equation (5), i.e. that A + sBe 2'. This proves that the
z0)
To do this, consider three points: a point M (x0,
lying "above"
z0
lying in the plane it, a point M' (x0,
z0 c) lying "below" it
the plane it (c > 0), and a point M" (x0,
(Fig. 15). Since Me it, we have z0 = kx0 + 1Yo + p and hence
zo + 8> kx0 +
zo 8
+p
This shows that the coordinates of the point M' satisfy the
strict inequality
z> kx + ly + p
and the coordinates of the point M" satisfy the strict inequality
z < kx + ly + p
16
Thereby M' and M" do not belong to 2'. This proves that 2'
coincides with the plane it. In addition, it follows from our
arguments that the set of all points satisfying the inequality
ax + by + cz + d ?
is
one of the two half-planes (either the "upper" or the "lower" one)
1M'(x0,g0,z0*s)
x9,y0,z9)
Fig. 16
Fig. 15
x and y be given
a2x+b2y+c2?0
)
(1)
The region
is called the feasible region of the system (1).
Note from the outset that a feasible region is not always bounded:
Fig. 17
Fig. 18
ture on linear inequalities this term does not designate the line itself but
all the points of a plane which are spanned by it (i. e. lie inside or on
the line itself). It is in the sense accepted in the literature that the term
"polygon" will be understood in what follows
18
Proof Let ir1 and Jr2 be two convex sets and let Jr be
their intersection. Consider any two points A and B belonging
Be
and the set
is convex,
to ir (Fig. 20). Since A e
the segment AB belongs to
Similarly, the segment AB
belongs to Jr2. Thus the segment AB simultaneously belongs to
and
and hence to their intersection Jr.
both sets
Fig. 19
Fig. 20
(2)
0)
Fig 21
Fig. 22
Fig. 23
the surface rather than to the surface itself, the set of course including
the surface itself but only as its part
ax + by + cz + d ? 0
ax
which
by
cz
ax + by + cz + d =0
The latter gives a plane it in space. The remaining inequalities (2)
Fig. 24
the region ir
the region ir
21
Fig. 25
Fig. 26
It is fairly clear that by that time the bag will have taken shape
of a convex polyhedron with vertices at some of the points A1, A2, ..., A,,.
22
s,,
the point A,, with the points of the set <A1, ...,
Proof: For notational convenience, denote the set <A1, ...,
A,,> by .H,,.
by J41 and the set <A1
...
+ s,,.
A,,_ i>
i+
where
s1, ...,
Si +...+Sp=1
If = 0, then A ed/,, ; thus the set /1,, then A = A,,; thus the point A,, belongs to
is
a part of 1/,,. If
So
.4
1,
contains
and the point A,,. We now show that any segment A'A,, where
A' e. - belongs wholly to
If A is a point of such a segment, then
A==tA'+sA,,
(t,
t+s= 1)
t1+...+t,,1=1)
hence
Setting tt1 =
tt,,_1 = s,,_1, s = s,,, we have (1), (2). This proves
that A E
So any of the above segments belongs wholly to .
23
+...+s,,_1)[3
-
1,
1,
+...
+
sPA,,
Since the coefficients of A' and A,, are also nonnegative and their
sum is one, the point A lies on the segment A'A,,. This completes
the proof of the lemma.
Now it is not difficult to see that the visual definition of the
convex hull given at the beginning of this section and the strict
definition which follows it are equivalent. Indeed, whichever of the
two definitions of the convex hull may be assumed as the basis,
in either case going over from the convex hull of the system A1,
A,,_1 to that of the system A1, ..., A,,_1,
follows one and the
same rule, namely the point A,, must be joined by segments to
..., A,,_1
(this rule is
B=t1A,+t2A2+...+t,,A,,
where
Si,
...,
t1, ...,
0
(3)
24-
C=sA+tB
where
s+t=1
s,
which gives
C = s(s1A1 + ... +
= (ss1
(4)
means that the point C belongs to the set <A1, A2, ..., A,,>, i.e.
this set is convex.
At the same time it is easy to see that <A1, A2, ..., A,,> is the
least of all convex sets that contain the starting points A1, A2, ..., A,,,
viz, that it is contained in any of these sets. This statement follows
directly from the lemma proved above and from the definition of
a convex set.
This explains the name "convex hull" and at the same time
provides yet another explanation of the fact why the set <A1, A2,
A,,> can be obtained using the method described at the beginning of
the section. For the set spanned by a rubber ioop (or film) after
the latter has tightened to its limit around the system of the points
A1, A2, ..., A,, is exactly the least convex set that contains the
indicated points.
ax + by + cz = 0
(the absolute term in the equation must be equal to zero, otherwise
(0, 0, 0) will not satisfy the equation). Thus a convex p025
lyhedral cone with the vertex at the origin is the feasible region
of a system of homogeneous inequalities:
+ hmY + CmZ
amx
Fig. 27
(a)
(b)
(1)
Fig. 28
qf the ray.
8. A point. This "cone" can be obtained by taking the common
part of a ray and the corresponding half-space (Fig. 28, h).
Of course, the enumerated examples 1-8 diverge (to a greater
or lesser extent) from the usage of the, word "cone", but we are
compelled to reconcile ourselves to this if we are to preserve the
general definition of a convex polyhedral cone given at the beginfling of this section.
We now try to show in a few words that the sets pointed out
above exhaust all polyhedral convex cones in space.
27
tiBi+t2B2+...+tqBq
where t
t2
What is the geometric meaning of the set (B1, B2, ..., Bq)?
It is clear from the definition that it is the sum of the sets
(B1), (82)
(B,); we must first see, therefore, what is the
geometric meaning of the set (B), i.e. of the set of points
of the form tB, where t is any nonnegative number and B a fixed
Bq
3;
ight line or a ray. Now it becomes clear that there must exist
a close relation between the sets (B1), (B2), ..., (Be,) and convex
polyhedral cones. Such a relation does in fact exist. For greater
intelligibility we shall formulate the corresponding propositions as
two theorems.
Theorem 1. The set (B1, B2, ...,
of space or
else
is
the origin.
(C)
(a)
(b)
Fig 29
(B1, B2, B4), (B1, B3, B4), (B2, B3, B4) is an infinite trihedral
pyramid with vertex at the origin. The set (B1, B2, B3, B4) apparently
fact, prove it
to be valid
);
Fig. 31
Fig. 30
...,
(Fig. 31).
Apparently,
B2'k2B2,
where k1, k2,
..., kq
...,
and hence
B' = S1B1' +
Sq
30
...
kB
(k> 0), we
find that
B=t1B1 +t2132+...+tqBq
where
(i=1,2,..., q)
1thus we have shown that any point B of the pyramid .K belongs
Bq). The converse (i.e. that any point of the
to the set (B1, B2
belongs to
is obvious. So
coincides
set (B1, B2, ...,
with (B1, B2, ...,
to 8 can be
The case where Jr is one of the exceptional sets
proved without much trouble and is left to the reader.
1
,the fact that the number of unknowns is not large (only two),
we shall try to carry out the analysis of these systems from
general positions, i.e. so that the results obtained may be easily
extended to systems in a larger number of unknowns.
The solution of any system of linear inequalities is in the long
run reduced to the solution of a number of systems of linear equations.
+ b,,,y + Cm
it the
a2x + b2y
0
(2)
0
amx +
a2x+b2y=0
=0
We
(3)
and
let B be an
+ b2yA + c2 ? 0
a2xB
+ b2yB
and
+ bmYA + Cm 0
amxB + bmYB
Adding each inequality written on the left to the corresponding
inequality on the right we have
OmXA
al(xA+xB)+bl(yA+yB)+cl
XB)
Jljj)
Cm
if, however, one wants to record the fact that one set is a part of another,
the symbol c
32
is
used.
then P A e
(2)
If a straight
line belongs
wholly
to .K and A, P are
L--.
0
Fig. 32
A+sB
(4)
al(xA+sxil)+bl(J'A+syB)+cl
a2(xA
am(x4 + sx8) +
(5)
33
b2yB
+ b,,y8 0
whence we see that the point B belongs to the set
a?,xB
The proof of (2) is carried out in a similar way. The straight line
under consideration consists of points of the form (4) where s is
an arbitrary number. Therefore inequalities (5) are valid for any
value of s. Hence it follows that in each of these inequalities the
total coefficient of s must be equal to zero, i. e.
+ blyB 0
+ b2y8 = 0
=
+
Therefore B e 2'. Thus the lemma is proved.
It is easy to see that Lemmas I and 2 are valid for systems
a,flxB
It is not difficult to show that a compatible system of inequaliand only if its feasible region
contains no
ties is normal
straight lines.
Indeed, if the system is normal, i.e. the set 2' contains only
the origin of coordinates, then the region
does not contain any
straight lines, which follows directly from the second statement
34
is
set
contains at least one point B different from the origin of coordinates. Of course, all points of the form kB, where k is any number,
also belong to 2' But in this case, whatever the point PEit"
(and such a point is sure to exist for the system is compatible
and the region it" is therefore not empty), the set of all points of the
form P + kB (where k is any number) belongs, according to Lemma
is
(6)
Fig 33
a
have its beginning or end at this point (Fig. 33, a and h, where
the point A is one of the vertices; in Fig. 33, b the region
it" is a segment).
We now explain at greater length why the convex set it" we are
35
segments and the beginning of any of the rays will evidently be the
vertices of
One can find the vertices of the region .K without much trouble.
Notice first of all that in the vOv coordinate plane the ith inequality
of the system (1) corresponds to the half-plane whose boundary
line
is given by the equation
+ h,r + = 0 (i = 1, 2, ..., tn)
Evidently, the point A of the region .( is a vertex if and only
if it belongs to two different boundary lines.
Let us agree to call regular any subsystem of two equations of
the system
a1x + h1t' + c1 = 0
a2x + h2v + = 0
(lix +
(6)
=0
From the above description of the vertex now results the following
the
x+
x
2x
On solving the subsystems
1),
(1. 2).
(2, 0)
of which only the second and the third satisfy all the inequalities
given. So it is the points
A,(1, 2) and A2(2, 0)
A2
A,,
different
from the vertices of the region. The line A1 P intersects the convex
either along a segment A1A (Fig. 34) or along a ray
with origin at A1 (Fig 35). In the second case P A1
(Lemma 2), hence Pe A1 +
In the first case we reason as
follows, however. If the point A lies on the bounded edge A.A of
the region .K (as in Fig. 34), then P belongs to the convex
region
Jr0.
Among the boundary lines there are in this case at least two
different ones (for the system (2) is normal !). Hence
either
coincides with the origin of coordinates (v = 0,' y 0), or is a ray
with vertex at the origin of coordinates or an angle smaller
with vertex at origin. If we know two points B1 and
than
lying on different sides of the angle (Fig. 37). then all the
Fig. 34
Fig 35
A2
Fig 37
Fig. 36
points of thc
(8)
where t1 and
38
B=tB1
(9)
4,
Fig. 38
Example
2.
x+
(10)
2x
as well as the region
of the solutions of the system in Example 1.
Solution. The system (10) is normal; for the unique solution of
the corresponding homogeneous system of equations
x+
(11)
2x v==0)
is (0, 0).
Choose a point satisfying the first equation of (11) (but different
from (0. 0)), the point C( 1, 1), for example. We make sure by a
simple check that the point C does not satisfy all the inequalities
(10), therefore, neither the point itself nor any point of the ray OC
(different from the origin 0) belongs to
On considering the
point C (i.e. the point (1, 1)) we find that it belongs to
So B1 = (1, 1). The second equation is satisfied by the point
(2. 1): it is also a solution of the system (10), so that B, (2, 1).
39
The region
consists
i1B1 + i,B,
1)
+ 212.
where
Lv
Fig. 40
Fig 39
r consists
So
=(2s+t1+2t2,2st1+t2)
where s is any number in the interval [0, 1] and t1,
t2 are
nonnegative numbers.
x+
Proceeding as in Example 2 we find only a single ray:
(t?0)
(Fig. 40).
'i'
+1
3v +
40
any
+=0
3x + y = 0
has solutions (except (0, 0)) which would satisfy all the given inequalities. The region
consists of a single point (0, 0), the
origin of coordinates.
$9
Fig. 41
-fP.
is a1v + h1 r = 0. One of
The equation of the straight line
of this equation is different from zero;
the coefficients ai or
let h1 0, for example. One can then take as a straight line r not
the v-axis (its equation is
0). In this case the set
parallel to
is the part of the v-axis contained
.K ,- now denote it by -K
by .K. To find this set, one should put v = 0 in the system (1).
+ Cm
may
in one unknown y which is easy to solve Notice that the set
be either an empty set (then ,K is also empty) or a point, a segment or
a ray (but it cannot be the whole of the y-axis, for otherwise ir would
be the whole of the plane, which is impossible). On finding this set we
x+
2x + 2y + 3 0
It is easily seen that the system is not normal and that
is the
straight line
x+y=
y1
0
0
to 2.
*
Notice
42
Fig. 42
(14)
The augend of the sum is the convex hull of a system of points A1,
A2
t1B1 + t2B2 +
numbers.
...
+ tqBq, where
is
(why?).
..., Bq)
in the plaile iS either tlit' itIiok' p'ane or some conrev pOIi(JO!ia! region
in it.
0
Fig 43
either the whole plane (this is the case if Jr0 is the whole plane)
or some convex polygonal region in it
44
+ (1Z + (11
+ hrnV + CrnZ
('m
as in Section 5.
we
+ cr,:
(2)
.)
and
am +
= 0j
We shall again denote the feasible region of the system (1) by .V.
that of the system (2) by .11 and that of the system (3) by
is some ConTo use the previously introduced terminology.
vex polyhedral region in space and .110 is a convex polyhedral cone.
+ e,,1z + il,,, = 0
45
2x+ y+
x+2y+
(5
x+
y+
2x+ y+ z=O
x+2y+ z=O
x+ y+2z=O
x+ y+ z=O
Solving it we see that the unique solution
is (0,
0, 0); so the
2x+ y+ z1=O)
x+2y+
2x+ y+ z1=0
x+2y+ z1=O
x+y+2z1=0)
x+y+z1=O
2x+ y+ z1=0)
x+2y+ z1==O
y+2z1=O
z1=O
x+
x+ y+
tems are regular and that their solutions are the points
(1/4,1/4,1/4), (0, 0, 1), (0, 1, 0), (1, 0, 0)
46
of which the first does not satisfy the system (5) and the remaining
three do. Therefore, the vertices of the region
are the following:
A1(l, 0, 0), /42(0, 1, 0), A3(0, 0, 1)
2O. The normal homogeneous system of inequalities (2). Each of the
inequalities (2) determines a half-space whose boundary plane passes
= (B1,
where B1, B2
Bq are some points chosen one at a time on
each edge of the cone
(see Theorem 2 of Section 4). One can
find such points from the following considerations. Each of them
(i) belongs to
i.e. satisfies the system (2), and (ii) belongs to the
line of intersection of two different faces, i.e. satisfies disproportionate * equations of the system (3).
2x+ y+
x+2y+
x+
y+
and, further, the region .K of the solutions of the system in Exampie 1.
Notice first of all that the system (6) is connected with the system of
inequalities (5) of Example I; namely (6) is a homogeneous system
corresponding to (5). Hence the system (6) is normal.
* We call the equations
+ by + ez = 0 and a'x + b') + c'z = 0 "disproportionate" if at least one of the equalities a/a' = h/h' = c/c' fails to hold;
in this case the corresponding planes intersect along a straight line.
47
2x+ y+
z=05
v+2y+ :=O1
x+ y+ z=OJ
x+y+2z=O
z=O
For each of these six systems, choose two nonzero solutions: (x, y, z)
and ( x,
v,
1, 1)
and (
3,
1) for the first system: inequalities (6) are satisfied only by the
first of these solutions. Hence we have the point B1 = (3, 1, I)
Proceeding in similar fashion with the remaining five systems we
find the points B2 = ( 1, 3, I) and B3 ( 1, 1, 3). So the
consists of points of the form
region
t1B1 + t2B2 + t3B3 =
t2 + 313)
r3,
=
t3, t1 +
1,
are
where
The case where the system of inequalities (1) is not normal. This
48
Fig. 44
0)
b1y + ciz + d1
b,,1y + c,,,z
+ ci,,, ? OJ
2x+ y+
3x
* y + 4z
x2y+3z
(9)
* his easily seen that the system (7) is now normal (see a similar
footnote on page 42)
49
2x+ y+ z=0)
(10)
x2y+3z=0j
2x+y+ z=0
and
3x
+ 4z = 0
intersect.
Choose some point B on the straight line 2' different from the
origin of coordinates. To do this, it is enough to find some
three numbers x, y, z (not vanishing all together) which satisfy
the first two equations of the system (10). Take 1, 1, 1, for
example. So 2' is the straight line OB, where B = (1, 1, 1).
It is easy to see that the straight line 2' is not parallel to, say,
the yOz coordinate plane. Setting x = 0 in the system (9), we have
the system
y+
feasible region
is a set consisting of a
single point A(3/5, 2/5) (in the yOz plane). Hence the sought region
consists
A
+ tB = (o,
+ t(l,
1,
1) = (t,
+ t,
+ t)
coordinate axes. Assume, for example, that the z-axis is not parallel
to 2'; take it to be .3. To find the set .W, the part of the
z-axis contained in
50
c1z + d1 ? 0
(11)
0
c,,1z4
Fig. 45
Remark.
is found to
be empty, so
is .Jr.
+2
(13)
xy+zO
x + y z = 0
* The system (11) is normal.
(14)
therefore, the
given by the equation
+ 0
point and hence is not parallel to the z-axis. Find the set .K.
Setting x = 0, r = 0 in the system (13). we have the system
of the form
<A1,
in
is either the
Bq)
region in ii.
The proofs of both theorems follow almost word for word those
of the corresponding theorems in the two-dimensional case. Their
elaboration is left to the reader.
numbers
v1, X,
called the coordinates of
s1M1 +s2M2
where
arc any two nonnegative numbers whose sum is one.
Going from three-dimensional space to n-dimensional space we adopt
s'M' + s"M" =
(s'x'1 +
any two
s' = 1, s" =
s" are
one.
When
we
have
the
point
M'.
when
53
where at least one qf the numbers a1, a2,..., a,, (coefficients of the
from zero, is called a hyperplane. For n = 3 the
equation (2) assumes the form aixi + a2x2 + a3x3 + h = 0, which
unknowns) is
a1x1+a2x2+...+a,,x,,+b?O
holds and the region where we have
(4)
x1 =
x2 =
+ (1
+ (1
x,,
54
+ (1
1)
is equal to
0 and (1
s)
0).
(3),
efficient,
coincides with the convex hull of' somne finite s vstern 0/ pOints.
56
t1B1 +t2B2+...+tqBq
are any nonnegative numbers. A convex polyhedral
cone is defined as the intersection of a finite number of half-spaces
whose boundary hyperplanes have a common point (the vertex of the
cone). The validity of Theorem 2 in three-dimensional space was
established in Section 4 (Theorem of Section 4).
Theorem 2' Any convex polyhedra! cone with Lertev at origin can
B,,).
be represented as (B1, B2
The validity of the theorem for the three-dimensional case was
proved in Section 4 (Theorem 2 of Section 4).
region can be represented as
Theorem 3. Any convey
where t1,
the sum
+ (B1,
<A1, A2,...,
B2,...,
B,1)
Theorem 3'. Any sum of the indicated form is either the whole
qf space or some convex polyhedral region in it.
2x3i+
x
+2:=5
4i :=5
then, on solving each of the equations for :,
(1)
we have
2x+
(1')
4y_5)
2x+3y 1
2x+3y1=
(2
4y5J
(2'
y= 2x+4J
after which we get the equation
2x+4
involving now one unknown x. On solving the equation we finc
x = 1, after which from the system (2') we find v = 2, and
from the system (1') we have z = 3.
It turns out that something of the kind can be done with an
system of linear inequalities. Solving a system of inequalities in i
unknoWns x1,..., x,,_ x,, is thus reduced to solving a system c'
inequalities in n I unknowns x1,..., x,,_ , then the resulting systen
can be reduced to a system in ii 2 unknowns x1,...,
2 and
ar
on and on till we come to a system in one unknown
58
to the
In the case of > 0, transfer all the summands but
right-hand side of the inequality and divide both sides by a,, to
get the inequality
So on multiplying each of the inequalities of the original system
by an appropriate positive number we get the equit'alent system
of the form
Pi
P2
(fl
R,.
Qq. R1
*Ofcow.se, if the system (S) has no inequality in which a,, <0. then the
;ystem (1) will not have the first block. Similarly, if there are no inequali-
'es with a,,> 0, then the system (7) will not have the second block.
if (S) has no inequalities with a,, = 0, then the third block will
missing from
59
where
1.
p. J3
2,.... q, and
is any of the numbers 1. 2,.... r.
Along with the system ('1) consider the system
(where
numbers
in
(S')
0j
n I
1,
respect
we get a
The
of the unknowns satisfies the system (S), then it does the system (T) too
but then all the inequalities of the system (S') hold for the same set too).
= x?,...,x,,
P1,,
Q?....,
R1
for them
being
1.
p. and
any of the
If the first or the second block is found to be missing from the system (1),
then (S') will consist only of the inequalities R1 0 If the third block is
missing from (T), then (S) will have only inequalities
Q1y
60
r).
Tile first group of the above inequalities (i.e. (4)) shows that each
of the numbers
P?,...,
13
nn
p
4
pO
p0pO
1
Fig. 46
'n--1
of the unknowns
Xfl = '(fl
a solution of the system (T) and hence of (S). Thus the theorem
is proved.
The following two additions to the theorem will play an important
role below.
I
A system (S) of linear inequalith's is compatible if a;uI onli if so
is
is
theorem proved.
2. A ii the solutions of the original system (S) can he obtained in
the flu/owing vay. One must adjoin to each solution
...,
of
the concomItant system (S') any of the miumbers
lying between all
the numbers Q?
and all the numbers P?
As a matter
of fact this proposition was proved during the proof of the theorem.
in two unknowns
(S) tile feasible region of
y. Denote by
the system (S) (it is a certain set of points in space) and by .K
(S') the feasible region of the system (S') (a set of points in the plane)
In geometric terms the theorem proved has the following meaning.
The region .11 (5') is the projection of the region
(S) on the
xOv coordinate plane.
(S') involving the unknowns x1, x2 ,...,X,,i, and for the system
(S') one can in turn construct the concomitant system (S") (involving the unknowns
x2
for the latter system one can
construct the concomitant system (S") and so on.
Continuing the process we shall, after a number of steps, come
compatible or not.
Suppose that the system is compatible. Then the problem arises
of solving the system or, speaking at greater length, of indicating
all of its solutions (all sets of values of the unknowns which satisfy
the inequalities of the given system). We shall adopt the point of
view (though it may at first appear to the reader strange) that the
system (S) is solved
We shall presently see what explains such a point of view but first
we shall introduce a
DefInition. A set of values of the first k unknowns
o
o
0
X1,
X2,
..., Xk
(5t11
1)));
3t))
so on.
62
...,
1))
x y
2x+3y+
5x y+
On solving each of the inequalities for z
system
in the form
7
2x3y+1
()6
y2
The concomitant system
is of the form
2x3y+l
5x+ y2
2x3y+1
5x+ y2
or, after the cancelling of the terms,
?0
+6
4x 2i'
2x + 3
3
8
X+
(7)
63
9.
So far
systems are not only of interest in themselves, but also give key
to the understanding of a variety of important facts. Thus the main
linear programming theorem (the duality theorem, see Section 14) is
in the long run derived from some properties of incompatible systems.
Consider an arbitrary system of linear inequalities. For notational
convenience consider for the lime being that the number of unknowns
is three, although everything stated above equally applies to systems
a,x + h2v +
0
(1)
+
+ k2h2 + ... + k,,h,,)y +
+ (k1 Ci + k2c2 + ... +
k 1(/1 + k2c12 + ... +
0
(2)
Ox x+O x y+O x
(3)
65
in our system.
When n = 1 the system is of the form
a1x + b1
a2x + b2
0
(4)
(l??X +
are nonzero.
One can suppose that all the coefficients a1, a2
Indeed, if for example, a1 = 0, then the first inequality is of the form
x + h1 ? 0; if the number h1 is nonnegative, then such an inequality may be discarded; if b1 is negative, however, then the first
inequality of our system is inconsistent and there is nothing to prove.
is zero.
So let us suppose that none of the numbers a1, a2
It is easy to see that these numbers are sure to contain both positive
and negative numbers. Indeed, if all of the indicated numbers had the
o x
same sign, were they positive, for example, then the system (4) would
a'
)
would hence be compatible.
Assume for definiteness that the first k numbers of a1, 02 ak are
positive and the remaining in k are negative. Then the system (4)
and
66
1,
-'4
a,,,
h1
--->
a1
(6)
ii,,,
a<
(7)
(it should be remembered that a1 > 0 and a,,, <0). If the first inequality of (4) is now multiplied by a positive number a,,, and
the last inequality by the positive number a1, then after addition
we have the inequality
0 x x + (h,,,a1 h1a,,,) 0
3*
67
x,,
to introduce clarity into the concept of combination. A combination of equations is constructed in the same way as a combination
of inequalities, the only difference being that one is allowed to
multiply the given equations by any, not only nonnegative, numbers.
Especially important is one particular case of the theorem on incompatible systems of inequalities, namely, the case where a given
system contains the inequalities
...,
Denoting the rest of the system by (S), one can say that the problem
is to find all iwnnegat ire solutions of the system (S) (i.e. the solutions satisfying conditions (10)). If the problem has no solutions, then
according to the theorem just proved a certain combination of inequalities of the system (S),
a certain combination of inequalities of
(10),
0
+ k2x2 + ... +
nonnegative)
(k1, k2,..., k,,
k,,
0,
a<0
with in the present section: the general case (i.e. the case of a
69
two or
There is no need here to confine ourselves to the case
three unknowns, from the outset we shall consider a system of any
number of inequalities in any number of unknowns. For the reader's
convenience our account is broken down into a number of points.
10. A linear function of n arguments. The general form of a homogeneous inequality in n unknowns is
of n-dimensional space.
short:
L(X)
L(kX) = kL(X)
where X is any point and k is any number.
2.
L(X+
Y
Property
follows in
a1(kx1) + a2(kx2) +
...
X=
70
(x1,
and
Then
c,,x,,
where
X = (xi,
establish the following two propositions.
a solution of the system (2) and k is any
nonnegative number, then the point hX is again a solution of the
We
system (2).
(for k
0 and LAX)
0) as well as
+ Y) =
(for L1(X)
0 and L1( Y)
0).
consequence
X2... . X are solutions of mite system (2).
if certain points X
...
k1,X1)
k2X2
xi, x2,...,
of the homogeneous system (2) is said to be the fundamental set
of solutions if any solution X of the system can be given by the
formula
X = k1X1+ k2X2 +
...
k,,X1,
where k1, k2,..., k1, are nonnegative numbers. It follows that in this
case formula (4) in which k1. k2
k,, are any nonnegative numbers
gives a review of all solutions of the system (2). Hence it is clear
set of solutions for any system (2) to be found using quite simple
operations is what we set ourselves as the final object.
40 Construction of the fundamental sd for a system concicting of
a1, a2,...,
x,, =
Setting x1 =
=
that
1,
(6')
x2 = O,...,x,,
,0,
-
= (o. o,...,
"p-
1,
Now let
i,
Ot,,
ri,,)
(7)
/ aj\ + /
a1\
--
0, ...,
1,
...,
;-
(o, o,
...,
i,
"')
(9)
Indeed, let
numbers
X,.
X=
L(X) = L(X')
x1=
= X+
recall that the point X is a nonnegative combination of the
points
74
2x1
4x2 + x3
2x1
4x2
+ x3 =
X3=(X1+X2)=(l, 1, 6)
and finally
2x1
4x2
+ x3 =
X=
1,
1,
6)
k4(O,O, 1)
or
x1 =
=
k1 k3
k2 k1
x3=2k1 +4k26k3+k4
where k1, k2, k3, k4 are arbitrary nonnegative numbers.
50
The rearrangement of the structure of the fundamental set of
solutions when another inequality is added to the system. In order to
learn how to construct fundamental sets of solutions, first consider
a problem like this.
Suppose we are given the homogeneous system
L1(X)
(12)
L,,(X)
set of solutions
x1, x2,
...,
to (12).
The solutions of the system (12) are precisely all nonnegative
combinations of the points X1, X2,.... X1,. We must pick among
these combinations those which would satisfy inequality (13) and,
what is more, constitute the fundamental set of solutions for the
All points X1, X2, .., X,, can be divided into three groups
with respect to the function L(X), the left-hand side of inequality (13):
the points for which L(X)> 0, the points for which L(X) < 0, and
finally the points for which L(X) = 0. We designate the points of
the first group as
group as X X
vU /12
VU
vO
It1,
Thus
y+
y+
I'
and
y--
yO
yO
combination
+
so that it should satisfy the condition L(X) =
0.
To do this one
should take
a = -- L(X1).
b = L(Xfl
76
)L(X;) + L(X;)L(X14)
Denote the point (14), where a and b have the values indicated
above, by
(16)
the fundamental
set
and
points X,
of the
X
the point
a
h
and il/h. If the first is greater than the second,
the
where
> 0, we have
X =
are nonnegative.
is obviously
nothing to prove. Suppose therefore that there are strictly positive
where all the coefficients a1 ,...,
h1
h1, c1,...,
L(X) = h1L(Xfl +
...
+ b1L(Xfl + ciL(X?) +
+ c5L(X.?)
Xr
1.
If at
it
is
found that not all of the coefficients of Xj,..., XI are zero in the
newly obtained expression for X, then we again replace one of the
sums of the form
a nonnegative combination of
the points X,
or that of the points
as a result the
number of nonzero coefficients of
again
decreases at least by 1. We continue in this way till an expression
is obtained for the point X in which all the coefficients of Xj
X1 are zero. We then come to the equality of the form
X
= X + ... +
+ >2
+ ciX'? + ... +
where all the coefficients on the right are greater than or equal to
zero. But this is exactly the required representation for X. Thus
the theorem is proved.
The existence am! the method for the construction of the fundamental set of solutions. Let us consider an arbitrary system of homo-
5
construct the
fundamental set of solutions for the system consisting of the first
two inequalities. We then adjoin the third inequality and so on until
we have the fundamental set of solutions for the whole of the original
system of inequalities. This proves the existence, and at the same time
points out the method for the construction, of the fundamental set
of solutions.
78
for which one can at once point out the fundamental set of solutions, then that subsystem should be taken as the starting point;
successively adjoining to it the remaining inequalities one will arrive
(19)
0J
all nonnegative solutions, i.e. all solutions which satisfy the conditions
x1
0
0
(20)
It is easily seen that for the system (20) the fundamental set of
solutions will be the set of the points
= (1, 0, 0, 0), X2 = (0, 1. 0, 0)
X3 = (0, 0, 1, 0), X4 = (0, 0, 0, 1)
cz4) of the system (20) can be repre(indeed, any solution (oh,
sented in the form
+
+ x4X4). Adjoin the first ine+
quality of (19) to the system (20) and, using the theorem of sect.
construct the fundamental set of solutions for the system obtained in
this way. For computational convenience make the following table:
L1(X)
X1
X2
X3
X4
79
is
type arc X .
X2.
that
and
= 3X3 + 5X1 =
(5, 0, 3, 0)
5,4,0)
= 6X3 + 5X4 =
(0,
0, 6, 5)
for the system consisting of (20) and the first inequality of (19).
Adjoin the second inequality of (19) to that system and make
the following table:
L2(Y)
0
Y2
Y4
13
is now
is
It
played h
We find points of
=
=
=
1"2
13Y1
+ V3 =
+3 =
+ V4 =
80
and that
type:
(15,
(
0.
10.
15, 15,
0) =
5( 3.
0.
2. 0)
0)
5 ( 0,
3.
3, 0)
45,
5) = 5(13,
0,
9.
0.65,70,
15)=5( 0,
0.
(65,
0,
form
1)
13. 14,3)
of solutions for the system (19), (20). The general solution is of the form
X=aY1
+ bY2
+ eY?4
k6,
we have
(21)
a,,x,,
(1)
+
... +
a,,x,, at
0)
b,,x,, ht
ct
(2)
in n + unknowns x1, x2
v,. t.
There is a relation of certain kind between solutions of the systems
1
is
.,
(3)
we have
satis-
of the
Thus, in order to find all the solutions
system (1) one should find all the solutions of the system (2)
for which t> 0, and transform each of them according to formulas
(3).
3x1 4x2
+ 5x3 ? 6
x1
x2
2x1+3x23x3+
xl
x2 ?: 0
0
82
x2=5k2+3k4+
= 3k1 + 4k2 +
t = k5 + 3k(,
where k1,
find the general solution of the system (4) from the formulas
_5k1+3k3+13k5
x2
X3
'k5+3k6
+ 3k4 + 13k6
3k1
since
+ 3k6
(i =
1.
2, 3, 4)
=
the numbers
and
k
6
k5
3k6
X2, \3) =
0. 3) +
3, 3) +
5, 4) +
0, 9) +
0, 2) +
(6)
83
X5
3).
X2 =
(0, 5. 4).
(13, 0. 9),
X3 = (3. 0. 2)
x = (o.
as on
all the possible variants over to find the best. The situation
is
changing now. Over the past few decades the complexity of production has increased to such an extent that it has become impossible
just to look the variants over. The factors influencing decisions have
the first type and so many items of the second per diem) which
would ensure the largest profit for the factory.
84
The desired production plan is specified by means of two nonnegative integers x, y (x being the number of items of the first type and
y that of the second) which must satisfy the following conditions*:
2x + v
is
maximal
x+v
150
one should pick the one imparting the largest value to the linear
function
1= 2x + y
Given an
Fig 47
The first condition comes horn the assembly shop Indeed, it can turn
out three items of the second type instead of one of the first Hence, in
teirns of second-type items the shop's entire production is
-i- v articles:
this number must not exceed 1O()
85
From the same figure one can see that the point P (75, 75), one
of the vertices of the polygon, is the solution of the problem.
Indeed, consider a straight line 2x + y = c where c is a certain
number. Denote the line by
As the number c increases, the line
shifts "upwards" (at the same time remaining parallel to its initial
still has
of n variables
f= c1x1 + c2x2 +
...
on condition that all these variables obey a given system of linear inequalities (which must without fail contain the nonnegativity
constraints on the variables: x1 0
0). In some problems (such
as in the example above) an additional requirement is imposed on the
variables, that they should be integers.
Example 2 (diet problem). There are several kinds of food at one's
disposal. It is necessary to compose a diet which would, on the one
hand, satisfy the minimal human body requirements for nutrients
(proteins, fats, carbohydrates, vitamins, etc.) and involve the smallest
expenditures, on the other.
Consider a simple mathematical model of this problem.
Suppose there are two kinds of food, F1 and F2, which contain
the nutrients A, B. C. It is known how much nutrient of one or another
kind is contained in 1 lb of F1 or F2; this information is presented
in the following table.
of
In addition te
body
El.
lb
a1
12,
lb
'2
data we
1.
B.
C2
c1x1 + c2x2
(2)
clxi
+ c2X2
CJ
87
x12,
From D1
From D2
Amount
delivered
ToC1
ToC2
ToC3
x11
x12
x13
a1
x23 -
a2
x21
b2
1)
The total amount of coal shipped
must be a1; hence
we have the condition
x11 + x12 + x13 = ai
A similar condition must hold for D2:
x21 + x22 + x23
We
x13 + x23 =
88
X11 +X12+X13=a1
x21 + x22 + x23 =
a2
(3)
x12 + x22 =
x13 + x23 =
h3
X22, X23
of the system (3) such a solution that the function S attains its
smallest value (is minimized).
a,,i x1 +
a
linear function
f= C1X1 +
(II)
0,
'2
(III)
of the system (I) such that the function f assumc's the minimal value (is
minimized).
The equations of (I) are called the constraints of a given problem.
connect-
dxl + c2x2
x5
mm
( 1)
linear programming problem. Given a certain system of linear equaand a certain linear function f, it
tions in a unknowns v ,
is required to find among the
solutions of the given
system such a solution which minimizes the function f.
To begin working by the simplex method it is necessary that the
=+
+ 135x5
+75X5
where
=C+
+ C5X5
set the nonbasic unknowns equal to zero
.1
C4X4
\4O,
*
The
92
and find from the system (I) the values of the basic unknowns:
The
method falls into a number of steps. Each step consists in going over
from a given basis B to another basis B' so that the value of / should
decrease or at least not increase:
The new basis B' is derived
from the old basis B in a very simple way: one of the unknowns is
eliminated from B and another (a former nonbasic) unknown intro-
x4+2x5
x2=2+2x4
x5
x3=33x4
x5
/=x4x5
basic solution is
(1,
2, 3, 0, 0)
x5 =
we
+ 2x4
X2
x1=1 x4+2(2+2x4x2)
x3=33x4 (2+2x4x2)
x4 (2+2x4x2)
1=
Thus the problem is reduced to the form
x1=5+3x42x2
x5=2+2x4 x2
x3=l5x4+ x2
1=
x4 + x2
Let us see
if it is
equation to prevent it, and it is seen from the third that x4 can
be increased to 1 / 5 (at most, for otherwise x3 will become negative).
x2 = 0, we have
= 28/5, x5 = 12/5, x3 = 0. As a
Setting x4 =
0, 1/5, 12/5).
The new basis now consists of x1. x5, x4. The equation for x3 (the
94
x4 =
x1= 28
I'
and
1=
11
+4
2.
x3=1+x1 x2
x4 = 2
x1 + 2x2
x2
Here the basis is formed by the unknowns x3. x4. The basic
solution is of the form
(0, 0, 1, 2)
the corresponding value of the function I being equal to 0.
In the expression for f the coefficient of x1 is negative, therefore
we try to increase x1 (without changing .v2 = 0). The first equation
95
does not prevent us from doing so. and it is seen from the second
= 3,
equation that x1 can be increased only to 2, which yields
= 0. The new feasible solution is (2. 0, 3. 0): for this solution
X=
x2
x1= 2x4+2x2
/ = 2 + x1 3x2
The new basic solution is
(2, 0, 3, 0)
the corresponding value of f being equal to 2.
In the last expression for I the coefficient of v2 is negative. We
want to see how much we can decrease the value of f by increasing
(while keeping
0).
+ Xr+ I +
+ x, + +
...
+
+
XI
= T + Y1 + X, + +
...
X1 = Ct +
13 +
X2, ...,
Ct,
13,
..
(3)
7x5
Xi+3X22x3+x4+3x5_
it is required to solve it for
Let
41
5f
V2 (one
=4(2x1 +2x22x3+x4+7x5)
Y2=S( x1 +3x22x3+x4+3x5)
Obviously all solutions of the original system (4) can be obtained
by taking all the solutions of the system (5) which satisfy the
conditions Yi =0, Y2 =0, and keeping only the values of x1
x5
in each of those solutions.
The unknowns Yi' J2 form the basis in the system (5) Suppose
that starting from this basis we succeed in going over to another
*
97
should be able to use the simplex method for this purpose too;
namely, using this method we shall work out the problem of minimizing the function
F = Yt
+ Y2
0, ..., x5
0.
0, Y2
0.
Here the premises under which the simplex method begins to work
are satisfied since the system (5) is of the required form (it has
the original basis Yi. Y2). In some steps we arrive at the desired
minimum. Since F 0 and hence mm F 0, two cases are possible:
1. mm F> 0. This means that the system (5) has no nonnegative
solutions for which Yi =0,
=0 (otherwise mm F = 0). Therefore
the original system (4) will have no nonnegative solutions in this case.
Hence it will follow that any linear programming problem with such
a system of constraints is insoluble.
2. mm F
is an
=9+3x1 5x2+4x32x4--10x5
the expression for F the coefficient of x4 is negative,
we there-
x4=4+2x1
Y2=' X1
+4x5+y1(
X2
F=1 x1 x2
(6)
+4x5+2y1
x4=64x2+2x3+
x1=1
+4x5+y1 Y2r
x2
(7)
F=y1+ Y2
Thus as a result of the second step the artificial unknowns Yi' Y2 have
x4=64x2+2x3+ x5
x1=1 x2
+4x5
This means that we have separated a basis in the original system
(4). Thus the problem is solved.
As already noted, when applying the simplex method to finding
the minimum of the function F equal to the sum of the artificial
unknowns, it may happen that by the time the process is over (i.e.
when mm F = 0 has already been attained) some of the artificial
unknowns still make part of the basis. These unknowns can also
be transformed into nonbasic ones by simple methods which we
shall demonstrate using an example.
Suppose, for example, that after a certain number of steps the
problem has assumed the following form
X2
11
v2= x1
+
99
5
-
v1 -t
11x4
-
5x
72x4 -r 34'C5
in the equation
and
12x1 72x4-i-34x5=-0
Since here all the coefficients of the unknowns have the same
sign, this must yield
Xc
14.
+ ... +
x,,
+ C1
+ C2
0
0
(1')
0
+ ... + an,,,ym +
+
of n linear inequalities in in unknowns and the linear function
=
+ b2y2 + ... +
a1 nYl
max
values x?,...,
..
a1
...
+ h,,
+ .. + a,,,1 y, + c1
+ a,,,
I
+ h1
+ c,x,
h1
...
v1
+
- h,,y,,
+ (,,
the numbers x?,..., x?, give a solution to problem A and the numbers
a solution to problem A', with = Po.
y?,
tion requirements. In fact, the solution of the system (S) (in the
range of nonnegative values of unknowns) is of course not a bit
easier than the solution of the original linear programming problem
(problem
however, the very possibility of such a reduction is
very curious.
Now we prove the above statement. First of all, it is clear that
are nonnegative and satisfy the system (1); simithe numbers x?
are nonnegative and satisfy (1'). Morelarly the numbers
over, these numbers satisfy the inequality
103
i. e.
maxf=min p.
suppose that the system (S) has no nonnegative solutions. For this
case we have the corollary of the theorem on incompatible systems
(Section 9). True, this corollary refers to a system consisting of
0 type and our system (S) has inequalities
inequalities of the
of the 0 type. But this is easy to mend by writing the system
(S) in the form
amixi + ... +
1x1 + ... +
+ bm
C1
... amiym
...
h1y1 ...
104
(2)
a,,1k1
...
+ ens
a11!1 ...
aH,l!l
0)
b1s
...
oJ
0)
biki+...+bmkmCili...Cnln<O
We first of all show that the number
i.e.
a11?J1 +
...
0
+
If we now add to each inequality of (1), where x1 ,...,
amiAJi +
...
x,, are
the corresponding inequality of (5), we
substituted for by x?
find that the numbers
x? + Ala,...,
The numbers k1
Ii
1,,, s are
we multiply the first, second, ..., (in + ii + 1)th inequalities of the system
(S') to get (after addition) the inconsistent inequality a1x1 + ... + a,x, +
0 and c/is a negative
+ .. + aj,, + d 0. where a
+ a'1
a'1, , a
number.
+ ... +
X (c1
11
+ ... +
indefinitely with
Thus s
k1/s,..., k,,,/s
But this is contrary to the lemma. So on supposing that the system (S)
1=
2x2 + 12x3
provided the variables x1, x2, x3 are nonnegative and satisfy the
inequalities
x1x2-
x1 x2 4x3 +
SolutIon. Let us call the set problem problem A. The dual problem
(problem A') must be stated as follows: find the minimal value of
the function
= 2r1 + v,
provided the variables
ties
Yi
yr
Problem
Y2
1'
(6)
Y10Y2 coordinate plane the feasible region of the system (6). This
is done in Fig. 48. it is seen from the same figure that the
106
Fig. 48
A1, A2,...,
at point A.(i =
1,
expected at point
(I = 1,
It is required to develop
an optimal shipping schedule, i.e. to compute what quantity of goods
ThB3
From A1
x11
From A2
x21
x12
X33
h1
x14
a1
x23
From
Requirements
x13
Stocks
h2
-________
a2
X34
a3
b4
X11 +x12+x13+x14a1)
X21 +x22+x23+x24=a2
+ x32 + x33 + x34 = a1)
108
(1)
x134-x23+x33=b3
x14 + x24 + x34 =
Notice that equations (I) and (2) are very easy to remember.
Indeed, the ith equation of the system (1) implies that the sum
of the unknowns in the ith row of the transportation table is
equal to a,: we shall call equations (1) horizontal for this reason.
Similarly, the jth equation of the system (2) records the fact that
the sum of the unknowns in the fth column of the transportation
table is equal to h1; because of this we shall call the equations
S=
(3)
put hefoie
where the
signifies that the quantities
must be summed over all I = 2. 3. and all = 1. 2, 3, 4 (there
will be 12 summands in all).
We thus come to the following linear programming problem.
Given the .system of equations (1), (2) and the linear function
of the system such thai
(3). Find among the nonnegative
,r
= 60, a2
12=60,
80,
(13
= 100
b3=80,
1,4=60
Table 2
82
B3
Stocks
B4
40
60
80
A2
100
40
60
80
60
40
20. Note
20
of the point
It
B2
A1
B4
20
20
80
A2
100_
80
60
60
We enter the number 20, this being the maximum of what can
be transported from A1 to B2, in the cell x1 The requirements
of B2 will be reduced to
40 and the stocks of A1 will be found
fully exhausted. In virtue of this the A1 column of Table 3 may be
temporarily removed. Thus we come to Table 4 which has only two
source points, A2 and A3, and three destination points, B2, B3, B4,
now.
Table 4
B2
40
B3
B4
80
100
40
80
60
* When
On
=60
x23
= 40,
x33 = 40,
A2
A3
B1
B2
40
20
40
B4
B3
40
40
60
is
equal to the
every step we fill exactly one cell, after which one row or one
column is deleted from the table; there is one exception, where the
table consists of a single cell on filling which we exclude at once
both the row and the column. Since the number of all rows is m
and the number of all columns is n, it is clear that the number
of all steps and hence the number of filled (basic) cells is m + n 1.
We now show that the system of constraints (1), (2) can be solved
for the unknowns corresponding to the basic cells. Hence it will follow
that these unknowns may be assumed to be basic and the remaining
.-
X13
X11
x21
x22
X31
X32
X24
-.
X33
X34
x12,
This does not complete the solution of the set problem since
the expressions for the basic unknowns in terms of the nonbasic
ones still remain to be found. But these expressions will not be
required in explicit form at the next stages of solving the transportation problem.
3. Solving the problem by the method of potentials. The finding
of the first basis but a preparatory stage of solving the problem.
When this stage is over, all the unknowns are found broken down
into two groups:
nonbasic unknowns
xkl, basic unknowns, and
No actual expressions for the basic unknowns in terms of nonbasic
function S as
S=
we
where
is
a ton of goods from a point Ak to a point B1. The set of the equations of the form (6) worked out for all basic unknowns Xkl forms
a system of linear equations. This system has in + n 1 equations
(as many as there are basic unknowns) and rn + ii unknowns ctk, J3,
(as many as there are source points and destination points taken
together). It can be shown that this system is always compatible,
it being possible to arbitrarily specify the value of one of the unknowns and then unambiguously find the values of the remaining
unknowns from the system.
Let us fix some one solution (x1,..., cc,,
of the system of
equations (6) and then calculate the sum cx,, + I3q for each nonbasic
unknown xpq. We denote this sum by c,q:
+ 13q =
and call it indirect cost (in contrast to the real cost Cpq). It then
turns out that in the expression (5) that interests us the coefficients
of nonbasic unknowns are equal to
spq = cpq Cq
(7)
Formula (7) will be proved in the next section.
If all the variables Spq are nonnegative, then the initial basic
solution will be optimal. If, however, there are negative variables,
say
among them, then we go over to the next basis. it will
be recalled that this step begins with reasoning as follows. We shall
increase Xp0q0 (keeping the other nonbasic unknowns zero). If in
doing so we reach a moment when one of the basic unknowns,
say xkl, vanishes, we go over to new basis by eliminating the
from the old basis and introducing instead x
unknown Xk
the new basis completes one step of the
Going over
method.
114
So far we have presented only some of the considerations necessary for the understanding of the method of potentials. As
to the actual solution of the problem, it can be best explained
by considering a specific example.
Example. We are given three source points, A1, A2, A3, with the
stocks
a1=60, a2=80,
a3=100
and four destination points, B1, B2, B3, B4, having the requirements
b1 =
40,
b2 =
60,
b3 =
80,
b4 = 60
The values
(costs of transportation from A, to B3 per ton of
goods) are given by Table 6.
Table 6
1
20
40
40
40
this solution by
rollows for it:
60
equations
= c1i =
cx1 +
+132=c12=2
=3
+ T32 =
133 =
= c34 =
ot,=2,
1330.
+ 134 = 0'
11
C24
C31
C32
=
=
1'
P1
+ 134
'V
1
=
=
2'
3'
116
Table 9
Table 10
o
2'
1'
2'
3'
0'
s13=31'=
s14=4O=
s21=42'=
2,
4,
2,
1' = 1,
s31 =02'= 2,
s32 = 2 3' = I
S24 = 0
(while
2O+p
p L.
J40.-p
60
and vertical links: the broken line is dotted in Table 11. One
of the vertices of the broken line is in a nonbasic cell (x31 in
this case) and the rest are in basic cells (not necessarily in all of them;
80
40
the
118
60
60
(jn (8) we set x31 = p =40 and assume the values of the remaining
ttnknowns to be zero). Going over to the new basis completes one
'ep in the procedure. We have succeeded in decreasing the value of
The function S by 80 units.
(4) finding the new basic solution and the new value of the
function S.
would have meant that the basic solution X was optimal Then
stages (3), (4) would have become unnecessary. Thus the nonnegativity
of all d?fferences Spq is the criterion for terminating the procedure.
3'
2'
0'
1'
1'
1'
S24 = 0
iIl =
1' =
s32=21'=
s13 = 3 3' = 0
S14 = 4
2'
S21 = 4
0'
S=
=
=
s22=
119
60
80p
o+p
60p
40
60,
60
x
20
3
40
60
60
2
1
3'
1'
S13 =
3 3'
S14 = 4
1'
=
=
S22 =
S32 = 2
3 1'
1'
0'
=2
=1
s21=40'=4 s34=10'=l
120
Now the differences Sf,( are all nonnegative. This means that the
x12=x24=x33=60, x23=20,
x31=40;
the remaining
are zero.
In conclusion we show once again, but now only schematically,
the plan of solving the transportation problem:
and
Choosing a negative difference
constructing the recalculation cycle for
the corresponding vacant cell
Note. If the differences spq are all nonnegative, then the last basic
solution is optimal.
4. Justifying the method of potentials. The formula
Spq =
Cpq
Cq
(7)
L..
20+p
J 4.0...p
60
which contains the recalculation cycle for the cell x31. We used
this table to find p and thereby go over to a new basic solution.
But the table allows still another problem to be solved: it helps
to determine what are the coefficients of the unknown x31 in the
expressions for the basic unknowns.
We first of all make the following remark. If the values of nonbasic
unknowns are indicated, then this unambiguously determines the
1 in x11
+ I in x12
1 in x22
+ I in x23
1 in x33
x31
122
for S
S=
(i
1,
2, 3; j = 1, 2, 3, 4)
S31=C31C11+C12C22+C23C33
Hence
+132)+
or
S31
But
C31
that is just formula (7) (for the given case). The reasoning that
consumers are large enough (and this is what happens in appliwhere these numbers are of an order of hundreds), resorting to a computer is inevitable. Algorithms computer programmes
based on considerations other than those underlying the method
rof potentials. The interested reader can find more about these
in special literature devoted to the transportation
TO THE READER
Mir Publishers would be grateful for your comments on
the content, translation and design of this book
We would also be pleased to receive any other suggestions
This book explains in a popular form the methods of approximation, solutions of algebraic, trigonometric, model and
other equations.
Intended for senior schoolchildren, polytechnic students, mathematics teachers and for those who encounter solutions of
equations in their practical work In the course of exposition
elementary ideas about higher mathematics are
introduced in the book. About 20 solved exercises are
included in the appendix.
some
V. DYACHENKO, D.SC.
L. KALUZHNIN, D.SC.
INDUCTION IN GEOMETRY
L. GOLOVINA, D. SC. and I. YAGLOM, D. sc.
The book tells about the relation of systems of linear inequalities to convex polyhedra, gives a description of the set of all
solutions of a system of linear inequalities, analyses the questions of compatibility and incompatibility; finally, it gives an
Mir Publishers
Moscow