Discrete Applied Mathematics 14 (1986) 47-56 47

North-Holland

A SYSTEM OF LINEAR EQUATIONS RELATED TO THE

TRANSPORTATION PROBLEM WITH APPLICATION TO
PROBABILITY THEORY

T . F . LIN*
Department of Business Mathematics, Soochow University, Taipei, Taiwan, China

Received 2 October 1984

Revised 14 August 1985

Let P = [Pij] be a m × n matrix and let C be the coefficient matrix of ~ = 1 PijXij = Ui' 1 <--i<__m,
~i=m1 PijXij
=Oj, 1 <_j<_n. The relation between the reducibility of P and the rank of C is in-
vestigated. An application to martingale extension is given.

Keywords. Transportation problem, cell, loop, conditional expectation, martingale, optional

sample theorem.

1. Introduction

Let P = [Pij] be a m × n matrix. C o n s i d e r a system of (m + n) linear e q u a t i o n s

j!lPijX~j=u~, l <_i<_m,
(1.1)
~i~=lPijXij=oj, l <_j<_n,

where x = ( X l l . . . . . xln,x21 . . . . . Xzn . . . . . Xml . . . . . Xm,) is the u n k n o w n vector. If

P O - - 1 , then (1.1) becomes the constraints o f a t r a n s p o r t a t i o n p r o b l e m (see [2, p.
205]). We shall study the r a n k o f the coefficient m a t r i x of (1.1) a n d give a solvability
c o n d i t i o n for the system (1.1). T o this end, let us define

Definition 1.1. A m × n matrix M = [ao] is reducible if there exist S C { 1, 2 . . . . . m},

T C {1,2 . . . . . n}, at least one o f t h e m is a n o n e m p t y proper subset, such that aij = 0
for i e S c, j e T a n d i e S, j e T c. M is irreducible if it is not reducible.

R e m a r k 1.2. A matrix M is reducible if a n d only if it can be reduced, u n d e r row

exchanges a n d c o l u m n exchanges, to a block matrix of a n y o f the following three
forms

'6
* This research was partially supported by National Science Council of ROC under grant no. NSC
75-0208-M031-01.
0166-218X/86/$3.50 © 1986, Elsevier Science Publishers B.V. (North-Holland)
48 T.F. Lin

A large part o f this article is devoted to the discussion of the relation between the
reducibility of the matrix P and the rank of the coefficient matrix C o f the system
o f equations (1.1). The central result is T h e o r e m 2.8 which says that the rank o f C
is m + n - 1 if P is irreducible. T h e n it can be shown that (1.1) is consistent if

ui = ~ oj, (1.2)
i=1 j=l

P is irreducible. (1.3)

In the derivation o f T h e o r e m 2.8, we need the notions o f 'cell' and ' l o o p ' (see
[2, p. 209]). For each 1 <_i<_m, 1 <_j<_n, (i,j) is called a cell. A sequence o f cells
f o r m s a loop if this sequence is o f the f o r m (k, p), (k, r), (s, r), (s, q) .... , (u, v), (u, p).
In Section 2, we shall construct a m a x i m u m cell set which has (m + n - 1) elements
and contains no loop. This a m o u n t s to say that the rank o f C is m + n - 1. In the
last section, we give an application to conditional expectation (see [1], [3]) in a finite
discrete probability space.

2. Consistency

Let us rewrite equation (1.1) into

C'-PllXll + ... +PlnXln =U 1

P 2 1 X 2 1 + . . . +P2nX2n = U2

P m l X m l + ... + PmnXmn = U m
P11X11 +P21X21 + -.. +PmlXml =01

~_ PlnXln + P2nX2n + ... + PmnXmn = On

(2.1)

and let C denote the coefficient matrix o f the system o f equations (2.1). F r o m the
left hand side o f (2.1), it is easy to check that

R a n k ( C ) < m + n - 1. (2.2)

Let E k, 1 _<k < m + n, denote the k-th column vector of the identity matrix I(m + n)
and let Cij, 1 < i < m , 1 <j<_n, denote the ( ( i - 1)n + j ) - t h column vector o f C. T h e n

C 0 =Pij(Ei +Em +j). (2.3)

Every column vector Cij corresponds to a cell (i,j). F r o m (2.3) and T h e o r e m 10-2
o f [2, p. 210], we have

L e m m a 2.1. Let W be a set o f nonzero column vectors o f C. Then, elements o f W

 Linear equations related to the transportation problem 49

are linearly dependent if and only if the corresponding cells, or a subset o f them,
can be arranged into a sequence which forms a loop.

We shall show that under the condition (1.3), there exists a maximum cell set con-
taining (m + n - 1) cells such that no subset of them can form a loop. Therefore,
Lemma 2.1 implies that Rank(C) > m + n - 1. To see this, we first define

Definition 2.2. Let R be a set of cells. We say that two cells (i,j), (k, l) are rowmate
(columnmate) in R if they belong to R and if i - - k (respectively, j = l).

Remark 2.3. Let R be a set of cells. If elements of R form a loop, then every cell
has at least one rowmate and at least one columnmate in R.

From Remark 2.3, we can prove easily, by mathematical induction, the following
Lemma.

L e m m a 2.4. let Ri, 1 <_i<_k, be sets o f cells, R*= U~:I Rj, 1 <_i<_k. Suppose that
they satisfy the following conditions.
For odd i, each element o f R i has no columnmate in R*. (2.4)
For even i, each element o f R i has no rowmate in R*. (2.5)
Then R~ contains no loop.

Suppose that P satisfies the condition (1.3). Let S l = { 1 }, T 2 = { j : l<_j<_n,

Plj--/:O}. Note that T2~0 from Remark 1.2 and assumption (1.3). Define induc-
tively, for k > l,
SEk-1 = {i: 1 <_i<_m,pij--/:O for some j ~ T 2 k _ 2 } ,

I T2k= {j : 1 <_j<_n,Pij*O for some i~$2~_1}. (2.6)

It is clear that {$2k_1}, {T2k} are increasing sequences of subsets of {1,2 ..... m}
and { 1,2 ..... n} respectively. Therefore, there exists a smallest natural number k 0
such that Sk0=Sk0+2 (if k 0 is odd) or Tk0= Tko÷2 (if k o is even). We have

L e m m a 2.5. Let k o be the smallest natural number such that S k o = S k o + 2 o r Tk0=

Tko+Z. Then the following inclusions are proper:

I SICS3C'''CSk°" i f k o is odd, ko->-3, (2.7)

T2C T 4 C ' " C Tko+I,
SIcS3c...CSko+I, if k o is even, ko>_2,

And,
I T2C T4C...C Tko, if k o is even, ko>_3. (2.8)

I S k o ----S k o + 2i, Tk o + 1 = Tk o + 1 + 2i, i >-- 1, if k o is odd,

(2.9)
S k o + 1 = S k o + 1 + 2i, Tk o = Tk o + 2i, i >_1, if k o is even.
50 T.F. Lin

Let us construct a sequence o f cell sets {Ri} as following. Let R l = {(1,j) : j e T2}.
Then R 1 satisfies condition (2.4). F o r every i E S 3 - S 1 , (2.6) implies that
{ J : J ~ T2,Pij ~ 0} ¢ 0. Pick an arbitrary element f r o m thi set and denote it by f(i).
Let R 2 = {(i,j) : i e S 3 - S l , j = f ( i ) ~ T2}. Then R l N R 2 = 0 and R 2 satisfies condition
(2.5). Similarly, for each j e T4- T2, {i :ieS3,Pij~O } 4:0. Select any element f r o m
this set and denote it by g(j). Let R 3 = { ( i , j ) : j e T4 - 7"2, i = g ( j ) e S 3 } . Then Ri,
1 _< i < 3, are disjoint and R 3 satisfies condition (2.4). If we continue this construct-
ing procedure, we obtain a sequence {Ri} o f disjoint sets o f cells. F r o m (2.7), (2.8)
and (2.9), we see that

L e m m a 2.6. Let {Ri} be constructed as above. Then

(a) Ri#:O f o r i < k o and Ri=O f o r i > k o.
(b) R i, 1 <_i<_ko, are mutually disjoint.
(c) R i, 1 < i <_ko, satisfy conditions (2.4) and (2.5).
(d) N(RI)=N(T2), N ( R 2 ) = N ( S 3 -$1), N ( R 3 ) = N ( T 4 - T2). . . . .

N(Rgo)=N(Tko+l- Tko_l), i l k 0 is odd,

N(Rko) = N(Sko+ l - Sko_ l) i l k 0 is even,
where N(. ) denotes the number o f elements o f the set (.).
L e m m a 2.7. Let R~o = [,-Ji=
ko 1 ei" Then R~o contains no loop and N(R~o) > m + n - 1.

P r o o f . By L e m m a 2.6(c) and L e m m a 2.4, Rff0 contains no loop. To see that

N(R~o) _>m + n - 1, we first suppose that N(R~o ) < m + n - 1. L e m m a 2.6(b), (d)
and the fact that N(SI) = 1 imply that

N(Rff0) = ~N(Tko+l)+N(Sko ) - 1, if k 0 is odd, (2.10)

(N(Sko+l)+N(Tko)- 1, if k0 is even.
Let S=Sko, T= Tko+l if k 0 is odd, S=Sko+l, T= Tko if k 0 is even. Then, S C
{1,2 .... ,m}, T C { 1 , 2 . . . . . n}. According to (2.6) and (2.9), pij=O for i e S , j ~ T c
and for i ~ S c, j e T. We check f r o m (2.10) that N(S) + N ( T ) < m + n. Hence, either
N ( S ) < m or N ( T ) < n . Thus at least one o f S, T is a proper subset (both o f S, T are
nonempty). Therefore, P is reducible according to Definition 1. I. This contradicts
to the assumption (1.3). Therefore N(R~o ) >_m + n - 1. This completes the p r o o f o f
L e m m a 2.7.

Let H = {C O : (i,j) e Rff0}. F r o m (2.3) and the construction o f R i, 1 < i<_ k0, H is
a collection o f n o n z e r o column vectors o f the coefficient matrix C. F r o m L e m m a
2.7, Rff0 contains no loop, therefore, L e m m a 2.1 implies that vectors o f H are
linearly independent. Hence (see [4, T h e o r e m 24, p. 114])

R a n k ( C ) >_N(H) = N(R~o) _>m + n - 1. (2.11)

F r o m (2.2) and (2.11), we obtain

 Linear equations related to the transportation problem 51

Theorem 2.8. I f P is irreducible, then the rank o f the coefficient matrix C o f the
system o f equations (1.1) is m + n - 1.

Theorem 2.9. The system o f equations (1.1)/s consistent under the conditions (1.2)
and (1.3).

Proof. Let C* denote the augmented matrix o f (2.2). Then condition (1.2) implies
that the sum of the first m equations is equal to the sum of the last n equations.
Hence, at least one o f the equation in (2.1) is redundant. Therefore,

Rank(C)_< Rank(C*) ~ m + n - 1.

Now Theorem 2.9 follows f r o m the above inequality and Theorem 2.8.

R e m a r k 2.10. I f P is irreducible, then R a n k ( C ) = m + n - 1 . Therefore, R[0 is a

m a x i m u m cell set which contains no loop. However, if P is reducible, say, like the
last case in R e m a r k 1.2, then the union of the m a x i m u m cell sets, which contain no
loop, corresponding to M 3 and M 4, respectively, is a m a x i m u m cell set of P which
contains no loop.

F r o m the Remark 2.10, we obtain

Theorem 2.11. Let P be a m x n matrix and C be the coefficient matrix of(2.1). Sup-
pose that, under row changes and column changes, P is equivalent to a block matrix
[Mij], l <_i<_s, l <_j<_t, such that
(a) Mij is a m i x nj matrix f o r each i and j ( ~ = 1 mi = m, ~ = l nj = n),
(b) Mij is a zero matrix f o r i ¢ j ,
(c) Mii is irreducible f o r 1 <_i<_sAt (=min(s,t)).
Then Rank(C) = Y,s~tl(m i + n i - 1).

Corollary 2.12. Let P, C, [Mij] be as in Theorem 2.11 but with s = t. Let a (respec-
tively, fl) be the permutation o f { 1, 2 . . . . . m} (respectively, { 1, 2 . . . . . n}) correspond-
ing to the row (respectively, column) changes in Theorem 2.11. I f f o r each i=
1,2,...,s,

too+ ... + m i no+ . " +hi

2 Uu(j)= 2 Off(j), (mo =no=O), (2.12)
j=mo+...+mi-i+l j=no+...+ni 1+1

then the system o f equation (2.1) is consistent.

Proof. Condition (2.12) implies that there are at least s redundant equations in (2.1).
Thus, the rank of the augmented matrix C* is smaller than or equal to m + n - s .
This fact together with Theorem 2.11 imply that Rank(C*) -- Rank(C) = m + n - s.
Therefore, the system of equations (2.1) is consistent.
52 T.F. Lin

3. An application

Let (12, ~ P r ) be a finite discrete probability space and let T = { 1 , 2 , 2 ' } , 1 < 2 ,
1 <2', be a partially ordered set. Assume that (Xt, ~ t ) t ~ r is a martingale and that
r < y are stopping times with respect to ( J t ) t e T (see [3],[5]). Will the optional
sampling theorem

E(Xyi 3~) = X r (3.1)

hold? The answer is negative as the following example borrowed from [5, p. 677]
shows. Let t2={co, co'}, g = p o w e r set of f2, Pr(o))=Pr(co')=½; J l = { 0 , t 2 } ,
,¢2=,~2,=g; Xl--0, X2((-o)=l=X2,((.o'), X2((.o')=-I=X2,((.D), 2"--1, y(09)=2',
y(o/) = 2. Then (Xt, ~ ) t e r is a martingale and

E(XyIJ~) = - 1 :~0 = X r.
However, if the martingale (At, ~ ) ~ r can be extended to a martingale (Xt, <~)t~ r'
where T' = { 1, 2, 2', 3 }, 2 < 3, 2 ' < 3, then the optional sampling theorem (3.1) holds
(see [5, Lemma 2.3, p. 6761). Therefore, the question becomes whether we can make
the above martingale extension. We now give a necessary and sufficient condition
for the above martingale extension.

Lemma 3.1. L e t Y, Z be r. v. "s a n d ~, S be sub-a-fields such that Y is ~-measurable

and Z is S-measurable. Then

e(r i ~ AS) =E(ZI~ n s ) (3.2)

is the necessary and sufficient condition that there exists a a( ~ U S ) - m e a s u r a b l e r. v.
X such that

E ( X I ~ ) = Y, E(XIS)=Z. (3.3)

Proof. The necessary part is trivial. For the sufficient part, let F e ~ O S be an
atom. There exist atoms A I , A 2 . . . . . A m o f ~ and atoms B I , B 2. . . . . B n of S such
that

(3.4)
i=1 j=l
Suppose that Y = a i on Ai, l <i<_m, and Z = b j on Bj, l < j < n . Let
ui=aiPr(Ai), l <i<m,
I oj=bjPr(Bj), l <j<__n.
(3.5)

Then (3.4), (3.5) and condition (3.2) imply that

i ui =E(Y.IF)=E{E(Y. 1FI ~ n~)} =E{IFE(Yi ~ ns)}

i=I
= E { I F E ( Z I ~ n s)} = E { E ( Z . IF I ~ n S ) }
 Linear equations related to the transportation problem 53
n
=E(Z. Ip) = ~ Oy, (3.6)
j=l

where I F denotes the indicator function of the set F. The equality (3.6) is exactly
the condition (1.2). Let

pij=Pr(AiNBj), l < i < m , l <j<_n. (3.7)

Then the fact that F is an atom of ,~Yf ) , ~ implies that the m × n matrix P = [Pij] is
irreducible. Hence condition (1.3) is satisfied. Therefore, according to Theorem 2.9,
equation (1.1) is consistent. Let xij(F), 1 <_i<m, 1 < j < n , be a solution of equation
(1.1). Similarly, condition (3.2) will enable us to do the same fro every other atom
of ~jc,f) y .
Let X be a r.v. defined by

X=xij(F) (3.8)
on A i O B j C F and similarly defined on other atoms. It is clear that X is a(,~~L I ~ ) -
measurable. On each A i C F , 1 <_i<m,

E ( X I ~¢) = E(XIA)/Pr(Ai) = Ij~= l PijXij(F)I /Pr(Ai)

= ui/Pr(Ai) = ai = Y. (3.9)

The first equality of (3.9) follows from the fact that the value of E(X I ~5) on the
atom A i E fgJ equals to the mean value of X o v e r A i (see [1, Ex. 3, p. 305]), the
second equality follows from (3.8) and (3.7), the third equality follows from the fact
that xij(F), 1 <_i<m, 1 <_j<_n, is a solution of (1.1), the fourth equality follows
from (3.5) and the last equality is an assumption. Therefore,

E(X[ fY): Y (3.10)

on the atom F of ~ N.~ Since X is similarly defined on other atoms of ~ CI~,

(3. I0) also holds on other atoms of ,~'fq ,~. Therefore (3.10) holds on f2. Similarly,
we can prove that E(XIH)=Z. Therefore (3.3) is true. This proves the sufficient
part of Lemma 3.1.

Theorem 3.2. Let T= {1,2,2'}, 1 < 2 , 1 <2', be a partially ordered set and let
(Xt, ,~)t~ r be a martingale. Then the optional sampling theorem (3.1) holds if and
only if E(x I J2,) =E(Xz,)l J2n @).
Proof. Let T ' = T t . J { 3 } , 2 < 3 , 2 ' < 3 , and let J 3 = a ( ~ U g 2 , } . By Lemma 3.1,
there exists a g3-measurable r.v. X 3 such that E(x31 ~2)=x2, E(xsl @)=x2,.
This means that (X t, ~t)t~ T' is a martingale. Then the optional sampling theorem
(3.1) holds for (X t, J,)teT' (see [5, Lemma 2.3, p. 676]). Since T C T', (3.1) also
holds for (Xt, Jt)t~r. This proves the 'if' part of Theorem 3.2. The 'only if' part
is trivial.
54 T.F. Lin

A conjecture. The author believes that Theorem 3.2 is a special case of the following
conjecture: Let T be a countable, partially ordered set and ( X t, ~t)teT be a mar-
tingale. Then the optional sampling theorem (3.1) holds if and only if, for every
subset SC TM, the conditional expectations E ( X t I N s ~ s ~ ) , t ~ S, are equal, where
TM = {t~ T: t is a maximal element of T}.

Acknowledgement

The author wishes to thank the referee who gave the following simple p r o o f of
Theorem 2.9 by using circulation theorem (see [4]).

Appendix

Let P = [Pij] be a real m × n matrix. Define a matrix D = [do] by

dij=I~ ifPij=O'
if Pij ~ O.
Consider the following system of m + n linear equations

. . . . . u i, l <_i<_m,
J (A. 1)

Lemma A1. The system (A. 1) has a solution i f and only i f the system (1.1) has a
solution. Moreover, the matrix P is irreducible if and only i f the matrix D is
irreducible.

Proof. The p r o o f is obvious - the relation between both the solutions is given by
yij=pijxij for dij~eO and yij, x O. arbitrary for diy =0.

For the matrix D and the numbers ui, vj, consider the network G=
(N, A1UA2UA3) with
g = { 1 , 2 . . . . . m,i,~2 . . . . . & q } ,

A I = {(q,r),r= 1, 2 . . . . . m} with lqr = Cqr = Ur,

A 2 = {(g,q),g= i,52 . . . . . ~} with 1,q = Csq = Us,

A3= {(r,g),drs~O } with l r s = - M , Crs=M,

where lij are lower bounds, cij are capacities, and M is a large number (e.g. let
M = ]~i luil+ ~j rojl).
 Linear equations related to the transportation problem 55

Lemma A2. The system (A. 1) has a solution if and only if there ex&ts a feasible cir-
culation in the network G.

Proof. Let F = fro.] be a feasible circulation in G. If yo.=fo, for dij¢O and yo.=O
otherwise, then [yo.] satisfies the system (A.1).

Let [yo.] be a solution of (A.1). Define

Ur if (i,j)=(q,r)~Al,
fo.= Os if (i,j)=(g,q)~A2,
Yrs if (i,j)=(r,g)~A 3.
Clearly f/j satisfy 10.<fo.<co.. The conservation law in nodes i,2 ..... m and
i,2 ..... n is satisfied due to the system (A.I). The existence of the solution of (A.1)
implies ~i ui = ~j oj which guarantees the conservation law in the node q.

L e m m a A3. If D & irreducible and ~ i Ui = ~ j Oj, then there ex&ts a feasible circula-
tion in the network G.

Proof. Suppose there is no such circulation. We shall use H o f f m a n ' s theorem:

" I n a network with lower bounds and capacities a feasible circulation exists if and
only if

E lij~ E co.
i~Q,j~iQ i~Q,j~Q
for arbitrary 0 ~ Q c N . "
It follows that there exists O#:QCNin G such that

E 10.> E co.. (A.2)

i~Q, jq.Q i~Q,j~Q

Moreover, Q--g{q} and (N-Q)--#{q} due to Ziui=~jvj. Without loss of

generality, we may suppose that q e Q (otherwise change the roles of 16'-*cO in the
proof).
Let us show that no arc of A 3 can 'enter' or 'leave' Q:

(i,j)~A3=(i,j~Q or i,j~Q). (A.3)

Indeed, suppose (r,g)eA3, re Q, and g~Q. Since
m
E
i~Q,j~Q
lij< E lUll-M,
i=1
E
i~Q,j~Q
Cij>-- E Iojl,
j=l

it follows that

E co.- ~ lo.>-~ ] o j l - ~ lui[+M>-O,

i~Q, j a Q i~Q, jCQ j i
which contradicts (A.2). (The case rE Q, ~ Q is similar.)
56 T.F. Lin

Now set

R = Q O { 1 , 2 ..... m}, R ' = ( N - Q ) O { 1 , 2 ..... m},

S = Q O { i , 2 . . . . . m}, S ' = ( N - Q ) N { i , 2 ..... m}.
Since Q~O, Q ~ { q } , ( N - Q ) ~ O , and ( N - Q ) ~ { q } , it follows that

R = O ~ S~=O,

R ' = 0 = S'4:0.

Since dij=O if and only if (i,j)~A 3, it follows from (A.3) that

(r~R and g~S') = drs=O,

(r~R' and g ~ S ) = drs=0.
Hence the reducibility of the matrix D is proven, which completes the proof.

Note that if there is no feasible circulation inG, then in both the reducts S, R and
S',R' the equations Ziui = Zj vj are violated.

Theorem A4. I f P is irreducible and ~iUi ~- ~.jOj, then the system (1.1) has a
solution.

Proof. The p r o o f is a composition of Lemmas A1-A3.

R e m a r k A5. Moreover, considering the note at the end of the p r o o f of L e m m a A3,

we can prove also an 'opposite' implication: If the system (1.1) does not have a solu-
tion, then
(a) either 2iml Hi--[=Z jn= l Oj,
(b) or P is reducible and in one of the reducts R, S the equality ~ieR Hi = ~jES Oj
is violated.

References

[1] H. Bauer, Probability Theory and Elements of Measure Theory (Holt, Rinehart and Winston, New
York, 1972).
[2] L. Cooper and D. Steinberg, Methods and Applications of Linear Programming (Saunders,
Philadelphia, 1974).
[3] J.L. Doob, Stochastic Processes (Wiley, New York, 1953).
[4] A.J. Hoffman, Some recent applications of the theory of linear inequalities to extremal com-
binatorial analysis, Proc. Symp. Appl. Math. 10 (Amer. Math. Soc. Providence, RI, 1960) 113-127.
[5] K. Hoffman and R. Kunze, Linear Algebra (Prentice-Hall, Englewood, N J, 1971).
[6] T.B. Kurtz, The optional sampling theorem for martingale indexed by directed sets, Ann. Prob. 8
(1980) 675-687.