MINIMAL (MAX,+) REALIZATION OF CONVEX SEQUENCES

ST

EPHANE GAUBERT

, PETER BUTKOVI

, AND RAYMOND CUNINGHAME-GREEN

SIAM J. CONTROL OPTIM. c 1998 Society for Industrial and Applied Mathematics
Vol. 36, No. 1, pp. 137147, January 1998 007
Abstract. We show that the minimal dimension of a linear realization over the (max,+) semiring
of a convex sequence is equal to the minimal size of a decomposition of the sequence as a supremum
of discrete ane maps. The minimal-dimensional realization of any convex realizable sequence can
thus be found in linear time. The result is based on a bound in terms of minors of the Hankel matrix.
Key words. max plus algebra, minimal realization, rational sequences, discrete-event systems
AMS subject classications. 93B20, 15A15
PII. S036301299528534X
1. Introduction. A classical problem consists in studying innite sequences
h
0
, h
1
, . . . with values in a semiring (S, , ), generated by a nite device. The
simplest and most studied class is probably that of realizable or recognizable sequences,
obtained as the scalar output of nite-dimensional recurrent S-linear systems:
x
0
= b, x
k+1
= Ax
k
, h
k
= cx
k
, k = 0, 1, . . . , (1.1)
where A S
nn
, b, x
0
, x
1
, . . . S
n1
, c S
1n
for some positive integer n, and where
concatenation denotes the matrix product as usual.
1
Equivalently,
h
k
= cA
k
b, k = 0, 1, . . . . (1.2)
The triple (A, b, c) is called a linear realization or representation of the sequence h, and
n is called the dimension of the realization (A, b, c). By the KleeneSch utzenberger
theorem [3], realizable sequences coincide with rational sequences (sequences of coef-
cients of rational series). The theory of these sequences is much developed in the
case of elds (particularly S = R or C). The case of realizable sequences over the
semiring of nonnegative reals (R
+
, +, ) is also much studied in connection with prob-
ability measures and Markov chains [13, 22]. Here, we are concerned with realizable
sequences over the (max,+) semiring R
max
def
= (R{}, , ), with max as addi-
tion (a b
def
= max(a, b)) and + as product (a b
def
= a +b). The interest in realizable
sequences over R
max
arises from at least two elds.
a) In discrete-event systems theory, it is known that an important subclass of
man-made systems with synchronization features (manufacturing systems, transporta-
tions networks, etc.) can be modeled by input-output variants of the dynamics (1.1),
namely,
x
k+1
= Ax
k
bu
k+1
, y
k
= cx
k
, k Z, (1.3)
where u
k
, y
k
R {}. In the case of manufacturing systems, typically, the input
u
k
represents the availability date of the kth unit of raw material, and y
k
represents

Received by the editors April 28, 1995; accepted for publication (in revised form) October 17,
1996.
http://www.siam.org/journals/sicon/36-1/28534.html

INRIA, Domaine de Voluceau, BP 105, 78153 Le Chesnay cedex, France (Stephane.

[email protected]).

School of Mathematics and Statistics, University of Birmingham, Birmingham B15 2TT, UK

([email protected], [email protected]).
1
E.g., (Ax
k
)
i
=
n
j=1
A
ij
(x
k
)
j
, cx
k
=
n
j=1
c
j
(x
k
)
j
, etc.
137
138 S. GAUBERT, P. BUTKOVI

C, AND R. CUNINGHAME-GREEN
the availability date of the kth nished part, while the vector x
k
represents the dates
of completion of internal events. It is not dicult to see that the minimal output
y of (1.3) corresponding to the earliest behavior of the system is given by the sup-
convolution
y
k
=

pN
h
p
u
kp
= sup
pN
[h
p
+u
kp
], (1.4)
so that the realizable sequence h determines the input-output relation of (1.3). As in
the case of conventional linear systems, the sequence h
0
, h
1
, . . . is called the impulse
response of the system, for it coincides with the output y associated with the impulse
input u: u
k
= 0 if k = 0, u
k
= otherwise. See [1] for a complete presentation.
b) In dynamic programming, the simplest stationary deterministic Markovian
decision problem with nite state space Q = {1, . . . , n}, transition reward A : QQ
R {}, initial reward c : Q R {}, nal reward b : Q R {}, and
horizon k, writes
max
q0,...,q
k
_
c(q
0
) +
k

i=1
A(q
i1
, q
i
) +b(q
k
)
_
. (1.5)
Identifying A, b, c with matrices of appropriate sizes, it is immediately seen that the
optimal reward in horizon k, given by (1.5), coincides with h
k
= cA
k
b.
In this paper, we are concerned with the minimal realization problem, which, given
a sequence h, consists in nding a (linear) realization (A, b, c) with minimal dimension.
For instance, in the Markov decision context, the minimal realization problem asks
whether or not there exists another decision problem (A

, b

, c

) of type (1.5), with

state space Q

of strictly smaller cardinality but the same reward history h

0
, h
1
, . . .
as (A, b, c). In the context of discrete-event systems, this is a natural problem, which
consists in nding a minimal internal realization of the system (1.4), known only by
its input-output relation u y. This has often interesting practical interpretations:
loosely speaking, the nonminimality of the triple (A, b, c) arises from the existence
of nontrivial temporal relations between the dierent physical events in the system.
Particularly, nonminimalities occur when some component of the system (a particular
machine or process) which plays a physical role in the production process is invisible
from the performance evaluation point of view, i.e., when the normal functioning of
this process will never delay the output dates due to the existence of margins. This
striking phenomenon is illustrated on the cover page of the book [1], to which the
reader is referred for more motivation.
Unlike in the elds case, the minimal realization problem over R
max
is not solved
by rank arguments. It is indeed very much analogous to that of the nonnegative
realization (over the usual algebra) [14] for which only partial solutions are known. We
refer the reader to Olsder [29, 28], Cuninghame-Green [6], Qi and Chen [31], Gaubert
[15, Chap. VI], De Schutter and De Moor [10, 11, 12, 9] for existing results (realization
procedures, bounds, heuristics, reduction of the partial realization problem to an
extended linear complementarity problem). See also [1, sections 1.3 and 9.2.3]. In the
present paper, we characterize the minimal dimension of a realization for the subclass
of convex realizable sequences, extending a result given by Cuninghame-Green and
Butkovic [7] for the strictly convex case. The proof requires the minor bound given
by Gaubert in [15, Chapter VI; 16], together with a classical majorization result [26].
It is worth noting that the convexity assumption, although restrictive, is natural
with respect to a subclass of discrete-event systems. Input-output systems (1.4) with
MINIMAL (MAX,+) REALIZATION OF CONVEX SEQUENCES 139
ane realizable impulse response, h
k
= + k, can be interpreted as (delayed)
ow limiters [1, section 6.2.2], i.e., as periodic systems, with minimal interevent delay
and transfer delay . When building complex discrete-event systems from simple
ones, one uses in particular the synchronization (or parallel composition) operation
[1], which corresponds to the pointwise max of the impulse responses. Since a convex
map is the upper envelope of its tangent lines, it is not dicult to see that realizable
convex responses correspond exactly to parallel composition of nitely many delayed
ow limiters.
Finally, we refer the reader interested in an overview of the theory and applications
of the (max,+) semiring to [5, 19, 1, 27, 21]. General results on semirings can be found
in [18].
2. Statement of the result. A sequence h
0
, h
1
, . . . R is convex if
k 1 h
k+1
h
k
h
k
h
k1
.
It follows from the well-known periodicity property of (max,+) realizable sequences
(see [5], [1, Theorem 3.112], [17, Theorem 7]) that a realizable convex sequence is
eventually periodic of period one; that is, there exists N N and R such that
k N h
k+1
= +h
k
. (2.1)
In what follows, N will always stand for the least natural number satisfying (2.1) and
will be called the length of the transient of the sequence. A max-polynomial [8] is the
function
p(x) = max
1ir
(
i
+
i
x), (2.2)
with
i
R,
i
R. The name polynomial refers to the notation of the semiring
R
max
: a b = max(a, b), a b = a + b, and, for n N, a
n
= a a
. .
n times
(= n a).
Then, (2.2) becomes
p(x) =
r

i=1

i
x
i
. (2.3)
Note that we extend the exponent notation and use x
i
for x
i
, even when
i
N:
unlike in the conventional case, maxpolynomials may have real (nonintegral) expo-
nents. We say that p is a polynomial realization of h if for all nonnegative integers k,
p(k) = h
k
. We denote by mpr(h) the minimal number of monomials of a polynomial
realization of h (i.e., the minimal value of r). By convention, mpr(h) = + when h
does not admit a polynomial realization. Denote by mlr(h) the minimal dimension of
a linear realization (with mlr(h) = + if h is not realizable). The main result of this
paper is the following characterization which solves the minimal realization problem
for convex sequences.
THEOREM 2.1. For every convex sequence h there holds
mpr(h) = mlr(h). (2.4)
Note that the theorem states in particular that the existence of a polynomial
realization is equivalent to that of a linear one, for convex sequences.
Ecient computation of mpr(h) is not dicult: given a convex sequence h with
the length of the transient N, the algorithm given in the Appendix provides a minimal
polynomial realization in time O(N).
140 S. GAUBERT, P. BUTKOVI

C, AND R. CUNINGHAME-GREEN
The inequality mlr(h) mpr(h) is immediate: if p is a polynomial realization of h
of type (2.3), then h
k
= c diag()
k
b, where b denotes the r1 matrix with entries 0, c
the 1 r matrix with entries
1
, . . . ,
r
, and diag() the r r matrix with diagonal
entries
1
, . . . ,
r
and o-diagonal elements equal to . The proof of the reverse
inequality will use a bound in terms of determinants and Hankel matrices, which we
introduce next.
3. Minor bound for the minimal dimension of realization. Recall that
the Hankel matrix [13, 3] associated with the sequence h
0
, h
1
, . . . is the NN-matrix
H = (H
ij
), H
ij
= h
i+j
for all i, j = 0, 1, . . . .
A classical result for the minimal realization problem over elds states that the min-
imal dimension of any realization of a sequence h is equal to the rank of its Hankel
matrix, which can be dened equivalently as the cardinality of a basis of the vector
space generated by the rows (or columns) of H or as the maximal size of a square
submatrix of H with nonzero determinant.
Over a general (commutative) semiring S, several nonequivalent rank notions
exist,
2
which do not characterize the minimal dimension of realization but only pro-
vide bounds. Here, we will need the rank notion originating from determinants and
bideterminants over semirings.
Given a positive integer n, let S
+
n
, S

n
, respectively, denote the sets of even and
odd permutations on {1, . . . , n} (we use the concepts of even and odd permutations
in the conventional sense [4]). The positive and negative determinants of an n n
matrix A with entries from a (commutative) semiring S are dened as follows:
det
+
A =

S
+
n
n

i=1
A
i(i)
,
det

A =

n
n

i=1
A
i(i)
.
The bideterminant [20] or, equivalently, the determinant in the symmetrized semiring
S
2
[30, 15] of a square matrix A is the ordered pair
det A
def
= (det
+
A, det

A).
We say that the determinant is balanced if det
+
A = det

A, otherwise it is unbal-
anced. Let I, J N denote (possibly innite) sets of row and column indices. Given
an I J matrix A and two subsets I

= {i
1
, . . . , i
r
} I, J

= {j
1
, . . . , j
s
} J, with
i
1
< < i
r
, j
1
< < j
s
, we denote by A[I

|J

] or A[i
1
, . . . , i
r
|j
1
, . . . , j
s
] the r s
submatrix (A
imjt
)
1mr,1ts
. Let X denote the cardinality of a set X. The
minor rank rk
m
A of a matrix A is the supremum of the order of the nite square
submatrices of A with unbalanced determinant
rk
m
A = sup
_
r > 0 ;
I

I, J

J, I

=J

= r
and det
+
A[I

|J

] = det

A[I

|J

]
_
,
2
Row rank, column rank, Schein rank are distinct standard notions for Boolean matrices [24],
which can also be dened in general semirings. Moreover, in the R
max case, other rank notions have
been used in relation to the uniqueness of solutions of linear systems [5], [15, Chapter 3, section 10].
MINIMAL (MAX,+) REALIZATION OF CONVEX SEQUENCES 141
and rk
m
A = 0 if no submatrix of A with unbalanced determinant exists. The following
result taken from [15, 16] is a semiring weak version of a well-known result over elds
[13, 3].
THEOREM 3.1 (minor bound). The dimension of a linear realization of a sequence
h is not less than the minor rank of its Hankel matrix.
Hence,
rk
m
H mlr(h). (3.1)
This result holds in an arbitrary commutative semiring S (and not only in R
max
). It
is purely combinatorial in nature. We will prove it as a consequence of the following
semiring version [15, 16] of the classical BinetCauchy identity [25, section 2.4.14].
LEMMA 3.2 (BinetCauchy formula). Let S be an arbitrary commutative semiring.
Let A S
nr
, B S
rp
. For all subsets I {1, . . . , n}, J {1, . . . , p} of k elements,
we have
det
+
(AB)[I|J]

_
det
+
A[I|K

] det

B[K

|J] det

A[I|K

] det
+
B[K

|J]
_
(3.2)
= det

(AB)[I|J]

K
_
det
+
A[I|K] det
+
B[K|J] det

A[I|K] det

B[K|J]
_
;
where the sums are taken over all the k-element subsets K, K

{1, . . . , r}. By
convention, these two sums are equal to the zero element of S if k > r.
More generally, a folklore transfer principle [15, Chapter 1] asserts that usual
ring identities admit semiring analogues whenever written without minus sign. Such
semiring analogues can be obtained by direct combinatorial means (Zeilberger [33]
proves the case n = p = r of (3.2); the general case can be proved along the same
lines). They can also be deduced formally from their classical ring versions, following
an algebraic argument due to Reutenauer and Straubing [32], which we reproduce
here for the sake of completeness. Note also that a dierent BinetCauchy identity in
the (max,+) semiring (valid for permanents) has been given by Bapat [2].
Proof of Lemma 3.2. Let X = {a

ij
, b

kl
; 1 i n; 1 j, k r; 1 l p}
denote a family of distinct commuting indeterminates, and consider the semiring of
(formal) commutative polynomials with coecients in N and indeterminates x X:
S

= N[X] Z[X]. Introduce the two matrices A

= (a

ij
), B

= (b

kl
) with entries
in S

. We rst note that the identity (3.2) holds for A

and B

. Indeed, using the

invertibility of the addition of Z[X], the identity (3.2) for A

and B

is equivalent to
the conventional BinetCauchy identity which is known to be valid in the ring Z[X]:
det A

[I|J] =

K
det A

[I|K] det B

[K|J]. (3.3)
Now, there is a unique morphism of semirings : S

S, such that i, j, k, l, (a

ij
) =
a
ij
, (b

kl
) = b
kl
. This morphism transforms the identity (3.2), applied to the matrices
A

, B

with entries in S

, to the required identity for the matrices A, B with entries

in S.
Proof of Theorem 3.1. Consider a linear realization (A, b, c) of h of dimension
r. Let O = [c, cA, cA
2
, . . .]
T
and C = [b, Ab, A
2
b, . . .], and consider two nite subsets
I, J N. Applying the BinetCauchy identity to the nite size factorization H[I|J] =
O[I|1 . . . r]C[1 . . . r|J] following from H = OC, we get det
+
(H[I|J]) = det

(H[I|J]) if
|I| = |J| > r. Hence, rk
m
H r.
142 S. GAUBERT, P. BUTKOVI

C, AND R. CUNINGHAME-GREEN
4. Preliminary majorization results. The proof of Theorem 2.1 will use stan-
dard convexity results that we recall here. The following famous result, due to Hardy,
Littlewood, and P olya, states the equivalence of two possible denitions of the relation
of majorization.
THEOREM 4.1 (see [23, section 2.20], [26]). Let
1

2

k
and

k
be real nonnegative numbers. The following two statements are equiv-
alent:
1.
1
+ +
k
=

1
+ +

k
and

+ +
k

+ +

k
for all , 2 k.
2. There exists a doubly stochastic
3
matrix P such that

= P where
= (
1
, . . . ,
k
)
T
,

= (

1
, . . . ,

k
)
T
.
We write

and say that

is majorized by when these two equivalent state-

ments hold.
The following majorization inequality is standard [23, 26]. We single out the
strict-inequality case for further use.
THEOREM 4.2. Let g be a convex function R R.
1. If

, then

i
g(

i
)

i
g(
i
). (4.1)
2. Take P = (P
jm
) such that

= P as in Theorem 4.1. If for some j the

restriction of g to the set {
m
| P
jm
= 0} does not coincide with an ane
function, then the strict inequality holds in (4.1).
Proof. Classically, (4.1) is obtained by summing up the convexity inequalities
j, g(

j
)

m
P
jm
g(
m
). (4.2)
The strict inequality in (4.1) follows from the strict inequality in (4.2), as soon as j
satises condition 4.2 of the theorem.
5. Proof of Theorem 2.1. Suppose that the sequence h
0
, h
1
, . . . is convex. To
prove Theorem 2.1, it remains to show that mlr(h) mpr(h). We will assume that
mlr(h) < (otherwise, there is nothing to prove). Then, the existence of a polynomial
realization follows readily from the convexity of h together with (2.1). Let p be such
a polynomial realization satisfying
x R, p(x) = max
1ir

i
(x) = max
1ir
(
i
+
i
x), (5.1)
with

1
<
2
< <
r
and r = mpr(h). (5.2)
We set
I
0
= [x
0
, z
0
]
def
= {x N | p(x) =
1
(x)} and
I
i
= [x
i
, z
i
]
def
= {x N | p(x) =
i+1
(x) >
i
(x)} for i = 1, . . . , r 1,
(5.3)
3
Doubly stochastic refers, as usual, to a matrix with nonnegative entries in which both the
row and column sums are equal to 1.
MINIMAL (MAX,+) REALIZATION OF CONVEX SEQUENCES 143
with x
i
, z
i
N, except the last value z
r1
= . (The fact that I
i
is nonempty
follows from the minimality of r. The fact that it is an interval is immediate due
to the convexity of max
i

i
.) Note that
1
>
i
for i = 2, . . . , r (otherwise, using
(5.2), we get
1
(k)
i
(k)k 0, which contradicts the minimality or r). Hence,
p(0) =
1
=
1
(0), and thus x
0
= 0. To summarize,
0 = x
0
z
0
< x
1
z
1
z
r2
< x
r1
. (5.4)
The following elementary lemma states the existence of a minimal polynomial real-
ization in which each line passes through at least two consecutive points.
LEMMA 5.1. There exists a minimal polynomial realization (5.1) such that
x
i+1
x
i
+ 2 for i = 0, . . . , r 2. (5.5)
The proof of the lemmas is at the end of this part.
COROLLARY 5.2. If p is the polynomial realization in Lemma 5.1, then p does not
coincide with an ane function on [x
i
, x
i+1
] for i = 0, . . . , r 2.
Without loss of generality, we suppose that the polynomial realization p satises
the conditions of Lemma 5.1 and Corollary 5.2. We set
u
i
def
= x
i
i for i = 0, . . . , r 1. (5.6)
We get from (5.5) that
u
0
< u
1
< < u
r1
.
LEMMA 5.3. Let M = H[0, 1, . . . , r 1|u
0
, u
1
, . . . , u
r1
]. We have
det
+
M =
r1

i=0
p(x
i
) > det

M. (5.7)
Thus, Hcontains an rr submatrix with unbalanced minor, i.e., rk
m
H r, which
together with the minor bound gives mlr(h) rk
m
H r = mpr(h), and Theorem 2.1
follows.
Proof of Lemma 5.1. We show that there exists a minimal polynomial realization
of the form (5.1) with
z
i
x
i
+ 1 for i = 0, . . . , r 1. (5.8)
We start from an arbitrary minimal realization (5.1). Let i
0
= min{i | x
i
= z
i
}. By
replacing
i0+1
with the ane map passing through the two points (x
i0
, h
xi
0
), (x
i0
+
1, h
xi
0
+1
), we obtain a new decomposition of the form (5.1) with i
0
< i

0
= min{j |
x

j
= z

j
}, where x

j
, z

j
are points dened by (5.3) for the new polynomial realization.
Indeed, x

i
= x
i
, z

i
= z
i
for all i < i
0
and x

i0
= x
i0
, z

i0
x
i0
+ 1 > x
i0
= x

i0
. Note
that =
i0+2
; otherwise
i0+1
could be removed from p and the arising function
would still be a polynomial realization of h which contradicts the minimality of p.
After a nite number of such replacements, we get z
i
> x
i
for all i, so that (5.5)
becomes satised.
Proof of Lemma 5.3. From the denition of M we have
M = (h
i+uj
) = (p(i +u
j
))
i,j=0,...,r1
.
144 S. GAUBERT, P. BUTKOVI

C, AND R. CUNINGHAME-GREEN
We prove that for all permutations of 0, . . . , r 1 such that = Id ,
r1

i=0
p(i +u
i
) >
r1

i=0
p(i +u
(i)
). (5.9)
Clearly, it is sucient to show that for all elementary cycles c = (i
1
, . . . , i
k
) (k 2),
p(i
1
+u
i1
) p(i
2
+u
i2
) p(i
k
+u
i
k
)
> p(i
1
+u
i2
) p(i
2
+u
i3
) p(i
k
+u
i1
) (5.10)
or with the conventional notation
p(i
1
+u
i1
) +p(i
2
+u
i2
) + +p(i
k
+u
i
k
)
> p(i
1
+u
i2
) +p(i
2
+u
i3
) + +p(i
k
+u
i1
). (5.11)
Let
1
, . . . ,
k
, with
1
< <
k
, denote the sequence obtained by reordering the
x
l
= i
l
+ u
i
l
(since x
t
< x
s
t < s u
t
< u
s
, i
l
and u
i
l
are ordered in the
same way), and let

1
, . . . ,

k
with

k
denote the sequence obtained by
reordering the i
l
+u
i
l+1
.
We claim that

. Indeed, condition 1 of Theorem 4.1 is satised, because

+ +
k
is equal to the sum of the k +1 highest possible values of i
l
and u
i
l
;
hence, it is greater than

+ +

k
which is also the sum of k + 1 values of i
l
and u
i
l
.
Moreover, take P such that

= P as in Theorem 4.1. Since P is doubly

stochastic, there is at least one j {1, 2, . . . , k} such that P
jk
= 0. Since

k
<
k
, we have P
jk
= 1; thus there is at least one m {1, 2, . . . , k 1} such that
P
jm
= 0. It remains to apply Corollary 5.2 together with the strict inequality case in
Theorem 4.2 to get (5.11).
Example 1. The function p(x) = max(0, 3+x, 8+2x, 22+4x) is a polynomial
realization of the sequence h = 0, 0, 0, 0, 1, 2, 4, 6, 10, 14, . . . . From (5.3) we have
x
0
= 0, x
1
= 4, x
2
= 6, x
3
= 8,
u
0
= 0, u
1
= 3, u
2
= 4, u
3
= 5,
det H[0, 1, 2, 3|0, 3, 4, 5] = det
_

_
0 0 1 2
0 1 2 4
0 2 4 6
0 4 6 10
_

_
= (15, 14),
which is unbalanced. Hence, mlr(h) = 4 and a minimal linear realization of h is
(A, b, c), where A = diag(0, 1, 2, 4), c = (0, 3, 8, 22), b = (0, 0, 0, 0)
T
. Note that
this minimal realization is not unique.
6. Appendix. Now we develop a method for nding a polynomial realization of
the minimal dimension.
Suppose that h
0
, h
1
, . . . is a convex sequence satisfying (2.1). Consider the points
P
j
= [j, h
j
], j = 0, 1, . . . , M, in the plane with Cartesian coordinate system. If
M = N + 1, then p is a polynomial realization of h i p(j) = h
j
for j = 0, . . . , M,
MINIMAL (MAX,+) REALIZATION OF CONVEX SEQUENCES 145
so that we may restrict our investigation to a method for nding a polynomial re-
alization of the nite sequence P
0
, . . . , P
M
. It is evident that to every polynomial
realization (which in general may contain redundant monomials) we can assign an-
other polynomial realization of no greater dimension in which every line (monomial)
passes through at least two points of the set {P
0
, . . . , P
M
}. Indeed, in what follows
we consider only polynomial realizations which possess this property.
A subset T of the set
S = {[j, h
j
] | j = 0, 1, . . . , M}
is called aligned if T 3 and there exists a line q such that
1. T q,
2. (S T) q = .
Note that an aligned subset of S may not exist and that two dierent aligned subsets
are either disjoint or have exactly one common point.
Let T be a xed aligned set of points lying on a line q. Since T 3, [j
1, h
j1
], [j, h
j
], [j + 1, h
j+1
] q for some j.
Let s(t) represent a line of an arbitrary polynomial realization which passes
through [j, h
j
]. Since
h
j1
s(j 1), h
j+1
s(j + 1),
and [j 1, h
j1
], [j, h
j
], [j + 1, h
j+1
] are collinear, we have that q coincides with s.
Hence every polynomial realization contains each line passing through all points of an
aligned subset, and thus in the construction of the minimal polynomial realization we
must always include these lines. This concerns also the two special lines: one covering
[0, h
0
], [1, h
1
] and the other passing through [M 1, h
M1
], [M, h
M
], which by trivial
reasons must be involved too.
The point [j, h
j
] (0 < j < M) is called a breaking point if [j 1, h
j1
], [j, h
j
], [j +
1, h
j+1
] are not collinear. Clearly, the rst and last points of every aligned set are
breaking points (except [0, h
0
], [M, h
M
]).
Consider a xed set
B = {[r, h
r
], [r + 1, h
r+1
], . . . , [s, h
s
]}
of consecutive breaking points which is maximal; i.e., both [r, h
r
] belongs to an aligned
set or r = 1 and [s, h
s
] belongs to an aligned set or s = M 1. Hence both [r, h
r
]
and [s, h
s
] can be assumed to belong already to a line of the realization. Clearly, a
line cannot pass through more than two consecutive breaking points. If B consists of
k points, then the minimal number of lines joining pairs of consecutive points which
contain all the points in B (except for the extreme points [r, h
r
], [s, h
s
]) is
_
k
2
_
1.
A self-evident strategy to achieve this lower bound is to take alternatively every other
line consecutively joining the pairs of adjacent points starting by the line passing
through
[r + 1, h
r+1
], [r + 2, h
r+2
].
The foregoing discussion justies the following algorithm, which starts from [0, h
0
].
(Note that (P, Q) reads, line determined by the points P and Q.)
146 S. GAUBERT, P. BUTKOVI

C, AND R. CUNINGHAME-GREEN
Algorithm. MINIMAL POLYNOMIAL REALIZATION (MPR)
Input: Finite sequence h
0
, . . . , h
M
of real numbers, with M 3.
Output: Polynomial realization of h of minimal dimension represented by the lines

1
,
2
, . . .; no if h is not convex.
(1) r := 1, s := 0, j := 0, P
i
= [i, h
i
], (i = 0, 1, . . . , M).
(2) Accept
r
(P
j
, P
j+1
).
(3) r := r + 1.
(4) Let j (j > s) be the least index for which
r1
does not pass through P
j
(if
no such exists then stop).
(5) If P
j
is below
r1
then stop (no).
(6) If j = M then go to (8).
(7) s := j and go to (2).
(8) Accept
r
(P
j1
, P
j
), stop.
Example 2. Input: h = 1, 0, 1, 1, 0, 2, 5, 8. (M = 7)
P
0
= (0, 1), P
1
= (1, 0), P
2
= (2, 1), P
3
= (3, 1), P
4
= (4, 0), P
5
= (5, 2),
P
6
= (6, 5), P
7
= (7, 8).
The algorithm MPR produces the following:
r = 1, s = 0
accept
1
(P
0
, P
1
) accept
2
(P
3
, P
4
) accept
3
(P
5
, P
6
)
r = 2 r = 3 r = 4
j = 3 j = 5
s = 3 s = 5
stop
p
1
(t) = t + 1 (line
1
)
p
2
(t) = t 4 (line
2
)
p
3
(t) = 3t 13 (line
3
)
A = diag(1, 1, 3), c = (1, 4, 13), b = (0, 0, 0)
T
.
REFERENCES
[1] F. BACCELLI, G. COHEN, G. OLSDER, AND J. QUADRAT, Synchronization and Linearity, John
Wiley, New York, 1992.
[2] R. B. BAPAT, Permanents, max algebra and optimal assignment, Linear Algebra Appl., 226/228
(1995), pp. 7386.
[3] J. BERSTEL AND C. REUTENAUER, Rational Series and Their Languages, Springer-Verlag,
New York, 1988.
[4] G. BIRKHOFF AND S. MACLANE, A Survey of Modern Algebra, Macmillan, New York, 1965.
[5] R. CUNINGHAME-GREEN, Minimax Algebra, Lecture Notes in Econom. and Math. Systems
166, Springer-Verlag, New York, 1979.
[6] R. CUNINGHAME-GREEN, Algebraic realization of discrete dynamic systems, in Proc. of the
1991 IFAC Workshop on Discrete Event System Theory and Applications in Manufacturing
and Social Phenomena, Shenyang, China, June 1991.
[7] R. CUNINGHAME-GREEN AND P. BUTKOVI

C, Discrete-event dynamic systems: The strictly

convex case, Ann. Oper. Res., 57 (1995), pp. 4563.
[8] R. CUNINGHAME-GREEN AND P. MEIJER, An algebra for piecewise-linear minimax problems,
Discrete Appl. Math., 2 (1980), pp. 267294.
[9] B. DE SCHUTTER, Max-Algebraic System Theory for Discrete-Event Systems, Ph.D. thesis,
Katholieke Univ. Leuven, Feb. 1996.
[10] B. DE SCHUTTER AND B. DE MOOR, The characteristic equation and minimal state space
realization of siso systems in the max algebra, in Proc. of the 11th Conf. on Anal. and
Opt. of Systems: Discrete-Event Systems, Lect. Notes. in Control and Inform. Sci. 199,
Springer-Verlag, New York, 1994.
[11] B. DE SCHUTTER AND B. DE MOOR, Minimal realization in the max algebra is an extended
linear complementarity problem, Sys. Control Lett., 25 (1995), pp. 103111.
MINIMAL (MAX,+) REALIZATION OF CONVEX SEQUENCES 147
[12] B. DE SCHUTTER AND B. DE MOOR, Minimal state space realization of mimo systems in the
max algebra, in Proc. of the 3rd European Control Conference, Roma, Italy, Sept. 1995,
pp. 411416.
[13] M. FLIESS, Matrices de Hankel, J. Math. Pures. Appl., 53 (1974), pp. 197222.
[14] M. FLIESS, Series rationnelles positives et processus stochastiques, Ann. Inst. H. Poincare, XI
(1975), pp. 121.
[15] S. GAUBERT, Theorie des syst`emes lineaires dans les diodes, th`ese,

Ecole des Mines de Paris,
1992.
[16] S. GAUBERT, On Rational Series in One Variable over Certain Dioids, Rapport de Recherche
2162, INRIA, Le Chesnay, France, Jan. 1994.
[17] S. GAUBERT, Rational series over dioids and discrete-event systems, in Proc. of the 11th Conf.
on Anal. and Opt. of Systems: Discrete-Event Systems, Lect. Notes. in Control and Inform.
Sci. 199, Springer-Verlag, New York, 1994.
[18] J. S. GOLAN, The Theory of Semirings with Applications in Mathematics and Theoretical
Computer Science, Pitman Monogr. Surveys Pure Appl. Math. 54, Longman Scientic and
Technical Publishers, Harlow, U.K., 1992.
[19] M. GONDRAN AND M. MINOUX, Graphes et algorithmes, Eyrolles, Paris, 1979. Engl. transl.
Graphs and Algorithms, John Wiley, New York, 1984.
[20] M. GONDRAN AND M. MINOUX, Linear algebra in dioids: A survey of recent results, Ann.
Discrete Math., 19 (1984), pp. 147164.
[21] J. GUNAWARDENA, ED., Idempotency, Publications of the Newton Institute, Cambridge Uni-
versity Press, London, 1996.
[22] G. HANSEL AND D. PERRIN, Mesures de probabilites rationnelles, in Mots, M. Lothaire, ed.,
Hermes, Paris, 1990.
[23] G. HARDY, J. LITTLEWOOD, AND G. P

OLYA, Inequalities, Cambridge University Press, London,

1934, reprinted 1973.
[24] K. KIM, Boolean Matrix Theory and Applications, Marcel Dekker, New York, 1982.
[25] M. MARCUS AND H. MINC, A Survey of Matrix Theory and Matrix Inequalities, Allyn and
Bacon, Boston, 1964.
[26] A. MARSHALL AND I. OLKIN, Inequalities: Theory of Majorization and Its Applications, Aca-
demic Press, New York, 1979.
[27] V. MASLOV AND S. SAMBORSKI

I, EDS., Idempotent analysis, Adv. in Soviet Math. 13, AMS,

Providence, RI, 1992.
[28] G. OLSDER, On the characteristic equation and minimal realizations for discrete-event dynamic
systems, in Analysis and Optimization of Systems, A. Bensoussan and J. Lions, eds., in
Lecture Notes in Control and Inform. Sci. 83, Springer-Verlag, New York, 1986, pp. 189
201.
[29] G. OLSDER, Some results on the minimal realization of discrete-event systems, in 25th IEEE
Conf. on Decision and Control, Athens, Greece, 1986.
[30] M. PLUS, Linear systems in (max, +)-algebra, in Proceedings of the 29th Conference on Deci-
sion and Control, Honolulu, Dec. 1990.
[31] X. QI AND W. CHEN, The minimal realization of discrete-event systems, in Proc. of the 1991
IFAC Workshop on Discrete Event System Theory and Applications in Manufacturing and
Social Phenomena, Shenyang, China, June 1991, pp. 2933.
[32] C. REUTENAUER AND H. STRAUBING, Inversion of matrices over a commutative semiring, J.
Algebra, 88 (1984), pp. 350360.
[33] D. ZEILBERGER, A combinatorial approach to matrix algebra, Discrete Math., 56 (1985),
pp. 6172.