Weakly linear systems for matrices over the

max-plus quantale and their applications

Miroslav Ćirić
Department of Computer Science, Faculty of Sciences
University of Niš, Serbia
miroslav.ciric@pmf.edu.rs

joint work with

Aleksandar Stamenković, Dragan Djurdjanović

Weighted Automata: Theory and Applications – WATA 2020/2021, 10th International Workshop
April 19–23, 2021, CIRM@Marseille, France

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 1 / 16
Weakly linear systems – origins
Origins
⋆ Weakly linear systems (abbr. WLS’s) – particular systems of matrix inequations

⋆ emerged in Fuzzy Set Theory – in the study of fuzzy finite automata (abbr. FFAs)

Use
⋆ state reduction for FFAs

Ćirić, Stamenković, Ignjatović, Petković, J COMPUT SYST SCI 76 (2010) 609–633

Stamenković, Ćirić, Ignjatović, INF SCI 275 (2014) 168–198

⋆ simulations and bisimulations for FFAs

Ćirić, Ignjatović, Damljanović, Bašić, FUZZY SETS SYST 186 (2012) 100–139
Ćirić, Ignjatović, Jančić, Damljanović, FUZZY SETS SYST 208 (2012) 22–42

⋆ simulations and bisimulations for NFAs

Ćirić, Ignjatović, Bašić, Jančić, INF SCI 261 (2014) 185–218

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 2 / 16
Weakly linear systems – the fuzzy context
complete residuated lattices – L = (L, ∨, ∧, ⊗, →, 0, 1)
⋆ (L, ∨, ∧, 0, 1) – complete lattice, 0 – the least element, 1 – the greatest element
⋆ (L, ⊗, 1) – commutative monoid
⋆ residuation property – x⊗y6z ⇔ x6y→z
fuzzy matrices – matrices with entries in L
⋆ order on L pointwise extended to fuzzy matrices – to each Lm×n
⋆ ∨-⊗-matrix multiplication – product of matrices over the semiring (L, ∨, ⊗, 0, 1)
WLS’s for fuzzy matrices
Ai X 6 XBi , i ∈ I1
XBi 6 Ai X , i ∈ I2
X ⊤ Ai 6 Bi X ⊤ , i ∈ I3
⊤ ⊤
Bi X 6 X Ai , i ∈ I4
m×m n×n
{Ai }i∈I ⊂ L , {Bi }i∈I ⊂ L , I = I1 ∪ I2 ∪ I3 ∪ I4 , X – unknown matrix taking values in Lm×n

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 3 / 16
Weakly linear systems – homogeneous and heterogeneous

⋆ homogeneous WLS - m = n (square matrices) and Ai = Bi for each i ∈ I

used in the state reduction of FFAs
general properties of homogeneous WLS’s
Ignjatović, Ćirić, Bogdanović, FUZZY SETS SYST 161 (2010) 3081–3113

⋆ heterogeneous WLS – rectangular matrices

used in the study of simulations and bisimulations for FFAs
general properties of heterogeneous WLS’s
Ignjatović, Ćirić, Damljanović, Jančić, FUZZY SETS SYST 199 (2012) 64–91

⋆ other applications
⋆ Social Network Analysis – positional analysis of fuzzy social networks (FSNs), simulations
and bisimulations for FSNs
⋆ Multivalued Modal Logics – simulations and bisimulations for multivalued Kripke models
Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 4 / 16
Solving WLS’s with fuzzy matrices

⋆ residuation in L w transferred to matrices

m×n
A\C = ⊤{X ∈ L | AX 6 C}, C/B = ⊤{Y ∈ Lm×n | YB 6 C}, A ∈ Lm×m , B ∈ Ln×n , C ∈ Lm×n
w reduces the WLS to an inequation of the form X 6 φ(X), where φ is a particular isotone
function on Lm×n

⋆ completeness of L w transferred to matrices

w creates conditions for using the Knaster-Tarski fixed-point theorem
w provides the existence of the greatest solution to the WLS

⋆ computing the greatest solution

w by means of a modification of the Kleene fixed-point theorem

⋆ application to FFAs  compatibility of the matrix ordering with matrix multiplication

⋆ compatibility  positivity of the ordering in L

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 5 / 16
Weakly linear systems – for matrices over a semiring

⋆ How to extend WLS’s to the context of matrices over a semiring?

⋆ How to apply them to general WFAs over a semiring ?

⋆ Residuation?

⋆ additively idempotent semirings  relative residuation  Boolean residuals –

Boolean matrices

⋆ Boolean simulations and bisimulations for WFAs over additively idempotent semirings

Damljanović, Ćirić, Ignjatović, THEOR COMPUT SCI 534 (2014) 86–100

⋆ Boolean simulations and bisimulations for max-plus automata

Komenda, Lahaye, Boimond, IFAC PAPERS ONLINE 51-7 (2018) 192–197

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 6 / 16
Weakly linear systems – for max-plus matrices

⋆ How to extend WLS’s to the context of max-plus matrices ?

w How to compute general (non-Boolean) solutions?

⋆ How to apply WLS’s to max-plus automata ?

w How to get general (non-Boolean) simulations and bisimulations?

⋆ for FFAs

w general fuzzy solutions provide better state reductions than Boolean (crisp) solutions

w there are cases when there exist general fuzzy simulations and bisimulations but there
are no any Boolean (crisp) simulations and bisimulations

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 7 / 16
Quantales – max-plus quantale

⋆ Quantale – an algebra Q = (Q, ∨, ∧, ⊗, 0, 1) where

(Q1) (Q, ∨, ∧, 0, 1) is a complete lattice
(Q2) (Q, ⊗) is a semigroup
_  _ _  _
(Q3) infinite distributive laws are satisfied: a ⊗ as = (a ⊗ as ), as ⊗ a = (as ⊗ a)
s∈I s∈I s∈I s∈I

w each quantale is residuated

⋆ Unital quantale – Q = (Q, ∨, ∧, ⊗, 0, 1, e), (Q, ⊗, e) is a monoid with identity e (in general e , 1)

⋆ max-plus algebra (also: max-plus semiring) – Rmax = (R ∪ {−∞}, ∨, +, −∞, 0)

⋆ max-plus quantale (also: complete max-plus algebra, complete max-plus semiring) –
commutative unital quantale R∞ = (R ∪ {−∞, +∞}, ∨, ∧, +, −∞, +∞, 0) where




 a+b if a, b ∈ R,

a+b=  −∞ if a = −∞ or b = −∞,

 +∞

if a = +∞, b , −∞ or b = +∞, a , −∞.

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 8 / 16
Max-plus quantale – residuation
⋆ max-plus algebra does not have residuation , but max-plus quantale does
⋆ in the general case  two different residuum operations in a quantale
⋆ due to commutativity  only one residuum operation in the max-plus quantale
. 



 b − a if a, b ∈ R,

a→b=  −∞ if b = −∞ or a , −∞,

 +∞

if b = +∞ or a = b = −∞.

m×n n×k m×k

⋆ residuum operations on matrices  A ∈ R∞ , B ∈ R∞ , C ∈ R∞
m
^ k
^
(A\C)(j, l) = A(i, j) → C(i, l), (C/B)(i, j) = B(j, l) → C(i, l).
i=1 l=1

⋆ residuation property: AB 6 C ⇔ A 6 C/B ⇔ B 6 A\C

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 9 / 16
Weakly linear systems over R∞ – the greatest solutions

Stamenković, Ćirić, Djurdjanović, DISCRETE EVENT DYNAMIC SYSTEMS (accepted)

m×n
Theorem 1 (The greatest solution in R∞ )
The set of all solutions of a WLS over R∞ is a complete lattice.
m×n
Consequently, there exists the greatest solution of this WLS in R∞ .

m×n
Theorem 2 (The greatest solutions in Rmax )
m×n
It is not necessary that there is the greatest solution of the WLS in Rmax .
m×n
However, for each X0 ∈ Rmax there exists the greatest solution of the WLS less than or equal to X0 .

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 10 / 16
Weakly linear systems over R∞ – the equivalent form

m×n m×n
⋆ φ : R∞ → R∞
       
^  ^  ^
⊤  ^
⊤ 
⊤ ⊤
φ(C) =  Ai \(CBi ) ∧  (Ai C)/Bi ) ∧ 
    (Bi C )/Ai  ∧  Bi \(C Ai ) 
i∈I1 i∈I2 i∈I3 i∈I4

Theorem 3 (The equivalent form)

(1) The WLS is equivalent to a single inequation X 6 φ(X).
(2) φ is isotone and ω-continuous.

m×n
⋆ ω-continuous – preserves infima of non-increasing sequences in R∞

⋆ solving WLS  computing the greatest post-fixed point of φ

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 11 / 16
Weakly linear systems over R∞ – computing the greatest solution

m×n
⋆ sequence of matrices  given X0 ∈ R∞ we define
^
X1 = X0 , Xn+1 = φ(Xn ) ∧ Xn , S(X0 ) = Xn .
n∈N

Theorem 3 (Computing the greatest solution)

S(X0 ) is the greatest solution to X 6 φ(X), X 6 X0 .
(the greatest solution to the WLS contained in X0 )

⋆ this does not necessarily hold in the fuzzy context

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 12 / 16
Weakly linear systems over R∞ – finitely computable solutions
⋆ Problem: How to efficiently compute the greatest solution?

⋆ S(X0 ) is finitely computable if {Xn }n∈N converges in a finite number of steps

⋆ there is n ∈ N such that Xn+1 = Xn  then S(X0 ) = Xn
⋆ number of computational steps  the smallest such n

m×n m×n
⋆ d : R∞ × R∞ → R+ ∪ {0, +∞}  extended metric


 inf{λ ∈ R+ | (−λ)B 6 A 6 B} if {λ ∈ R+ | (−λ)B 6 A 6 B} , ∅

d(A, B) = 
+∞ otherwise

d(A, B) = max |A(i, j) − B(i, j)| (Chebyshev distance)
i,j

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 13 / 16
Weakly linear systems over R∞ – finitely computable solutions
⋆ finite solutions  solutions that are finite matrices  entries only from R

Theorem 4 (Finitely computable solutions)

Let the WLS has a finite solution Y and suppose that there is λ ∈ R such that λX0 , λAi ’s and λBi ’s
take entries in Z ∪ {−∞, +∞}.
Then S(X0 ) is finitely computable and the corresponding number of computational steps does not
exceed 2mn · d(λY, λX0 ) .

Consequences
(1) If X0 , Ai ’s and Bi ’s take entries in Z ∪ {−∞, +∞}, then S(X0 ) is finitely computable and the corres-
ponding number of computational steps does not exceed 2mn · d(Y, X0 ).
(2) If X0 , Ai ’s and Bi ’s take entries in Q ∪ {−∞, +∞}, then S(X0 ) is finitely computable and the corres-
ponding number of computational steps does not exceed 2mn · d(λY, λX0 ) , where λ is the least
common multiple of all entries of X0 , Ai ’s and Bi ’s.

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 14 / 16
Our ongoing research

⋆ Applications of WLS’s to max-plus automata

w simulations and bisimulations for max-plus automata

w state reduction for max-plus automata

w determinization of max-plus automata

⋆ for simulations and bisimulations

w we expect much better algorithms than the existing ones, which are based on Boolean
solutions to WLS’s

Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 15 / 16
Miroslav Ćirić, Aleksandar Stamenković, Dragan Djurdjanović Weakly linear systems for matrices over the max-plus quantale 16 / 16