• Aucun résultat trouvé

Lov´ asz extension

N/A
N/A
Protected

Academic year: 2021

Partager "Lov´ asz extension"

Copied!
27
0
0

Texte intégral

(1)

.

...

Quasi-Lov´ asz extensions and their symmetric counterparts

Miguel Couceiro

Joint work with Jean-Luc Marichal

University of Luxembourg

(2)

Order simplexes

Letσbe a permutation on [n] ={1,. . .,n} (σSn) Rnσ =

{

x= (x1,. . .,xn)Rn : xσ(1)6· · ·6xσ(n)

}

[0, 1]nσ = Rnσ[0, 1]n

Example : n=2 (2!=2 permutations 2 simplexes!)

- 6

x16x2

x1>x2

In general: The hypercube [0, 1]n has exactlyn!simplexes, and

(3)

Order simplexes

Letσbe a permutation on [n] ={1,. . .,n} (σSn) Rnσ =

{

x= (x1,. . .,xn)Rn : xσ(1)6· · ·6xσ(n)

}

[0, 1]nσ = Rnσ[0, 1]n

Example : n=2 (2!=2 permutations 2 simplexes!)

- 6

x16x2

x1>x2

In general: The hypercube [0, 1]n has exactlyn!simplexes, and each simplex [0, 1]nσ has exactlyn+1 vertices.

(4)

Order simplexes

Letσbe a permutation on [n] ={1,. . .,n} (σSn) Rnσ =

{

x= (x1,. . .,xn)Rn : xσ(1)6· · ·6xσ(n)

}

[0, 1]nσ = Rnσ[0, 1]n

Example : n=2 (2!=2 permutations 2 simplexes!)

- 6

x16x2

x1>x2

In general: The hypercube [0, 1]n has exactlyn!simplexes, and

(5)

Lov´ asz extension

Letφ:{0, 1}nR be a (pseudo-Boolean) function s. t. φ(0) =0.

.Definition (Lov´asz, 1983) ..

...

TheLov´asz extension of φ:{0, 1}n R is the function fφ: Rn R whose restriction to eachRnσ is the unique linear function which coincides withφat then+1 vertices of the simplex[0, 1]nσ

In particular: fφ|{0,1}n =φ

(6)

Lov´ asz extension

Example :

- 6

r r

r r

φ(0, 0) =0 φ(1, 0) =1 φ(0, 1) =5 φ(1, 1) =3

x1>x2 fφ(x1,x2) =x1+2x2 x16x2 fφ(x1,x2) =2x1+5x2

OnR2 :

fφ(x1,x2) = x1+5x23 min(x1,x2)

(7)

Representations of Lov´ asz extensions

In general: fφ can always be written in the form fφ(x) =

S⊆[n]

aφ(S) min

i∈Sxi (xRn)

where the coefficientsaφ(S) are given by theM¨obius transformof φ

Consequence: fφ is always piecewise linear and continuous!

(8)

Representations of Lov´ asz extensions

... and on each order simplexRnσ ?

In general :

fφ(x) =xσ(1)φ(1) +

n i=2

(xσ(i)xσ(i−1))φ(1{σ(i),...,σ(n)}) (xRnσ)

(9)

Choquet integral

f:RnRis aLov´asz extensionif there is φ:{0, 1}nRs.t. f =fφ.

.Definition ..

...AChoquet integralis a nondecreasing Lov´asz extension (vanishing at 0).

(10)

Generalization: Quasi-Lov´ asz extensions

LetI be a real interval containing 0.

.Definition ..

...

Aquasi-Lov´asz extensionis a functionf:InRdefined by f = L◦ϕ = L◦(ϕ,. . .,ϕ),

where .

1.. L:RnRis a Lov´asz extension .

2.. ϕ:I Ris a nondecreasing function satisfying ϕ(0) =0.

In DMU: ϕis a utility function andf an overall preference functional.

(11)

Symmetric Lov´ asz extension

ForxRn, set x+ =x0 and x= (x)+

.Definition ..

...

Thesymmetric Lov´asz extensionof φ:{0, 1}nRis defined by fˇφ(x) = fφ(x+) fφ(x)

For a (nonsymmetric) Lov´asz extension:

fφ(x) = fφ(x+) fφd(x)

where φd(1A) =φ(1)φ(11A) =φ(1)φ(1[n]\A).

(12)

Symmetric Lov´ asz extension

ForxRn, set x+ =x0 and x= (x)+

.Definition ..

...

Thesymmetric Lov´asz extensionof φ:{0, 1}nRis defined by fˇφ(x) = fφ(x+) fφ(x)

For a (nonsymmetric) Lov´asz extension:

fφ(x) = fφ(x+) fφd(x)

where φd(1A) =φ(1)φ(11A) =φ(1)φ(1[n]\A).

(13)

Symmetric Lov´ asz extension

ForxRn, set x+ =x0 and x= (x)+

.Definition ..

...

Thesymmetric Lov´asz extensionof φ:{0, 1}nRis defined by fˇφ(x) = fφ(x+) fφ(x)

Immediate consequences:

fˇφ is piecewise linear and continuous

fˇφ(cx) = cfˇφ(x)for everyc∈R

Sipoˇˇ s (1979): Asymmetric Choquet integralis a nondecreasing symmetric Lov´asz extension

(14)

Symmetric Lov´ asz extension

Example :

- 6

r r

r r

φ(0, 0) =0 φ(1, 0) =1 φ(0, 1) =5 φ(1, 1) =3

fφ(x ,x ) =x +5x 3 min(x ,x ) fˇφ(x) =fφ(x+)fφ(x)

(15)

Symmetrizing quasi-Lov´ asz extensions

LetI be a real interval centered at 0: xI wheneverxI.

.Definition ..

...

Asymmetric quasi-Lov´asz extensionis a functionf:InRdefined by f = Lˇ◦ϕ = Lˇ(ϕ,. . .,ϕ)

.

1.. Lˇ:RnRis a symmetric Lov´asz extension .

2.. ϕ:I Ris a nondecreasingoddfunction.

(16)

Comonotonic modularity

x,x0 Inarecomonotonicif x,x0 Iσn = InRnσ for some σ∈Sn. .Definition

..

...

f:InRiscomonotonically modular if for all comonotonicx,x0In f(x) +f(x0) = f(xx0) +f(xx0)

.

1.. If f(0) =0, then

f(x) = f(x+) + f(x) (takex0 =0)

.

2.. If f is odd, then

f(x) = f(x+) f(x)

(17)

Comonotonic modularity

x,x0 Inarecomonotonicif x,x0 Iσn = InRnσ for some σ∈Sn. .Definition

..

...

f:InRiscomonotonically modular if for all comonotonicx,x0In f(x) +f(x0) = f(xx0) +f(xx0)

.

1.. If f(0) =0, then

f(x) = f(x+) + f(x) (takex0 =0)

.

2.. If f is odd, then

f(x) = f(x+) f(x)

(18)

Comonotonic modularity

x,x0 Inarecomonotonicif x,x0 Iσn = InRnσ for some σ∈Sn. .Definition

..

...

f:InRiscomonotonically modular if for all comonotonicx,x0In f(x) +f(x0) = f(xx0) +f(xx0)

.

1.. If f(0) =0, then

f(x) = f(x+) + f(x) (takex0 =0)

.

2.. If f is odd, then

f(x) = f(x+) f(x)

(19)

Complete description of comonotonically modular functions

For 0I R, let I=I∩R and I+=I R+

.Theorem: For anyf:InR s.t. f(0) =0, T.F.A.E.:

..

. ...

.

1.. f is comonotonically modular.

.

2.. There are comonotonically modularg:I+n Randh:In R s.t.

f(x) = g(x+) +h(x).

Furthermore: For everyxIσn s.t. xσ(p)<06xσ(p+1)

g(x+) =

p+16i6n

(g(xσ(i)1{σ(i),...,σ(n)}) g(xσ(i)1{σ(i+1),...,σ(n)}))

h(x) =

16i6p

(h(xσ(i)1{σ(1),...,σ(i)}) h(xσ(i)1{σ(1),...,σ(i−1)}))

In this case: We can choose g =f|I+n and h=f|In.

(20)

Complete description of comonotonically modular functions

For 0I R, let I=I∩R and I+=I R+

.Theorem: For anyf:InR s.t. f(0) =0, T.F.A.E.:

..

.

1.. f is comonotonically modular.

.

2.. There are comonotonically modularg:I+n Randh:In R s.t.

f(x) = g(x+) +h(x).

Furthermore: For everyxIσn s.t. xσ(p)<06xσ(p+1)

g(x+) =

p+16i6n

(g(xσ(i)1{σ(i),...,σ(n)}) g(xσ(i)1{σ(i+1),...,σ(n)}))

h(x) =

16i6p

(h(xσ(i)1{σ(1),...,σ(i)}) h(xσ(i)1{σ(1),...,σ(i−1)}))

In this case: We can choose g =f| n and h=f| n.

(21)

Complete description of comonotonically modular functions

For 0I R, let I=I∩R and I+=I R+

.Theorem: For anyf:InR s.t. f(0) =0, T.F.A.E.:

..

...

.

1.. f is comonotonically modular.

.

2.. There are comonotonically modularg:I+n Randh:In R s.t.

f(x) = g(x+) +h(x).

Furthermore: For everyxIσn s.t. xσ(p)<06xσ(p+1)

g(x+) =

p+16i6n

(g(xσ(i)1{σ(i),...,σ(n)}) g(xσ(i)1{σ(i+1),...,σ(n)}))

h(x) =

16i6p

(h(xσ(i)1{σ(1),...,σ(i)}) h(xσ(i)1{σ(1),...,σ(i−1)}))

In this case: We can choose g =f|I+n and h=f|In.

(22)

Comonotonic modularity comonotonic separability

.Corollary:

..

...

f is comonotonically modular iff it iscomonotonically separable:

for every σ∈Sn, there are fiσ:I R, i∈[n], s.t.

f(x) =

n i=1

fiσ(xσ(i)) for xInRnσ.

(23)

Axiomatization of symmetric quasi-Lov´ asz extensions

.Definition ..

...

f:InRisoddly homogeneousif there is a nondecreasing odd function ϕ:I R s.t. for every x∈I andA[n]

f(x1A) = ϕ(x)f(1A)

.Theorem: Assume thatI is centered at 0 with[1, 1]I R...

..

...

Iff:InRis nonconstant, then T.F.A.E:

.

1.. f is symmetric quasi-Lov´asz with f(1A)6=0 for someA⊆[n]. .

2.. f is comonotonically modular and oddly homogeneous.

Moreover: f =Lˇf|

{0,1}n◦ϕf where ϕf(x) = ff(x1(1A)

A).

(24)

Axiomatization of symmetric quasi-Lov´ asz extensions

.Definition ..

...

f:InRisoddly homogeneousif there is a nondecreasing odd function ϕ:I R s.t. for every x∈I andA[n]

f(x1A) = ϕ(x)f(1A)

.Theorem: Assume thatI is centered at 0 with[1, 1]I R...

..

...

Iff:InRis nonconstant, then T.F.A.E:

.

1.. f is symmetric quasi-Lov´asz with f(1A)6=0 for someA⊆[n]. .

2.. f is comonotonically modular and oddly homogeneous.

Moreover: f =Lˇf|

{0,1}n◦ϕf where ϕf(x) = ff(x1(1A)

A).

(25)

Axiomatization of symmetric quasi-Lov´ asz extensions

.Definition ..

...

f:InRisoddly homogeneousif there is a nondecreasing odd function ϕ:I R s.t. for every x∈I andA[n]

f(x1A) = ϕ(x)f(1A)

.Theorem: Assume thatI is centered at 0 with[1, 1]I R...

..

...

Iff:InRis nonconstant, then T.F.A.E:

.

1.. f is symmetric quasi-Lov´asz with f(1A)6=0 for someA⊆[n]. .

2.. f is comonotonically modular and oddly homogeneous.

Moreover: f =Lˇf|

{0,1}n◦ϕf where ϕf(x) = ff(x1(1A)

A).

(26)

Final remarks

.

1.. Fornonsymmetricquasi-Lov´asz extensions “odd homogeneity” is replaced by: there is ϕ:I Rnondecreasing s.t.

f(x1A) = sign(x)ϕ(x)f(sign(x)1A)

.

2.. ForsymmetricandnonsymmetricLov´asz extensions by:

f(x1A) =x f(1A) and f(x1A) =sign(x)x f(sign(x)1A)

.

3.. Conditionf(0) =0 can be dropped off: f0 = f f(0)

(27)

Thank you for your attention!

Références

Documents relatifs

In this section, we present optimization methods for minimizing convex objective functions regular- ized by the Lov´asz extension of a submodular function.. These lead to

Theorem ( Gr¨otschel-Lov´asz-Schrijver, Fujishige-Fleicher-Iwata, Schrijver ) The minimum value of a submodular function can be found

symmetric quasi-Lov´ asz extensions) that are weakly homogeneous (resp. oddly homogeneous). In Section 4 we define the concepts of comonotonic modular- ity, invariance under

Indeed, the rich structure of submodular functions and their link with convex analysis through the Lov´asz extension [135] and the various associated polytopes makes them par-

In Section 6, we consider separable optimization problems associated with the Lov´asz extension; these are reinterpreted in Section 7 as separable optimization over the submodular

In the present section we show that, for a class of intervals I, comonotonic ad- ditivity is equivalent to horizontal min-additivity (resp. horizontal max-additivity) and we

We describe the function class axiomatized by these properties and we show that, up to certain regularity conditions (based on those we usually add to the Cauchy functional equation

We propose a machine learning approach to automatically classify newly submitted vocabularies with statistical models which take texts describing the subjects of the vocabularies