28.05.2013 Views

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

tel-00703797, version 2 - 7 Jun 2012<br />

Chapitre 3. Calcul d’équilibre <strong>de</strong> Nash généralisé<br />

points zk+1 = zk − V −1<br />

k F (zk). If (zk)k converges to z ⋆ then (zk)k converges superlinearly to<br />

z ⋆ and F (z ⋆ ) = 0 is equivalent to ∃Wk ∈ ∂bF (zk),<br />

with sk = zk+1 − zk.<br />

||(Vk − Wk)sk||<br />

lim<br />

= 0,<br />

k→+∞ ||sk||<br />

Local convergence of the generalized Levenberg-Marquardt method is studied by Facchinei<br />

and Kanzow (1997), cf. Theorem 6.<br />

Theorem. Let z ⋆ a solution of F (z ⋆ ) = 0, with F a locally Lipschitzian, semismooth at z ⋆<br />

and ∀J ⋆ ∈ ∂BF (z ⋆ ) are <strong>non</strong>singular. If the Levenberg-Marquardt parameters (λk) converge to<br />

0, then the (generalized) Levenberg-Marquardt converges superlinearly to z ⋆ . If in addition, F<br />

is strongly semismooth and locally directionnally differentiable at z ⋆ , and λk = O(||Jkgk||) or<br />

λk = O(||gk||), then the sequence converges quadratically.<br />

We turn our attention to the assumption analysis of preceding theorems. All theorems<br />

require the root function to be semismooth at a solution z ⋆ . This is verified for our function<br />

Φ <strong>de</strong>fined in Equation (3.3) as long as the complementarity function φ is semismooth. Furthermore,<br />

strong semismoothness improves the convergence rate. In the previous subsection,<br />

we have seen this requires for φ to be strongly semismooth, as e.g., the Fischer-Burmeister<br />

function.<br />

The <strong>non</strong>singularity condition is required only for elements J ⋆ of the limiting Jacobian at<br />

a solution z ⋆ . As analyzed in Facchinei et al. (2009), the limiting Jacobian (3.20) might have<br />

some i<strong>de</strong>ntical rows at the bottom part. Let us investigate first this issue. We recall first the<br />

expression of generalized Jacobian<br />

<br />

Jac<br />

JΦ(z) =<br />

xD(x, λ) diag ∇xigi (x) <br />

i .<br />

−Da(z)Jac xg(x) Db(z)<br />

As only terms (ξij, ζij) in diagonal matrices D a and D b change between the generalized and<br />

the limiting Jacobian of Φ, we study the <strong>non</strong>singularity condition directly on the generalized<br />

Jacobian. In a <strong>de</strong>tailed form, the bottom part has the following structure<br />

⎛<br />

⎜<br />

⎝<br />

−D a 1 (x, λ1 )Jac x1 g1 (x) . . . −D a 1 (x, λ1 )Jac xN g1 (x)<br />

.<br />

.<br />

D b 1 (x, λ1 ) 0<br />

. ..<br />

−Da N (x, λN )Jac x1gN (x) . . . −Da N (x, λN )Jac xN gN (x) 0 Db N (x, λN )<br />

(3.23)<br />

where Da i and Db i are mi × mi diagonal matrices. In the following, we <strong>de</strong>note by Da-part and<br />

Db-part the left and right parts of Equation (3.23).<br />

Assume the generalized Jacobian has two i<strong>de</strong>ntical rows, say for players i and ĩ and components<br />

ji and jĩ . The, the Db-part requires that the jith row of Db i and the jĩth row of Dbĩ<br />

equals zero<br />

⎞<br />

⎟<br />

⎠ ,<br />

b i j(x, λ i j) = b ĩ jĩ (x, λĩj ) = 0, (3.24)<br />

ĩ<br />

with b i j (x, λi j ) = φ′ b (−gi j (x), λi j ). I<strong>de</strong>ntical rows in the Da -part is equivalent to the n dimen-<br />

sional equation<br />

156<br />

a i j(x, λ i j)Jac xg i j(x) = a ĩ jĩ (x, λĩ jĩ )Jac xg ĩ j (x). (3.25)<br />

ĩ

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!