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 />

3.2. Methods to solve <strong>non</strong>linear equations<br />

In our GNEP context, the root function Φ is given in Equation (3.3) and the merit function<br />

is <strong>de</strong>fined as<br />

f(z) = 1<br />

<br />

<br />

<br />

D(x, λ) 2<br />

2 <br />

.<br />

φ◦(−g(x), λ)<br />

The gradient of f can be expressed as<br />

<br />

Jac<br />

∇f(z) =<br />

xD(x, λ) diag ∇xigi (x) T<br />

<br />

D(x, λ)<br />

i<br />

.<br />

−Da(z)Jac xg(x) Db(z) φ◦(−g(x), λ)<br />

using Equation (3.20). As mentioned in Dreves et al. (2011), the gradient ∇f is single-valued<br />

since only the bottom part of the generalized Jacobian ∂Φ(z) contains multi-valued expressions<br />

(Da and Db when gi j (x) = λij = 0) and the bottom part of Φ(x, λ) has zero entries in that<br />

case. Hence, f is C1 as long as the objective and constraint functions are C2 . Therefore, the<br />

line-search is well <strong>de</strong>fined for our GNEP reformulation.<br />

The <strong>non</strong>singularity assumption of elements in the limiting Jacobian ∂BΦ was already<br />

studied in the previous subsection for local convergence. Furthermore, Dreves et al. (2011)’s<br />

theorem 3 analyzes the additional conditions for a stationary point of f to be a solution of<br />

the <strong>non</strong>smooth equation, hence of the original GNEP. The condition is that Jac xD(x, λ) is<br />

<strong>non</strong>singular and the matrix<br />

M = Jac xg(x)Jac xD(x, λ) −1 diag ∇xigi (x) <br />

(3.28)<br />

is a P0-matrix ∗ . The <strong>non</strong>singularity condition (3.28) for local convergence requires that<br />

Db(z) + MDa(z) is <strong>non</strong>singular. If diagonal matrices Da and Db have <strong>non</strong> zero terms, then M<br />

is P0-matrix implies that Db(z) + MDa(z) is <strong>non</strong>singular. Thus, Conditions (3.28) and (3.26)<br />

are closely related.<br />

3.2.4 Specific methods for constrained equations<br />

This subsection aims to present methods specific to solve constrained (<strong>non</strong>linear) equations,<br />

first proposed by Dreves et al. (2011) in the GNEP context. The KKT system can be<br />

reformulated as a constrained equation, see Equation (3.4) of Subsection 3.1.1. Techniques to<br />

solve such equation may provi<strong>de</strong> good alternatives to standard optimization procedures. In a<br />

VI problem context, Facchinei and Pang (2003) <strong>de</strong>votes a chapter to interior-point methods<br />

for solving such constrained equations (CE). Here, we focus on the method of Monteiro and<br />

Pang (1999) providing a general framework for CE problems.<br />

A constrained equation is <strong>de</strong>fined as<br />

2<br />

H(z) = 0, z ∈ Ω, (3.29)<br />

where Ω is a closed subset of Rn . Generally, the constraint set Ω has a simple structure, e.g.<br />

the <strong>non</strong>negative orthant Ω = Rn + or a hypercube Ω = [l, u]. As mentioned in Wang et al.<br />

(1996), in practice, the root function H has also a structured form<br />

<br />

F (z)<br />

H(z) = .<br />

G(z)<br />

∗. A m × m square matrix M is a P0-matrix if for all subscript set α ⊂ {1, . . . , m} the <strong>de</strong>terminant<br />

<strong>de</strong>t(Mαα) ≥ 0.<br />

i<br />

159

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

Saved successfully!

Ooh no, something went wrong!