The document discusses algorithms for determining whether a spectrahedron is polyhedral or not. A spectrahedron generalizes the concept of a polyhedron and encapsulates the eigenvalue role in semidefinite programming. The authors present an algorithm that constructs an outer polyhedral approximation of the spectrahedron and determines containment to verify polyhedrality. They show that the faces of a spectrahedron can be characterized by a joint invariant subspace of a matrix map representation. The algorithm exploits finite generation of polyhedral cones to determine polyhedrality by checking if the generators of the outer approximation are contained within the original spectrahedron.
The document discusses algorithms for determining whether a spectrahedron is polyhedral or not. A spectrahedron generalizes the concept of a polyhedron and encapsulates the eigenvalue role in semidefinite programming. The authors present an algorithm that constructs an outer polyhedral approximation of the spectrahedron and determines containment to verify polyhedrality. They show that the faces of a spectrahedron can be characterized by a joint invariant subspace of a matrix map representation. The algorithm exploits finite generation of polyhedral cones to determine polyhedrality by checking if the generators of the outer approximation are contained within the original spectrahedron.
The document discusses algorithms for determining whether a spectrahedron is polyhedral or not. A spectrahedron generalizes the concept of a polyhedron and encapsulates the eigenvalue role in semidefinite programming. The authors present an algorithm that constructs an outer polyhedral approximation of the spectrahedron and determines containment to verify polyhedrality. They show that the faces of a spectrahedron can be characterized by a joint invariant subspace of a matrix map representation. The algorithm exploits finite generation of polyhedral cones to determine polyhedrality by checking if the generators of the outer approximation are contained within the original spectrahedron.
The document discusses algorithms for determining whether a spectrahedron is polyhedral or not. A spectrahedron generalizes the concept of a polyhedron and encapsulates the eigenvalue role in semidefinite programming. The authors present an algorithm that constructs an outer polyhedral approximation of the spectrahedron and determines containment to verify polyhedrality. They show that the faces of a spectrahedron can be characterized by a joint invariant subspace of a matrix map representation. The algorithm exploits finite generation of polyhedral cones to determine polyhedrality by checking if the generators of the outer approximation are contained within the original spectrahedron.
1) Context, def - Spectrahedron: generalizes polyhedron (why?) and encapsules the eigenvalue role in semidefinite programming - Ramana 1998, computational complexity: decision pb is NP-Hard (il existe algo pour la verification d’une solution en temps polynomial mais pas pour trouver une solution) + if S polyhedron, relevant linear factors appear in a block diagonal form of A(x). What form? - Motivation : have a practical algo for exploration, experimentation, hypothesis testing with spectrahedral - Find an appropriate matrix form, joint invariant space - A face of a polyhedron is supported by a linear form - Face of codim 1, supporting hyperplane is in the alg boundary of S, yields a linear factor of f=detA(x). alg boundary=closure of dS in Zariski Topo, contained in the vanishing locus of f. - Ramana-Goldman : Faces of spectra : F_U={x | U \incl ker(A(x)) }, faces of a spectra are always exposed - Ramana: original problem from graph theory. A perfect graph = clique number of any induced subgraph equals the colouring number
Recognition complexity of perfect graphs reduce to that of polyhedrality of projected
spectrahedral ! M. Grotschel, L. Lovasz and A. Schrijver, Geometric algorithms and combinatorial optimization, Springer-Verlag, Berlin, 1988. - - Choice of spectra cone to work in linear setting (prop1.1 show it does change anything) - 2) Main theorems - Full dim spectra (in Ramana, exemple of non full dim) - Linear matrix map presentation (preimage in R^d) via a parametrization with sym matrices A0,…Ad. Une matrice symmetrique est vue par ses coordonnees suivant une collection fixe de matrice symmetriques. Dans le cadre lineaire, on a la condition x_d>=0 en plus de A(x)>=0 - Assume A(x) full rank ie there exists p st A(p)>0. Phi: x -> A(x), Rd ->Sym+(n) surjective if the Ai are lin indep if all comb of Ai =0 imply coef=0 - Codim 1 faces of S : a face is a subset F of S such that if a point of segment lie in F, the extreme points lie also in F. A face of codim 1 is such that adding a direction give the entire space. Kernel of a linear form. - Recall characterization of faces of spectrahedron (Ramana Goldman) - OBS! A polyhedral cone has all its inclusion maximal faces of codimension 1 - Thm2.2 and coro 2.3 : matrix map under block diago form where the bottom right is diagonal and pictures all codim1 faces of S - For polyhedral cone, the diagonal block allows to recover S - Q proper, A(x) in normal form - A(x) unital (the map) if A(p_0)=Id for some p_0 - Prop 2.7 if A unital, there exists an orthogonal matrix U st UAUt is in normal form (prevue comprehensible). Allows to compute normal form: seek the JOINT INVARIANT SUBSPACE - Here generic points refer to points not satisfying a certain polynomial condition - Thm2.9 A unital, p, q generic pts, N smallest subspace containing all eigenvalues common to A(p) and A(q). Then N is invariant under any matrix in the image of A(x) and N^orthogonal is the largest invariant subspace on which A(x) restricts to a proper matrix map. (preuve floute) - Generic point : point at which “all generic pties of the set are true”. (generic pty=true for almost every point)
3) Algo - Description of an algo for recognizing polyhedrality of a spectrahedral cone - Approximation : construct an outer polyhedral approximation S^ from A(x) that coincides with S when S is polyhedral - Containmaint: co-NP hard. Verify if S^=S. ie if S^ Included in S. - We present S as intersection of a spectahedron without codim 1 faces (Q(x)>=0 ) and a polyhedron (D(X) >=0) - Prop3.1 S^= {x st D(x) >=0} is a polyhedral cone containing S - Prop 3.2 how to get rid of the full dim condition on S. Use linear hull of S and consider S in a lower dimensional ev where it will be full dim. Need to compute a point in relative interior of S. - How to find points in relative interior? Interior point methods. If S is strictly feasible (ie there exists p with A(p)>0) IPalgo finds a point arbitrarily close to the analytic center (optimal solution of a certain pb) of a suitable dehomogenization of S. S is not strictly feasible if the linear subspace only meets the boundary of {X>=0} - OBS ! the computation of a point p ∈ relint S is potentially numerically delicate and has to be handled with care - Thm3.3 simultaneous diagonalization by intersecting commutators - Containment: exploit the FINITE GENERATION OF POLYHEDRAL CONES - Thm3.4 a polyhedral cone admits a finite number of generator(S^) - Consequence: if R(S^) included in S, S is polyhedral - Lineality space: largest linear subspace contain in S is given by the kernel of the linear map A(x) - Certificate of the algo : it returns the collection R(S^) - Word about complexity: joint invariant subspace of a matrix map calculated in pol time. Transfo of A(x) in unital more demanding - 4) Further - Recently, the conjecture that every convex semialgebraic set would be the linear projection of a spectrahedron (“Helton-Nie conjecture”) has been disproven by Scheiderer [34].
- About algo : step 4 vs step 5, Abar(z) est simult diago sur N. Et on peut mettre Abar(z) en forme diagonal par bloc suivant une decompo adaptée à N. Maintenant cette decompo ne donne pas a priori une forme diagonale pour le bloc inférieur droit D’(z). C’ess=t pourquoi on rediagonalise à l’étape 5.
