Exact algorithms: from Semidefinite to Hyperbolic programming

Transcription

1 Exact algorithms: from Semidefinite to Hyperbolic programming Simone Naldi and Daniel Plaumann PGMO Days 2017 Nov EDF Lab Paris Saclay

2 1 Polyhedra P = {x R n : i, l i (x) 0} Finite # linear inequalities : l i (x) 0, i = 1,..., m One can associate a diagonal matrix : A(x) = diag(l i (x)) f = l 1 (x)l 2 (x) l m (x) = det A(x) Feasibility Problem: Is P =? ( x R n s.t. l i (x) 0, i) Linear Programming (LP): Compute inf l(x) s.t. x P Solutions are rational Combinatorics of boundary active constraints l i = 0

3 2 Spectrahedra Start with a symmetric linear matrix : A(x) = A 0 + x 1 A x n A n Infinite # linear ineq. y T A(x)y 0 Finite # Non-linear inequalities : {Principal Minors of A(x)} 0 S = {x R n : A(x) 0} f = det A(x) Feasibility Problem (LMI): x R n s.t. A(x) 0 Semidefinite Progr. (SDP): Compute inf l(x) s.t. x S Irrational solutions - Algebraic degree = δ(size,var,rank) Combinatorics of boundary co-rank of A(x)

4 3 This talk: Hyperbolicity cones C(f, e) hyperbolicity cone Input data : a polynomial and a vector f R[x 1,..., x n ] d, e R n Finite # Non-linear inequalities : Coeff. of t f(te x) f hyperbolic w.r. to e Feasibility Problem? By definition e Int(C(f, e))!! Hyperbolic Progr. (HP): Compute inf l(x) s.t. x C(f, e) Algebraic degree =?(deg,n,mult) ( irrational solutions) Combinatorics of boundary multiplicity of x

5 4 Hyperbolic polynomials Definition of hyperbolic polynomial f R[x] d is hyperbolic w.r.t. e = (e 1,..., e n ) R n if f(e) 0 a R n t ch a (t) := f(t e a) has only real roots If such e exists, f is called a hyperbolic polynomial. Fundamental examples: General case (here d = 4) : (1) f = x 1 x d ch a (t) = i(te i a i ) (2) f = det A(x), A(x) sym. ch a (t) = det(ti d A(a)) R 3 P 2 Brändén (2010): hyperb. polynomials without determinantal representations

6 5 Cones and Multiplicity Hyperbolicity cone The hyperbolicity cone of f R[x] d (w.r.t. e) is C(f, e) = {a R n : ch a (t) = 0 t 0} Multiplicity: For a R n, we define mult(a) := multiplicity of 0 as root of ch a (t) = f(te a) Multiplicity set: For m d, Γ m = {a R n : mult(a) m} Remark: The set Γ m is real algebraic. Indeed, if ch a (t) = t d + g 1 (a)t d g d 1 (a)t + g d (a) then Γ m = {a : g i (a) = 0, i d m + 1}

7 6 Recap Cone Polynomial Optimization Boundary Hyperbolic HP Multiplicity f = det A(x) SDP Co-rank of A(x) f = l i (x) LP Active constraints Two different approaches: Interior approach (e.g. classical interior-point methods) Boundary approach (more suitable for algebraic methods) General goal: certification of information on the solution

8 7 Motivation in real algebraic geometry Generalized Lax conjecture Every hyperbolicity cone is a spectrahedron, that is A 1,..., A n such that C(f, e) = {x R n : A 1 x A n x n 0} Helton-Vinnikov, Lewis-Parrilo-Ramana True for n = 3, with the stronger result that f hyperbolic f = det(x 1 A 1 + x 2 A 2 + x 3 A 3 ) By Brändén s counterexamples, the Lax conjecture cannot be proved by proving that every hyperbolic polynomial admits a determinantal representation. Determinantal representations : Leykin, Netzer, Plaumann, Sinn, Sturmfels, Thom, Vinzant... with many different techniques Kummer (2015) Partial results towards Lax conjecture, for hyperbolic polynomials without real singularities.

9 8 This work: Algebraic approach to HP We prove that every hyperbolic programming problem is equivalent to linear optimization over some multiplicity locus. Moreover, computing max mult(x) over C(f, e) is equivalent to the computation of witness points on connected components of the multiplicity loci (classical problem in R.A.G.) All the reduced problems have simply exponential complexity with respect to the number n of variables. Representation of the solution: x i = q i (t)/q 0 (t), q(t) = 0 with q i Q[t]

10 9 A test on a hyperbolic quartic Optimizing random linear forms yields solutions of multiplicity 1 for 64% of the times multiplicity 2 for 36% of the times One can get rational solutions (multiplicity 2) : x 1 = 0 x 2 = 1/2 x 3 = 0 or irrational smooth boundary solutions (mult 1) : x 1 [ , ] x 2 [ , ] x 3 [ , ]

11 10 Renegar s derivative cones Based on the remark that f hyperbolic w.r. to e D e f = i e i f x i still hyperbolic This gives a nested sequence of convex hyperbolicity cones: C(f, e) C(D e f, e) C(D (d 1) e f, e) (the last one being a half-space), giving a sequence of lower bounds for the linear function to optimize: inf l(a) C(f,e) inf l(a) C(D e f,e) inf C(D (d 1) e f,e) l(a)

12 11 The quartic Quartic hyperbolic polynomial

13 11 The quartic + First derivative

14 12 Renegar s derivative cones (continued) Why is Renegar s method useful from an effective viewpoint? At each step of the relaxation, the degree of the polynomial decreses by 1 The multiplicity decreases: the solution becomes smoother at each step One of the C(D e (j) f, e) could be a section of the PSD cone (solution set of a LMI), in which case a lower bound can be computed by solving a single SDP: Sanyal (2013) : spectrahedral repr. for derivative cones of polyhedra Saunderson (2017) : spectrahedral repr. for first derivative of PSD cone

15 13 Planar 3 ellipse Here I minimize a linear function l(x) over the 3-ellipse C(f, e) = {P R 2 : d(p, P 1 ) + d(p, P 2 ) + d(p, P 3 ) c} The boundary is given by f = 9x 8 72x 7 z + 36x 6 y 2 96x 6 yz... Deriv. x Mult. l(x ) deg(q(t)) Alg.Deg. of x 0 (0.750, 0.000, 0.250) (0.759, 0.018, 0.258) (0.797, 0.051, 0.250) (0.862, 0.116, 0.254) (0.981, 0.254, 0.273) (1.336, 0.762, 0.426) Still OK for 4 ellipse, becoming hard for 5 ellipse

16 14 Summary Hyperbolic polynomials (resp. HP) is a rich class of real polynomials, defining highly-structured optimization problems. It is important to develop an effective approach to hyperbolic polynomials, independent on their determinantal representability to hyperbolicity cones, independent on Lax conjecture Algebraic alternatives to interior-point methods for polynomial optimization Complexity of HP: polynomial in the number of variables with fixed degree? True for generic SDP

17 Thanks!

18 Thanks!

19 Thanks!