Optimization over Polynomials with Sums of Squares and Moment Matrices

Transcription

1 Optimization over Polynomials with Sums of Squares and Moment Matrices Monique Laurent Centrum Wiskunde & Informatica (CWI), Amsterdam and University of Tilburg Positivity, Valuations and Quadratic Forms Konstanz, October 2009 Optimization over Polynomials with Sums of Squares and Moment Matrices p.

2 Polynomial optimization problem (P) Minimize a polynomial function p over a basic closed semi-algebraic set K p min := inf x K p(x) where K := {x R n h 1 (x) 0,..., h m (x) 0} p, h 1,..., h m R[x] are multivariate polynomials Optimization over Polynomials with Sums of Squares and Moment Matrices p.

3 Unconstrained polynomial minimization: K = R n p min := inf x R n p(x) p min 0 p 0 on R n Example: The partition problem. A sequence a 1,..., a n N can be partitioned if i I a i = i [n]\i a i for some I [n], i.e. if p min = 0, where p(x) = ( n i=1 a ix i ) 2 + n i=1 (x2 i 1)2 E.g., the sequence 1, 1, 2, 2, 3, 4, 5 can be partitioned. NP-complete problem Optimization over Polynomials with Sums of Squares and Moment Matrices p.

4 Example: Testing matrix copositivity M R n n is copositive if x T Mx 0 x R n + i.e. if p min = 0, where p(x) = n i,j=1 M ijx 2 i x2 j co-np-complete problem C A = C A C A is copositive B A is copositive Optimization over Polynomials with Sums of Squares and Moment Matrices p.

5 0/1 Linear programming min c T x s.t. Ax b, x 2 i = x i (i = 1,..., n) Example: The stability number α(g) of a graph G = (V, E) can be computed via any of the programs: α(g) = max i V x i s.t. x i +x j 1 (ij E), x 2 i = x i (i V ) 1 α(g) = min xt (I + A G )x s.t. i V x i = 1, x i 0 (i V ) (P) is NP-hard for linear objective and quadratic constraints, or for quadratic objective and linear constraints Optimization over Polynomials with Sums of Squares and Moment Matrices p.

6 Strategy Approximate (P) by a hierarchy of convex (semidefinite) relaxations Shor (1987), Nesterov, Lasserre, Parrilo (2000-) Such relaxations can be constructed using representations of nonnegative polynomials as sums of squares of polynomials and the dual theory of moments Optimization over Polynomials with Sums of Squares and Moment Matrices p.

7 Underlying paradigm Testing whether a polynomial p is nonnegative is hard but one can test whether p is a sum of squares of polynomials efficiently via semidefinite programming Optimization over Polynomials with Sums of Squares and Moment Matrices p.

8 Plan of the talk Role of semidefinite programming in sums of squares SOS/Moment relaxations for (P) Main properties: (1) Asymptotic/finite convergence via SOS representation results for positive polynomials (2) Optimality criterion via results for the moment problem (3) Extract global minimizers by solving polynomial equations Application to unconstrained polynomial optimization Optimization over Polynomials with Sums of Squares and Moment Matrices p.

9 A beautiful monograph about positive polynomials... Alexander Prestel Charles N. Delzell Positive Polynomials From Hilbert s 17th Problem to Real Algebra Optimization over Polynomials with Sums of Squares and Moment Matrices p.

10 Some notation R[x] = R[x 1,..., x n ]: ring of polynomials in n variables R[x] d : all polynomials with degree d p R[x] d p(x) = α N n α d p α x α 1 1 xα n }{{ n} x α = α N n α d p α x α p(x) = p T [x] d after setting and p = (p α ) α : vector of coefficients [x] d = (x α ) α : vector of monomials Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

11 What is semidefinite programming? Semidefinite programming (SDP) is linear optimization over the cone of positive semidefinite matrices LP SDP vector variable matrix variable x R n X Sym n [symmetric matrix] x 0 X 0 [positive semidefinite] sup X C, X s.t. A j, X = b j (j = 1,..., m) X 0 There are efficient algorithms to solve semidefinite programs Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

12 A small example of SDP max (X 13 + X 31 )/2 such that X 0, X R 3 3 X 11 = 1, X 12 = 1 X 23 = 1, X 33 = 2 2X 13 + X 22 = 3 max c such that X = 1 1 c 1 3 2c 1 0 c 1 2 One can check that max c = 1 and min c = 1 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

13 The Gram-matrix method to recognize sums of squares [cf. Powers-Wörmann 1998] Write p(x) = α 2d p αx α R[x] 2d as a sum of squares: p(x) = k j=1 (u j(x)) 2 = k j=1 [x]t d u j u j T [x] d = [x] d ( k = α 2d j=1 u j u j T } {{ } =: U 0 x α( β, γ d β+γ=α )[x] d = U β,γ }{{} =p α ) β, γ d x β x γ U β,γ Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

14 Recognize sums of squares via SDP p(x) = α 2d p α x α is a sum of squares of polynomials The following semidefinite program is feasible: β, γ d β+γ=α U 0 U β,γ = p α ( α 2d) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

15 Example: Is p = x 4 + 2x 3 y + 3x 2 y 2 + 2xy 3 + 2y 4 SOS? Solution: Try to write p(x, y) (x 2 xy y 2 ) Equating coefficients: a b c b d e c e f }{{} U x 2 xy with U 0 y 2 x 4 = x 2 x 2 x 3 y = x 2 xy x 2 y 2 = xy xy = x 2 y 2 xy 3 = xy y 2 y 4 = y 2 y 2 1 = a 2 = 2b 3 = d + 2c 2 = 2e 2 = f Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

16 Example continued 1 1 c Hence U = 1 3 2c c 1 c ( ) For c = 1, U = p = (x 2 + xy y 2 ) 2 + (y 2 + 2xy) For c = 0, U = p = (x 2 + xy) (xy + y2 ) (xy y2 ) 2 2 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

17 Which nonnegative polynomials are SOS? Hilbert [1888] classified the pairs (n, d) for which every nonnegative polynomial of degree d in n variables is SOS: n = 1 d = 2 n = 2, d = 4 Σ n,d P n,d for all other (n, d) Motzkin polynomial: x 4 y 2 + x 2 y 4 3x 2 y lies in P 2,6 \ Σ 2,6 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

18 How many nonnegative polynomials are sums of squares? [Blekherman 2003]: Very few! At fixed degree 2d and large number n of variables, there are significantly more nonnegative polynomials than sums of squares: c n d 1 2 ( vol( Pn,2d ) vol( Σ n,2d ) ) 1 D C n d 1 2 P n,2d := {p P n,2d p homogeneous, deg(p) = 2d, S p(x)µ(dx) = 1} n 1 D := ( n+2d 1 2d ) 1, the dimension of the ambient space Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

19 How many nonnegative polynomials are sums of squares? Many! The SOS cone is dense in the cone of nonnegative polynomials (allowing variable degrees): [Lasserre 2004]: If p 0 on R n, then ǫ > 0 k N s.t. p + ǫ ( k h=0 n i=1 x 2h i h! ) is SOS [Lasserre-Netzer 2006]: If p 0 on [ 1, 1] n, then ǫ > 0 k N s.t. p + ǫ ( 1 + n i=1 x 2k i ) is SOS Optimization over Polynomials with Sums of Squares and Moment Matrices p. 1

20 Artin [1927] solved Hilbert s 17th problem [1900] ( pi ) 2, where p i, q i R[x] p 0 on R n = p = i q i That is, p q 2 is SOS for some q R[x] Sometimes, the shape of the common denominator is known: Pólya [1928] + Reznick [1995]: For p homogeneous p > 0 on R n \ {0} = p ( n i=1 x 2 i ) r SOS for some r N Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

21 An example [Parrilo 2000] M := p is not SOS p := But ( 5 i=1 x2 i )p is SOS 5 i,j=1 M ij x 2 i x2 j This is a certificate that p 0 on R 5, i.e., that M is copositive Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

22 SOS certificates for positivity on a semi-algebraic set K Let K = {x R n h 1 (x) 0,..., h m (x) 0} Set h 0 := 1 Quadratic module: M(h) := Preordering: T(h) := { m j=0 s j h j s j Σ n } { } s e h e1 1 he m m s e Σ n e {0,1} m p M(h) = p T(h) = p 0 on K Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

23 The Positivstellensatz [Krivine 1964] [Stengle 1974] Not an equivalence: K = {x R (1 x 2 ) 3 0} p = 1 x 2 Then, p 0 on K, but p T(h) The Positivstellensatz characterizes equivalence: p 0 on K pf = p 2m + g for some f, g T(h) and some m N But this does not yield tractable relaxations for (P)! Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

24 The Positivstellensatz [Krivine 1964] [Stengle 1974] K = {x R n h 1 (x) 0,..., h m (x) 0} (1) p > 0 on K pf = 1 + g for some f, g T(h) (2) p 0 on K pf = p 2m + g for some f, g T(h), m N (3) p = 0 on K p 2m T(h) for some m N (2) = solution to Hilbert s 17th problem (3) = Real Nullstellensatz But this does not yield tractable relaxations for (P)!!! Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

25 SOS certificates of positivity on K compact Schmüdgen [1991]: Assume K is compact. p > 0 on K = p T(h) Putinar [1993]: Assume the following Archimedean condition holds: (A) p R[x] N N N ± p M(h) Equivalently: N N N n i=1 x2 i M(h) p > 0 on K = p M(h) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

26 SOS relaxations for (P) [Lasserre 2001, Parrilo 2000] p min = inf x K p(x) = sup λ s.t. p λ 0 on K Relax p λ 0 on K [hard condition] by p λ M(h) [SOS but unbounded degrees...] by p λ M(h) 2t [tractable SDP!] { m } M(h) 2t := j=0 s jh j s j Σ n, deg(s j h j ) 2t Relaxation (SOSt): p sos t := sup λ s.t. p λ M(h) 2t p sos t+1 p min Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

27 Asymptotic convergence If (A) holds for K, then lim p sos t t = p min Proof: p p min + ǫ > 0 on K = t p p min + ǫ M(h) 2t = p sos t p min ǫ Note: A representation result valid for p 0 on K" gives finite convergence: p sos t = p min for some t [Nie-Schweighofer 2007]: p min p sos t where c = c(h) and c = c(p, h) c c log(t/c) for t big, Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

28 Dual moment relaxations for (P) [Lasserre 2001] min = inf x K p(x)= inf µ probability measure on K K p(x)dµ(x) = inf L R[x] L(p) s.t. L comes from a probability measure µ on K = inf L R[x] L(p) s.t. L(f) 0 f 0 on K [Haviland] The following are equivalent for L R[x] : L comes from a nonnegative measure on K, i.e., L(f) = K f(x)dµ(x) f R[x] L(p) 0 if p 0 on K Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

29 Dual moment relaxations for (P) [continued] p min = inf L R[x] L(p) s.t. L(f) 0 f 0 on K Relax by by L(f) 0 f 0 on K L(f) 0 f M(h) L(f) 0 f M(h) 2t Relaxation (MOMt): p mom t := inf L R[x] 2t L(p) s.t. L 0 on M(h) 2t Weak duality: p sos t p mom t p min Equality: p sos t = p mom t e.g. if int(k) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 2

30 The dual relaxation (MOMt) is again an SDP L R[x] 2t M t(l) := (L(x α x β )) α, β t M t (L) is the moment matrix of L (of order t) Lemma: L(f 2 ) 0 f R[x] t M t (L) 0 Proof: L(f 2 ) = f T M t (L) f Can express L 0 on M(h) 2t, i.e., L(f 2 h j ) 0 f j with deg(f 2 h j ) 2t as SDP conditions Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

31 Optimality criterion Observation: Let L be an optimum solution to (MOMt). If L comes from a probability measure µ on K, then (MOMt) is exact: p mom t = p min. Question: How to recognize whether L has a representing measure on K? Next: Sufficient condition of Curto-Fialkow for the moment problem Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

32 A sufficient condition for the moment problem Theorem [Curto-Fialkow 1996] Let L R[x] 2t. If M t (L) 0 and (RC) rankm t (L) = rankm t 1 (L), then L has a representing measure. Corollary [Curto-Fialkow 2000] + [Henrion-Lasserre 2005] Set d := max j deg(h j )/2. Let L be an optimum solution to (MOMt) satisfying rankm t (L) = rankm t d (L). Then, p mom t = p min and V (KerM t (L)) { global minimizers of p on K} with equality if rankm t (L) is maximum. Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

33 Remarks Compute V (KerM t (L)) with the eigenvalue method. If the rank condition holds at a maximum rank solution, then (P) has finitely many global minimizers. But the reverse is not true! The rank condition holds in the finite variety case: K = {x R n h 1 (x) = 0,..., h k (x) = 0, h }{{} k+1 (x) 0,...} ideal I with V R (I) < Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

34 Finite convergence in the finite variety case K = {x R n h 1 (x) = 0,..., h k (x) = 0, h }{{} k+1 (x) 0,...} ideal I Theorem: [La 2002] [Lasserre/La/Rostalski 2007] (i) If V C (I) <, p sos t = p mom t = p min for some t (ii) If V R (I) <, p mom t = p min for some t Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

35 The flat extension theorem Theorem [Curto-Fialkow 1996] Let L R[x] 2t. If rank M t (L) = rank M t 1 (L), then there is an extension L R[x] of L for which rank M( L) = rank M t (L). Main tool: KerM( L) is an ideal. [La-Mourrain 2009] The flat extension theorem can be generalized to matrices indexed by a set C of monomials (connected to 1) and its closure C + = C x 1 C... x n C, satisfying rank M C = rank M C +. Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

36 The finite rank moment matrix theorem Theorem: [Curto-Fialkow 1996] Let L R[x]. M(L) 0 and rank M(L) =: r < L has a (unique) r-atomic representation measure µ. [La 2005] Simple proof for = : I := KerM(L) is a real radical ideal I is 0-dimensional, as dim R[x]/I = r Hence: V (I) = {x 1,..., x r } R n Verify: L is represented by µ = r i=1 L(p2 i )δ x i, where the p i s are interpolation polynomials at the x i s Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

37 Implementations of the SOS/moment relaxation method GloptiPoly by Henrion, Lasserre (incorporates the optimality stopping criterion and the extraction procedure for global minimizers) SOSTOOLS by Prajna, Papachristodoulou, Seiler, Parrilo YALMIP by Löfberg SparsePOP by Waki, Kim, Kojima, Muramatsu Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

38 Example 1 min p = 25(x 1 2) 2 (x 2 2) 2 (x 3 1) 2 (x 4 4) 2 (x 5 1) 2 (x 6 4) 2 s.t. (x 3 3) 2 + x 4 4, (x 5 3) 2 + x 6 4 x 1 3x 2 2, x 1 + x 2 2, x 1 + x 2 6, x 1 + x 2 2, 1 x 3 5, 0 x 4 6, 1 x 5 5, 0 x 6 10, x 1, x 2 0 order t rank sequence bound p mom t solution extracted unbounded none (5, 1, 5, 0, 5, 10) d = 1 The global minimum is found at the relaxation of order t = 2 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

39 Example 2 min p = x 1 x 2 s.t. x 2 2x 4 1 8x x x 2 4x x x2 1 96x x 1 3, 0 x 2 4 order t rank sequence bound p mom t solution extracted none none (2.3295, ) d = 2 The global minimum is found at the relaxation of order t = 4 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 3

40 An example where (RC) cannot hold Perturbed Motzkin form: p = x 2 1 x2 2 (x2 1 + x2 2 3x2 3 ) + x6 3 + ǫ(x6 1 + x6 2 + x6 3 ) K = {x 3 i=1 x2 i But (RC) never holds 1} (0, 0) is the unique minimizer as p M(h) and p sos t = p mom t < p min = 0 order t rank sequence bound p mom t val. moment vect d = 3, ǫ = 0.01 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

41 Application to Unconstrained Polynomial Minimization p min = inf x R n p(x) where deg(p) = 2d As there is no constraint, the relaxation scheme just gives one bound: p sos t = p mom t = p sos d = pmom d p min for all t d with equality iff p(x) p min is SOS How to get better bounds? Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

42 Idea: Transform the Unconstrained Problem into a Constrained Problem If p has a minimum: p min =p grad := inf x V R grad p(x) where V R grad := {x Rn p(x) = 0 (i = 1,..., n)} If, moreover, a bound R is known on the norm of a global minimizer: p min =p ball := inf R 2 P i x2 i 0 p(x) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

43 When p attains its minimum The ball approach : Convergence of the SOS/MOM bounds to p min = p ball The gradient variety approach: [Demmel, Nie, Sturmfels 2004]: p > 0 on V R grad = p M grad p 0 on V R grad = p M grad if I grad radical M grad := M(± p/ x i ) = Σ n + n i=1 R[x] p/ x i } {{ } I grad Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

44 Convergence Result [Demmel, Nie, Sturmfels 2004] Asymptotic convergence of the SOS/MOM bounds to p grad Finite Convergence to p grad when the gradient ideal I grad is radical Hence: When p attains its minimum, we have a converging hierarchy of SDP bounds to p min Example: p = x 2 + (xy 1) 2 does not attain its minimum p min = 0 < p grad = 1 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

45 What if p is not known to have a minimum? Strategy 1: Perturb the polynomial p [Hanzon-Jibetean 2003] [Jibetean-Laurent 2004] ( n ) p ǫ (x) := p(x) + ǫ x 2d+2 i for small ǫ > 0 i=1 p ǫ has a minimum and lim ǫ 0 (p ǫ ) min = p min The global minimizers of p ǫ converge to global minimizers of p as ǫ 0 The gradient variety of p ǫ is finite finite convergence of (p ǫ ) sos t, (p ǫ ) mom t to (p ǫ ) min Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

46 Example: Perturb the polynomial p = (xy 1) 2 + x 2 inf p(x) s.t. p ǫ / x 1 = 0, p ǫ / x 2 = 0 ǫ order t rank sequence p mom t extracted solutions ±(0.4729,1.3981) ±(0.4060,1.9499) ±(0.3287,2.6674) ±(0.2574,3.6085) ±(0.1977,4.8511) ±(0.1503,6.4986) ±(0.1136,8.6882) ±(0.0856, ) ±(0.0643, ) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

47 When p does not have a minimum: Algebraic/analytical approach of Schweighofer [2005] Strategy 2: Minimize p over its gradient tentacle If p min >, then p min = inf x K grad p(x) where K grad := {x R n p(x) 2 x 2 1} V R grad p = ( p/ x i ) n i=1 Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

48 Representation result on the gradient tentacle K grad [Schweighofer 2005]: Assume p min > and p has only isolated singularities at infinity (*) (e.g. n = 2). Then, p 0 on R n p 0 on K grad ǫ > 0 p + ǫ M(1 p(x) 2 x 2 ) Convergent SOS/moment bounds to p min = inf Kgrad p(x) (*): The system p d (x) = 0, p d 1 (x) = 0 has finitely many projective zeros, where p = p d + p d p 0 and p i is the homogeneous component of degree i Tools: Algebra (extension of Schmüdgen s theorem) + analysis (Parusinski s results on behaviour of polynomials at infinity) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

49 When p does not have a minimum: The tangency variety approach of Vui-Son [2008] Strategy 3: Minimize p over its truncated tangency variety : ( ) { Γ p := x R n p(x) rank 1} x = {x g ij := x i p/ x j x j p/ x i = 0 i, j n} Γ 0 p := Γ p {x p(x) p(0)} [Vui-Son 2008]: For p R[x] such that p min > p 0 on R n p 0 on Γ 0 p ǫ > 0 p + ǫ M(p(0) p, ±g ij ) Convergent SOS/moment bounds to p min = inf Γ 0 p p(x) Optimization over Polynomials with Sums of Squares and Moment Matrices p. 4

50 Further directions Exploit structure (equations, sparsity, symmetry, convexity,...) to get smaller SDP programs [Kojima, Grimm, Helton, Lasserre, Netzer, Nie, Riener, Schweighofer, Theobald, Parrilo,...] Application to the generalized problem of moments, to approximating integrals over semi-algebraic sets,... [Henrion, Lasserre, Savorgnan,...] Extension to NC variables [Helton, Klep, McCullough, Schmüdgen, Schweighofer,...] Optimization over Polynomials with Sums of Squares and Moment Matrices p. 5