Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Nonnegative polynomials modulo their gradient ideal
About the article
"Minimizing Polynomials via Sum of Squares over the Gradient Ideal"
from Demmel, Nie and Sturmfels.
Richard Leroy
6th May 2005
Universität Konstanz
Nonnegative polynomials modulo their gradient ideal
Plan
Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction
Polynomials over their gradient varieties
Unconstrained optimization
What’s next ?
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction : motivation and notations
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction : motivation and notations
◮ Goal : Use semidefinite programming (SDP) to solve the
problem of minimizing a real polynomial over R n .
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction : motivation and notations
◮ Goal : Use semidefinite programming (SDP) to solve the problem of minimizing a real polynomial over R n .
◮ Idea : If u ∈ R n is a minimizer of a polynomial f ∈ R [X ] := R [X 1 , . . . , X n ], then ∇ f (u) = 0, i.e.
∀ i = 1, . . . , n, ∂f
∂x i (u) = 0
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction : motivation and notations
◮ Goal : Use semidefinite programming (SDP) to solve the problem of minimizing a real polynomial over R n .
◮ Idea : If u ∈ R n is a minimizer of a polynomial f ∈ R [X ] := R [X 1 , . . . , X n ], then ∇ f (u) = 0, i.e.
∀ i = 1, . . . , n, ∂f
∂x i (u) = 0
◮ Tools :
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction : motivation and notations
◮ Goal : Use semidefinite programming (SDP) to solve the problem of minimizing a real polynomial over R n .
◮ Idea : If u ∈ R n is a minimizer of a polynomial f ∈ R [X ] := R [X 1 , . . . , X n ], then ∇ f (u) = 0, i.e.
∀ i = 1, . . . , n, ∂f
∂x i (u) = 0
◮ Tools :
◮
(Real) algebraic geometry : Sos representation of a
nonnegative polynomial modulo its gradient ideal
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Introduction : motivation and notations
◮ Goal : Use semidefinite programming (SDP) to solve the problem of minimizing a real polynomial over R n .
◮ Idea : If u ∈ R n is a minimizer of a polynomial f ∈ R [X ] := R [X 1 , . . . , X n ], then ∇ f (u) = 0, i.e.
∀ i = 1, . . . , n, ∂f
∂x i (u) = 0
◮ Tools :
◮
(Real) algebraic geometry : Sos representation of a nonnegative polynomial modulo its gradient ideal
◮
SDP : duality theory (sos representation / moment
approach)
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Polynomials over their gradient varieties
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Polynomials over their gradient varieties
◮ Notations :
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Polynomials over their gradient varieties
◮ Notations :
◮
Gradient varieties :
V grad (f ) := {u ∈ C n : ∇f (u) = 0} ⊂ C n V grad
R(f ) := {u ∈ R n : ∇f (u) = 0} ⊂ R n
◮
Gradient ideal :
I grad (f ) := h∇f (X)i = ∂f
∂x
1, . . . , ∂f
∂x n
⊂ R [X ]
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Polynomials over their gradient varieties
◮ Notations :
◮
Gradient varieties :
V grad (f ) := {u ∈ C n : ∇f (u) = 0} ⊂ C n V grad
R(f ) := {u ∈ R n : ∇f (u) = 0} ⊂ R n
◮
Gradient ideal :
I grad (f ) := h∇f (X)i = ∂f
∂x
1, . . . , ∂f
∂x n
⊂ R [X ]
◮ Theorem 1
f ≥ 0 on V grad R (f ) I grad (f ) radical
⇒ f sos modulo I grad (f ) :
∃q i , φ j ∈ R [X ], f = X s
i=1
q i 2 + X n
j =1
φ j ∂f
∂x j
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Proof
The proof is based on the following two lemmas :
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Proof
The proof is based on the following two lemmas :
◮ lemma 1.1
V 1 , . . . , V r pairwise disjoint varieties in C n
⇓
∃ p 1 , . . . , p r ∈ R [X ], ∀ i, j , p i (V j ) = δ ij
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Proof
The proof is based on the following two lemmas :
◮ lemma 1.1
V 1 , . . . , V r pairwise disjoint varieties in C n
⇓
∃ p 1 , . . . , p r ∈ R [X ], ∀ i, j , p i (V j ) = δ ij
◮ lemma 1.2
W irreducible subvariety of V grad (f ) s.t. W ∩ R n 6= ∅
⇓
f ≡ const on W
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
In case I grad ( f ) is not radical
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
In case I grad ( f ) is not radical
◮ Example :
f (x , y , z) := x 8 + y 8 + z 8 + x 4 y 2 + x 2 y 4 + z 6 − 3x 2 y 2 z 2
| {z }
Motzkin polynomial M(x,y,z)
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
In case I grad ( f ) is not radical
◮ Example :
f (x , y , z) := x 8 + y 8 + z 8 + x 4 y 2 + x 2 y 4 + z 6 − 3x 2 y 2 z 2
| {z }
Motzkin polynomial M(x,y,z)
◮
Fact 1 : f ≡ 1
4 M (mod I grad (f ))
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
In case I grad ( f ) is not radical
◮ Example :
f (x , y , z) := x 8 + y 8 + z 8 + x 4 y 2 + x 2 y 4 + z 6 − 3x 2 y 2 z 2
| {z }
Motzkin polynomial M(x,y,z)
◮
Fact 1 : f ≡ 1
4 M (mod I grad (f ))
◮
Fact 2 : M is not a sos in R [x, y , z ]/ I grad (f )
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
In case I grad ( f ) is not radical
◮ Example :
f (x , y , z) := x 8 + y 8 + z 8 + x 4 y 2 + x 2 y 4 + z 6 − 3x 2 y 2 z 2
| {z }
Motzkin polynomial M(x,y,z)
◮
Fact 1 : f ≡ 1
4 M (mod I grad (f ))
◮
Fact 2 : M is not a sos in R [x, y , z ]/ I grad (f )
◮
Fact 3 : Ask Claus Scheiderer for more details
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
In case I grad ( f ) is not radical
◮ Example :
f (x , y , z) := x 8 + y 8 + z 8 + x 4 y 2 + x 2 y 4 + z 6 − 3x 2 y 2 z 2
| {z }
Motzkin polynomial M(x,y,z)
◮
Fact 1 : f ≡ 1
4 M (mod I grad (f ))
◮
Fact 2 : M is not a sos in R [x, y , z ]/ I grad (f )
◮
Fact 3 : Ask Claus Scheiderer for more details
◮ Theorem 2
f > 0 on V grad R (f ) ⇒ f sos modulo I grad (f )
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
◮ Notations :
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
◮ Notations :
◮
deg(f ) = d even
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
◮ Notations :
◮
deg(f ) = d even
◮
f i := ∂f
∂x i
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
◮ Notations :
◮
deg(f ) = d even
◮
f i := ∂f
∂x i
◮
∀k ≥ d , f ∈ R [X ] k : f = P
f
αx
α! f ∈ R
νn,kwhere ν n,k =
n + k k
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
◮ Notations :
◮
deg(f ) = d even
◮
f i := ∂f
∂x i
◮
∀k ≥ d , f ∈ R [X ] k : f = P
f
αx
α! f ∈ R
νn,kwhere ν n,k =
n + k k
◮
∀N , mon
N( x ) =
t( 1 , x
1, . . . , x
n, x
12, x
1x
2, . . . , x
nN) ∈ R
νn,NNonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
Unconstrained optimization
◮ Notations :
◮
deg(f ) = d even
◮
f i := ∂f
∂x i
◮
∀k ≥ d , f ∈ R [X ] k : f = P
f
αx
α! f ∈ R
νn,kwhere ν n,k =
n + k k
◮
∀N , mon
N( x ) =
t( 1 , x
1, . . . , x
n, x
12, x
1x
2, . . . , x
nN) ∈ R
νn,N◮ Restrictive hypothesis (H ) :
f attains its infimum f ∗ over R n
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
◮ Primal SDP : moment formulation
(P ) :
f N,mom ∗ := inf
y
t fy = P f α y α
s .t.
∀ i, M N −
d2
(f i ∗ y ) = 0 M N (y ) 0
y 0 = 1
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
◮ Primal SDP : moment formulation
(P ) :
f N,mom ∗ := inf
y
t fy = P f α y α
s .t.
∀ i, M N −
d2
(f i ∗ y ) = 0 M N (y ) 0
y 0 = 1
◮ Dual SDP : sos formulation
(D) :
f N,grad ∗ := sup
γ ∈R
γ
s .t.
f − γ = σ + P n j =1
φ j
∂f
∂x j
σ ∈ P
( R [X ] N ) 2
φ j ∈ R [X ] 2N−d+1
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
◮ Theorem 3
Under the assumption (H), the following holds :
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
◮ Theorem 3
Under the assumption (H), the following holds :
◮
lim
N f N,grad
∗= lim
N f N,mom
∗= f
∗Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
◮ Theorem 3
Under the assumption (H), the following holds :
◮
lim
N f N,grad
∗= lim
N f N,mom
∗= f
∗◮
I grad (f ) radical ⇒ ∃N
0, f N
∗0,grad= f N
∗0,mom= f
∗Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?
SDP relaxations
◮ Theorem 3
Under the assumption (H), the following holds :
◮
lim
N f N,grad
∗= lim
N f N,mom
∗= f
∗◮
I grad (f ) radical ⇒ ∃N
0, f N
∗0,grad= f N
∗0,mom= f
∗◮ Extracting solutions
In practice, Lasserre and Henrion’s technique :
If, for some N , and some optimal primal solution y ∗ , we have
rank M N (y ∗ ) = rank M N−d/2 (y ∗ )
then we have reached the global minimum f ∗ , and one can extract global minimizers (implemented in
Gloptipoly).
Nonnegative polynomials modulo their gradient ideal
Plan Introduction Polynomials over their gradient varieties Unconstrained optimization What’s next ?