complex structure of the optimal power flow problem · 2016-01-26 · complex structure of the...

121
Complex Structure of the Optimal Power Flow Problem edric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign November 16 th 2015 edric JOSZ, Daniel K. MOLZAHN [email protected]

Upload: others

Post on 10-Feb-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex Structure of the Optimal Power FlowProblem

Cedric JOSZ, Daniel K. [email protected]

Talk at the University of Illinois at Urbana-Champaign

November 16th 2015

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 2: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

My research project

Page du projet MISTIS http://mistis.inrialpes.fr/

1 sur 1 12/07/11 13:28

Université Pierre et Marie CURIE - Sciences et Médecine - ... http://www.upmc.fr/fr/index.html

1 sur 1 12/07/11 12:19

Cedric Josz University of Paris VI (UPMC)

Jean Charles GilbertFrench national institute

in scientific computing (INRIA)

Jean Maeght, Patrick PanciaticiFrench transmission

system operator (RTE)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 3: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

French high-voltage network: 400 and 225 kV

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 4: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Benchmark network

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 5: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Underlying graph

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 6: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Optimal power flow

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 7: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Motivations

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 8: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Time (s)

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035

Vo

lta

ge

(kV

)

-500

-400

-300

-200

-100

0

100

200

300

400

500

Two Voltages in Steady State

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 9: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Re(V)-1 -0.5 0 0.5 1

Im(V

)

-1

-0.5

0

0.5

1

Voltages in Complex Plane: Local Optimum of 9241-bus European Network

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 10: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Re(V)-1 -0.5 0 0.5 1

Im(V

)

-1

-0.5

0

0.5

1

Voltages in Complex Plane: Local Optimum of 9241-bus European Network

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 11: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Two-bus network

“g” = conductance

“b” = susceptance

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 12: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Power loss minimization

Minimize

g |v1|2 − g v1v2 − g v2v1 + g |v2|2

over v1, v2 ∈ C subject to

−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2 = 0

b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2 = 0

|v1|2 6 (vmax1 )2

|v2|2 6 (vmax2 )2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 13: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quadratically-constrained quadratic programming

QCQP-C : infz∈Cn

zHH0z s.t. zHHiz 6 hi , i = 1, . . . ,m

(H0, . . . ,Hm are Hermitian matrices)

(h1, . . . , hm ∈ R)

which can also be written:

infz∈Cn

trace(H0zzH) s.t. trace(Hizz

H) 6 hi , i = 1, . . . ,m

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 14: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quadratically-constrained quadratic programming

QCQP-C : infz∈Cn

zHH0z s.t. zHHiz 6 hi , i = 1, . . . ,m

(H0, . . . ,Hm are Hermitian matrices)

(h1, . . . , hm ∈ R)

which can also be written:

infz∈Cn

trace(H0zzH) s.t. trace(Hizz

H) 6 hi , i = 1, . . . ,m

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 15: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Non-commutative diagram

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 16: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Conversion from complex to real numbers

Ring homomorphism:

Λ : (Cn×n,+,×) −→ (R2n×2n,+,×)

Z 7−→(<Z −=Z=Z <Z

)

Useful properties

1 trace[HiZ ] = 12 trace[Λ(HiZ )] = 1

2 trace [Λ(Hi )Λ(Z )]

2 Z < 0 ⇐⇒ Λ(Z ) < 0

3 Z < 0 and rank Z = 1 ⇐⇒ Λ(Z ) < 0 and rank Λ(Z ) = 2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 17: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Conversion from complex to real numbers

Ring homomorphism:

Λ : (Cn×n,+,×) −→ (R2n×2n,+,×)

Z 7−→(<Z −=Z=Z <Z

)

Useful properties

1 trace[HiZ ] = 12 trace[Λ(HiZ )] = 1

2 trace [Λ(Hi )Λ(Z )]

2 Z < 0 ⇐⇒ Λ(Z ) < 0

3 Z < 0 and rank Z = 1 ⇐⇒ Λ(Z ) < 0 and rank Λ(Z ) = 2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 18: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Preprocessed data

Test Number of Number ofCase Complex EdgesName Variables in Graph

GB-2224 2,053 2,581PL-2383wp 2,177 2,651PL-2736sp 2,182 2,675PL-2737sop 2,183 2,675PL-2746wop 2,189 2,708PL-2746wp 2,192 2,686PL-3012wp 2,292 2,805PL-3120sp 2,314 2,835PEGASE-89 70 185PEGASE-1354 983 1,526PEGASE-2869 2,120 3,487PEGASE-9241 7,154 12,292PEGASE-9241R 7,154 12,292

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 19: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Shor relaxation (generation cost minimization)

Case SDP-R SDP-CName Val. ($/hr) Time (sec) Val. ($/hr) Time (sec)

GB-2224 1,928,194 10.9 1,928,444 6.2PL-2383wp 1,862,979 48.1 1,862,985 23.0PL-2736sp* 1,307,749 35.7 1,307,764 22.0PL-2737sop* 777,505 41.7 777,539 19.5PL-2746wop* 1,208,168 51.1 1,208,182 22.8PL-2746wp 1,631,589 43.8 1,631,655 20.0PL-3012wp 2,588,249 52.8 2,588,259 24.3PL-3120sp 2,140,568 64.4 2,140,605 25.5PEGASE-89* 5,819 1.5 5,819 0.9PEGASE-1354 74,035 11.2 74,035 5.6PEGASE-2869 133,936 38.2 133,936 20.6PEGASE-9241 310,658 369.7 310,662 136.1PEGASE-9241R 315,848 317.2 315,731 95.9

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 20: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Second-order conic programming (generation cost min.)

Case SOCP-R SOCP-CName Val. ($/hr) Time (sec) Val. ($/hr) Time (sec)

GB-2224 1,855,393 3.5 1,925,723 1.4PL-2383wp 1,776,726 8.5 1,849,906 2.4PL-2736sp 1,278,926 4.8 1,303,958 1.7PL-2737sop 765,184 5.5 775,672 1.6PL-2746wop 1,180,352 5.1 1,203,821 1.7PL-2746wp 1,586,226 5.5 1,626,418 1.7PL-3012wp 2,499,097 5.9 2,571,422 2.0PL-3120sp 2,080,418 6.2 2,131,258 2.2PEGASE-89 5,744 0.5 5,810 0.4PEGASE-1354 73,102 3.4 73,999 1.5PEGASE-2869 132,520 9.0 133,869 2.7PEGASE-9241 306,050 35.3 309,309 10.0PEGASE-9241R 312,682 36.7 315,411 5.4

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 21: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

We need better relaxations!

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 22: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Moment/sum-of-squares hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 23: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Moment/sum-of-squares hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 24: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Moment/sum-of-squares hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 25: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Power loss minimization

Minimize

g |v1|2 − g v1v2 − g v2v1 + g |v2|2

over v1, v2 ∈ C subject to

−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2 = 0

b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2 = 0

|v1|2 6 (vmax1 )2

|v2|2 6 (vmax2 )2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 26: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex polynomial optimization

Minimize

f (z) :=∑α,β

fα,β zαzβ (where zα := zα1

1 . . . zαnn )

over z ∈ Cn subject to

gi (z) :=∑α,β

gi ,α,β zαzβ > 0 , i = 1, . . . ,m

All functions are real-valued complex polynomialsa.k.a.

Hermitian symmetric polynomials

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 27: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Stone-Weierstrass theorem

Real version

Any continuous function on a compact set can be approximated by∑α

fn,αxα uniform convergence−−−−−−−−−−−−→

n−→+∞f ∈ C(K ⊂ Rn,R)

Complex version

Any continuous function on a compact set(((

(((((((

((hhhhhhhhhhhhcan be approximated by∑

α,β

fn,αzα

������

���XXXXXXXXX

uniform convergence−−−−−−−−−−−−→n−→+∞

f ∈ C(K ⊂ Cn,C)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 28: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Stone-Weierstrass theorem

Real version

Any continuous function on a compact set can be approximated by∑α

fn,αxα uniform convergence−−−−−−−−−−−−→

n−→+∞f ∈ C(K ⊂ Rn,R)

Complex version

Any continuous function on a compact set(((

(((((((

((hhhhhhhhhhhhcan be approximated by∑

α,β

fn,αzα

������

���XXXXXXXXX

uniform convergence−−−−−−−−−−−−→n−→+∞

f ∈ C(K ⊂ Cn,C)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 29: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Stone-Weierstrass theorem

Real version

Any continuous function on a compact set can be approximated by∑α

fn,αxα uniform convergence−−−−−−−−−−−−→

n−→+∞f ∈ C(K ⊂ Rn,R)

Complex version

Any continuous function on a compact set can be approximated by∑α,β

fn,α,β zαzβ

uniform convergence−−−−−−−−−−−−→n−→+∞

f ∈ C(K ⊂ Cn,C)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 30: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Moment approach

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 31: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Non-convex optimization

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 32: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Global optimum and value

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 33: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Variable = point

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 34: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Variable = interval

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 35: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Average on the interval

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 36: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

The optimal interval

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 37: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Weighted average

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 38: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

variable = probability distribution

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 39: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Optimal probability distribution

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 40: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Real moment hierarchy (Lasserre 2000)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 41: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex moment hierarchy (J. and Molzahn 2015)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 42: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Dual point of view

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

Very simple observation:

ε > 0 =⇒ f − f opt︸ ︷︷ ︸>0

+ ε > 0 on the feasible set

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 43: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Dual point of view

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

Very simple observation:

ε > 0 =⇒ f − f opt︸ ︷︷ ︸>0

+ ε > 0 on the feasible set

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 44: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Positivstellensatz

Real version (Putinar 1993)

f (x) > 0 for all x ∈ Rn s.t. gi (x) > 0 and 1− x21 − . . .− x2

n > 0

=⇒

f (x) = σ0(x) +∑m

i=1 σi (x)gi (x) + σm+1(x)(1− x21 − . . .− x2

n )

(where σi ’s are sums of squares)

Example:

x + 2 > 0 for all x ∈ R s.t. 1− x2 > 0

=⇒

x + 2 =(

x√2

+ 1√2

)2+(

1√2

)2(1− x2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 45: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Positivstellensatz

Real version (Putinar 1993)

f (x) > 0 for all x ∈ Rn s.t. gi (x) > 0 and 1− x21 − . . .− x2

n > 0

=⇒

f (x) = σ0(x) +∑m

i=1 σi (x)gi (x) + σm+1(x)(1− x21 − . . .− x2

n )

(where σi ’s are sums of squares)

Example:

x + 2 > 0 for all x ∈ R s.t. 1− x2 > 0

=⇒

x + 2 =(

x√2

+ 1√2

)2+(

1√2

)2(1− x2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 46: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Positivstellensatz

Real version (Putinar 1993)

f (x) > 0 for all x ∈ Rn s.t. gi (x) > 0 and 1− x21 − . . .− x2

n > 0

=⇒

f (x) = σ0(x) +∑m

i=1 σi (x)gi (x) + σm+1(x)(1− x21 − . . .− x2

n )

(where σi ’s are sums of squares)

Example:

x + 2 > 0 for all x ∈ R s.t. 1− x2 > 0

=⇒

x + 2 =(

x√2

+ 1√2

)2+(

1√2

)2(1− x2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 47: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Positivstellensatz

Complex version (D’Angelo and Putinar 2008)

f (z) > 0 for all z ∈ Cn s.t. gi (z) > 0 and 1− |z1|2 − ..− |zn|2 = 0

=⇒

f (z) = σ0(z) +∑m

i=1 σi (z)gi (z) + pm+1(z)(1− |z1|2 − . . .− |zn|2)

(where σi ’s are Hermitian sums of squares)

Example:

z + z + 3 > 0 for all z ∈ C s.t. 1− |z |2 = 0

=⇒

z + z + 3 = |1|2 + |z + 1|2 + 1× (1− |z |2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 48: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Positivstellensatz

Complex version (D’Angelo and Putinar 2008)

f (z) > 0 for all z ∈ Cn s.t. gi (z) > 0 and 1− |z1|2 − ..− |zn|2 = 0

=⇒

f (z) = σ0(z) +∑m

i=1 σi (z)gi (z) + pm+1(z)(1− |z1|2 − . . .− |zn|2)

(where σi ’s are Hermitian sums of squares)

Example:

z + z + 3 > 0 for all z ∈ C s.t. 1− |z |2 = 0

=⇒

z + z + 3 = |1|2 + |z + 1|2 + 1× (1− |z |2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 49: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Positivstellensatz

Complex version (D’Angelo and Putinar 2008)

f (z) > 0 for all z ∈ Cn s.t. gi (z) > 0 and 1− |z1|2 − ..− |zn|2 = 0

=⇒

f (z) = σ0(z) +∑m

i=1 σi (z)gi (z) + pm+1(z)(1− |z1|2 − . . .− |zn|2)

(where σi ’s are Hermitian sums of squares)

Example:

z + z + 3 > 0 for all z ∈ C s.t. 1− |z |2 = 0

=⇒

z + z + 3 = |1|2 + |z + 1|2 + 1× (1− |z |2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 50: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

ε > 0 =⇒ f−f opt + ε > 0 on the feasible set

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 51: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

ε > 0 =⇒ f − (f opt − ε) > 0 on the feasible set

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 52: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

ε > 0 =⇒ f − (f opt − ε)︸ ︷︷ ︸λ

> 0 on the feasible set

f opt = supλ,σi

λ subject to f − λ = σ0 +m∑i=1

σigi

(under the assumption in the Positivstellensatz)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 53: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with semidefinite programming

Real version

σ(x) =∑

α σαxα is a sum of squares

∑k(∑

γ pk,γxγ)2

⇐⇒

Matrix (σα+β)α,β is positive semidefinite

Complex version

σ(z) =∑

α,β σα,β zαzβ is a Hermitian sum of squares∑k |∑

γ pk,γzγ |2

⇐⇒

Matrix (σα,β)α,β is Hermitian positive semidefinite

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 54: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with semidefinite programming

Real version

σ(x) =∑

α σαxα is a sum of squares

∑k(∑

γ pk,γxγ)2

⇐⇒

Matrix (σα+β)α,β is positive semidefinite

Complex version

σ(z) =∑

α,β σα,β zαzβ is a Hermitian sum of squares∑k |∑

γ pk,γzγ |2

⇐⇒

Matrix (σα,β)α,β is Hermitian positive semidefinite

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 55: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Sum-of-squares hierarchy

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

supλ,σi

λ subject to

{f − λ = σ0 +

∑mi=1 σigi

deg(σ0), deg(σigi ) 6 2d

Real version

One of the constraints is a ball x21 + . . .+ x2

n 6 R2

=⇒GLOBAL convergence when d −→ +∞

Complex version

One of the constraints is a sphere |z1|2 + . . .+ |zn|2 = R2

=⇒GLOBAL convergence when d −→ +∞

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 56: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Sum-of-squares hierarchy

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

supλ,σi

λ subject to

{f − λ = σ0 +

∑mi=1 σigi

deg(σ0), deg(σigi ) 6 2d

Real version

One of the constraints is a ball x21 + . . .+ x2

n 6 R2

=⇒GLOBAL convergence when d −→ +∞

Complex version

One of the constraints is a sphere |z1|2 + . . .+ |zn|2 = R2

=⇒GLOBAL convergence when d −→ +∞

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 57: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Sum-of-squares hierarchy

f opt := inf f subject to gi > 0 , i = 1, . . . ,m

supλ,σi

λ subject to

{f − λ = σ0 +

∑mi=1 σigi

deg(σ0), deg(σigi ) 6 2d

Real version

One of the constraints is a ball x21 + . . .+ x2

n 6 R2

=⇒GLOBAL convergence when d −→ +∞

Complex version

One of the constraints is a sphere |z1|2 + . . .+ |zn|2 = R2

=⇒GLOBAL convergence when d −→ +∞

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 58: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Power loss minimization

Minimize

g |v1|2 − g v1v2 − g v2v1 + g |v2|2

over v1, v2 ∈ C subject to

−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2 = 0

b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2 = 0

|v1|2 6 (vmax1 )2

|v2|2 6 (vmax2 )2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 59: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Slack variable

Minimize

g |v1|2 − g v1v2 − g v2v1 + g |v2|2

over v1, v2, v3 ∈ C subject to

−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2 = 0

b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2 = 0

|v1|2 6 (vmax1 )2

|v2|2 6 (vmax2 )2

|v1|2 + |v2|2 + |v3|2 = (vmax1 )2 + (vmax

2 )2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 60: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Actually works in practice!

D’Angelo and Putinar (2008):

1

18= inf

z∈C1− 4

3|z |2 +

7

18|z |4 s.t. 1− |z |2 > 0

→ complex hierarchy yields −0.3333 at second and third orders

J. and Molzahn (2015):

1

18= inf

z1,z2∈C1− 4

3|z1|2 +

7

18|z1|4 s.t. 1− |z1|2 − |z2|2 = 0

→ complex hierarchy yields 0.0556 ≈ 118 at second order

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 61: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Actually works in practice!

D’Angelo and Putinar (2008):

1

18= inf

z∈C1− 4

3|z |2 +

7

18|z |4 s.t. 1− |z |2 > 0

→ complex hierarchy yields −0.3333 at second and third orders

J. and Molzahn (2015):

1

18= inf

z1,z2∈C1− 4

3|z1|2 +

7

18|z1|4 s.t. 1− |z1|2 − |z2|2 = 0

→ complex hierarchy yields 0.0556 ≈ 118 at second order

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 62: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Successful sum of squares decomposition

1− 4

3|z1|2 +

7

18|z1|4 − 0.0556

=

0.2780|z2|2 + 0.2776|z1z2|2 + 0.6667|z2|4

+

(0.9444− 0.3889|z1|2 + 0.6665|z2|2)(1− |z1|2 − |z2|2)

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 63: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Another example: an ellipse

Putinar and Scheiderer (2012):

1 = infz∈C

3− |z |2 s.t. |z |2 − 1

4z2 − 1

4z2 − 1 = 0

→ complex hierarchy is unbounded at second and third orders

J. and Molzahn (2015):

1 = infz1,z2∈C

3− |z1|2 s.t.

{|z1|2 − 1

4z21 − 1

4 z21 − 1 = 0

3− |z1|2 − |z2|2 = 0

→ complex hierarchy yields 0.6813 at second order and 0.9699 atthird order

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 64: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Another example: an ellipse

Putinar and Scheiderer (2012):

1 = infz∈C

3− |z |2 s.t. |z |2 − 1

4z2 − 1

4z2 − 1 = 0

→ complex hierarchy is unbounded at second and third orders

J. and Molzahn (2015):

1 = infz1,z2∈C

3− |z1|2 s.t.

{|z1|2 − 1

4z21 − 1

4 z21 − 1 = 0

3− |z1|2 − |z2|2 = 0

→ complex hierarchy yields 0.6813 at second order and 0.9699 atthird order

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 65: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Nonnegative slack variable

Putinar and Scheiderer (2012):

1 = infz∈C

3− |z |2 s.t. |z |2 − 1

4z2 − 1

4z2 − 1 = 0

→ complex hierarchy is unbounded at second and third orders

J. and Molzahn (2015):

1 = infz1,z2∈C

3− |z1|2 s.t.

|z1|2 − 1

4z21 − 1

4 z21 − 1 = 0

3− |z1|2 − |z2|2 = 0

iz2 − iz2 = 0

z2 + z2 > 0

→ complex hierarchy yields 0.6813 at second order and 1.0000 atthird order

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 66: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex plane

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 67: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Ellipse

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 68: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Nonnegative slack variable

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 69: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Sphere constraint

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 70: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Projection on semi-sphere

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 71: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Edge of chips

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 72: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex vs. real hierarchy

Size: asymptotically 2d−1 times smaller at order d

Bound quality: poorer or equal at each order

Global optimality check: rank of moment matrix

Convergence guarantee: if feasible set bounded by radius R, add

x21 + . . .+ x2

n 6 R2 to real hierarchy

zn+1 and |z1|2 + . . .+ |zn+1|2 = R2 to complex hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 73: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex vs. real hierarchy

Size: asymptotically 2d−1 times smaller at order d

Bound quality: poorer or equal at each order

Global optimality check: rank of moment matrix

Convergence guarantee: if feasible set bounded by radius R, add

x21 + . . .+ x2

n 6 R2 to real hierarchy

zn+1 and |z1|2 + . . .+ |zn+1|2 = R2 to complex hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 74: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex vs. real hierarchy

Size: asymptotically 2d−1 times smaller at order d

Bound quality: poorer or equal at each order

Global optimality check: rank of moment matrix

Convergence guarantee: if feasible set bounded by radius R, add

x21 + . . .+ x2

n 6 R2 to real hierarchy

zn+1 and |z1|2 + . . .+ |zn+1|2 = R2 to complex hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 75: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex vs. real hierarchy

Size: asymptotically 2d−1 times smaller at order d

Bound quality: poorer or equal at each order

Global optimality check: rank of moment matrix

Convergence guarantee: if feasible set bounded by radius R, add

x21 + . . .+ x2

n 6 R2 to real hierarchy

zn+1 and |z1|2 + . . .+ |zn+1|2 = R2 to complex hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 76: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Oscillatory polynomial optimization

Lemma for ϕ(z) =∑

α,β ϕα,β zαzβ

∀z ∈ Cn, ∀θ ∈ R, ϕ(e iθz) = ϕ(z)

⇐⇒

∀α, β, (|α| − |β|)ϕα,β = 0

Example:ϕ(z) = z2z2 with z ∈ C is oscillatoryϕ(z) = z3z2 with z ∈ C is not oscillatory

In an oscillatory problem

minimum order of the complex hierarchy=

minimum order of the real hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 77: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Oscillatory polynomial optimization

Lemma for ϕ(z) =∑

α,β ϕα,β zαzβ

∀z ∈ Cn, ∀θ ∈ R, ϕ(e iθz) = ϕ(z)

⇐⇒

∀α, β, (|α| − |β|)ϕα,β = 0

Example:ϕ(z) = z2z2 with z ∈ C is oscillatoryϕ(z) = z3z2 with z ∈ C is not oscillatory

In an oscillatory problem

minimum order of the complex hierarchy=

minimum order of the real hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 78: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Oscillatory polynomial optimization

Lemma for ϕ(z) =∑

α,β ϕα,β zαzβ

∀z ∈ Cn, ∀θ ∈ R, ϕ(e iθz) = ϕ(z)

⇐⇒

∀α, β, (|α| − |β|)ϕα,β = 0

Example:ϕ(z) = z2z2 with z ∈ C is oscillatoryϕ(z) = z3z2 with z ∈ C is not oscillatory

In an oscillatory problem

minimum order of the complex hierarchy=

minimum order of the real hierarchy

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 79: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Exploiting sparsity

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 80: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Moment/sum-of-squares hierarchy (power loss min.)

Case MSOS-R MSOS-CName Val. (MW) Time (sec) Val. (MW) Time (sec)

PL-2383wp 24,990 583.4 24,991 53.9PL-2736sp 18,334 44.0 18,335 17.8PL-2737sop 11,397 52.4 11,397 25.7PL-2746wop 19,210 2,662.4 19,212 124.3PL-2746wp 25,267 45.9 25,269 18.5PL-3012wp 27,642 318.7 27,644 141.0PL-3120sp 21,512 386.6 21,512 193.9PEGASE-1354 74,043 406.9 74,042 1,132.6PEGASE-2869 133,944 921.3 133,939 700.8

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 81: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Do we need a sphere constraint to ensure convergence?

Minimizeg |v1|2 − g v1v2 − g v2v1 + g |v2|2

over v1, v2 ∈ C subject to

−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2 = 0

b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2 = 0

|v1|2 6 (vmax1 )2

|v2|2 6 (vmax2 )2

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 82: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Instrinsic property of electricity?

p(v1, v2)

(−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2

)+

q(v1, v2)

(b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2

)=

|v1|2 + |v2|2 + Hermitian SOS + constant ???

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 83: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Instrinsic property of electricity?

p(v1, v2)

(−g − ib

2v1v2 −

g + ib

2v2v1 + g |v2|2 + pdem

2

)+

q(v1, v2)

(b + ig

2v1v2 +

b − ig

2v2v1 − b |v2|2 + qdem

2

)=

|v1|2 + |v2|2 + Hermitian SOS + constant ???

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 84: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

References

RTE. “Carte du reseau de transport 400 000 et 225 000 volts,”Gestionnaire du Reseau de Transport, Centre National d’ExpertiseReseau, Depot legal : Mai 2011. [link]

University of Washington, Electrical Engineering, Power SystemsTest Case Archive. [link]

N.Z. Shor, Quadratic Optimization Problems, Sov. J. Comput.Syst. Sci., 25 (1987), pp. 1–11.

J. Lavaei and S.H. Low, Zero Duality Gap in Optimal Power FlowProblem, IEEE Trans. Power Syst., 27 (2012), pp. 92–107.

J. B. Lasserre, Global Optimization with Polynomials and theProblem of Moments, SIAM J. Optim., 11 (2001), pp. 796–817.

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 85: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

References

P. A. Parrilo, Structured semidefinite programs and semialgebraicgeometry methods in robustness and optimization, DoctoralThesis, California Institute of Technology, 2000.

D. K. Molzahn and Ian A. Hiskens, Sparsity-ExploitingMoment-Based Relaxations of the Optimal Power Flow Problem,IEEE Transaction on Power Systems, vol. 30, no. 6, pp.3168-3180, November 2015.

J. P. D’Angelo and M. Putinar, Polynomial Optimization onOdd-Dimensional Spheres, in Emerging Applications of AlgebraicGeometry, Springer New York, 2008.

M. Putinar and C. Scheiderer, Quillen Property of Real AlgebraicVarieties, to appear in Munster J. Math.

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 86: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

References

M. Putinar and C. Scheiderer, Hermitian Algebra on the Ellipse,Illinois J. Math., 56 (2012), pp. 213–220.

C. J. and D. K. Molzahn, Moment/Sum-of-Squares Hierarchy forComplex Polynomial Optimization, submitted to SIAM J. Optim.

M. Schweighofer, Optimization of Polynomials on CompactSemialgebraic Sets, SIAM J. Optim., 15 (2005), pp. 805–825.

E. J. Anderson and P. Nash, Linear Programming inInfinite-Dimensional Spaces, Theory and Applications, Wiley Int.Ser. Disc. Math. Optim., 1987.

C. D. Aliprantis and K. Border, Infinite Dimensional Analysis, AHitchhiker’s guide, Second Edition, Springer-Verlag BerlinHeidelberg, 1999.

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 87: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Images

Ecociel France, Panneaux Solaires Photovoltaıques. [link]

Autolib’ va s’exporter a Lyon, L’Argus, Actualites auto, May 30th

2013. [link]

Transmission Line Monitor and Dynamic Line Rating System,Lindsey. [link]

Dispatching national, Rte France. [link]

Paint, Version 6.1, Microsoft Windows, 2009.

Jean Bernard Lasserre, Tuchan, Photograph. [link]

POV-Ray, The Persistence of Vision Raytracer, Version 3.7.

Kellogg’s achete les chips Pringles a Procter & Gamble, Le Figaro,December 15th 2012. [link]

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 88: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Images

Earth at night from NASA’s Suomi National Polar-orbitingPartnership Satellite, Earth shimmers at Christmas as billions offairy lights become visible from space, The Telegraph, ScienceNews. [link]

Perry Babin, Oscilloscope, Basic Car Audio Electronics. [link]

MRI Image of Knee, Medical Media Images, Precision Color MedialImages. [link]

David Ratledge, Orion Nebula taken with a modified Canon 40D(Baader filter) from lighted polluted Lancashire!, Digital SLRImaging. [link]

Radar, Technology, Telecommunications and Media, Speyside,Corporate Relations. [link]

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 89: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Oscillatory phenomena in physical systems

Schrodinger equation: HΨ = i~∂Ψ

∂t

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 90: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Future directions

Enhance tractability of complex hierarchy

When are real and complex hierarchies equal?

Do power flow equations possess the Quillen property?

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 91: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Thank you for your attention!

Feel free to me contact at

[email protected]

for questions or suggestions.

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 92: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Backup slides

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 93: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Moment/sum-of-squares hierarchy

Name Active power minimization (MW)Case MSOS-R MSOS-C Matpower

PL-2383wp 24,990 24,991 24,991PL-2736sp 18,334 18,335 18,336PL-2737sop 11,397 11,397 11,397PL-2746wop 19,210 19,212 19,212PL-2746wp 25,267 25,269 25,269PL-3012wp 27,642 27,644 27,646PL-3120sp 21,512 21,512 21,513PEGASE-1354 74,043 74,042 74,043PEGASE-2869 133,944 133,939 133,945

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 94: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Nonnegative slack variable

Putinar and Scheiderer (2012):

1 = infz∈C

3− |z |2 s.t. |z |2 − 1

4z2 − 1

4z2 − 1 = 0

→ complex hierarchy is unbounded at second and third orders

J. and Molzahn (2015):

1 = infz1,z2∈C

3− |z1|2 s.t.

|z1|2 − 1

4z21 − 1

4 z21 − 1 = 0

3− |z1|2 − |z2|2 = 0

iz2 − iz2 = 0

z2 + z2 > 0

→ complex hierarchy yields 0.6813 at second order and 1.0000 atthird order

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 95: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

3rd order complex moment matrix

(0,0

)

(1,0

)

(0,1

)

(2,0

)

(1,1

)

(0,2

)

(3,0

)

(2,1

)

(1,2

)

(0,3

)

(0, 0) 1 0 1 2 0 1 0 2 0 1(1, 0) 0 2 0 0 2 0 4 0 2 0(0, 1) 1 0 1 2 0 1 0 2 0 1(2, 0) 2 0 2 4 0 2 0 4 0 2(1, 1) 0 2 0 0 2 0 4 0 2 0(0, 2) 1 0 1 2 0 1 0 2 0 1(3, 0) 0 4 0 0 4 0 8 0 4 0(2, 1) 2 0 2 4 0 2 0 4 0 2(1, 2) 0 2 0 0 2 0 4 0 2 0(0, 3) 1 0 1 2 0 1 0 2 0 1

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 96: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Certificate of global optimality

(0,0

)

(1,0

)

(0,1

)

(2,0

)

(1,1

)

(0,2

)

(3,0

)

(2,1

)

(1,2

)

(0,3

)

(0, 0) 1 0 1 2 0 1 0 2 0 1(1, 0) 0 2 0 0 2 0 4 0 2 0(0, 1) 1 0 1 2 0 1 0 2 0 1(2, 0) 2 0 2 4 0 2 0 4 0 2(1, 1) 0 2 0 0 2 0 4 0 2 0(0, 2) 1 0 1 2 0 1 0 2 0 1(3, 0) 0 4 0 0 4 0 8 0 4 0(2, 1) 2 0 2 4 0 2 0 4 0 2(1, 2) 0 2 0 0 2 0 4 0 2 0(0, 3) 1 0 1 2 0 1 0 2 0 1

rank M1(y) = rank M3(y) = 2 so there are two global solutions

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 97: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex polynomial optimization

Minimize

f (z) :=∑α,β

fα,β zαzβ (where zα := zα1

1 . . . zαnn )

over z ∈ Cn subject to

gi (z) :=∑α,β

gi ,α,β zαzβ = 0 , 1 6 i 6 m

hj(z) :=∑α,β

hj ,α,β zαzβ > 0 , 1 6 j 6 p

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 98: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If R2 − |z1|2 − . . .− |zn|2 ∈ S , then: f|XM> 0 =⇒ f ∈ M

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 99: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If R2 − |z1|2 − . . .− |zn|2 ∈ S , then: f|XM> 0 =⇒ f ∈ M

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 100: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If R2 − |z1|2 − . . .− |zn|2 ∈ S , then: f|XM> 0 =⇒ f ∈ M

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 101: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If R2 − |z1|2 − . . .− |zn|2 ∈ S , then: f|XM> 0 =⇒ f ∈ M

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 102: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If R2 − |z1|2 − . . .− |zn|2 ∈ S , then: f|XM> 0 =⇒ f ∈ M

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 103: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If R2 − |z1|2 − . . .− |zn|2 ∈ S , then: f|XM> 0 =⇒ f ∈ M

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 104: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Quillen property of semi-algebraic varieties

Ideal generated by equality constraints:

I := Ch[z , z ]g1 + . . .+ Ch[z , z ]gm

Semiring defined by ideal:

S := I + Σ[z ] where Σ[z ] := {∑k

|pk |2 | pk ∈ C[z ] }

Module defined by equality and inequality constraints:

M := S + Σ[z ]h1 + . . .+ Σ[z ]hp

Feasible set of complex polynomial optimization:

XM := { z ∈ Cn | ∀ϕ ∈ M, ϕ(z) > 0 }

Putinar and Scheiderer (to appear in Munster J. Math.)

If S is Archimedean, then M has Quillen’s property.

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 105: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

Lemma

Quillen property ⇐⇒ inf{f (z) | z ∈ XM} = sup{λ | f − λ ∈ M}

Duality bracket:〈., .〉 : Ch[z , z ]×H −→ R

(ϕ, y) 7−→∑

α,β ϕα,βyα,β

Continuous operator:A : Ch[z , z ] −→ Ch[z , z ]

ϕ 7−→ ϕ− ϕ0,0

Convex cone: M = I + Σ[z ] + Σ[z ]h1 + . . .+ Σ[z ]hp

Constants: δ0,0 ∈ H and b := Af

Linear program of infinite dimension

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 106: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

Lemma

Quillen property ⇐⇒ inf{f (z) | z ∈ XM} = sup{λ | f − λ ∈ M}

Duality bracket:〈., .〉 : Ch[z , z ]×H −→ R

(ϕ, y) 7−→∑

α,β ϕα,βyα,β

Continuous operator:A : Ch[z , z ] −→ Ch[z , z ]

ϕ 7−→ ϕ− ϕ0,0

Convex cone: M = I + Σ[z ] + Σ[z ]h1 + . . .+ Σ[z ]hp

Constants: δ0,0 ∈ H and b := Af

Linear program of infinite dimension

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 107: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

Lemma

Quillen property ⇐⇒ inf{f (z) | z ∈ XM} = sup{λ | f − λ ∈ M}

Duality bracket:〈., .〉 : Ch[z , z ]×H −→ R

(ϕ, y) 7−→∑

α,β ϕα,βyα,β

Continuous operator:A : Ch[z , z ] −→ Ch[z , z ]

ϕ 7−→ ϕ− ϕ0,0

Convex cone: M = I + Σ[z ] + Σ[z ]h1 + . . .+ Σ[z ]hp

Constants: δ0,0 ∈ H and b := Af

Linear program of infinite dimension

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 108: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

Lemma

Quillen property ⇐⇒ inf{f (z) | z ∈ XM} = sup{λ | f − λ ∈ M}

Duality bracket:〈., .〉 : Ch[z , z ]×H −→ R

(ϕ, y) 7−→∑

α,β ϕα,βyα,β

Continuous operator:A : Ch[z , z ] −→ Ch[z , z ]

ϕ 7−→ ϕ− ϕ0,0

Convex cone: M = I + Σ[z ] + Σ[z ]h1 + . . .+ Σ[z ]hp

Constants: δ0,0 ∈ H and b := Af

Linear program of infinite dimension

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 109: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

Lemma

Quillen property ⇐⇒ inf{f (z) | z ∈ XM} = sup{λ | f − λ ∈ M}

Duality bracket:〈., .〉 : Ch[z , z ]×H −→ R

(ϕ, y) 7−→∑

α,β ϕα,βyα,β

Continuous operator:A : Ch[z , z ] −→ Ch[z , z ]

ϕ 7−→ ϕ− ϕ0,0

Convex cone: M = I + Σ[z ] + Σ[z ]h1 + . . .+ Σ[z ]hp

Constants: δ0,0 ∈ H and b := Af

Linear program of infinite dimension

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 110: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with optimization

Lemma

Quillen property ⇐⇒ inf{f (z) | z ∈ XM} = sup{λ | f − λ ∈ M}

Duality bracket:〈., .〉 : Ch[z , z ]×H −→ R

(ϕ, y) 7−→∑

α,β ϕα,βyα,β

Continuous operator:A : Ch[z , z ] −→ Ch[z , z ]

ϕ 7−→ ϕ− ϕ0,0

Convex cone: M = I + Σ[z ] + Σ[z ]h1 + . . .+ Σ[z ]hp

Constants: δ0,0 ∈ H and b := Af

Linear program of infinite dimension

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 111: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with semidefinite programming

Primal problem

ϕ ∈ Σ[z ] ⇐⇒ (ϕα,β)α,β < 0

Dual problem

M∗ =(∑

i Ch[z , z ]gi + Σ[z ] +∑

j Σ[z ]hj

)∗=

⋂i (|C[z ]|2gi )⊥ ∩ (|C[z ]|2)∗ ∩

⋂j(|C[z ]|2hj)∗

For example, if φ ∈ C[z ], then:

〈|φ|2hj , y〉 =∑α,β

φαφβ〈zαzβhj(z), y〉 > 0

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 112: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with semidefinite programming

Primal problem

ϕ ∈ Σ[z ] ⇐⇒ (ϕα,β)α,β < 0

Dual problem

M∗ =(∑

i Ch[z , z ]gi + Σ[z ] +∑

j Σ[z ]hj

)∗=

⋂i (|C[z ]|2gi )⊥ ∩ (|C[z ]|2)∗ ∩

⋂j(|C[z ]|2hj)∗

For example, if φ ∈ C[z ], then:

〈|φ|2hj , y〉 =∑α,β

φαφβ〈zαzβhj(z), y〉 > 0

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 113: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Relationship with semidefinite programming

Primal problem

ϕ ∈ Σ[z ] ⇐⇒ (ϕα,β)α,β < 0

Dual problem

M∗ =(∑

i Ch[z , z ]gi + Σ[z ] +∑

j Σ[z ]hj

)∗=

⋂i (|C[z ]|2gi )⊥ ∩ (|C[z ]|2)∗ ∩

⋂j(|C[z ]|2hj)∗

For example, if φ ∈ C[z ], then:

〈|φ|2hj , y〉 =∑α,β

φαφβ〈zαzβhj(z), y〉 > 0

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 114: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex moment/sum-of-squares hierarchy

f opt := infz∈Cn f (z) s.t. gi (z) = 0 and hj(z) > 0

Semidefinite programming hierarchy

infy∈Hd〈f , y〉

s.t. y0,0 = 1 and Md(y) < 0Md−d(gi )(giy) = 0, i = 1, . . . ,mMd−d(hj )(hjy) < 0, j = 1, . . . , p

supλ,r ,σ λ

s.t. f − λ = σ +∑m

i=0 rigi +∑p

j=0 σjhjλ ∈ R, σ ∈ Σd [z ], ri ∈ Rd−d(gi )[z , z ], σj ∈ Σd−d(hj )[z ]

Md−k(ϕy) is a Hermitian matrix indexed by |α|, |β| 6 d − k and:

( (Md−k(ϕy) )α,β :=∑γ,δ

ϕγ,δ yα+γ,β+δ

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 115: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex moment/sum-of-squares hierarchy

f opt := infz∈Cn f (z) s.t. gi (z) = 0 and hj(z) > 0

Semidefinite programming hierarchy

infy∈Hd〈f , y〉

s.t. y0,0 = 1 and Md(y) < 0Md−d(gi )(giy) = 0, i = 1, . . . ,mMd−d(hj )(hjy) < 0, j = 1, . . . , p

supλ,r ,σ λ

s.t. f − λ = σ +∑m

i=0 rigi +∑p

j=0 σjhjλ ∈ R, σ ∈ Σd [z ], ri ∈ Rd−d(gi )[z , z ], σj ∈ Σd−d(hj )[z ]

Md−k(ϕy) is a Hermitian matrix indexed by |α|, |β| 6 d − k and:

( (Md−k(ϕy) )α,β :=∑γ,δ

ϕγ,δ yα+γ,β+δ

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 116: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Complex moment/sum-of-squares hierarchy

f opt := infz∈Cn f (z) s.t. gi (z) = 0 and hj(z) > 0

Semidefinite programming hierarchy

infy∈Hd〈f , y〉

s.t. y0,0 = 1 and Md(y) < 0Md−d(gi )(giy) = 0, i = 1, . . . ,mMd−d(hj )(hjy) < 0, j = 1, . . . , p

supλ,r ,σ λ

s.t. f − λ = σ +∑m

i=0 rigi +∑p

j=0 σjhjλ ∈ R, σ ∈ Σd [z ], ri ∈ Rd−d(gi )[z , z ], σj ∈ Σd−d(hj )[z ]

Md−k(ϕy) is a Hermitian matrix indexed by |α|, |β| 6 d − k and:

( (Md−k(ϕy) )α,β :=∑γ,δ

ϕγ,δ yα+γ,β+δ

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 117: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Algebraic geometry and measure theory

Quillen property:

∀ϕ ∈ Ch[z , z ], ϕ|XM> 0 =⇒ ϕ ∈ M

Strong moment property:

∀y ∈ H, y ∈ M∗ =⇒ yα,β =∫XM

zαzβdµ

Putinar and Scheiderer (to appear in Munster J. Math.)

Quillen’s property =⇒ Strong moment property

Converse true if M is Archimedean

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 118: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Algebraic geometry and measure theory

Quillen property:

∀ϕ ∈ Ch[z , z ], ϕ|XM> 0 =⇒ ϕ ∈ M

Strong moment property:

∀y ∈ H, y ∈ M∗ =⇒ yα,β =∫XM

zαzβdµ

Putinar and Scheiderer (to appear in Munster J. Math.)

Quillen’s property =⇒ Strong moment property

Converse true if M is Archimedean

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 119: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Optimization

Linear program

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Quillen property ⇐⇒ val(P) = f0,0 − f opt

Strong moment property ⇐⇒ val(D) = f0,0 − f opt

Weak duality

val(P) > val(D) > f0,0 − f opt

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 120: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Optimization

Linear program

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Quillen property ⇐⇒ val(P) = f0,0 − f opt

Strong moment property ⇐⇒ val(D) = f0,0 − f opt

Weak duality

val(P) > val(D) > f0,0 − f opt

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign

Page 121: Complex Structure of the Optimal Power Flow Problem · 2016-01-26 · Complex Structure of the Optimal Power Flow Problem C edric JOSZ, Daniel K. MOLZAHN cedric.josz@rte-france.com

Optimization

Linear program

(P) : infϕ∈Ch[z,z] 〈ϕ, δ0,0〉 s.t. Aϕ = b and ϕ ∈ M(D) : supy∈H 〈b, y〉 s.t. δ0,0 − A∗y ∈ M∗

Quillen property ⇐⇒ val(P) = f0,0 − f opt

Strong moment property ⇐⇒ val(D) = f0,0 − f opt

Weak duality

val(P) > val(D) > f0,0 − f opt

Cedric JOSZ, Daniel K. MOLZAHN [email protected] Talk at the University of Illinois at Urbana-Champaign