Download Here
Download
Report
Transcript Download Here
Difference of Convex (DC)
Decomposition of
Nonconvex Polynomials with
Algebraic Techniques
Georgina Hall
Princeton, ORFE
Joint work with
Amir Ali Ahmadi
Princeton, ORFE
7/13/2015
MOPTA 2015
1
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Difference of Convex (DC) programming
β’ Problems of the form
min π0 (π₯)
π . π‘. ππ π₯ β€ 0
where:
β’ ππ π₯ β ππ π₯ β βπ π₯ , π = 0, β¦ , π,
β’ ππ : βπ β β, βπ : βπ β β are convex.
2
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Concave-Convex Computational Procedure (CCCP)
β’ Heuristic for minimizing DC programming problems.
β’ Has been used extensively in:
β’ machine learning (sparse support vector machines (SVM), transductive SVMs,
sparse principal component analysis)
β’ statistical physics (minimizing Bethe and Kikuchi free energies).
β’ Idea:
Input
πβ0
xπ₯0 , initial point
ππ = ππ β βπ ,
π = 0, β¦ , π
Convexify by linearizing π
xπππ π = ππ π₯ β (βπ π₯π + π»βπ π₯π
convex convex
πππ π
ππ (π)
π
Solve convex subproblem
π₯ β π₯π )
affine
π βπ+1
Take π₯π+1 to be the solution of
min π0π π₯
π . π‘. πππ π₯ β€ 0, π = 1, β¦ , π
3
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Concave-Convex Computational Procedure (CCCP)
β’ Toy example: min π π₯ , where π π₯ β π π₯ β β(π₯)
π₯
Convexify π π₯ to
obtain π 0 (π₯)
Initial point: π₯0 = 2
Minimize π 0 (π₯) and
obtain π₯1
Reiterate
π₯π₯π₯β43π₯π₯2 1
π₯0
4
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
CCCP for nonconvex polynomial optimization problems (1/2)
CCCP relies on input functions being given as a difference of convex
functions.
What if we donβt have access to such a decomposition?
We will consider polynomials in π variables and of degree π.
β’ Any polynomial can be written as a difference of convex polynomials.
β’ Proof by Wang, Schwing and Urtasun
β’ Alternative proof given later in this presentation, as corollary of stronger
theorem
5
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
CCCP for nonconvex polynomial optimization problems (2/2)
π π₯ = π π₯ β β(π₯)
β’ In fact, for any polynomial, β an infinite number of decompositions.
Example
xπ(π₯) = π₯ 4 β 3π₯ 2 + 2π₯ β 2
Possible decompositions
π π₯ = π₯4,
β π₯ = 3π₯ 2 β 2π₯ + 2
π π₯ = π₯ 4 + ππ ,
β π₯ = 3π₯ 2 + ππ β 2π₯ + 2,
Which one would be a natural choice for CCCP?
etc.
6
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Picking the βbestβ decomposition (1/2)
Algorithm
Linearize π π around a point π₯π to obtain convexified version of π(π)
Idea
Pick β π₯ such that it is as close as possible to affine
Mathematical translation
Minimize curvature of β (π»β is the hessian of β)
At a point π
Over a region π
min ππππ₯ (π»β π )
min max ππππ₯ (π»β π₯ )
s.t. π = π β β
π, β convex
s.t. π = π β β,
π, β convex
g,h
π,β
π₯βΞ©
7
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Picking the βbestβ decomposition (2/2)
Theorem: Finding the βbestβ decomposition of a degree-4 polynomial
over a box is NP-hard.
Proof idea: Reduction via testing convexity of quartic polynomials is
hard (Ahmadi, Olshevsky, Parrilo, Tsitsiklis).
The same is likely to hold for the point version, but we have been unable
to prove it.
How can we efficiently find such a decomposition?
8
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Convex relaxations for DC decompositions (1/6)
SOS, DSOS, SDSOS polynomials (Ahmadi, Majumdar)
β’ Families of nonnegative polynomials.
Type
Characterization
Testing
membership
Sum of squares (sos)
βππ , polynomials, s.t. π(π₯) = βππ2 (π₯)
SDP
2
β
Scaled diagonally dominant
sum of squares (sdsos)
p= βπ πΌπ ππ2 + βπ,π π½π+ ππ + πΎπ+ ππ + π½πβ ππ β πΎπβ ππ
ππ , ππ monomials, πΌπ β₯ 0
Diagonally dominant
sum of squares (dsos)
+
β
p= βπ πΌπ ππ2 + βπ,π π½π,π
ππ + ππ + π½π,π
ππ β ππ
+,β
ππ , ππ monomials, πΌπ , π½ππ
β₯0
2
2
SOCP
2
LP
9
β
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Convex relaxations for DC decompositions (2/6)
DSOS-convex, SDSOS-convex, SOS-convex polynomials
π(π₯)
convex
β π»π π₯ β½ 0, βπ₯ β
π¦ π π»π π₯ π¦ β₯ 0,
βπ₯, π¦ β βπ
β
π¦ π π»π π₯ π¦
sos/sdsos/dsos
Definitions:
β’ π is dsos-convex if π¦ π π»π π₯ π¦ is dsos. LP
β’ π is sdsos-convex if π¦ π π»π π₯ π¦ is sdsos. SOCP
β’ π is sos-convex if π¦ π π»π π₯ π¦ is sos. SDP
10
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Convex relaxations for DC decompositions (3/6)
Comparison of these sets on a parametric family of polynomials:
π π₯1 , π₯2 = 2π₯14 + 2π₯24 + ππ₯13 π₯2 + ππ₯12 π₯22 + ππ₯1 π₯23
π = β0.5
π=0
π=1
π
π
π
π
dsos-convex
π
sdsos-convex
π
sos-convex=convex
11
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Convex relaxations for DC decompositions (4/6)
How to use these concepts to do DC decomposition at a point π?
Original problem
min ππππ₯ (π»β π )
s.t. π = π β β
π, β convex
β
Relaxation 1:
sos-convex
min π‘
s.t. π»β π βΌ π‘πΌ
π =πββ
π, β sos-convex
Relaxation 2:
sdsos-convex
min π‘
s.t. π»β π βΌ π‘πΌ
π =πββ
π, β sdsos-convex
Relaxation 3:
dsos-convex
min π‘
s.t. π»β π βΌ π‘πΌ
π =πββ
π, β dsos-convex
SDP
SOCP + βsmallβ SDP
LP + βsmallβ SDP
min π‘
s.t. π»β π βΌ π‘πΌ
π =πββ
π, β convex
Relaxation 4:
sdsos-convex+sdd
min π‘
s.t. ππ° β π―π π sdd (**)
π =πββ
π, β, sdsos-convex
SOCP
Relaxation 5:
dsos-convex + dd
min π‘
s.t. ππ° β π―π π dd (*)
π =πββ
π, β, dsos-convex
LP
β π is diagonally dominant (dd) β βπ πππ < πππ , βπ
ββ π is sdd β βπ· > 0 diagonal, s.t. π·ππ· dd.
12
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Convex relaxations for DC decompositions (5/6)
Can any polynomial be written as the difference of two dsos/sdsos/sos
convex polynomials?
Lemma about cones: Let πΎ β πΈ a full dimensional cone (πΈ, any vector
space). Then any π£ β πΈ can be written as π£ = π1 β π2 , π1 , π2 β πΎ.
=: πβ²
Proof sketch:
K
E
π
π
πβ²
β πΌ < 1 such that 1 β πΌ π£ + πΌπ β πΎ
1
πΌ
β²
βπ£=
π β
π
1βπΌ
1βπΌ
π1 β πΎ
π2 β πΎ
13
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Convex relaxations for DC decompositions (6/6)
Theorem: Any polynomial can be written as the difference of two dsosconvex polynomials.
Corollary: Same holds for sdsos-convex, sos-convex and convex.
Proof idea:
β’ Need to show that dsos-convex polynomials is full-dimensional cone.
β’ βObviousβ choices (i.e., π π₯ =
π/2
2
(βπ π₯π ) )
Induction on π: for π = 2, take
π π₯1 , π₯2 = π0 π₯1π + π1 π₯1πβ2 π₯22 + β― +
2 πβ2
π
π0 >
+
ππ
π(π β 1) 4(π β 1) 4
π1 = 1
do not work.
π π
ππ π₯12 π₯22
4
ππ+1
+ β― + π1 π₯12 π₯2πβ2 + π0 π₯2π
π β 2π
π
=
π , π = 1, β¦ , β 1
2π + 2 π
4
14
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Comparing the different relaxations (1/4)
β’ Impact of relaxations on solving
for random π (π = 4).
Type of relaxation
min π‘
π‘,π,β
s.t. π‘πΌ β π»β π psd/sdd/dd
π = π β β,
π, β s/d/sos-convex
π=π
π = ππ
π = ππ
Time (s)
Opt value
Time (s)
Opt Value
Time (s)
Opt value
dsos-convex + dd
1.05
17578.54
2.79
21191.55
20.80
168327.89
dsos-convex + psd
1.19
15855.77
3.19
19426.13
25.36
146847.73
sdsos-convex + sdd
1.21
1089.41
5.17
1962.64
34.66
7936.57
sdsos-convex + psd
1.21
1069.79
5.29
1957.03
39.43
7935.72
sos-convex + psd MOSEK
2.02
193.07
93.74
317.63
+β
------------------
sos-convex + psd SEDUMI
11.48
193.06
10324.12
317.63
+β
------------------
Computer:
8Gb RAM,
2.40GHz
processor
15
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Comparing the different relaxations (2/4)
β’ Iterative decomposition algorithm implemented for unconstrained π.
Decompose π = π β π,
using one of the
relaxations at point π₯π
Minimize convexified ππ , using an SDP
subroutine [Lasserre; de Klerk and
Laurent]
DSOS PSD SDSOS SDD SDSOS PSD
SOS PSD
0
DSOS DD
-50000
-100000
-150000
-200000
β’
β’
β’
β’
β’
β’
Value of the objective after 3 mins.
Algorithm given above.
5 different relaxations used
π random with π = 9, π = 4
Average over 25 iterations
Solver: Mosek
-250000
16
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Comparing the different relaxations (3/4)
β’ Constrained case: min π(π₯) , where π΅ = π₯ βπ π₯π2 β€ π
2 }.
π₯βπ΅
Single decomposition
vs
Iterative decomposition
vs
One min-max decomp.
Relaxation: min π‘
Decompose
s.t. π»β ππ =
βΌπ
π‘πΌβ π,
once
π
= πatβπ₯β0
π, β sdsos convex
Relaxation: min π‘
Decompose
s.t. π»β ππ =
βΌπ
π‘πΌβ π,
atπ a=point
πβπ₯
βπ
π, β sdsos convex
Decompose
π =to
π use?
βπ
What
relaxation
over B
Minimize convexified ππ
Minimize convexified ππ
Minimize convexified ππ
Second
relaxation:
Equivalent
formulation:
First relaxation:
Original
problem:
min
min
π‘π‘
π‘,π,β
min maxπ‘,π,β
ππππ₯π (π»β π₯π )
π,β
π₯βΞ©
ππ° π
β
π―
π
β½
π
π₯β π©
π΅ β ππ°
π‘πΌ βπΉπ―
π»βπβπ₯πβπβ½π 0ππ(π)
s.t.
= π β β,
πππ»ππ(π)π
= π β sos
β
π,π β=convex
πββ
β convex
π, ππ,sdsos-convex
π, β sdsos-convex
17
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Comparing the different relaxations (4/4)
β’ Constrained case: single decomposition vs. iterative decomposition
vs. min-max decomposition
4000
2000
0
-2000
-4000
-6000
-8000
-10000
-12000
Single
decomp
Iter decomp
Min max
β’
β’
β’
β’
Value of the objective after 3 mins.
Algorithms described above.
π random with π = 10, π = 4
Radius π
random integer between
100 and 400.
β’ Average over 200 iterations
-14000
-16000
18
DC Decomposition of Nonconvex Polynomials with Algebraic Techniques
Main messages
β’ To apply CCCP to polynomial optimization, a DC decomposition is
needed. Choice of decomposition impacts convergence speed.
β’ Not computationally tractable to find βbestβ decomposition.
β’ Efficient convex relaxations based on the concepts of dsos-convex
(LP), sdsos-convex (SOCP), and sos-convex (SDP) polynomials.
β’ Dsos-convex and sdsos-convex scale to a larger number of variables.
19
Thank you for listening
Questions?
20