a generalization of sage certificates for constrained optimization · 2020-02-06 ·...
TRANSCRIPT
![Page 1: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/1.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
A Generalization of SAGE Certificates forConstrained Optimization
Riley Murray
California Institute of Technology
July 11, 2019
Joint work with Venkat Chandrasekaran and Adam Wierman (Caltech).
arXiv:1907.00814 Riley Murray 1
![Page 2: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/2.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Nonnegativity and optimization
Given a function f : Rn → R and a set X ⊂ Rn, we have
f?X = inf{f(x) : x in X}
= sup{γ : f(x) ≥ γ for all x in X}
= sup{γ : f − γ is nonnegative over X}.
How do we test if f (or f − γ) is nonnegative over X?
This is usually NP-Hard, even for X = Rn.
Desire sufficient conditions that f is nonnegative over X.
Different sufficient conditions for different function classes.
arXiv:1907.00814 Riley Murray 2
![Page 3: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/3.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Signomials and polynomials
Parameters αi in Rn, ci in R.
x 7→m∑i=1
ci exp(αi · x)
Uncountable basis.
Complexity measured bynumber of terms m.
Nonnegativity certificates
SAGE
Parameters αi in Nn, ci in R.
x 7→m∑i=1
ci
n∏j=1
xαij
j
Countable basis.
Complexity measured by degreemaxij αij .
Select nonnegativity certificates
SOS
SONC
SAGE
arXiv:1907.00814 Riley Murray 3
![Page 4: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/4.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The signomial nonnegativity cones
Write f = Sig(α, c) to mean f(x) =∑m
i=1 ci exp(αi · x).
Define the nonnegativity cone for signomials over exponents α:
CNNS(α).= { c : Sig(α, c)(x) ≥ 0 for all x in Rn}.
These nonnegativity cones exhibit affine-invariance:
CNNS(α) = CNNS(α− 1u) = CNNS(αV )
for all row vectors u in Rn and all invertible V in Rn×n.
arXiv:1907.00814 Riley Murray 4
![Page 5: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/5.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The signomial nonnegativity cones
Write f = Sig(α, c) to mean f(x) =∑m
i=1 ci exp(αi · x).
Define the nonnegativity cone for signomials over exponents α:
CNNS(α).= { c : Sig(α, c)(x) ≥ 0 for all x in Rn}.
These nonnegativity cones exhibit affine-invariance:
CNNS(α) = CNNS(α− 1u) = CNNS(αV )
for all row vectors u in Rn and all invertible V in Rn×n.
arXiv:1907.00814 Riley Murray 4
![Page 6: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/6.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The signomial nonnegativity cones
Write f = Sig(α, c) to mean f(x) =∑m
i=1 ci exp(αi · x).
Define the nonnegativity cone for signomials over exponents α:
CNNS(α).= { c : Sig(α, c)(x) ≥ 0 for all x in Rn}.
These nonnegativity cones exhibit affine-invariance:
CNNS(α) = CNNS(α− 1u) = CNNS(αV )
for all row vectors u in Rn and all invertible V in Rn×n.
arXiv:1907.00814 Riley Murray 4
![Page 7: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/7.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
SAGE certificates for signomials
Definition. A nonnegative signomial with at most one negativecoefficient is an “AM/GM Exponential,” or an “AGE function.”
For each k, have cone of coefficients for AM/GM Exponentials
CAGE(α, k).= {c : c\k ≥ 0 and c in CNNS(α)}.
We take sums of AGE cones to obtain the SAGE cone
CSAGE(α) =m∑k=1
CAGE(α, k).
Crucial question: How to represent the AGE cones?
arXiv:1907.00814 Riley Murray 5
![Page 8: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/8.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
SAGE certificates for signomials
Definition. A nonnegative signomial with at most one negativecoefficient is an “AM/GM Exponential,” or an “AGE function.”
For each k, have cone of coefficients for AM/GM Exponentials
CAGE(α, k).= {c : c\k ≥ 0 and c in CNNS(α)}.
We take sums of AGE cones to obtain the SAGE cone
CSAGE(α) =m∑k=1
CAGE(α, k).
Crucial question: How to represent the AGE cones?
arXiv:1907.00814 Riley Murray 5
![Page 9: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/9.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
SAGE certificates for signomials
Definition. A nonnegative signomial with at most one negativecoefficient is an “AM/GM Exponential,” or an “AGE function.”
For each k, have cone of coefficients for AM/GM Exponentials
CAGE(α, k).= {c : c\k ≥ 0 and c in CNNS(α)}.
We take sums of AGE cones to obtain the SAGE cone
CSAGE(α) =
m∑k=1
CAGE(α, k).
Crucial question: How to represent the AGE cones?
arXiv:1907.00814 Riley Murray 5
![Page 10: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/10.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
SAGE certificates for signomials
Definition. A nonnegative signomial with at most one negativecoefficient is an “AM/GM Exponential,” or an “AGE function.”
For each k, have cone of coefficients for AM/GM Exponentials
CAGE(α, k).= {c : c\k ≥ 0 and c in CNNS(α)}.
We take sums of AGE cones to obtain the SAGE cone
CSAGE(α) =
m∑k=1
CAGE(α, k).
Crucial question: How to represent the AGE cones?
arXiv:1907.00814 Riley Murray 5
![Page 11: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/11.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The convex duality behind AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0.
Does c belong to CNNS(α)?
Appeal to affine invariance of CNNS(α), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
infx∈Rn
Sig(α\k − 1αk, c\k)(x) ≥ −ck
holds if and only if there exists ν in Rm−1 satisfying
D(ν, c\k)− νᵀ1 ≤ ck and [α\k − 1αk]ν = 0.
arXiv:1907.00814 Riley Murray 6
![Page 12: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/12.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The convex duality behind AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0.
Does c belong to CNNS(α)?
Appeal to affine invariance of CNNS(α), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
infx∈Rn
Sig(α\k − 1αk, c\k)(x) ≥ −ck
holds if and only if there exists ν in Rm−1 satisfying
D(ν, c\k)− νᵀ1 ≤ ck and [α\k − 1αk]ν = 0.
arXiv:1907.00814 Riley Murray 6
![Page 13: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/13.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The convex duality behind AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0.
Does c belong to CNNS(α)?
Appeal to affine invariance of CNNS(α), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
infx∈Rn
Sig(α\k − 1αk, c\k)(x) ≥ −ck
holds if and only if
there exists ν in Rm−1 satisfying
D(ν, c\k)− νᵀ1 ≤ ck and [α\k − 1αk]ν = 0.
arXiv:1907.00814 Riley Murray 6
![Page 14: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/14.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The convex duality behind AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0.
Does c belong to CNNS(α)?
Appeal to affine invariance of CNNS(α), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
infx∈Rn
Sig(α\k − 1αk, c\k)(x) ≥ −ck
holds if and only if there exists ν in Rm−1 satisfying
D(ν, c\k)− νᵀ1 ≤ ck and [α\k − 1αk]ν = 0.
arXiv:1907.00814 Riley Murray 6
![Page 15: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/15.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The convex duality behind AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0.
Does c belong to CNNS(α)?
Appeal to affine invariance of CNNS(α), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
infx∈Rn
Sig(α\k − 1αk, c\k)(x) ≥ −ck
holds if and only if there exists ν in Rm−1 satisfying
D(ν, c\k)− νᵀ1 ≤ ck and [α\k − 1αk]ν = 0.
arXiv:1907.00814 Riley Murray 6
![Page 16: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/16.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
SAGE certificates for signomials
The preceding slides established that
CAGE(α, k) = {c : c\k ≥ 0, and there exists ν in Rm−1 satisfying
D(ν, c\k)− νᵀ1 ≤ ck and [α\k − 1αk]ᵀν = 0}.
This allows us to optimize over
CSAGE(α) =
m∑k=1
CAGE(α, k)
with a convex program of size O(m2).
arXiv:1907.00814 Riley Murray 7
![Page 17: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/17.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Handling conditional nonnegativity
Want sufficient conditions that f is nonnegative over X.
Typical approach: search for a function L for which we can prove
0 ≤ L(x) for all x in Rn and L(x) ≤ f(x) for all x in X
Concretely, we fix a representation X = {x : g(x) ≥ 0}, and use
L = f −∑
i λigi
with dual variables λi : Rn → R+. Apply your favorite (tractable!)proof system to ensure that λi and L are nonnegative functions.
We will be taking a different route.
arXiv:1907.00814 Riley Murray 8
![Page 18: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/18.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Handling conditional nonnegativity
Want sufficient conditions that f is nonnegative over X.
Typical approach: search for a function L for which we can prove
0 ≤ L(x) for all x in Rn and L(x) ≤ f(x) for all x in X
Concretely, we fix a representation X = {x : g(x) ≥ 0}, and use
L = f −∑
i λigi
with dual variables λi : Rn → R+. Apply your favorite (tractable!)proof system to ensure that λi and L are nonnegative functions.
We will be taking a different route.
arXiv:1907.00814 Riley Murray 8
![Page 19: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/19.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Handling conditional nonnegativity
Want sufficient conditions that f is nonnegative over X.
Typical approach: search for a function L for which we can prove
0 ≤ L(x) for all x in Rn and L(x) ≤ f(x) for all x in X
Concretely, we fix a representation X = {x : g(x) ≥ 0}, and use
L = f −∑
i λigi
with dual variables λi : Rn → R+. Apply your favorite (tractable!)proof system to ensure that λi and L are nonnegative functions.
We will be taking a different route.
arXiv:1907.00814 Riley Murray 8
![Page 20: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/20.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE Signomials
arXiv:1907.00814 Riley Murray 9
![Page 21: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/21.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The signomial X-nonnegativity cones
Define the X-nonnegativity cone for signomials over exponents α:
CNNS(α, X).= { c : Sig(α, c)(x) ≥ 0 for all x in X}.
These nonnegativity cones exhibit affine-invariance:
CNNS(α, X) = CNNS(α− 1u, X) = CNNS(αV ,V−1X)
for all row vectors u in Rn, and all invertible V in Rn×n.
arXiv:1907.00814 Riley Murray 10
![Page 22: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/22.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The signomial X-nonnegativity cones
Define the X-nonnegativity cone for signomials over exponents α:
CNNS(α, X).= { c : Sig(α, c)(x) ≥ 0 for all x in X}.
These nonnegativity cones exhibit affine-invariance:
CNNS(α, X) = CNNS(α− 1u, X) = CNNS(αV ,V−1X)
for all row vectors u in Rn, and all invertible V in Rn×n.
arXiv:1907.00814 Riley Murray 10
![Page 23: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/23.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE certificates for signomials
Definition. A signomial which is nonnegative over X and which hasat most one negative coefficient is called an “X-AGE function.”
For each k, have cone of coefficients for X-AGE functions
CAGE(α, k,X).= {c : c\k ≥ 0 and c in CNNS(α, X)}.
We take sums of X-AGE cones to obtain the X-SAGE cone
CSAGE(α, X) =
m∑k=1
CAGE(α, k,X).
This definition assumes nothing of the set X!
arXiv:1907.00814 Riley Murray 11
![Page 24: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/24.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE certificates for signomials
Definition. A signomial which is nonnegative over X and which hasat most one negative coefficient is called an “X-AGE function.”
For each k, have cone of coefficients for X-AGE functions
CAGE(α, k,X).= {c : c\k ≥ 0 and c in CNNS(α, X)}.
We take sums of X-AGE cones to obtain the X-SAGE cone
CSAGE(α, X) =
m∑k=1
CAGE(α, k,X).
This definition assumes nothing of the set X!
arXiv:1907.00814 Riley Murray 11
![Page 25: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/25.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE certificates for signomials
Definition. A signomial which is nonnegative over X and which hasat most one negative coefficient is called an “X-AGE function.”
For each k, have cone of coefficients for X-AGE functions
CAGE(α, k,X).= {c : c\k ≥ 0 and c in CNNS(α, X)}.
We take sums of X-AGE cones to obtain the X-SAGE cone
CSAGE(α, X) =
m∑k=1
CAGE(α, k,X).
This definition assumes nothing of the set X!
arXiv:1907.00814 Riley Murray 11
![Page 26: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/26.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Basic properties
Conditional SAGE cones ...
are order-reversing in the set-argument.
If X2 ⊂ X1, then CSAGE(α, X1) ⊂ CSAGE(α, X2).
preserve sparsity.
c ∈ CSAGE(α, X)⇔ c ∈∑
i:ci<0CAGE(α, i,X).
respect boundedness of the set-argument.
sup{γ : f − γ is X-SAGE} > −∞ for all bounded X.
are tractable whenever X is a tractable convex set.
arXiv:1907.00814 Riley Murray 12
![Page 27: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/27.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Basic properties
Conditional SAGE cones ...
are order-reversing in the set-argument.
If X2 ⊂ X1, then CSAGE(α, X1) ⊂ CSAGE(α, X2).
preserve sparsity.
c ∈ CSAGE(α, X)⇔ c ∈∑
i:ci<0CAGE(α, i,X).
respect boundedness of the set-argument.
sup{γ : f − γ is X-SAGE} > −∞ for all bounded X.
are tractable whenever X is a tractable convex set.
arXiv:1907.00814 Riley Murray 12
![Page 28: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/28.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Basic properties
Conditional SAGE cones ...
are order-reversing in the set-argument.
If X2 ⊂ X1, then CSAGE(α, X1) ⊂ CSAGE(α, X2).
preserve sparsity.
c ∈ CSAGE(α, X)⇔ c ∈∑
i:ci<0CAGE(α, i,X).
respect boundedness of the set-argument.
sup{γ : f − γ is X-SAGE} > −∞ for all bounded X.
are tractable whenever X is a tractable convex set.
arXiv:1907.00814 Riley Murray 12
![Page 29: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/29.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Basic properties
Conditional SAGE cones ...
are order-reversing in the set-argument.
If X2 ⊂ X1, then CSAGE(α, X1) ⊂ CSAGE(α, X2).
preserve sparsity.
c ∈ CSAGE(α, X)⇔ c ∈∑
i:ci<0CAGE(α, i,X).
respect boundedness of the set-argument.
sup{γ : f − γ is X-SAGE} > −∞ for all bounded X.
are tractable whenever X is a tractable convex set.
arXiv:1907.00814 Riley Murray 12
![Page 30: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/30.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0. Convex X ⊂ Rn.
Appeal to affine invariance of CNNS(α, X), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
inf{Sig(α\k − 1αk, c\k)(x) : x in X
}≥ −ck
holds if and only if there exists ν in Rm−1, λ in Rn satisfying
σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck, and
[α\k − 1αk]ᵀν + λ = 0.
arXiv:1907.00814 Riley Murray 13
![Page 31: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/31.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0. Convex X ⊂ Rn.
Appeal to affine invariance of CNNS(α, X), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
inf{Sig(α\k − 1αk, c\k)(x) : x in X
}≥ −ck
holds if and only if there exists ν in Rm−1, λ in Rn satisfying
σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck, and
[α\k − 1αk]ᵀν + λ = 0.
arXiv:1907.00814 Riley Murray 13
![Page 32: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/32.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0. Convex X ⊂ Rn.
Appeal to affine invariance of CNNS(α, X), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
inf{Sig(α\k − 1αk, c\k)(x) : x in X
}≥ −ck
holds if and only if
there exists ν in Rm−1, λ in Rn satisfying
σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck, and
[α\k − 1αk]ᵀν + λ = 0.
arXiv:1907.00814 Riley Murray 13
![Page 33: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/33.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0. Convex X ⊂ Rn.
Appeal to affine invariance of CNNS(α, X), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
inf{Sig(α\k − 1αk, c\k)(x) : x in X
}≥ −ck
holds if and only if there exists ν in Rm−1, λ in Rn satisfying
σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck, and
[α\k − 1αk]ᵀν + λ = 0.
arXiv:1907.00814 Riley Murray 13
![Page 34: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/34.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
Fix α in Rm×n, and c in Rm satisfying c\k ≥ 0. Convex X ⊂ Rn.
Appeal to affine invariance of CNNS(α, X), and rearrange terms:
Sig(α, c)(x) ≥ 0 ⇔ Sig(α− 1αk, c)(x) ≥ 0
Sig(α\k − 1αk, c\k)(x) ≥ −ck.
Appeal to convex duality. The nonnegativity condition
inf{Sig(α\k − 1αk, c\k)(x) : x in X
}≥ −ck
holds if and only if there exists ν in Rm−1, λ in Rn satisfying
σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck, and
[α\k − 1αk]ᵀν + λ = 0.
arXiv:1907.00814 Riley Murray 13
![Page 35: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/35.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
The preceding slides established that when X is convex,
CAGE(α, k,X) = {c : c\k ≥ 0, and there exist ν in Rm−1, λ in Rn
satisfying [α\k − 1αk]ᵀν + λ = 0,
and σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck}.
For implementations, suppose X = {x : Ax+ b ∈ K}. Then
σX(λ).= supx∈X
λᵀx ≤ inf{ bᵀη : Aᵀη + λ = 0, η ∈ K†},
and equality holds generically.
arXiv:1907.00814 Riley Murray 14
![Page 36: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/36.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
The preceding slides established that when X is convex,
CAGE(α, k,X) = {c : c\k ≥ 0, and there exist ν in Rm−1, λ in Rn
satisfying [α\k − 1αk]ᵀν + λ = 0,
and σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck}.
For implementations, suppose X = {x : Ax+ b ∈ K}.
Then
σX(λ).= supx∈X
λᵀx ≤ inf{ bᵀη : Aᵀη + λ = 0, η ∈ K†},
and equality holds generically.
arXiv:1907.00814 Riley Murray 14
![Page 37: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/37.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing X-AGE cones
The preceding slides established that when X is convex,
CAGE(α, k,X) = {c : c\k ≥ 0, and there exist ν in Rm−1, λ in Rn
satisfying [α\k − 1αk]ᵀν + λ = 0,
and σX(λ) +D(ν, c\k)− νᵀ1 ≤ ck}.
For implementations, suppose X = {x : Ax+ b ∈ K}. Then
σX(λ).= supx∈X
λᵀx ≤ inf{ bᵀη : Aᵀη + λ = 0, η ∈ K†},
and equality holds generically.
arXiv:1907.00814 Riley Murray 14
![Page 38: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/38.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Dual SAGE relaxations
Let f = Sig(α, c) have α1 = 0.
The primal and dual SAGE relaxations for f?X are
fSAGEX = sup{ γ : c− γ(1, 0, . . . , 0) in CSAGE(α, X)}
= inf{cᵀv : v1 = 1 and v in CSAGE(α, X)†}.
When X is convex, the dual X-SAGE cone can be expressed as
CSAGE(α, X)† = cl{v : some z1, . . . ,zm in Rn satisfy
vk log(v/vk) ≥ [α− 1αk]zk
and zk/vk ∈ X for all k in [m]}.
Solution recovery? Consider vectors xk = zk/vk for k in [m].
arXiv:1907.00814 Riley Murray 15
![Page 39: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/39.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Dual SAGE relaxations
Let f = Sig(α, c) have α1 = 0.
The primal and dual SAGE relaxations for f?X are
fSAGEX = sup{ γ : c− γ(1, 0, . . . , 0) in CSAGE(α, X)}
= inf{cᵀv : v1 = 1 and v in CSAGE(α, X)†}.
When X is convex, the dual X-SAGE cone can be expressed as
CSAGE(α, X)† = cl{v : some z1, . . . ,zm in Rn satisfy
vk log(v/vk) ≥ [α− 1αk]zk
and zk/vk ∈ X for all k in [m]}.
Solution recovery? Consider vectors xk = zk/vk for k in [m].
arXiv:1907.00814 Riley Murray 15
![Page 40: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/40.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Dual SAGE relaxations
Let f = Sig(α, c) have α1 = 0.
The primal and dual SAGE relaxations for f?X are
fSAGEX = sup{ γ : c− γ(1, 0, . . . , 0) in CSAGE(α, X)}
= inf{cᵀv : v1 = 1 and v in CSAGE(α, X)†}.
When X is convex, the dual X-SAGE cone can be expressed as
CSAGE(α, X)† = cl{v : some z1, . . . ,zm in Rn satisfy
vk log(v/vk) ≥ [α− 1αk]zk
and zk/vk ∈ X for all k in [m]}.
Solution recovery? Consider vectors xk = zk/vk for k in [m].arXiv:1907.00814 Riley Murray 15
![Page 41: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/41.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
A Signomial Example
infx∈R3
f(x).= 0.5 exp(x1 − x2)− expx1 − 5 exp(−x2)
s.t. 100− exp(x2 − x3)− expx2 − 0.05 exp(x1 + x3) ≥ 0
expx− (70, 1, 0.5) ≥ 0
(150, 30, 21)− expx ≥ 0
Compute fSAGEX = −147.85713 ≤ f?X , recover feasible
x? = (5.01063529, 3.40119660,−0.48450710)
satisfying f(x?) = −147.66666. This is actually optimal!
arXiv:1907.00814 Riley Murray 16
![Page 42: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/42.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
A Signomial Example
infx∈R3
f(x).= 0.5 exp(x1 − x2)− expx1 − 5 exp(−x2)
s.t. 100− exp(x2 − x3)− expx2 − 0.05 exp(x1 + x3) ≥ 0
expx− (70, 1, 0.5) ≥ 0
(150, 30, 21)− expx ≥ 0
Compute fSAGEX = −147.85713 ≤ f?X , recover feasible
x? = (5.01063529, 3.40119660,−0.48450710)
satisfying f(x?) = −147.66666. This is actually optimal!
arXiv:1907.00814 Riley Murray 16
![Page 43: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/43.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE polynomials
arXiv:1907.00814 Riley Murray 17
![Page 44: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/44.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Geometric-form signomials
If x > 0, thenx 7→
∑mi=1 cix
αi
is defined for any real αi.
For X ⊂ Rn++ and α in Rm×n, define
CGEOMNNS (α, X) = {c :
∑mi=1 cix
αi ≥ 0 for all x in X}.
From a change of variables x 7→ expy, we have
CGEOMNNS (α, X) = CNNS(α, logX).
Thus we naturally define
CGEOMSAGE (α, X)
.= CSAGE(α, logX).
arXiv:1907.00814 Riley Murray 18
![Page 45: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/45.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Geometric-form signomials
If x > 0, thenx 7→
∑mi=1 cix
αi
is defined for any real αi.
For X ⊂ Rn++ and α in Rm×n, define
CGEOMNNS (α, X) = {c :
∑mi=1 cix
αi ≥ 0 for all x in X}.
From a change of variables x 7→ expy, we have
CGEOMNNS (α, X) = CNNS(α, logX).
Thus we naturally define
CGEOMSAGE (α, X)
.= CSAGE(α, logX).
arXiv:1907.00814 Riley Murray 18
![Page 46: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/46.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Geometric-form signomials
If x > 0, thenx 7→
∑mi=1 cix
αi
is defined for any real αi.
For X ⊂ Rn++ and α in Rm×n, define
CGEOMNNS (α, X) = {c :
∑mi=1 cix
αi ≥ 0 for all x in X}.
From a change of variables x 7→ expy, we have
CGEOMNNS (α, X) = CNNS(α, logX).
Thus we naturally define
CGEOMSAGE (α, X)
.= CSAGE(α, logX).
arXiv:1907.00814 Riley Murray 18
![Page 47: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/47.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The polynomial X-nonnegativity cones
Fix α in Nm×n. Write f = Poly(α, c) to mean
f(x) =
m∑i=1
cixαi , where xαi .=
n∏j=1
xαij
j .
The matrix α and the set X induce a nonnegativity cone
CNNP(α, X) = {c : Poly(α, c)(x) ≥ 0 for all x in X}.
arXiv:1907.00814 Riley Murray 19
![Page 48: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/48.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
The polynomial X-nonnegativity cones
Fix α in Nm×n. Write f = Poly(α, c) to mean
f(x) =
m∑i=1
cixαi , where xαi .=
n∏j=1
xαij
j .
The matrix α and the set X induce a nonnegativity cone
CNNP(α, X) = {c : Poly(α, c)(x) ≥ 0 for all x in X}.
arXiv:1907.00814 Riley Murray 19
![Page 49: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/49.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE certificates for polynomials
We call f = Poly(α, c) an X-AGE polynomial if
1 c belongs to CNNP(α, X), and
2 at most one “i” has cixαi < 0 for some x ∈ X.
In conic form, we can express CPOLYAGE (α, i,X) =
{c : Poly(α, c)(x) ≥ 0 for all x in X,
cj ≥ 0 if j 6= i and xαj > 0 for some x in X,
cj ≤ 0 if j 6= i and xαj < 0 for some x in X }.
Define the X-SAGE polynomial cone in the natural way:
CSAGE(α, X) =∑m
i=1CPOLYAGE (α, i,X).
arXiv:1907.00814 Riley Murray 20
![Page 50: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/50.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE certificates for polynomials
We call f = Poly(α, c) an X-AGE polynomial if
1 c belongs to CNNP(α, X), and
2 at most one “i” has cixαi < 0 for some x ∈ X.
In conic form, we can express CPOLYAGE (α, i,X) =
{c : Poly(α, c)(x) ≥ 0 for all x in X,
cj ≥ 0 if j 6= i and xαj > 0 for some x in X,
cj ≤ 0 if j 6= i and xαj < 0 for some x in X }.
Define the X-SAGE polynomial cone in the natural way:
CSAGE(α, X) =∑m
i=1CPOLYAGE (α, i,X).
arXiv:1907.00814 Riley Murray 20
![Page 51: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/51.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
X-SAGE certificates for polynomials
We call f = Poly(α, c) an X-AGE polynomial if
1 c belongs to CNNP(α, X), and
2 at most one “i” has cixαi < 0 for some x ∈ X.
In conic form, we can express CPOLYAGE (α, i,X) =
{c : Poly(α, c)(x) ≥ 0 for all x in X,
cj ≥ 0 if j 6= i and xαj > 0 for some x in X,
cj ≤ 0 if j 6= i and xαj < 0 for some x in X }.
Define the X-SAGE polynomial cone in the natural way:
CSAGE(α, X) =∑m
i=1CPOLYAGE (α, i,X).
arXiv:1907.00814 Riley Murray 20
![Page 52: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/52.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing CPOLYSAGE (α, X)
Consider the case when X is contained within a single orthant.
W.l.o.g, take X ⊂ Rn+.
If X ⊂ Rn+ is representable as
X = cl{x : 0 < x, H(x) ≤ 1}
for a continuous map H : Rn++ → Rr, then for
Y = {y : H(expy) ≤ 1},
we haveCPOLYSAGE (α, X) = CSAGE(α, Y ).
arXiv:1907.00814 Riley Murray 21
![Page 53: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/53.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing CPOLYSAGE (α, X)
Consider the case when X is contained within a single orthant.
W.l.o.g, take X ⊂ Rn+.
If X ⊂ Rn+ is representable as
X = cl{x : 0 < x, H(x) ≤ 1}
for a continuous map H : Rn++ → Rr,
then for
Y = {y : H(expy) ≤ 1},
we haveCPOLYSAGE (α, X) = CSAGE(α, Y ).
arXiv:1907.00814 Riley Murray 21
![Page 54: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/54.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing CPOLYSAGE (α, X)
Consider the case when X is contained within a single orthant.
W.l.o.g, take X ⊂ Rn+.
If X ⊂ Rn+ is representable as
X = cl{x : 0 < x, H(x) ≤ 1}
for a continuous map H : Rn++ → Rr, then for
Y = {y : H(expy) ≤ 1},
we haveCPOLYSAGE (α, X) = CSAGE(α, Y ).
arXiv:1907.00814 Riley Murray 21
![Page 55: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/55.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing CPOLYSAGE (α, X)
Define the set of signomial-representative coefficient vectors
SR(α, c) = {c : ci = ci whenever αi is in 2Nn, and
ci ≤ −|ci| whenever αi is not in 2Nn}.
If X admits the representation
X = cl{x : 0 < |x|, H(|x|) ≤ 1}
for a continuous map H : Rn++ → Rr, then for
Y = {y : H(expy) ≤ 1},
we have
CPOLYSAGE (α, X) = {c : SR(α, c) ∩ CSAGE(α, Y ) is nonempty }.
arXiv:1907.00814 Riley Murray 22
![Page 56: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/56.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing CPOLYSAGE (α, X)
Define the set of signomial-representative coefficient vectors
SR(α, c) = {c : ci = ci whenever αi is in 2Nn, and
ci ≤ −|ci| whenever αi is not in 2Nn}.
If X admits the representation
X = cl{x : 0 < |x|, H(|x|) ≤ 1}
for a continuous map H : Rn++ → Rr,
then for
Y = {y : H(expy) ≤ 1},
we have
CPOLYSAGE (α, X) = {c : SR(α, c) ∩ CSAGE(α, Y ) is nonempty }.
arXiv:1907.00814 Riley Murray 22
![Page 57: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/57.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Representing CPOLYSAGE (α, X)
Define the set of signomial-representative coefficient vectors
SR(α, c) = {c : ci = ci whenever αi is in 2Nn, and
ci ≤ −|ci| whenever αi is not in 2Nn}.
If X admits the representation
X = cl{x : 0 < |x|, H(|x|) ≤ 1}
for a continuous map H : Rn++ → Rr, then for
Y = {y : H(expy) ≤ 1},
we have
CPOLYSAGE (α, X) = {c : SR(α, c) ∩ CSAGE(α, Y ) is nonempty }.
arXiv:1907.00814 Riley Murray 22
![Page 58: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/58.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Working with CPOLYSAGE (α, X)
How can we formulate a problem to appeal to previous theorems?
Examples include
−a ≤ xj ≤ a, ‖x‖p ≤ a, |xαi | ≥ a, and x2j = a
where a > 0 is a fixed constant.
arXiv:1907.00814 Riley Murray 23
![Page 59: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/59.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
A polynomial example
Minimize
f(x) = −647∑i=1
∏j∈[7]\{i}
xj
over the box X = [−1/2, 1/2]7.
Symbolically, we have
fi(x).= 1− 64
∏j 6=i xi are X-AGE, hence fSAGE
X ≥ −7,
and f(1/2) = f(−1/2) = −7.
For numeric computation (with MOSEK, on a large workstation)
the X-SAGE relaxation takes 0.01 seconds to solve
the earliest tight SOS relaxation takes 90 seconds to solve.
arXiv:1907.00814 Riley Murray 24
![Page 60: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/60.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
A polynomial example
Minimize
f(x) = −647∑i=1
∏j∈[7]\{i}
xj
over the box X = [−1/2, 1/2]7.
Symbolically, we have
fi(x).= 1− 64
∏j 6=i xi are X-AGE, hence fSAGE
X ≥ −7,
and f(1/2) = f(−1/2) = −7.
For numeric computation (with MOSEK, on a large workstation)
the X-SAGE relaxation takes 0.01 seconds to solve
the earliest tight SOS relaxation takes 90 seconds to solve.
arXiv:1907.00814 Riley Murray 24
![Page 61: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/61.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
A polynomial example
Minimize
f(x) = −647∑i=1
∏j∈[7]\{i}
xj
over the box X = [−1/2, 1/2]7.
Symbolically, we have
fi(x).= 1− 64
∏j 6=i xi are X-AGE, hence fSAGE
X ≥ −7,
and f(1/2) = f(−1/2) = −7.
For numeric computation (with MOSEK, on a large workstation)
the X-SAGE relaxation takes 0.01 seconds to solve
the earliest tight SOS relaxation takes 90 seconds to solve.
arXiv:1907.00814 Riley Murray 24
![Page 62: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/62.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 63: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/63.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 64: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/64.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 65: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/65.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 66: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/66.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 67: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/67.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 68: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/68.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25
![Page 69: A Generalization of SAGE Certificates for Constrained Optimization · 2020-02-06 · arXiv:1907.00814 Riley Murray 2. Background X-SAGE signomials X-SAGE polynomialsDiscussion Signomials](https://reader035.vdocuments.mx/reader035/viewer/2022070916/5fb6304fd5993b2c2b2156c9/html5/thumbnails/69.jpg)
Background X-SAGE signomials X-SAGE polynomials Discussion
Discussion
The content of this presentation was a small fraction of
Signomial and Polynomial Optimization via Relative Entropy andPartial Dualization
Refer to that article for ...
Mixed X-SAGE and Lagrangian relaxations.
Hierarchies for signomial and polynomial optimization.
Twelve explicit examples, and 55+ problems in total.
Detailed solution recovery algorithms.
Check out sageopt, the “Gloptipoly3” of SAGE relaxations.
Source: https://github.com/rileyjmurray/sageopt/
User site: https://rileyjmurray.github.io/sageopt/
arXiv:1907.00814 Riley Murray 25