7/23/2019 Helton Talk
1/27
Various Developments in Control
Bill Helton UCSD
Department of Mathematics
7/23/2019 Helton Talk
2/27
Numerical Optimization 2
Unconstrained: Gradient descent, Newton - nds a localoptimum efficiently.
Constrained: is harderOld idea is interior point method .Example subject to g(x ) 0
minX
f (x 1 , . . . , x g ) subject to g (x ) 0
Solution use a barrier function. Example
b (x ) := ln g (x )
7/23/2019 Helton Talk
3/27
Numerical Optimization 3
Solution use a barrier function.Example b (x ) := ln y y = g(x ).
y
Set f (x ) := f (x ) ln g (x )
Solve minx
f (x ) with a unconstrained algorithm. Take
short steps. Send 0
7/23/2019 Helton Talk
4/27
INTERIOR POINT METHOD 4
Set f (x ) := f (x ) ln g (x ). Solve minx
f (x ) withan unconstrained algorithm.
Rough History:
1. Solution 197? SOL paper said is VERY unreliable,
since f (x
) is badly conditioned if solution x
is onboundary, g(x ) = 0.
2. Kamarkar, Bell labs 1985ish - showed works on
many examples.
3. Nestrov-Nimerovski 1990ish introduced LMIs, andnumerical solution
7/23/2019 Helton Talk
5/27
LINEAR MATRIX INEQUALITIES 5
GIVEN a linear pencil L (x ) := A 0 + A 1 x 1 + + A g x gsymmetric matricesFIND numbers (or matrices) X := {X 1 , X 2 , , X g }
making L (X ) Pos SemiDef .SOLN: Interior Pt using barrier b (X ) := n det L (X )
APPLICATIONS:Combinatorics 1993ish I Do not know this area,Control - S. Boyd 1993ish. A Large crowd convertedcontrol problems to LMI (about 2 dozen are clean).Elimination Method, Change of Variables ( CarstenScherer ). Also lots of tricks with Schur Complements.Computational RAG appl to Lin and Nonlin Sys. 2000 Lasserre and Parrilo
7/23/2019 Helton Talk
6/27
LINEAR MATRIX INEQUALITIES 6
1. Many many uses of LMIs in compromises.
2. Vast vast vast literature.
3. Bilinear Matrix Inequalities BMI
(contains most systems problems)
XAY + Y A T X + BX + XB + Q P osDef.
People here who are experienced in converting:Linear Sys Probs LMIs
7/23/2019 Helton Talk
7/27
LIN MATRIX INEQS vs CONVEX 7
Stabilize two linear systems A, B j with a state feedbackcontroller, as in gain scheduling.
SOLN:Lyapunov inequality: Find X which simultaneouslysolves
XA + AT
X XB 1 BT 1 X + Q P osDef
XA + A T X XB 2 B T 2 X + Q P osDef
Can do with an Matrix Ineq package for 150 150matrices. Not possible with a Riccati equation solver.
Solution identify if model 1 or 2 applies; then use the
controller u = B T j Xx with B 1 or B 2 .
7/23/2019 Helton Talk
8/27
Sums of Squares 8
POSITIVE POLYNOMIALS vs SUMS of SQUARESOF POLYNOMIALS
CONSTRUCTING A LYAPUNOV FUNCTION
FROM LALL- PARRILO SLIDESIn Appendices
7/23/2019 Helton Talk
9/27
LIN MATRIX INEQS vs CONVEX 9
LINEAR MATRIX INEQUALITIES LMIsGIVEN a linear pencil L (x ) := A 0 + A 1 x 1 + + A g x gsymmetric matricesFIND matrices X := {X 1 , X 2 , , X g } making L (X )Pos SemiDef.
CONVEX MATRIX INEQUALITIES CMIs
QUESTION: How much more general are ConvexMatrix Inequalities than Linear Matrix Inequalities?
Bill believes there are two different situations with verydifferent behavior
Dimension Free , Dimension Dependent
7/23/2019 Helton Talk
10/27
Systems problems Matrix Ineq 10
Given
Find
Many such problems Eg. H control
Example: Get Riccati expressions like
A X + X A T X BB T X + CC T 0
OR Linear Matrix Inequalities (LMI) like
A X + X A T + C T C X BB T X I
0
which is equivalent to the Riccati inequality.ncDimlessPart
7/23/2019 Helton Talk
11/27
DIMENSIONLESS FORMULAS 11
1. DIMENSIONLESS FORMULAS THIS TALKTopology is xed; but many systems . E.g.
+ S 1 S 2
S 3WANT FORMULAS: which hold regardless of thedimension of system S 1 , S 2 , S 3 . Then unknowns arematrices and formulas respect matrix multiplication.Eg. Most classical control text problems:
Control pre 1990: Zhou,Doyle,Glover.LMIs in Control: Skelton, Iwasaki,. Grigoriadis.
Get noncommutative formulas
7/23/2019 Helton Talk
12/27
Keeping Matrices Whole 12
Matrices Whole
A X + X A T + C T C X B
BT
X I
0 (1)
Looks the same regardless of system size.
7/23/2019 Helton Talk
13/27
Keeping Matrices Whole 12
Matrices Whole
A X + X A T + C T C X B
BT
X I
0 (1)
Looks the same regardless of system size.
Matrices Entry by Entry DisaggregatedIf dimensions of the matrices A, B, C, X are specied, wecan write formula (1) with matrices L 0 , . . . , L m as
m
j =0
L j X j 0
with the unknown numbers X j taken as entries of X .
7/23/2019 Helton Talk
14/27
Disaggregated Matrices 13
Example: If A R 2 2 , B R 2 1 , C R 1 2 , then X S 2and we would take
X =X 1 X 2
X 2 X 3 and the LMI becomes
3
j =0L j X j 0
where the 4 4 symmetric matrices L 0 , L 1 , L 2 , L 3 are:
L 0 :=C T C 0
0 I L 1 :=
2 a 11 a 21 b 11 b 12a 21 0 0 0b 11 0 0 0
b 12 0 0 0
L 2 :=
2 a 12 a 11 + a 22 b 21 b 22a 22 + a 11 2 a 21 b 11 b 12
b 21 b 11 0 0b 22 b 12 0 0
L 3 :=
0 0 0 a 120 0 0 a 220 0 0 b 21
a 12 a 22 b 21 2 b 22
Down with vec
7/23/2019 Helton Talk
15/27
+ and of Keeping Matrices Whole 14
+ not many variables
+ short formulas
Trouble is formulas are noncommutative.
+ NCAlgebra package does symbolic noncommutativealgebra.
7/23/2019 Helton Talk
16/27
1. INTRO NC RAG - SEMIRAG 15
R eal Algebraic G eometry = RAGStudy of Polynomials on R g
RAG Semialgebraic Geometry = poly inequalities.
What about NonCommutative Polys?Part I. A FEW PHENOMENA
NC Positive Polynomial
NC Convex Polynomials
introNCPolyPart
7/23/2019 Helton Talk
17/27
1. POSITIVE NC POLYS 16
p 0 Means What ?
p is a symmetric polynomial in non-commutativevariables
x = {x 1 , x 2 } with real coefficients, eg.
p(x ) = x 1 2 + ( x 1 2 ) T + x 2 T x 2
Dene MATRIX POSITIVE polynomial
Plug in n n matrices X j for x j in p
always get
p(X 1 , X 2 ) is a PosSD n n matrix.
MATRIX POSITIVE POLYNOMIALS
7/23/2019 Helton Talk
18/27
MATRIX POSITIVE POLYNOMIALS-Examples 17
Example
x = {x 1 , x 2 }
p ( x ) = (3 x 21 + x 2 ) T (3 x 21 + x 2 ) +( x 1 5 ) T ( x 1 5 )
p (X ) = PosSD + PosSD
This is a sum of squares (SoS). More generally
p(
x ) =
c
j =1L j (
x ) T L j (
x )
is clearly always Matrix Positive.
NON COM POLYS ARE BETTER
7/23/2019 Helton Talk
19/27
NON-COM POLYS ARE BETTERBEHAVED THAN COMMUTATIVE
POLYS 18
THM:Matrix positive non-commutative
polynomials are sums of squares
Can compute SoS, thus CAN TEST Matrix Positivity.
Proof: Operator theory techniques
See 2001 H preprint.
S. McCullough 2000 preprint - when all variables x jare unitary.
7/23/2019 Helton Talk
20/27
19
NC CONVEXPOLYNOMIALS
7/23/2019 Helton Talk
21/27
20
Function p of noncommutative variables x := ( x 1 , x 2 ) is
MATRIX CONVEX (geometric def.) 0 1
p(X + (1 )
Y ) p (
X ) + (1 ) p(
Y )
12
p(X ) + 1
2 p(
Y ) p( 1
2X + 1
2Y ) is Pos Def ?
Question: Consider the noncommutative polynomial
p(x ) := x 4 + ( x 4 ) T .
Is it matrix convex?
CONVEX POLYNOMIALS ARE
7/23/2019 Helton Talk
22/27
CONVEX POLYNOMIALS AREWIMPS 21
The riddle revealed: p(x ) := x 4 + ( x 4) T is not matrix convex.
THM: ( McC + H)Every symmetric noncommutative polynomialwhich is matrix convex has degree 2 or less.
COR A Convex NC Polys is the Schur complement of somelinear pencil. Proof:
1. NC Positive Polynomials
2. NC Second Derivatives
3. Put the two together
7/23/2019 Helton Talk
23/27
Convexity Algorithm 22
In[1] :=
7/23/2019 Helton Talk
24/27
SUMMARY 23
1. THM EVERY POSITIVE NC POLYNOMIAL ISA SUM OF SQUARES. COMPUTABLE.
2. THM EVERY CONVEX NC POLYNOMIAL HASDEGREE 2. Equiv. to an LMI.
3. ALGORTHIM CHECKS CONVEXITY OF NC
RATIONAL FUNCTION. TRY NCAlgebra
ALGORITHMS and
7/23/2019 Helton Talk
25/27
ALGORITHMS andIMPLEMENTATIONS 24
1. Convexity Checker - Camino, Skelton, H Public
2. Realization Builder: Convex Rational to LMI -Slinglend, Shopple in progress
3. Numerical matrix unknowns - Camino, Skelton, H
in house
4. LMI Producer (uses existing methods on specialproblems) de Oliveira, H in house (out soon)
Try NCAlgebra
SUMMARY CONVEXITY LMI
7/23/2019 Helton Talk
26/27
SUMMARY: CONVEXITY vs LMI 25
CONJ: FOR DIM INDEP LINEAR SYS PROBS
CONVEX MIs are equivalent to LINEAR MI.
FOR DIM DEPENDENT LINEAR SYS PROBSNOT EQUIVALENT: There is an extra constraint,the line test
WEAK CONJ: the line test is necessary and sufficient
Proved in two dimensions Vinnikov + H
THERE IS NEW FIELDNC REAL ALGEBRAIC GEOMETRY.
THERE IS NC computer algebra software.
END 2006
7/23/2019 Helton Talk
27/27
END 2006 26
END 2006