0% found this document useful (0 votes)
2 views

04 - Convex Optimization Problems

Uploaded by

mahe.velay
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
2 views

04 - Convex Optimization Problems

Uploaded by

mahe.velay
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 33

MGT-418 Convex Optimization:

Convex Optimization Problems


Daniel Kuhn
([email protected])

77
78

Optimization Problems in Standard Form

P: minimize f0 (x)

subject to fi (x)  0 8i = 1, . . . , m
<latexit sha1_base64="mFyRpzqHZQzYQ7r39j/hql0Aa54=">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</latexit>
hi (x) = 0 8i = 1, . . . , p

x 2 Rn optimization or decision variable


f0 : Rn ! ( 1, +1] objective or cost function
fi : Rn ! ( 1, +1] inequality constraint functions
<latexit sha1_base64="QzHwapusKgF3kIHRjl3/WQEsWrU=">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</latexit>
hi : R n ! R equality constraint functions

inf P = +1 () P is infeasible
inf P = 1
<latexit sha1_base64="OHO6CXEQhc9EUiFClMpi/olOTjc=">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</latexit>
() P is unbounded
79

Transforming Problems to the Standard Form

• “ <latexit sha1_base64="sak57wurpVJvxHb1PzCGdWiFDzs=">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</latexit>
”-inequalites:
f̂(x)
<latexit sha1_base64="F4eTQFYgg86DJXDay7rpeiHNm1Y=">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</latexit>
0 () f(x) = f̂(x)  0

• nonzero right hand sides:


f̂(x)  b () f(x) = f̂(x)
<latexit sha1_base64="nVk8MWWPWm5DvRdMaRyXYKsWe4A=">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</latexit>
b0

minimizingcost
• maximization problems: Maximizing profit

sup f̂0 (x) = inf f0 (x) where f0 (x) = f̂0 (x)


<latexit sha1_base64="gs2q2kdi9WYS+6bpHTut1haOosA=">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</latexit>
x2X x2X

Note: In principle, all equality constraints can be transformed to


inequality constraints:
f(x) = 0 () f(x)  0 and
<latexit sha1_base64="YwwNyHwrpmfujbOw+Vqe6D2BOJA=">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</latexit>
f(x)  0
80

Convex Optimization Problems


P: minimize f0 (x)

subject to fi (x)  0 8i = 1, . . . , m
<latexit sha1_base64="u1tWqqdQJkyWZ507Qb/x9OLdilA=">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</latexit>
a> i x = bi 8i = 1, . . . , p

The objective and inequality constraint functions are convex,


while all equality constraint functions are linear (affine). The
equality constraints can be written compactly as Ax = b. <latexit sha1_base64="Ae+vmritZq121t0rNPcx/mrSov0=">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</latexit>

Proposition: The feasible set of a convex optimization problem is


convex.
EEE
Proof: The feasible set of P is given by
Tex
p
dom(f0 ) \m
ËE.ae
>
f 0} a i x = b}.
ËËÏÎËÏÏ i=1 {x : i (x)  \ i=1 {x :
amour
FÉE
<latexit sha1_base64="O2DkIZbZkzJZCDS8F5MqqnsDd6c=">AAAEAnicbZPLbtNAFIadhEspl7awYMFmSoSUSFVkIxAXqVIQSLBgESRCK8WpGY+P01HmYmbGxZHlJRtehRViy4vwGLwBM06oGqdjeeb4//65nnGcMaqN7/9ptTtXrl67vnVj++at23d2dvfuftIyVwTGRDKpjmOsgVEBY0MNg+NMAeYxg6N4/trxozNQmkrx0SwymHI8EzSlBBsrRXut76GAr0RyjkVShmlVhrFkiV5w21Tba5ATSzk2pwRvoHiF4rhBcmZJLhJQbokNKM8qV13KRqAySzPbhPv148KGKWY5WFN9DmX90TAoSM65ixt4pgDEuWH5ZS0JpHZBS5EvViY1s5v0B8EB8gdPXRVYp4HClInkVS+N/H5IcBaV9DCoTnhYFi/TiPaKfsjgC/LD6gLNkMM4oiehkRkqDuOwGkS7XX/g1wVtBsEq6HqrMrLZ+xwmkuQchCEMaz0J/MxMS6wMJcwdaK4hw2SOZzCxocAc9LSs91WhR1ZJUCqVfYVBtXqxRxnztRFKzLW7F7any7RuMidexia5SZ9PSyqy3IAgy4nTnCEjkbuSKKEKiGELG2CiqF07IqdYYWLsxV2bpR47A7K+riIXlMgEGiozhVHYihoMx1S4XZavFK0vb63ZwZzYe0Nn1OiD9/afEAdvbbbn/f9Gm5OgmYHNYPx48GLgf3jSHQ5XydnyHngPvZ4XeM+8offOG3ljj7T+tu+3UXu/863zo/Oz82tpbbdWfe55a6Xz+x+dfF34</latexit>

It is convex as the domains and sublevel sets of convex functions


are convex, while hyperplanes are also convex. Intersections of
convex sets are convex.
81

Local Optima are Global Optima

Theorem: Every locally optimal solution of a convex optimization


problem is also globally optimal.

Proof: Suppose that x is not a global minimizer, and let y be a


global minimizer. By convexity, the decision θx + (1 θ)y is
feasible for every θ 2 [0, 1] , and its objective value satisfies
<latexit sha1_base64="4zty8m0s7l2USV/BlUAtUm+IwaY=">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</latexit>

<latexit sha1_base64="mEHxGjDmaSGFSlHVjud7kS3yF1A=">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</latexit>

f0 (θx + (1 θ)y)  θf0 (x) + (1 θ)f0 (y)


costofanypointontheline ix y thanfood < θf0 (x) + (1 θ)f0 (x)
<latexit sha1_base64="z+hQNs7eBYZZuqc4mdCg5HBuIkk=">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</latexit>
= f0 (x)
si ewhnwepproacytkliethcostfxisasavgt.pe
for all θ 2 (0, 1) . Therefore, x cannot be a local minimizer. By
contraposition, if x is a local minimizer, then it must be a global I
<latexit sha1_base64="fq0V1b0/svUusNKckt76fK55OOw=">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</latexit>

minimizer.
vrai
82

Optimality Criterion (Differentiable Objective)


optimizationprosien

Theorem: For P convex, x? is optimal iff it is feasible and <latexit sha1_base64="WPBVO4I8eBQm9/yojDaQbQ5DId4=">AAADvHicbVJdb9MwFHUWPkb52uCRl0CFNKSqSiYQ8EQRk+CBhyJRNqkpw3FuMqu2E9lOaRX1F/FreAT+DNdpmGg6R7GPzzm2r31vUgpubBj+8vb8a9dv3Ny/1bt95+69+weHD76YotIMJqwQhT5LqAHBFUwstwLOSg1UJgJOk/k7p58uQBteqM92VcJM0lzxjDNqkTo/9E5iBd9ZISVVaR1n6zpOCpGalcRh3dsSJUNVUnvB6I6UtFKSdJRKoFKpFLQLsSMWi7XrrtTGoEtUSxzix83nYMeUiArQ1LxD3Uw6Bg3ppe5wR841gLo0bGZoSSHDgDakXLUmneMlw2E0CMLhC9dF697ya2ws1ecH/XAYNi3YBVEL+qRtY3z3b3FasEqCskxQY6ZRWNpZTbXlTLinqAyUlM1pDlOEikows7qJaB08RSYNskLjr2zQsP+vqBO5tUNNpXEZxZUuR6arOfIqbVrZ7NWs5qqsLCi2OTirRGCLwBVTkHINzIoVAso0x9gDdkE1ZRZLbuuUZu8S2HZcy0pxVqTQYYVdWk2RNGAl5crdsn6reVN2DYebOfLohOfcmsFHrGY1eI95mj/7Z8ScRN0M7ILJ8fD1MPz0vD8atcnZJ4/IE3JEIvKSjMgHMiYTwrwf3k/vt/fHf+ODP/flxrrntWsekq3mL/4CozdFDQ==</latexit>

1ˢᵗaideredtin
rf0 (x? )> (x
<latexit sha1_base64="pXgHrb1O9xWfpMWrGaA+JjxLj/0=">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</latexit>
x? ) 0 for all feasible x. (⇤)2ndhandout
<latexit sha1_base64="1Xo++P6RblaxANjTcMRF/TPtIa8=">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</latexit>

costal 0 then α 90

Proof: Assume that (⇤) holds for some feasible x? and let x be <latexit sha1_base64="1Xo++P6RblaxANjTcMRF/TPtIa8=">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</latexit>
<latexit sha1_base64="WPBVO4I8eBQm9/yojDaQbQ5DId4=">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</latexit>

any feasible point. The 1st-order conditions for f0 imply <latexit sha1_base64="1W9V2WMMXfDiHk3JYbpGmDbTuEM=">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</latexit>

f0 (x)
<latexit sha1_base64="KJxpPmhCK9WR9MqyOE9uRgookJc=">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</latexit>
f0 (x? ) + rf0 (x? )> (x x? ) 8x 2 Rn .
since A 1ˢᵗ ordertaylor
sioundynay
Using (⇤), we conclude that f0 (x) f0 (x ) . Thus, x? is optimal.
?
<latexit sha1_base64="WPBVO4I8eBQm9/yojDaQbQ5DId4=">AAADvHicbVJdb9MwFHUWPkb52uCRl0CFNKSqSiYQ8EQRk+CBhyJRNqkpw3FuMqu2E9lOaRX1F/FreAT+DNdpmGg6R7GPzzm2r31vUgpubBj+8vb8a9dv3Ny/1bt95+69+weHD76YotIMJqwQhT5LqAHBFUwstwLOSg1UJgJOk/k7p58uQBteqM92VcJM0lzxjDNqkTo/9E5iBd9ZISVVaR1n6zpOCpGalcRh3dsSJUNVUnvB6I6UtFKSdJRKoFKpFLQLsSMWi7XrrtTGoEtUSxzix83nYMeUiArQ1LxD3Uw6Bg3ppe5wR841gLo0bGZoSSHDgDakXLUmneMlw2E0CMLhC9dF697ya2ws1ecH/XAYNi3YBVEL+qRtY3z3b3FasEqCskxQY6ZRWNpZTbXlTLinqAyUlM1pDlOEikows7qJaB08RSYNskLjr2zQsP+vqBO5tUNNpXEZxZUuR6arOfIqbVrZ7NWs5qqsLCi2OTirRGCLwBVTkHINzIoVAso0x9gDdkE1ZRZLbuuUZu8S2HZcy0pxVqTQYYVdWk2RNGAl5crdsn6reVN2DYebOfLohOfcmsFHrGY1eI95mj/7Z8ScRN0M7ILJ8fD1MPz0vD8atcnZJ4/IE3JEIvKSjMgHMiYTwrwf3k/vt/fHf+ODP/flxrrntWsekq3mL/4CozdFDQ==</latexit>

me
Conversely, assume that x? is optimal and let x be any feasible
<latexit sha1_base64="1Xo++P6RblaxANjTcMRF/TPtIa8=">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</latexit> <latexit sha1_base64="O4tqb6+tXbnKZNihxMBo3aEhXno=">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</latexit>

<latexit sha1_base64="WPBVO4I8eBQm9/yojDaQbQ5DId4=">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</latexit>

point. By convexity, the line segment [x? , x] is contained in the


feasible set. By optimality of x?, the function
<latexit sha1_base64="DYsl6DErFQj4eUmjEYCMkn+RZQY=">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</latexit>

linemoves from away


<latexit sha1_base64="WPBVO4I8eBQm9/yojDaQbQ5DId4=">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</latexit>

? VIH 4101
ψ(t) = f0 ((1 t)x + tx) so
sin4101 fixa
<latexit sha1_base64="b0VMm+B0zfT7WL2xviWaGvIv6vU=">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</latexit>

must satisfy ψ(t) <latexit sha1_base64="oBsllNX3KTsVmGdIBcc1OYaY0fk=">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</latexit>


ψ(0) 8t 2 [0, 1]. This is only possible if
ψ 0 (0) = rf0 (x? )> (x x? ) 0. ï the
dérivante should be so at
the
<latexit sha1_base64="U1/TejCRjm6PkKJdY/kIR98i++M=">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</latexit>
83

Optimality Criterion (Differentiable Objective)


Mein if we can go out ofthe feusable

f0 (x ) x? is optimal if rf0 (x? )


<latexit sha1_base64="WPBVO4I8eBQm9/yojDaQbQ5DId4=">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</latexit>

has an angle of more


<latexit sha1_base64="ppPjCNZHq7EdKtmeJfDaU2ypta4=">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</latexit>

x x than 90° with (x x? ) <latexit sha1_base64="T9lvPjIa/yBwEHxllMkqOtfxMN8=">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</latexit>

for every feasible x.

Special case:
Unconstrained problem:
<latexit sha1_base64="Kx7jE+jYghPFla0c0LjWRqZ9IWI=">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</latexit>
x? optimal () x? 2 dom(f0 ), rf0 (x? ) = 0
84

Equivalent Optimization Problems


Two problems P and P’ are equivalent (written as P () P’ ) if <latexit sha1_base64="dN+IemGp5YGR+AvplWT19xv4VDI=">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</latexit>

the solution of P’ is obtained from that of P via “elementary


transformations” and vice versa.

Epigraphical reformulation: The standard form convex


optimization problem is equivalent to ce minimizing
font
minimizing t
t
P’ : minimize (over x, t) t since foin
subject to f0 (x) t  0
fi (x)  0 8i = 1, . . . , m
<latexit sha1_base64="wuZ9YWgNL193hwMV4Yu6XgrMYbo=">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</latexit>
Ax = b
(i.e., one can always assume that the objective is linear.)
85

Equivalent Optimization Problems

Partial minimization: The problem

P: minimize (over x1 , x2 ) f0 (x1 , x2 )


subject to fi (x1 )  0 8i = 1, . . . , m
gi (x2 )  0 8i = 1, . . . , q
<latexit sha1_base64="fJtz8N4RQrTL+3l/FkzgI65K3Uo=">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</latexit>

l t
fif fiscal
is equivalent to see in exercise that
j fix Et
P’: minimize (over x1 ) f̂0 (x1 )
<latexit sha1_base64="5zJ1TrH0g0YwcFGK8SceoN79Wds=">AAAEZnicbVNdb9MwFM26AqN8bAMhHngxbIxOqqoEgfiQkIZAggceikTZpLkqjnPTmdlOZjujJcob/Eh+Ai/8Bq7TMtFsjmJfn3OufX2vHedSWBeGv1Zaq+1Ll6+sXe1cu37j5vrG5q3PNisMhyHPZGYOYmZBCg1DJ5yEg9wAU7GE/fj4jef3T8FYkelPbpbDSLGJFqngzCE03lz5QzV845lSTCclTauSxplM7EzhUHWWSMWRVcwdcXaOihdUHDeYQiJT6ASMD7FBZqeV7y7kBmByZHMc6P3682ZDFMsCUFTnoawnDYGB5Iz3doOeGAB9JpjPUJJAigHNQTVbiMwEDxn2ox4J+099F6EyhonQJTOGzapSSp8XB1NXDh69rHZoImwu2cy6mQRS5yfLy7lACS2U+A6k689PtqfjaHu3qsgOoUfMkXQcki5iu4TSzs7cxRbxV+COuMzL0rGoBVTCCQnpScESQns0zQyTkohXUY/KJHO2pyjgUechdsYbW2E/rBs5b0QLYytYtAHekC80yXihQDsumbWHUZi7Ea7nBJe+aIWFnPFjNoFDNDVTYEdlnbuKPEQkIRgT/tqRGv3fo4zV0golU9bfPfT02bJNzoMXcYeFS5+PSqHzwoHm843TQmKuiL/2JBEGUydnaDBuBMZO+BEzjDt8HEu71GvnwJfjmhZa8CyBBird1BmGoAWnmND+lOVrI+oHUmO4mAe7b8VEYDE+4LvTvXd4o453/wmxJlGzAueN4eP+i3748cnW3t6iOGvBveBB0A2i4FmwF7wPBsEw4K1hq2z9aP1c/d1eb99p351LWysLn9vBUmuTv72mgAw=</latexit>
subject to fi (x1 )  0 8i = 1, . . . , m

where

f̂0 (x1 ) = inf {f0 (x1 , x2 ) : gi (x2 )  0 8i = 1, . . . , q}.


<latexit sha1_base64="XWWTD/BEvnbr+n4ETlfupj2vcQ8=">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</latexit>
x2
86

Equivalent Optimization Problems

Example: The following problems are equivalent.

x1 P: minimize x21 + x22


1+x22
subject to x1 /(1 + x22 )  0
<latexit sha1_base64="jz0BaneW6smLIf+Wao4PI5+bBTo=">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</latexit>
(x1 + x2 )2 = 0
nFiearfunctia so Pnot
a optimization

P’: minimize x21 + x22


x2 x1
subject to x1  0
convex optimization
<latexit sha1_base64="PlNZHQxBgIYD9hsfT1onD1I5cwE=">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</latexit>
x1 + x2 = 0
problem
However, P’ is a convex optimization problem according to our
definition, but P is not. Indeed, P involves a nonconvex inequality
constraint function and a nonlinear equality constraint function.
87

Linear Program (LP)


minimize c> x + d
subject to Ax = b
<latexit sha1_base64="FhdLWAox5djmdzawE/fIgC006oQ=">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</latexit>
Cx  g

c
x Affine objective and
constraints, polyhedral
feasible set.

Problems with n variables and encoded in L input bits can be


solved in O(n3.5 L) arithmetic operations via interior point methods.
<latexit sha1_base64="VHQVAjRsm8samEFNjDZfse2QFqI=">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</latexit>
88

Transportation Problem
ask.fi o
Pm Pn
minimize i=1 j=1 cij xij transportation costs
Pn
subject to j=1 xij = ai 8i total shipment from source i
Pm
i=1 xij = bj 8j total shipment to destination j
xij 0 8i, j nonnegative shipments only
E
<latexit sha1_base64="Y3CTNdPNoh53VQM9+sPpXjalDAg=">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</latexit>

a1 a2 a3 ....... am
Problem feasible if total
supply = total demand: <latexit sha1_base64="M9of3Cb96VOv83YzoH93jlijU8s=">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</latexit>
<latexit

cmn m n
c11 X X
ai = bj
<latexit sha1_base64="Ttwvh1Ri665bbphKH+3EPQo4Rac=">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</latexit>
i=1 j=1

b1 ....... bn
89

Project Scheduling

Project: directed acyclic graph G = (V, E) <latexit sha1_base64="cHjJ7ZnREt34zFmSBJQfXKsQsdc=">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</latexit>

• tasks V = {1, . . . , n}
<latexit sha1_base64="JVlr3HzcGCeD/9Dk7orwj8pLtYg=">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</latexit>

2 3
• precedences E ✓ V ⇥ V <latexit sha1_base64="o9SqDVH47UuXHFc/iIUyrTkLees=">AAADynicbVJdb9MwFHUWPkb52uCRl0CFNKSqShAIeCsCBA8gFUS7SU01HOcms2o7wXbGqqhv/Cl+Ck+8wr/gOg0TTeco9vE5x/a1701KwY0Nw5/ejn/p8pWru9d612/cvHV7b//O1BSVZjBhhSj0UUINCK5gYrkVcFRqoDIRcJgsXjn98BS04YX6bJclzCXNFc84oxap433vU6zgGyukpCqt42xVx0khUrOUOKx6G6JkqEpqTxjdkpJWSpKOUglUKpWCdiF2xOJ05boLtTHoEtUSh/h+8znYMSWiAjQ171A3k45BQ3quO9yRcw2gzg3rGVpSyDCgNSmXrUnneMlwGA2CcPjUddGq9yY2VWLAwtdgGlsuwQTT3vFePxyGTQu2QdSCPmnbGLPwJU4LVklQlglqzCwKSzuvqbacCfcwlYGSsgXNYYZQUTxoXjfxrYKHyKRBVmj8lQ0a9v8VdSI3dqipNC6/uNJlzHQ1R16kzSqbPZ/XXJWVBcXWB2eVCGwRuNIKUq6BWbFEQJnmGHvATqimzGIBbpzS7F0C24zrrFKcFSl0WGHPrKZI4jtLypW7Zf1S86YIGw43c+TBa55zawbvsbbV4C1mbfHonxFzEnUzsA0mj4cvhuHHJ/3RqE3OLrlHHpADEpFnZETekTGZEOb98H55v70//gff+Eu/Xlt3vHbNXbLR/O9/AWzBSi4=</latexit>

• task durations ζ v , v 2 V <latexit sha1_base64="CmHFotEI9m0xeYMINxkVYT6aHhY=">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</latexit>


1 6
4 5
Early Start Policy:
• find task start times via an LP

minn sn + ζ n
s2R+

<latexit sha1_base64="KBDzhUlx2d8j2FHoxY1MvfHm/Mk=">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</latexit>
s.t. sv su + ζ u 8(u, v) 2 E
90

Chebyshev Center of a Polyhedron

The Chebyshev center of a polyhedron


<latexit sha1_base64="MqS+ML9SpzxjJO2qU+pA2au9LGo=">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</latexit>
P = {x : a>
i x  bi , i = 1, . . . , m}

is the center of the largest ball inside P <latexit sha1_base64="BvgCD+P1gtDYDpIlsjswSaHzXKk=">AAADunicbVJbb9MwFHYWLmPcNnjkJVAhDamqEgQCHpDKRYIHHopE2aS2Go5zkln1JbKdsSrqD+LX8Mr+DcdpmEg6R7E/f99n+9jnpKXg1sXxRbATXrt+4+burb3bd+7eu79/8OC71ZVhMGVaaHOcUguCK5g67gQclwaoTAUcpcsPXj86A2O5Vt/cqoSFpIXiOWfUIXVyELyfK/jJtJRUZfU8X9fzVIvMriQO672OKBmqkrpTRrektJXStKdUApVKZWB8iD1Rn619d6U2AVOiWuIwf9x8HvZMqagATc071M2kZzCQXeoe9+TCAKhLw2aGlgxyDGhDylVrMgVeMh4lwygevfRdgk7JosnJ/iAexU2LtkHSggFp2wRf/cc806ySoBwT1NpZEpduUVPjOBP+ISoLJWVLWsAMoaIS7KJu4llHT5HJolwb/JWLGvb/FXUqOzvUVFqfT1zpM2T7miev0maVy18vaq7KyoFim4PzSkROR76UoowbYE6sEFBmOMYesVNqKHNYcJ1Tmr1LYN24zivFmc6gxwp37gxF0oKTlCt/y/qd4U3RNRxu5snDj7zgzg6/YC2r4SfM0vLZPyPmJOlnYBtMn4/ejOKvLwbjcZucXfKIPCGHJCGvyJh8JhMyJSz4FfwO/gQX4duQhTxcbqw7QbvmIem00P0FnJ5DqQ==</latexit>


<latexit

maximize r
xc ,r
subject to
<latexit sha1_base64="vVOizli43ZHL7v87qXmZmUpj4xs=">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</latexit>
a>
i x  bi 8x 2 B(xc , r), i = 1 . . . , m

() a>
i (xc + u)  bi 8u : kuk2  r
() max a> i (xc + u)  bi
kuk2 r
()
<latexit sha1_base64="Fh2jCo9Fp+VGB50OU+C+8wenvyo=">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</latexit>
a>
i xc + rkai k2  bi
91

Quadratic Program (QP)


1 >
minimize 2 x Px + q> x + r
subject to Ax = b
<latexit sha1_base64="5Z9ZvfstKLjh5YYdmuQv2Cc1GKk=">AAAEMHicbVNNb9QwEE13+SjLVwtHDhgqqiKqVVKBgANSq4LggGCRWFqpWYrjTFJT20ltp2SJ8qu48Fd6Qlz5FYyzoWLTOoo9ee+NPZ6ZRLngxvr+yUKvf+HipcuLVwZXr12/cXNp+dYnkxWawZhlItO7ETUguIKx5VbAbq6BykjATnS47fidY9CGZ+qjneYwkTRVPOGMWoT2lxd+hgq+sUxKquIqTOoqjDIRm6nEpR7MkZIhK6k9YPQMFbVUFHWYQiBTqBi0C7FDZse1m87lRqBzZHNcwnvN48yOKBIFoKjJQ9V8dAQa4lPe2R061QDqVDD7QkkMCQY0A+W0FekUL+kPg3XiD5+4KUBlBClXFdWaTutKuLRYKG0lueKSf4earJIw0ZRVQV1t1KT8HNosJ6OSPDqameUjHYatlymir8AssZnz2yLliygMV8l2GQo4IukgBIx5dtZgsL+04g/9ZpCzRtAaK147RljrL2GcsUKCskxQY/YCP7cT3NByJlz6CwM5ZYc0hT00FZVgJlWThZo8QCQmSabxVZY06P8eVSTndqioNK6L0NP1helyDjyP2yts8mxScZUXFhSbHZwUApNCXAOTmGvMkZiiQZnmGDthBxRTbLHN505p9s6BzcdVFoqzLIYOKmxpNUXQgJWUK3fLakvzptUbDDdz4NpLnnJr1t/iH6TWX2NvHD78J8SaBN0KnDXGG8PnQ//D45XNzbY4i94d77635gXeU2/Te+ONvLHHend7r3rveu/7P/on/V/93zNpb6H1ue3Njf6fv6rlbwg=</latexit>
Cx  g

Convex quadratic objective ( P 2 Sn+) and affine constraints, i.e.,


a polyhedral feasible set.
<latexit sha1_base64="4K+Hnulu6e5JTz1wIk4NlfUeAow=">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</latexit>

Example: Least-squares problem:

minimize
<latexit sha1_base64="WJ/vNQyI+2S+p6RgUzb3Jej6Ecg=">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</latexit>
kAx bk22

If A has full column rank, we have

0 = rkAx bk22 = 2A> Ax 2A> b


=) x = (A> A) 1 A> b.
<latexit sha1_base64="UIJJdBl4l605VLxrDnHMImqVww8=">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</latexit>
92

Quadratically Constrained QP (QCQP)


1 > >
minimize 2 x P0 x + q 0 x + r0
1 > >
subject to
<latexit sha1_base64="6diLtb9gy7TEsdtq2e/MY7RGwbQ=">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</latexit>
2 x Pi x + q i x + ri  0 i = 1, . . . , m

Convex quadratic objective and constraints ( Pi 2 Sn+), i.e., the


feasible set is an intersection of ellipsoids (for Pi 0 ) and
<latexit sha1_base64="re4u77qFZc/aRNGFKyvBKVC1TTs=">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</latexit>

<latexit sha1_base64="P9zM34RsENbh/jMWFP4tsHRNflY=">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</latexit>

halfspaces (for Pi = 0). Every LP is a QCQP with Pi = 0 . <latexit sha1_base64="hpA3yPduGEF6LvlkLRwdiFoZ/+w=">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</latexit>


<latexit sha1_base64="hpA3yPduGEF6LvlkLRwdiFoZ/+w=">AAADvHicbVJdb9MwFHUWPkb52uCRl0CFNKSqSiYQ8IAoYhI88FAkyia1VXGcm86qPyLbKaui/iJ+DY/An+E6DRNN5yj28TnH9rXvTQvBrYvjX8FeeO36jZv7tzq379y9d//g8MFXq0vDYMS00OYspRYEVzBy3Ak4KwxQmQo4TRfvvX66BGO5Vl/cqoCppHPFc86oQ2p2GJxMFHxnWkqqsmqSr6tJqkVmVxKHdWdLlAxVSd05oztS2khp2lJKgUqpMjA+xJaol2vfXakNwRSoFjhMHtefhy1TKkpAU/0OVT1pGQxkl7rHLXluANSlYTNDSwY5BrQh5aoxmTleMu4nvSjuv/Bdsu4MZ/xNFHdmB924H9ct2gVJA7qkaUN892+TTLNSgnJMUGvHSVy4aUWN40z4pygtFJQt6BzGCBWVYKdVHdE6eopMFuXa4K9cVLP/r6hSubVDRaX1GcWVPke2rXnyKm1cuvzVtOKqKB0otjk4L0XkdOSLKcq4AebECgFlhmPsETunhjKHJbd1Sr13AWw7rotScaYzaLHCXThDkbTgJOXK37J6Z3hddjWHm3ny6ITPubO9T1jNqvcB87R49s+IOUnaGdgFo+P+6378+Xl3MGiSs08ekSfkiCTkJRmQj2RIRoQFP4Kfwe/gT/g2hHARyo11L2jWPCRbLVz+BRCdQ9A=</latexit>

Example: Distance between ellipsoids:


x2
x1
<latexit sha1_base64="pOKIQ3svE6c71d6cmGwm9PW84eE=">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</latexit>

<latexit sha1_base64="xhTrkX48bj0rhUMrk1LDA0VQIHA=">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</latexit>

minimize kx1 x2 k22


subject to (x1 μ 1 )> Σ1 1 (x1 μ1 )  1 constrain
laver quadratic
<latexit sha1_base64="yyyc2b2e/AENkf2XPZRwtzA/SGs=">AAAEb3icbVNNb9QwEE23C5Tlq4ULEhIyVKBWaldJBAJuRSDBgUMRLEWqd4PjzKamthNsp+yS5sYV/h9/gwtX7CStumkdxR6/98Yej8dxzpk2vv9nqbfcv3T5ysrVwbXrN27eWl27/UlnhaIwohnP1OeYaOBMwsgww+FzroCImMNefPjK8XtHoDTL5Eczz2EsSCrZlFFiLBStLf3DEr7TTAgikxJPqxLHGU/0XNihGiyQglpWEHNAyTkqbqk47jAFt0whE1AuxA6ZHVWuu5DbBZVbNrcDflB/zuyIYl6AFdV5KOtJR6AgOeWd3aFTBSBPBc3MShKY2oAaUMxbkUrtIf1hsIX84VPXBVYZQ8pkSZQi86rkLi0GZqYUTDLBfkCFHiN8PIuC7VkU4uMonIQYtxpdxF+BGmQyp9pwIiyKKNicYJPlCH9gqSBRMCm3g+oMizl8QwHGjU9Yo+GiT3jic8I2PgMM9sxNrINodd0f+nVD542gNda9tu3aUvmCk4wWAqShnGi9H/i5Gdv1DKPc3V6hISf0kKSwb01JBOhxWSexQo8skqBppuwvDarRsx5lLBZWKInQrgitpysr3eUceBG3X5jp83HJZF4YkLTZeFpwm2Xk6h8lTNmk87k1CFXMxo7oAVGEGvtKFnap186BLsY1KySjWQIdlJuZUcSCGowgTLpTli8Vq19KjdnFHLjxmqXM6K139gHKrTe2tA43T4T2ToLuDZw3RuHwxdB//2R9Z6e9nBXvnvfQ2/AC75m34731dr2RR3uT3s/er97v5b/9u/37fdRIe0utzx1vofU3/wO2i4Eo</latexit>
(x2 μ 2 )> Σ2 1 (x2 μ2 )  1
93

Second-Order Cone Program (SOCP)


secondordercone K ly.tl E R 114112Et

minimize f> x
subject to kAi x + bi k2  c>
i x + di i = 1, . . . , m
Fx = g
<latexit sha1_base64="sYwFgUpy0DHNAB0Qb3kETPXTx1I=">AAAEQ3icbVNNb9QwEE13+SjLVwtHLoaKqqirVbICAYdKrUDAgUORWFqpWYLtTFJT20kdp+yS7t/jwi/gP3BCXJEYZ0PFpnUUe/Lem7E9M2G5FIX1/R9Lne6ly1euLl/rXb9x89btldU7H4qsNBxGPJOZ2We0ACk0jKywEvZzA1QxCXvs6IXj907AFCLT7+00h7GiqRaJ4NQiFK0ufQ81fOGZUlTHVZjMqpBlMi6mCpdZb4FUHFlF7SGn5yjWUIy1mFIiU+oYjDtii8xOZm66kNsFkyOb4xLerx9ntkRMloCiOg9V/dESGIjPeGe36NQA6DPB/AslMSR4oDmopo3IpHhJfxD0iT944qYAlQxSoStqDJ3OKunSYmFiKyW0UOIrzMg6ST6GNsvJJAwbsijZZ+CW2MzR4elOJCabLBLhaTQkoYRjwiPROG3GkSDhcUljIraCfijjzBZ9hbHWyavJVtoLAe8x378Xraz5A78e5LwRNMaa14xdLP+nMM54qUBbLmlRHAR+bscYzwouXUXKAnLKj2gKB2hqqqAYV3ViZuQhIjFJMoOvtqRG//eomFqIUFFVuMZCT9cqRZtz4EXcQWmTZ+NK6Ly0oPl846SUmEDieprEwmA+5RQNyo3AsxN+SA3lFjt/YZc6dg588VyTUguexdBCpZ1YQxEswCoqtLtltWNE3f01hsEcuPFSpALL8hZ/Kt1/je1y9OifEGsStCtw3hgNB88H/rvHa9vbTXGWvXveA2/DC7yn3rb3xtv1Rh7vDDv7Hdph3W/dn91f3d9zaWep8bnrLYzun7+MC3aG</latexit>

LE Ier
(f 2 Rn , Ai 2 Rni ⇥n , bi 2 Rni , ci 2 Rn , di 2 R, F 2 Rp⇥n , g 2 Rp )
<latexit sha1_base64="eDXn8zq9KuAFmWwvK5G8elwZ74U=">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</latexit>

The inequalities are called second-order cone constraints as


the (ni + 1) -dimensional vector (Ai x + bi , c>
i x + di ) belongs to
the second-order cone in Rni +1 .
<latexit sha1_base64="6nrLpsNLHIWocOzEJNQMZ2D1Aew=">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</latexit>

<latexit sha1_base64="8AxU3GXzPlWVkxbKyOLAW6LdKAU=">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</latexit>

<latexit sha1_base64="xxwQSCZa0DdMCNCsHt7ioqdz/UA=">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</latexit>

Every LP is an SOCP with Ai = 0, bi = 0. Every SOCP with <latexit sha1_base64="ubSLwnW9I4lzXfyuhgifGCdK7wA=">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</latexit>

ci = 0 is a QCQP, and every QCQP is an SOCP.


<latexit sha1_base64="9sJ8SI0YYwkYswF+CsT4XiSzNdo=">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</latexit>

Pm
An ε-optimal solution can be found in O(( i=1 ni )3.5 log(ε 1
))
arithmetic operations via interior point methods.
<latexit sha1_base64="xTiT7moo2hY/kRwv5sYi0xrplII=">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</latexit>
EI R optimal when 11111112 4
s E 11 1 11 et can't write E on a computer
94

Grasping Force Optimization


Consider a rigid body held by N robotic fingers.
• Center of gravity at 0
• Contact points p1 , . . . , pN 2 R3
• Contact forces F1 , . . . , FN 2 R3
<latexit sha1_base64="Gv54QO/0pBki81UiUj7QtlPMZ4w=">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</latexit>

• inward pointing unit normal vectors v1 , . . . , vN 2 R3


<latexit sha1_base64="/SCnZmDyf0fEyyLRhGRwpGMcGiY=">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</latexit>

<latexit sha1_base64="HYT+G/mbhUpLHUtLZ6VyOG/rgxQ=">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</latexit>

Friction-cone constraints:
F1
<latexit sha1_base64="pwc+MN9R3K1oXZ10emLcmMrCgac=">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</latexit>
F3 <latexit sha1_base64="rRAcqWLErTzfhTDOxwqocrV7wpU=">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</latexit>

F2 v3
<latexit sha1_base64="wmj6MH2OPODzcuL3QrEhQS4iaF8=">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</latexit>
(I vi v>
i )Fi 2
 μ · v>
i Fi
p3
<latexit sha1_base64="cDNZF5AgOiGDhxdAHr6R2+6okl8=">AAADt3icbVLLjtMwFHUmPIbymoElm0CFNEhVlfAQj1URSLBgUSTKjNRWxXFuWqt+BNspU0X9Hr6GLYi/4TqNRjQdR7GPzzm2r31vWghuXRz/DQ7CK1evXT+80bl56/adu0fH975aXRoGI6aFNmcptSC4gpHjTsBZYYDKVMBpunzn9dMVGMu1+uLWBUwlnSuec0YdUrPjYDBR8INpKanKqkm+qSapFpldSxw2nR1RMlQldQtG96S0kdK0pZQClVJlYHyILVGvNr67VBuCKVAtcJg8rD8PW6ZUlICm+h2qetIyGMgudI9b8twAqAvDdoaWDHIMaEvKdWMyc7xk3E96Udx/4btk01nNns2OunE/rlu0D5IGdEnThvjms0mmWSlBOSaoteMkLty0osZxJvwzlBYKypZ0DmOEikqw06qOZhM9RiaLcm3wVy6q2f9XVKnc2aGi0vps4kqfH9vWPHmZNi5d/mpacVWUDhTbHpyXInI68oUUZdwAc2KNgDLDMfaILaihzGG57ZxS710A243rvFSc6QxarHDnzlAkLThJufK3rN4aXpdczeFmnjx5z+fc2d4nrGTV+4A5Wj5pjJiSpJ2AfTB62n/djz8/7w4GTW4OyQPyiJyQhLwkA/KRDMmIsOBn8Cv4HfwJ34TfwjxcbK0HQbPmPtlp4fd/5LtC8Q==</latexit>

<latexit sha1_base64="R6az3zyvsbwa8nDbYGRqR6UzF0Q=">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</latexit>

<latexit sha1_base64="Sn+gnsZwl2sUdeMkkzkfCgO0QDQ=">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</latexit>
p1 v1
<latexit sha1_base64="JPWRArtdbnbBiXsc/AN0ETf16pA=">AAADt3icbVLLjtMwFHUmPIbymoElm0CFNEhVlYxAPFZFIMGCRZEoM1JbFce5aa36EWynTBX1e/gatiD+hus0GtF0HMU+PufYvva9aSG4dXH8NzgIr12/cfPwVuf2nbv37h8dP/hqdWkYjJgW2pyn1ILgCkaOOwHnhQEqUwFn6fKd189WYCzX6otbFzCVdK54zhl1SM2Og8FEwQ+mpaQqqyb5ppqkWmR2LXHYdHZEyVCV1C0Y3ZPSRkrTllIKVEqVgfEhtkS92vjuSm0IpkC1wGHyuP48bJlSUQKa6neo6knLYCC71D1uyXMDoC4N2xlaMsgxoC0p143JzPGScT/pRXH/he+STWc1S2ZH3bgf1y3aB0kDuqRpQ3zz2STTrJSgHBPU2nESF25aUeM4E/4ZSgsFZUs6hzFCRSXYaVVHs4meIpNFuTb4KxfV7P8rqlTu7FBRaX02caXPj21rnrxKG5cufzWtuCpKB4ptD85LETkd+UKKMm6AObFGQJnhGHvEFtRQ5rDcdk6p9y6A7cZ1USrOdAYtVrgLZyiSFpykXPlbVm8Nr0uu5nAzT56853PubO8TVrLqfcAcLZ81RkxJ0k7APhid9l/348/Pu4NBk5tD8og8ISckIS/JgHwkQzIiLPgZ/Ap+B3/CN+G3MA8XW+tB0Kx5SHZa+P0f4JlC7w==</latexit>

v2 <latexit sha1_base64="sULLNcz1lJeEG/IswhEjn7hYQD0=">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</latexit>

p2 tangential friction normal


<latexit sha1_base64="c5D3S9ASDE+ZcwAeLpF5VRoJm5A=">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</latexit>

<latexit sha1_base64="Isg6w1F2u2006W8MAV7T3KvSSjU=">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</latexit>

force coefficient force

Static equilibrium constraints:


PN
i=1 Fi = Fext external force
<latexit sha1_base64="JBgz2igHvQdHcPZUHaZX4y20c9c=">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</latexit>

PN
<latexit sha1_base64="cUTPnHnO4J1I9xIm14ng4cFy0DM=">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</latexit>
i=1 pi ⇥ Fi = Text external torque
95

Grasping Force Optimization


Consider a rigid body held by N robotic fingers.
• Center of gravity at 0
• Contact points p1 , . . . , pN 2 R3
• Contact forces F1 , . . . , FN 2 R3
<latexit sha1_base64="Gv54QO/0pBki81UiUj7QtlPMZ4w=">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</latexit>

• inward pointing unit normal vectors v1 , . . . , vN 2 R3


<latexit sha1_base64="/SCnZmDyf0fEyyLRhGRwpGMcGiY=">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</latexit>

<latexit sha1_base64="HYT+G/mbhUpLHUtLZ6VyOG/rgxQ=">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</latexit>

F1 F3
<latexit sha1_base64="pwc+MN9R3K1oXZ10emLcmMrCgac=">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</latexit> <latexit sha1_base64="rRAcqWLErTzfhTDOxwqocrV7wpU=">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</latexit>

Minimize maximal normal force:


F2 v3
<latexit sha1_base64="wmj6MH2OPODzcuL3QrEhQS4iaF8=">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</latexit>

p3
<latexit sha1_base64="cDNZF5AgOiGDhxdAHr6R2+6okl8=">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</latexit>

<latexit sha1_base64="Sn+gnsZwl2sUdeMkkzkfCgO0QDQ=">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</latexit>
p1 v1
<latexit sha1_base64="JPWRArtdbnbBiXsc/AN0ETf16pA=">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</latexit>

v2 <latexit sha1_base64="sULLNcz1lJeEG/IswhEjn7hYQD0=">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</latexit>

minimize t
p2 t,Fi
<latexit sha1_base64="c5D3S9ASDE+ZcwAeLpF5VRoJm5A=">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</latexit>

subject to v>
i Fi  t 8i
<latexit sha1_base64="Isg6w1F2u2006W8MAV7T3KvSSjU=">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</latexit>

(I vi v>i )Fi 2
 μ · v>
i Fi 8i
PN
i=1 Fi = Fext
PN
<latexit sha1_base64="APh4RXMbMbg+3Nbpu//cGzVfcGE=">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</latexit>
i=1 pi ⇥ Fi = Text
96

“Hidden” SOCP Constraints


Quadratic constraints:

✓ ◆
2x
kxk22  t () t+1
<latexit sha1_base64="p/Jv8AS/LeVxYEJ6hGIHEkZUtpA=">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</latexit>
t 1 2

Hyperbolic constraints:

✓ ◆
2x
kxk22  st, s 0, t 0 ()  s + t, s 0
<latexit sha1_base64="bN96uQaV+x7A7hLiv+JNnG8gLGs=">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</latexit>
s t 2
11 11 Et 70
tt
it so

es

1 1 111 11 1 implicit fromformula


can'tbe negative
MI
t1 1
97

Problems with Generalized Inequalities


Standard form convex optimization problem with generalized
inequalities:

P: minimize f0 (x)
subject to fi (x) Ki 0 8i = 1, . . . , m
<latexit sha1_base64="SAsFhB35dZtuhPHiMG6l0ZvgGow=">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</latexit>
Ax = b

Assumptions:
• f0 : Rn ! R convex, fi : Rn ! Rki Ki -convex
<latexit sha1_base64="vOUk5DGJIKmzCv/8vE7tB5B3Qvg=">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</latexit>
<latexit sha1_base64="sMbYMv4i9RNMTwBJ9OjiIUckXQ4=">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</latexit>

• Ki ✓ Rki is a proper convex cone


<latexit sha1_base64="flCR1nBqWtuzH9NNb+HuyOCKfLM=">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</latexit>

Observations:
• P has a convex feasible set
• All local optima of P are global optima
98

Conic Form Problems


Special case where objective and constraints are affine in x.
P: minimize c> x
subject to Fx + g K 0
<latexit sha1_base64="ihxfKLJb03A0YPit7tT2t2PpVpg=">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</latexit>
Ax = b
The conic form problem reduces to an LP if K = Rm
+.

Note also that the SOCP


minimize f> x
subject to kAi x + bi k2  c>
i x + di i = 1, . . . , m
Fx = g <latexit sha1_base64="f8PIR5/StJua9ltA8pZnlkpI2Rc=">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</latexit>

is equivalent to
minimize f> x
subject to (Ai x + bi , c>
i x + di ) Ki 0 i = 1, . . . , m
<latexit sha1_base64="JFUf2iRmORj+gpDjMwvLvOUoDD0=">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</latexit>
Fx = g
where Ki = {(y, t) 2 Rni +1 : kyk2  t}. <latexit sha1_base64="FFD9MNv9D5+BgbqInPiVD+d+ZSc=">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</latexit>
99

Every Convex Problem is Conic


Theorem: Every convex optimization problem can be reformu-
lated as a conic form problem.
Proof idea: Consider an arbitrary convex problem inf x2C f0 (x) , <latexit sha1_base64="fBlVpqpC55jZRSLMTNgxZKM5Vto=">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</latexit>

where C is a convex feasible set. This problem is equivalent to


inf{t : f0 (x)  t, x 2 C}
() inf{t : (t, x) 2 B} B = {(t, x) : f0 (x)  t, x 2 C}
()
<latexit sha1_base64="tAVhVuzmgKZk5/k7htKAF4TRu5g=">AAAEk3icdZNda9RAFIbTuGqtX62CN70ZLZYthDURxQ8s1LagUIUKri10lnUyOUmHTibpzKTuEnLjr/FW/43/xpNkLW5aZ9nk5DzvvPNxZsJcCmN9//eCe6V39dr1xRtLN2/dvnN3eeXeF5MVmsOQZzLThyEzIIWCoRVWwmGugaWhhIPwZKfmB2egjcjUZzvNYZSyRIlYcGYxNV5xV6mCbzxLU6aiksZVScNMRmaa4qtamoMpR5oye8zZBRTOUBh2SCGRFCoCXU+xA7Ozqn5cyvZB50hzfNGHza8OO6JQFoCiZh/K5qMj0BCd8zru4EQDqHNB+4WSCGKcUJtMpzORTnCR/iDwiD94Xj8CVIaQCFUyrdm0KqXEfVknVKiYlvZ1PPb7kw0q4ZRYj3pkgoDsEFpRipo4JuvkXNu3HkqRbyNHsL1Jyyb3H5dLTcxmgIq+8RozU3AOp+M94reWe7Vly9rhnphmwNrU0AT9fVoNlijgxrQLGi+v+QO/aeRiEMyCNWfW9scrC19plPEiBWW5ZMYcBX5uR2hnBZd1hQsDOeMnLIEjDBVLwYzKZqMr8hgzEYkzjX9lSZP9t0cZpnMOJUtNfVCxZ330TJfVycvYUWHjl6NSqLywoHg7cFxIYjNS3xESCQ3cyikGjGuBcyf8mGnGLd6kuVEa7xz4/LwmhRI8i6CTlXZiNcOkAZsyrBmusnyrRXObmhya1cn+rkiENd4HvKTKe4fH72TjrxBrEnQrcDEYPh28Gvifnq1tbc2Ks+isOo+cvhM4L5wt572z7wwd7n53f7g/3V+9B703ve3ebit1F2Z97jtzrffxD6G+h5s=</latexit>
inf{t : s = 1, (s, t, x) ⌫K 0} K = {(s, t, x) : (t, x)/s 2 B, s 0}.
B
K
s s Note: K is
<latexit sha1_base64="LgEoDRFL0+mnjItdlIbJeC6uRik=">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</latexit>
<latexit

D D

1
B
1 a convex
cone (why?)
BC C

t
E
t
E

x
F xF
100

Semidefinite Program (SDP)


minimize c> x
subject to F1 x1 + · · · + Fn xn G
<latexit sha1_base64="KxiheYvt2aTf/SZHXgl8xzKIyh8=">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</latexit>
Ax = b

(c 2 Rn , F1 , . . . , Fn , G 2 Sk , A 2 Rm⇥n , b 2 Rm )
<latexit sha1_base64="TUfo08jzYjiPmC9W/PI2cA2g6xY=">AAAEAnicbVPLjtMwFE1bHkN5zcAGiU1mKqSOFFUJAgG7jkADCxblUWakplM5zk3Hqu1EtjNMFbU7VnwKK8SWH+E3+AKu06ii6TiKfXPOsX3tcxNlnGnj+38azda16zdu7txq375z99793b0HX3SaKwpDmvJUnUZEA2cShoYZDqeZAiIiDifR7LXlTy5AaZbKz2aewViQqWQJo8QgNNlrfA8lfKWpEETGRZgsijBKeaznAodFe4MUFFlBzDklW1RUUVFUY3KOTC5jUDbFGpleLGx3JTcAlSGb4RDul48Na6KI54Ci8h6K8qMmUBCveRvX6KkCkGvB6gslMSSY0AoU80qkpnhIvxd4rt97brtg0e7SkEk8vfvxTHrL40nghTxOjfaOJ9J7W3Gfzmbe8mgtLERomADtyoW3jNawOGxPdjt+zy+bux0EVdBxqjZA+yZhnNJcgDSUE61HgZ+ZcUGUYZTbG801ZITOyBRGGEqC+46L8mAL9wkisZukCl9p3BL9f0YRiY0VCiK0LQycaa3Wdc6CV3Gj3CQvxwWTWW5A0tXGSc5dk7q2Jt2YKaCGzzEgVDHM3aXnRBFqsHI3dinXzoBu5nWZS0bTGGooN5dGEQQ1GEGYtKcsjhQrq7fEcDELdt+wKUPT3uNPgbah3bPDSoiWBHUDtoPh096rnv/hWaffr7zZcR47B07XCZwXTt955wycoUMbf5uPmvvNg9a31o/Wz9avlbTZqOY8dDZa6/c/iAJc3w==</latexit>

The semidefinite constraint is a linear matrix inequality (LMI). Note


that several LMIs can be combined to a single LMI:
F̃1 x1 + · · · + F̃n xn
<latexit sha1_base64="MbDpX5mLkw4yRAqsGA22vkVL2rE=">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</latexit>
G̃ and F̂1 x1 + · · · + F̂n xn Ĝ
is equivalent to
✓ ◆ ✓ ◆ !
F̃1 0 F̃n 0 G̃ 0
x1 + · · · + xn .
0 F̂1 0 F̂n 0 Ĝ
<latexit sha1_base64="d/+XKnbiqVw92qOnU8XVUirPLNE=">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</latexit>

An ε-optimal solution can be found in O(n2 k2.5 log(ε <latexit sha1_base64="kmNpFQYoOXXO6UafZGfiD3zdlH0=">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</latexit>


1
)) arithmetic
operations via interior point methods.
101

Every SOCP is an SDP

The SOCP
minimize f> x
subject to kAi x + bi k2  c>
i x + di i = 1, . . . , m
<latexit sha1_base64="f8PIR5/StJua9ltA8pZnlkpI2Rc=">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</latexit>
Fx = g
is equivalent to an SDP as by Schur’s lemma

Arrow matrix:
kAi x + bi k2  c>
i x + di
✓ > >

ci x + di (Ai x + bi )
() > ⌫0
<latexit sha1_base64="q/SQAw85fwikNkQOhJCH1z0C6LU=">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</latexit>
Ai x + bi (ci x + di )I
102

Every LP is an SDP
We have already seen that every LP is an SOCP and every
SOCP is an SDP. Thus, every LP is an SDP.

Direct argument: The LP

minimize c> x + d
subject to Ax = b
<latexit sha1_base64="eGn5TAddK41VKczx4UIMFscNEX4=">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</latexit>
Cx  g

is equivalent to the SDP

minimize c> x + d
0 1
diag(b Ax)
subject to @ diag(Ax b) A ⌫ 0.
<latexit sha1_base64="ratAg4GNgh343ZZli57ULizxKww=">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</latexit>
diag(g Cx)
103

Eigenvalue Minimization
Minimize the largest eigenvalue of a matrix, i.e., solve
minimize λmax (A(x))
<latexit sha1_base64="dwhzt08s6YCcr4SP/yzTSB11dVY=">AAAD3XicbVLLjtMwFE0mPIbymoElm0CF1EpVlYxAwK4jkGDBokiUGampiuPcdKzaTmQ7Q0sUlqwQW76KD+A72ILEdRqNaDqOYh+fc2xf+94450ybIPjl7nlXrl67vn+jc/PW7Tt3Dw7vfdBZoShMaMYzdRoTDZxJmBhmOJzmCoiIOZzEy5dWPzkHpVkm35t1DjNBFpKljBKD1PzQTSMJn2gmBJFJGaVVGcUZT/Ra4FB1tkRBURXEnFGyI8WNFMctpeCoFDIBZUNsidl5ZbtLtTGoHNUch+hh/VnYMsW8ADTV71DWk5ZBQXKhW9ySFwpAXhg2M7QkkGJAG1KsG5Na4CWDYTjwg+FT24XoNLAypWCSCfYZqi8Rx6dPyLzhyarqHfdW/X5nftANhkHd/F0QNqDrNG2MmZlHSUYLAdJQTrSehkFuZiVRhlFuH6vQkBO6JAuYIpREgJ6VdcyV/xiZxE8zhb80fs3+v6KMxdYOJRHa5hxX2izqtmbJy7RpYdLns5LJvDAg6ebgtOC+yXxbbn7CFFDD1wgIVQxj9+kZUYQaLMqtU+q9c6Dbca0KyWiWQIvlZmUUQVKDEYRJe8vyWLG6MGsON7Nk7xVbMKMHb7He5eA1ZnLZb4yYkrCdgF0wORq+GAbvnnRHoyY3+84D55HTc0LnmTNy3jhjZ+JQ96f72/3j/vU+el+9b973jXXPbdbcd7aa9+MfWDBTPg==</latexit>

<latexit sha1_base64="eqxshVBpGebPRZ7nNdKZpTWKhKA=">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</latexit>
where A(x) = A0 + A1 x1 + · · · + An xn , Ai 2 Sk 8i = 0, . . . , n.
This problem is equivalent to the SDP

minimize {t : A(x) tI}.


<latexit sha1_base64="qDLSjGIrjcjOT5BxWhd/JVa54Hs=">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</latexit>
x,t

Consider X = RΛR> 2 Sk with Λ = diag(λ1 (X), . . . , λk (X))


and R orthogonal (R> = R 1 ).
<latexit sha1_base64="8uw5RF/VADPY8gtJjt5q+55g8mo=">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</latexit>

<latexit sha1_base64="pNfheM4AD9+DUpvOZqc1ZtR4bKw=">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</latexit>

λmax (X)  t () λi (X)  t 8i = 1, . . . , k


() Λ tI
<latexit sha1_base64="tVRtWqvUfUzQFjmC1ECK56EAmR0=">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</latexit>
() X = RΛR> tRIR> = tI
104

“Hidden” SDP Constraints


Upper bounds on inverse matrices:
✓ ◆
X I
X ⌫ Y 1 () ⌫0
I Y <latexit sha1_base64="Z/0xGcpprvPAhdGxQXBo1P347y4=">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</latexit>

Upper bounds on squared matrices:


✓ ◆
> X Y
X ⌫ YY () ⌫0
Y> <latexit sha1_base64="WvTmKSewDWd1wqjACZuFa36iodQ=">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</latexit>
I

Quadratic matrix inequalities:


(AXB)> AXB + CXD + (CXD)> Y <latexit sha1_base64="dR5GSPndgWxRxFi8c7QzZJNEgyA=">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</latexit>

✓ > >

Y CXD (CXD) (AXB)
() ⌫0
AXB I
<latexit sha1_base64="KbjAIuArNK9tA537RTysk/M2nCo=">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</latexit>

To prove these equivalences: use Schur!


105

Hierarchy of Convex Optimization Problems

Conic Form Problems


SDP
SOCP
QCQP
QP

LP

= all convex problems


106

Popular Solvers
CPLEX (https://www.ibm.com/analytics/cplex-optimizer):
• caters for LP, QP, SOCP and MILP (mixed-integer LP); free for
academic use
Gurobi (www.gurobi.com)
• caters for LP, QP, SOCP and MISOCP (mixed-integer SOCP); can be
deployed on the cloud; free for academic use
MOSEK (http://www.mosek.com)
• caters for LP, QP, SOCP, SDP and MISDP (mixed-integer SDP); ideal
for large-scale sparse problems; free for academic use
SDPT3 (http://www.math.cmu.edu/~reha/sdpt3.html):
• caters for SDP; free, runs on Matlab
IPOPT (https://github.com/coin-or/Ipopt)
• caters for general NLPs (nonlinear programs); free
CVXPY (https://www.cvxpy.org) is a modeling language for optimization
problems implemented as a free toolbox for Python.
A list of solvers that can be used with CVXPY is available from https://
www.cvxpy.org/tutorial/advanced/index.html#choosing-a-solver.
107

Main Take-Away Points

• Convex optimization problems: convex objective and


inequality constraints, linear equality constraints; feasible set
is always convex; every local minimizer is a global minimizer
• Optimality criterion: x? is optimal iff the gradient of the
<latexit sha1_base64="5LrFFfJnX63LA8pgrlaEwo9fUgY=">AAADu3icbVJdb9MwFHUWPkb52uCRl0CFNKSqShAIeEFFTIIHHopE2aSmVI5z05n6I7Kd0SrqL+LX8Aj8Ga7TaKLZHMU+PufYvva9WSm4dXH8O9gLr12/cXP/Vu/2nbv37h8cPvhqdWUYTJgW2pxm1ILgCiaOOwGnpQEqMwEn2fK910/OwViu1Re3LmEm6ULxgjPqkJofBsepgh9MS0lVXqfFpk4zLXK7ljhsejuiZKhK6s4YvSRlrZRlHaUSqFQqB+ND7Ij6fOO7K7UxmBLVEof0cfN52DFlogI0Ne9QN5OOwUB+oXvckRcGQF0YtjO05FBgQFtSrluTWeAl42EyiOLhS98lm97qW2odNfODfjyMmxZdBkkL+qRtY3z3eZprVklQjglq7TSJSzerqXGcCf8UlYWSsiVdwBShohLsrG4i2kRPkcmjQhv8lYsa9v8VdSZ3dqiptD6juNLnyHY1T16lTStXvJ7VXJWVA8W2BxeViJyOfDFFOTfAnFgjoMxwjD1iZ9RQ5rDkdk5p9i6B7ca1qhRnOocOK9zKGYqkBScpV/6W9TvDm7JrONzMk0fHfMGdHXzCalaDD5in5bPWiClJugm4DCbPh2+G8ecX/dGozc0+eUSekCOSkFdkRD6SMZkQFvwMfgV/gr/h2zAPv4dia90L2jUPyU4Lq39n7UT9</latexit>

objective makes an acute angle with every feasible direction


• Problem classes: definitions of LP, QP, QCQP, SOCP, SDP
and conic form problems; (strict!) subset relations between
problem classes
• Modeling tricks: use Schur’s lemma to express hyperbolic
constraints, bounds on inverse matrices and quadratic
matrix inequalities as explicit SOCP constraints or LMIs

You might also like