Professional Documents
Culture Documents
Iccopt PDF
Iccopt PDF
Iccopt PDF
Michael J. Todd
Weak duality
Strong duality
Algorithms
minx cT x
(P) Ax = b,
x ≥ 0.
maxy bT y
(D) AT y ≤ c.
maxy,s bT y
(D) AT y + s = c,
s ≥ 0.
minX C •X
(P) Ai • X = bi , i = 1, . . . , m,
X 0,
T P P
where S • Z := Trace S Z = i j sij zij for matrices of the same dimensions,
and X 0 means X is symmetric and positive semidefinite (psd). (We’ll also write
p×p
A B and B A for A − B 0.) We’ll write SR+ for the cone of psd real
matrices of order p. Note that, instead of the components of the vector x being
maxy bT y
P
(D) i yi A i C,
maxy,S bT y
P
(D) i yi A i + S = C,
S 0.
(P) A 1 x1 + . . . + A k xk = b,
1+nj
xj ∈ S2 , j = 1, . . . , k,
..
.
.
....
.....
.
..... .
... ..
... ..
..
.
.
x̄ ...
..
.....
.
..
.
.... ....
..... ....
{x := (ξ; x̄) ∈ <1+q : ξ ≥ kx̄k2 } .
....
.
. .
...
..
maxy bT y
1+nj
(D) cj − ATj y ∈ S2 , j = 1, . . . , k,
or
maxy,s1 ,...,sk bT y
(D) AT1 y + s1 = c1
..
.
ATk y + sk = ck
sj ∈ S 1+nj , j = 1, . . . , k.
minx hc, xi
(P) Ax = b,
x ∈ K,
where we have written hc, xi instead of cT x to emphasize that this can be thought
of as a general scalar/inner product. E.g., if our original problem is an SDP
involving X ∈ SRp×p , we need to embed it into <n for some n.
Even though our problem (P) looks very much like LP, it is important to note that
every convex programming problem can be written in the form (P).
Then we define
maxy,s hb, yi
(D) A∗ y + s = c,
s ∈ K ∗.
www.optimization-online.org/DB_HTML/2004/03/844.html.)
(Work by Shor, Parrilo, Lasserre, Bertsimas, Pena, ...: see, e.g., the Gloptipoly
home page at www.laas.fr/~henrion/software/gloptipoly/gloptipoly.html.)
(i) (ii)
T T T T T T
c x − b y = (A y + s) x − (Ax) y = s x ≥ 0.
both trivial.
T
X
C •X −b y = ( yi Ai + S) • X − ((Ai • X)m T
i=1 ) y
X X
= ( yi A i ) • X + S • X − yi (Ai • X)
(i) (ii)
= S • X ≥ 0.
X X X
cTj xj T
−b y = (ATj y T
+ s j ) xj − ( A j x j )T y
(i) X (ii)
= sTj xj ≥ 0.
(i) (ii)
∗
hc, xi − hb, yi = hA y + s, xi − hAx, yi = hs, xi ≥ 0,
where (i) follows by definition of the adjoint operator A∗ and (ii) by definition of the
dual cone K ∗ .
So in all cases we have weak duality, which suggests that it is worthwhile to
consider (P) and (D) together. In many cases, strong duality holds, and then it is
very worthwhile!
(S21+q )∗ = S21+q .
p×p ∗
(SR+ ) = SRp×p
+ .
Also,
p×p
for any u ∈ <p , so S ∈ SR+ .
Antonio Gramsci:
I’m a pessimist because of intelligence, but an optimist because of will.
control theory
Fermat-Weber problem
m
X
A(y) := A0 + yi A i
i=1
maxη,y −η
Pm
−ηI + i=1 yi A i −A0 ,
Hence the convex quadratic constraint (Ay + b)T (Ay + b) − cT y − d ≤ 0 holds iff
0 1
B I Ay + b C
@ A 0,
(Ay + b)T cT y + d
1 T
V̇ (x) = x (Y P + P T Y )x ≤ 0
2
maxη,Y −η
−ηI + Y 0,
−Y −I,
Y Pi + PiT Y 0, i = 1, . . . , m.
ei eTi • X = 1, i = 1, . . . , n,
X 0.
This gives a good bound and a good feasible solution (within 14%)
(Goemans and Williamson).
School of OR&IE – p.32/54
Extension of Chebyshev’s inequality
Suppose we have a random vector X ∈ <n and we know E(X) = x̄,
E(XX T ) = Σ. We wish to bound the probability that X ∈ C, with
maxY,y,η,ζ 1 − Σ • Y − 2x̄T y − η
2 3
6 Y − ζ i Ai y − ζ i bi 7
4 5 0, i = 1, . . . , m
(y − ζi bi )T η − 1 − ζi ci
2 3
6 Y y 7
4 5 0,
T
y η
ζi ≥ 0, i = 1, . . . , m.
School of OR&IE – p.33/54
Chebyshev’s inequality, cont’d
Suppose we have a feasible solution. Then, for any x ∈ < n ,
xT Y x + 2y T x + η ≥ 1 + ζi (xT Ai x + 2bTi x + ci ), i = 1, . . . , m,
To show that it is tight we use SDP duality (Vandenberghe, Boyd, and Comanor).
min η1 + · · · + η m
y,η
ηi ≥ ky − pi k2 , i = 1, . . . , m,
x1 + · · · + x m = 0,
kxi k2 ≤ 1, i = 1, . . . , m.
with optimal solution X = Diag(0; 0; 1) and optimal value 1, while its dual
0 1 0 1 0 1
B 1 0 0 C B 0 1 0 C B 0 0 0 C
B C B C B C
B C B C B C
max 2y2 , y1 B 0 0 0 C + y2 B 1 0 0 C B 0 C
B C B C B 0 0 C
@ A @ A @ A
0 0 0 0 0 2 0 0 1
Theorem Suppose (P) has a feasible solution and (D) a strictly feasible solution.
Then (P) has a nonempty bounded set of optimal solutions, and there is no duality
gap.
Corollary If both (P) and (D) have strictly feasible solutions, strong duality holds.
Henceforth, assume that both (P) and (D) have strictly feasible solutions, and
(wlog) that A has full row rank.
Mary Wollstonecraft:
What a weak barrier is truth when it stands in the way of an hypothesis!
Robert Frost:
My apple trees will never get across
And eat the cones under his pines, I tell him.
He only says, “Good fences make good neighbors.”
xk → x̄ ∈ ∂K ⇒ F (xk ) → +∞.
F (P) F (D)
F is a C 3 barrier for K;
F is ν-logarithmically homogeneous if
A∗ y + s = c, s ∈ int K ∗ ,
Ax = b, x ∈ int K,
µF 0 (x) + s = 0.
Similarly, (y, s) is optimal for (BDµ ) iff ∃x with the same first two equations and
x + µF∗0 (s) = 0.
within
√
O(ν ln(1/)) or O( ν ln(1/))
F 00 (w)x = s.
A∗ ∆y + ∆s = rd ,
A∆x = rp ,
F 00 (w)∆x + ∆s = rc .