Professional Documents
Culture Documents
John E. Mitchell
April 2018
Outline
1 SDP formulation
2 Simultaneous diagonalization
4 Strict complementarity
Primal SDP
We write our standard form semidefinite program as
minX C•X
subject to Ai • X = bi , i = 1, . . . , m (1)
X ⌫ 0
where:
C is an n ⇥ n symmetric matrix
Ai is an n ⇥ n symmetric matrix for i = 1, . . . , m
bi is a scalar for i = 1, . . . , m.
Dual SDP
Outline
1 SDP formulation
2 Simultaneous diagonalization
4 Strict complementarity
Simultaneous diagonalization
Lemma
If the n ⇥ n symmetric matrices M and R are simultaneously
diagonalizable then they commute.
1412=1211.
diagonal.
Proof of lemma A: a r e
eigenvalues.
entries
din,
Let the-columns of the orthogonal matrix P consist of the eigenvectors
of the matrices, so ftp.E-ptp
O F-fr...-in)
A
M = P⇤P T , R = P⌅P T
l
-
I
MR = P⇤P T P⌅P T
= P⇤⌅P T since P is orthogonal
T
= P⌅⇤P since diagonal matrices commute
I I
= P⌅P P⇤P Tl
T
since P is orthogonal
as required.
= RM
'f";D".. -11"":*.)
Mitchell Simultaneous Diagonalization 7 / 22
Simultaneous diagonalization
The converse
The converse also holds, so symmetric matrices commute if and only if
they are simultaneously diagonalizable.
Sketch of proof of converse:
ti¥Duu=u¥@Eu)
Assume symmetric R, M commute.
-
Assume R = UDU T , so diagonal entries of D are eigenvalues of R,
and columns of U are eigenvectors of R. Then
Outline
1 SDP formulation
2 Simultaneous diagonalization
4 Strict complementarity
Duality gap M i r C o x
m a x bly
st. Aiox--bi f t . S=C-Ey.-Ai
Xiao SEO
Let X be feasible in (1) and (y , S) be feasible in (2). The duality gap is
m
! n e
X
T
C•X b y = yi Ai + S • X bT y
i=1 i
m
X
= S•X + yi (Ai • X ) bT y
= S•X +b y = i=1
T
bT y
= S • X.
Assume (1) and (2) have optimal solutions and their values agree.
Let X ⇤ solve (1) and (y ⇤ , S ⇤ ) solve (2).
From the argument above, the Frobenius product
- trace(S ⇤ X ⇤ ) = S ⇤ • X ⇤ = 0. (3)
(5×44) i s z e r o .
of
Sun of alia, entries
Proof, part 1
Since S ⇤ 2 Sn+ it has an eigendecomposition,
S ⇤ = Q⇤Q T ,
indi
of S ⇤ . We now have trace
a:}
where the columns {qi , i = 1, . . . , n} of Q are the eigenvectors of S ⇤
and the diagonal entries of the diagonal matrix ⇤ are the eigenvalues
(ACE'X*Q)
=§(gAijCQTMQ7jiJ-
0 = S ⇤ • X→
⇤
= trace(Q⇤Q T X ⇤ )
EAiiCdx4QiiI@catx-
= trace(⇤Q T X ⇤ Q) from properties of trace
X n ⇣ ⌘
ealiiE.EQIu.x't.ie;
= ⇤ii Q T X ⇤ Q since ⇤ is diagonal
ii
i=1
Xn
= ⇤ii qiT X ⇤ qi
-
=§{QkiX*uQti
i=1
Proof, part 3
-
⇤
⇥
S = Q1 Q2
⇤ ⇤1 0
0 0
Q1T
Q2T
=eaQt
where the diagonal entries of ⇤1 are positive.
Note that-X ⇤ Q1 = 0, since each eigenvector qi of S ⇤ with positive
eigenvalue satisfies X ⇤ qi 1 = 0.(st) 4
gr) wilt r e ran
Proof, part 4
I
Thus,
m
cyan
⇤ ⇤
⇥
⇤
⇤ ⇤1 0 Q1T
X S = X Q1 Q2
- 0 0 Q2T
⇥ ⇤ ⇤1 0 Q1T
= 0 X ⇤ Q2
0 0 Q2T
①0 0
⇥ Q1T ⇤
=
Q2T
= 0, the n ⇥ n matrix of zeroes.
We also have
S ⇤ X ⇤ = (X ⇤ S ⇤ )T = 0T = 0,
as required. 54,#
since symmetric
Simultaneous diagonalization
Note that S ⇤ and X ⇤ commute, with both products- X ⇤ S ⇤ = S ⇤ X ⇤ = 0.
Hence, the matrices are simultaneously diagonalizable.
I n consisting of
In general, we can construct an orthonormal basis for R
three sets of vectors:
Eigenvectors of S ⇤ with positive eigenvalue that are in the
nullspace of X ⇤ . These eigenvectors comprise the columns of a
matrix Q1 .
A basis for the intersection of the nullspaces of X ⇤ and S ⇤ , which
we denote as the columns of a matrix Q̃2 , and
Eigenvectors of X ⇤ with positive eigenvalue that are in the
nullspace of S ⇤ . These eigenvectors comprise the columns of a
matrix Q̃3 .
Simultaneous diagonalization
Note that S ⇤ and X ⇤ commute, with both products X ⇤ S ⇤ = S ⇤ X ⇤ = 0.
Hence, the matrices are simultaneously diagonalizable.
I n consisting of
In general, we can construct an orthonormal basis for R
three sets of vectors:
Eigenvectors of S ⇤ with positive eigenvalue that are in the
nullspace of X ⇤ . These eigenvectors comprise the columns of a
matrix Q1 .
A basis for the intersection of the nullspaces of X ⇤ and S ⇤ , which
we denote as the columns of a matrix Q̃2 , and
Eigenvectors of X ⇤ with positive eigenvalue that are in the
nullspace of S ⇤ . These eigenvectors comprise the columns of a
matrix Q̃3 .
Simultaneous diagonalization
Note that S ⇤ and X ⇤ commute, with both products X ⇤ S ⇤ = S ⇤ X ⇤ = 0.
Hence, the matrices are simultaneously diagonalizable.
I n consisting of
In general, we can construct an orthonormal basis for R
three sets of vectors:
Eigenvectors of S ⇤ with positive eigenvalue that are in the
nullspace of X ⇤ . These eigenvectors comprise the columns of a
matrix Q1 .
A basis for the intersection of the nullspaces of X ⇤ and S ⇤ , which
we denote as the columns of a matrix Q̃2 , and
Eigenvectors of X ⇤ with positive eigenvalue that are in the
nullspace of S ⇤ . These eigenvectors comprise the columns of a
matrix Q̃3 .
and 2 32 T 3
h i 0 0 0 Q1
X⇤ = Q1 Q̃2 Q̃3 4 0 0 0 5 4 Q̃ T 5
2
0 0 ⇤3 Q̃3T
where the diagonal entries in ⇤1 and ⇤3 are positive.
Outline
1 SDP formulation
2 Simultaneous diagonalization
4 Strict complementarity
In the linear programming case, there exist optimal solutions that are
strictly complementary.
rank(X ⇤ ) + rank(S ⇤ ) = n,
O
2 3
1 0 0
X⇤ = 4 0 0 0 5 .
0 0 0
maxy ,S y1 2
-
y. fo,:{I-of
3
y1 y3 y2
subject to S = 4 y3 y2 0 5
y2 0 1 y3
S ⌫ 0.