Beruflich Dokumente
Kultur Dokumente
Introduction to matrices
Theorem 2.1 (Singular Value Decomposition Theorem). Let m, n N be given and let p = min{m, n}.
Let A Rmn be given. Then:
There exist orthogonal matrices U Rmm and V Rnn , and a diagonal matrix
DA = diag(1 , 2 , . . . , p ) Rmn
such that 1 2 p 0 and A = U DA V T .
Property 2.1. Let A Rmn . The following statements are equivalent:
1. The matrix A is left invertible.
2. m n and rank(A) = n.
3. The columns of A are linearly independent.
4. Ker(A) = {0}.
Property 2.2. Let A Rmn . The following statements are equivalent:
1. The matrix A is right invertible.
2. m n and rank(A) = m.
3. The rows of A are linearly independent.
4. Im(A) = Rm .
Theorem 2.2.
1. Let A Rmn be a left invertible matrix.
Then the collection of all left inverses of A are given by
.
1
1 .
A1
=
A
BA
A
.
B
P
2
1
1
L
where P is an m m permutationmatrix,
B Rn(mn) is an arbitrary matrix and A1 Rnn is
A1
a regular matrix for which P A =
.
A2
A similar result holds for the right inverse.
2. Let A Rmn be a right invertible matrix.
Then the collection of all right inverses of A are given by
" 1
#
A3 A1
3 A4 C
1
AR = Q
C
where Q is an n n permutation matrix,
C R(nm)m is an arbitrary matrix and A3 Rmm is
a regular matrix for which AQ = A3 A4 .
Property 2.3.
1. Let A Rmn be left invertible. Then:
1
Every left inverse A1
L of A satisfies: the matrix AAL is idempotent.
1 +
A , for
all R \ {0}.
Graphs
Property 3.1. Let G = (V, E) be an undirected graph and let A, B V with A 6= B. If there exists a
simple path from A to B (in G), then there exists a simple path from A to B in which no vertex occurs
repeatedly.
Property 3.2. Let V = {Pk : 1 k n} and let G = (V, E) be an undirected graph or multi-graph.
Then:
n
X
deg(Pk ) = 2|E|, where |E| is the number of edges in G.
k=1
Property 3.3. For every undirected graph or multi-graph, the number of vertices of odd degree is even.
Theorem 3.1. Let G = (V, E) be an undirected graph or multi-graph. Then:
G has an Euler circuit G is connected and the degree of every vertex of G is even.
Property 3.4. Let G = (V, E) be an undirected graph or multi-graph. Then:
G has an Euler path but no Euler circuit G is connected and the degree of two vertices is odd.
Non-negative matrices
A is irreducible for every (i, j) there exists some k N such that aij > 0 (1 i, j n).
Property 4.5. Every non-negative eigenvector of an irreducible non-negative n n matrix is positive.
Property 4.6. An irreducible non-negative n n matrix A has a positive eigenvalue A such that
A || for every eigenvalue of A and A has a positive eigenvector corresponding to A .
Property 4.7. Let A be a non-negative n n matrix A has a non-negative eigenvalue A such that
A || for every eigenvalue of A and A has a non-negative eigenvector corresponding to A .
Theorem 4.3. Let A 0 be an irreducible n n matrix and let A be the dominant eigenvalue of A.
Then:
1. A is a simple zero of the characteristic polynomial of A.
2. If Ax = A x then (the dominant eigenvector) x of A is a scalar multiple of a positive vector.
3. A has exactly one eigenvector in Rn+ (ignoring scalar multiples).
Property 4.8. Let A 0 be an n n matrix and assume that A is a simple dominant eigenvalue of A.
Then:
A is irreducible A and AT both have a positive eigenvector corresponding to A .
4
Property 4.9. Let A = [aij ]1i,jn Rnn be non-negative. Let A be the dominant eigenvalue of A.
n
X
Write Ri =
aij , 1 i n. Define Rmin = min {Ri : 1 i n} and Rmax = max {Ri : 1 i n}.
j=1
0
A12
0
0
0
0
0
A23 0
0
..
..
..
.
.
.
.
.
.
.
.
.
P T AP =
.
.
..
..
..
..
..
.
.
.
0
0
0 Ah1,h
Ah1
0
0
0
where Ai,i+1 (1 i h 1) and Ah1 are irreducible.
Theorem 4.6. Let A be a non-negative n n matrix and let GA be the directed graph corresponding
to the matrix A. Then:
A is irreducible GA is strongly connected.
Theorem 4.7. Let A Rnn be a matrix of which the entries outside the diagonal are non-positive.
Then:
A is a regular M -matrix the inverse A1 exists and A1 0.
Property 4.11. Let A, B Rnn be non-negative matrices and let Un be the n n matrix of which
each entry is equal to 1.
1. A is stochastic AUn = Un .
2. A is stochastic = for every eigenvalue of A, || 1.
3. A is stochastic Au = u , where = 1 is the dominant eigenvalue of A, and
T
u = 1 1 1 is the dominant eigenvector corresponding to .
Property 4.12. Let A Rnn be a non-negative matrix with positive dominant eigenvalue and a
positive dominant eigenvector. Then:
The matrix
Introduction to convexity
Theorem 5.1. Let V Rn be a non-empty set. Then: V is convex for all m N and all convex
m
m
X
X
combinations
k xk of points xk V (1 k m),
k xk V .
k=1
k=1
Theorem 6.1.
(a) Let V1 , V2 Rn be convex sets such that V1 V2 = . Then:
There exists a separating hyper-plane H of V1 and V2 .
(b) Let V1 , V2 Rn be convex sets such that V1 V2 = and V1 V2 is closed. Then:
There exists a separating hyper-plane H of V1 and V2 which has a non-empty intersection with at
most one of the sets V1 and V2 .
Theorem 6.2. Let V1 , V2 Rn be convex sets. If Int(V1 ) 6= and V2 Int(V1 ) = , there exists a
separating hyper-plane H of V1 and V2 .
Property 6.3. Let V Rn be a closed set. Let a Rn \ {0} and b R.
Let f : Rn R be defined by f (x) = (a, x). Then:
H = {x Rn : (a, x) = b} is a supporting hyper-plane of V such that (a, x) b for all x V
f assumes a global maximum on V and max{f (x) : x V } = b.
Property 6.4. Let V Rn be a non-empty closed convex set and let p Rn \ V . Then:
There exists a unique point q V for which kx pk kq pk for all x V and
H = {x Rn : (p q, x) = (p q, q)} is a separating hyper-plane of V and {p} and it is a supporting
hyper-plane of V for which (p q, x) (p q, q) for all x V and (p q, x) > (p q, q) for x = p.
Property 6.5. Let K Rn be a non-empty set. Then:
K is a convex cone For all x, y K and > 0, x + y K and x K.
Theorem 6.3. Let K be a non-empty set in Rn . Then:
K is a convex pointed cone all non-negative linear combinations of points of K belong to K.
Property 6.6. For every i N, let Ki Rn be a convex pointed cone. Then:
The intersection
i=1
Theorem 6.4. Let D Rn be an open convex set and let hk : D R (1 k N ) be given quasi-convex
differentiable functions such that:
There exists x V such that hk (x ) < 0 for all 1 k N . (Slater condition)
Define for p V the set Cp consisting of non-negative linear combinations of the gradients of the functions
hk at p by
(N
)
X
Cp =
k hk (p) : k 0, k = 0 if hk (p) < 0, where 1 k N
k=1
Then:
1. For every p V , the set Cp is a closed convex pointed cone and Cp is generated by
{hk (p) : hk (p) = 0 (1 k N )}.
2. If hk (p) < 0 for all k (1 k N ), then Cp = {0}.
3. If Cp 6= {0} then for a Cp \ {0}, the set Ha = {x Rn : (a, x) = (a, p)} is a supporting
hyper-plane of V through p and (a, x) (a, p) for all x V .
Extrema
n
X
j=1
k=1
m
X
k gk (p)
k=1
m
X
k=1
11