Math 371 - Alexiades
Norms on vectors and matrices
Let V be a vector space. A real-valued function
∥ . ∥ on V is called a norm on V if:
(i) ∥ x ∥ ≥ 0 for any x in V, and
∥ x ∥ = 0 iff x = 0
(ii) ∥ λ x ∥ = |λ| ∥x∥
for any scalar λ , any x in V
(iii) ∥ x + y ∥ ≤ ∥x∥ + ∥y∥
for any x, y in V (triangle inequality)
Several norms can be defined on vectors in ℝn.
Most useful :
p-norm, p≥1 : ∥ x ∥p =
( ∑ |xi|p )1/p
in particular:
1-norm: ∥x∥1 = ∑ |xi| ,
2-norm: ∥x∥2
= ( ∑ |xi|2 )1/2
= √x•x
= Euclidean norm
max-norm :
∥ x ∥∞ = max( |xi| )
(amazingly
limp→∞ ∥x∥p = ∥x∥∞ )
e.g. x = ( 1, −3, 2 ): ∥x∥1 = 6 ,
∥x∥2 = √14 ,
∥x∥∞ = 3
Theorem:
All vector norms on a finite-dimensional vector space are equivalent.
They define the same sense of convergence.
Matrix norm: Let A be m×n matrix.
The (operator) norm induced on A by a vector norm ∥ ∥ is
∥A∥ = max{ ∥Ax∥ / ∥x∥ }
= amplification factor under the action of A on any vector
= maximum stretching
∥ I ∥ = 1
Easiest to compute (and most used):
1-norm: ∥A∥1 = max absolute column sum
2-norm: ∥A∥2 = (no easy way)
= max |eigenvalue| = spectral norm
∞-norm: ∥A∥∞
= max absolute row sum
Frobenius-norm: ∥A∥F
= ( ∑1m ∑1n
|aij|2 )1/2
= tr( ATA )
e.g. A = [1 , −3 ; −2 , 2]:
∥A∥1 =max(3,5)=5 ,
∥A∥2 ≈ 4.1306 ,
∥A∥∞ = max(4,4)=4 ,
∥A∥F = (1+9+4+4)1/2
= √18 ≈ 4.2426
Very desirable property for a matrix norm
(which all induced norms satisfy):
∥Ax∥ ≤ ∥A∥ ∥x∥
,
∥AB∥ ≤ ∥A∥ ∥B∥
(but not all norms do this, and those that don't are not useful...)
Cauchy-Swartz Inequality:
|x • y| ≤ ∥x∥2 ∥y∥2 for any vectors x, y
Function norms
The set of functions f(x) defined on an interval [a, b] constitutes a
vector space (with usual pointwise addition and multiplication by a scalar).
Many useful norms can be defined, the most usual are:
p-norm, p≥1 : ∥ f ∥p =
( ∫ab |f(x)|p dx )1/p.
The space of functions with finite p-norm is
Lp(a, b).
in particular:
1-norm: ∥f∥1
= ∫ab |f(x)| dx.
The space of functions with finite 1-norm is
L1(a, b) of integrable functions on (a,b).
2-norm: ∥f∥2
= ( ∫ab |f(x)|2 dx )1/2.
The space of functions with finite 2-norm is
L2(a, b) of square-integrable
functions on (a, b).
The 2-norm is the only p-norm that comes from an
inner product:
<f , g> := ∫ab f(x)g(x) dx
(generalization of dot product),
∥f∥2 = √<f,f>
This makes L2(a, b) into a Hilbert space,
by far the most useful function space.
Orthogonality can be defined:
f ⊥ g means <f , g> = 0.
inf-norm or sup-norm or max-norm:
∥ f ∥∞ = sup{ |f(x)|: a≤x≤b }
The space of continuous functions with finite inf-norm is
ℭ[a, b], next most useful to L2 space.
It is a Banach space (complete normed space)
but not an inner product space.
All these generalize to functions defined in a multidimensional region
Ω by replacing the interval [a,b] by Ω.