Mathematical foundations - linear algebra

Save this PDF as:
 WORD  PNG  TXT  JPG

Size: px
Start display at page:

Download "Mathematical foundations - linear algebra"

Transcription

1 Mathematical foundations - linear algebra Andrea Passerini Machine Learning

2 Vector space Definition (over reals) A set X is called a vector space over IR if addition and scalar multiplication are defined and satisfy for all x, y, x X and λ, µ IR: Addition: associative x + (y + z) = (x + y) + z commutative x + y = y + x identity element 0 X : x + 0 = x inverse element x X x X : x + x = 0 Scalar multiplication: distributive over elements λ(x + y) = λx + λy distributive over scalars (λ + µ)x = λx + µx associative over scalars λ(µx) = (λµ)x identity element 1 IR : 1x = x

3 Properties and operations in vector spaces subspace Any non-empty subset of X being itself a vector space (E.g. projection) linear combination given λ i IR, x i X n λ i x i i=1 span The span of vectors x 1,..., x n is defined as the set of their linear combinations { n } λ i x i, λ i IR i=1

4 Basis in vector space Linear independency A set of vectors x i is linearly independent if none of them can be written as a linear combination of the others Basis A set of vectors x i is a basis for X if any element in X can be uniquely written as a linear combination of vectors x i. Necessary condition is that vectors x i are linearly independent All bases of X have the same number of elements, called the dimension of the vector space.

5 Linear maps Definition Given two vector spaces X, Z, a function f : X Z is a linear map if for all x, y X, λ IR: f (x + y) = f (x) + f (y) f (λx) = λf (x)

6 Linear maps as matrices A linear map between two finite-dimensional spaces X, Z of dimensions n, m can always be written as a matrix: Let {x 1,..., x n } and {z 1,..., z m } be some bases for X and Z respectively. For any x X we have: n n f (x) = f ( λ i x i ) = λ i f (x i ) f (x i ) = f (x) = i=1 m a ij z j j=1 i=1 n m m n λ i a ij z j = ( λ i a ij )z j = i=1 j=1 j=1 i=1 m µ j z j j=1

7 Linear maps as matrices Matrix of basis transformation a a 1n M IR m n =... a m1... a mn Mapping from basis coefficients to basis coefficients Mλ = µ

8 Matrix properties transpose Matrix obtained exchanging rows with columns (indicated with M T ). Properties: (MN) T = N T M T trace Sum of diagonal elements of a matrix tr(m) = n i=1 M ii inverse The matrix which multiplied with the original matrix gives the identity MM 1 = I rank The rank of an n m matrix is the dimension of the space spanned by its columns

9 Metric structure Norm A function : X IR + 0 is a norm if for all x, y X, λ IR: x + y x + y λx = λ x x > 0 if x 0 Metric A norm defines a metric d : X X IR + 0 : d(x, y) = x y Note The concept of norm is stronger than that of metric: not any metric gives rise to a norm

10 Dot product Bilinear form A function Q : X X IR is a bilinear form if for all x, y, z X, λ, µ IR: Q(λx + µy, z) = λq(x, z) + µq(y, z) Q(x, λy + µz) = λq(x, y) + µq(x, z) A bilinear form is symmetric if for all x, y X : Q(x, y) = Q(y, x)

11 Dot product Dot product A dot product, : X X IR is a symmetric bilinear form which is positive definite: x, x 0 x X A strictly positive definite dot product satisfies x, x = 0 iff x = 0 Norm Any dot product defines a corresponding norm via: x = x, x

12 Properties of dot product angle The angle θ between two vectors is defined as: cosθ = orthogonal Two vectors are orthogonal if x, z x z x, y = 0 orthonormal A set of vectors {x 1,..., x n } is orthonormal if x i, x j = δ ij where δ ij = 1 if i = j, 0 otherwise.

13 Eigenvalues and eigenvectors Definition Given a matrix M, the real value λ and (non-zero) vector x are an eigenvalue and corresponding eigenvector of M if Spectral decomposition Mx = λx Given an n n symmetric matrix M, it is possible to find an orthonormal set of n eigenvectors. The matrix V having eigenvectors as columns is unitary, that is V T = V 1 M can be factorized as: M = V ΛV T where Λ is the diagonal matrix of eigenvalues corresponding to eigenvectors in V

14 Eigenvalues and eigenvectors Singular matrices A matrix is singular if it has a zero eigenvalue Mx = 0x = 0 A singular matrix has linearly dependent columns: [ ] M1... M n 1 M n x 1. x n 1 = 0 x n

15 Eigenvalues and eigenvectors Singular matrices A matrix is singular if it has a zero eigenvalue Mx = 0x = 0 A singular matrix has linearly dependent columns: M 1 x M n 1 x n 1 + M n x n = 0

16 Eigenvalues and eigenvectors Singular matrices A matrix is singular if it has a zero eigenvalue Mx = 0x = 0 A singular matrix has linearly dependent columns: M n = M 1 x 1 x n + + M n 1 x n 1 x n

17 Eigenvalues and eigenvectors Singular matrices A matrix is singular if it has a zero eigenvalue Mx = 0x = 0 A singular matrix has linearly dependent columns: M n = M 1 x 1 x n + + M n 1 x n 1 x n Determinant The determinant M of a n n matrix M is the product of its eigenvalues A matrix is invertible if its determinant is not zero (i.e. it is not singular)

18 Eigenvalues and eigenvectors Symmetric matrices Eigenvectors corresponding to distinct eigenvalues are orthogonal: λ x, z = Ax, z = (Ax) T z = x T A T z = x T Az = x, Az = µ x, z Note An n n symmetric matrix can have at most n distinct eigenvalues.

19 Positive semi-definite matrix Definition An n n symmetrix matrix M is positive semi-definite if all its eigenvalues are non-negative. Alternative sufficient and necessary conditions for all x IR n x T Mx 0 there exists a real matrix B s.t. M = B T B

20 Functional analysis Cauchy sequence A sequence (x i ) i IN in a normed space X is called a Cauchy sequence if for all ɛ > 0 there exists n IN s.t. n, n > n x n x n < ɛ. A Cauchy sequence converges to a point x X if Banach and Hilbert spaces lim x n x 0 n A space is complete if all Cauchy sequences in the space converge (to a point in the space). A complete normed space is called a Banach space A complete dot product space is called a Hilbert space

21 Hilbert spaces A Hilbert space can be (and often is) infinite dimensional Infinite dimensional Hilbert spaces are usually required to be separable, i.e. there exists a countable dense subset: for any ɛ > 0 there exists a sequence (x i ) i IN of elements of the space s.t. for all x X min x i x < ɛ x i