Korean, Edit

Chapter 1. Vector Spaces(vector space)

Recommended post: 【Linear Algebra】 Table of Contents for Linear Algebra

1. Matrix Operations

2. Vector Spaces

1. Matrix Operations

⑴ Definition of a Matrix

① Define an m × n matrix A, its i-th row vector, and its j-th column vector as follows

② Zero matrix or null matrix : a matrix whose entries are all 0

③ Square matrix

⑵ Matrix Addition

⑶ Matrix Multiplication

① Let X[i][j] denote the entry in the i-th row and j-th column of matrix X. If A ∈ ℝl×m, B ∈ ℝm×n, C ∈ ℝl×n, and C = A × B, then the following holds

Property 1. In general, AB ≠ BA

Property 2. AB = O does not necessarily imply that A = O or B = O

Property 3. The following relationships hold between matrix multiplication and addition

○ A(B+C) = AB + AC, (A+B)C = AC + BC

○ A(BC) = (AB)C

○ c(A + B) = cA + cB

Theorem 1. Cayley–Hamilton theorem

○ By applying this theorem, An (where n ≥ 2) can be calculated in the form An = (A2 - (a + d)A + (ad - bc)E) Q(A) + kA + sE

⑥ Programming Code

Implementing Matrix Multiplication in C

○ Using Python numpy

○ Using Python sympy

⑷ Transposed Matrix

① Definition

② Symmetric matrix, sym : a square matrix A that is equal to its transpose

③ Given X ∈ ℝN×d (N pieces of d-dimensional data),

○ XtX ∈ ℝd×d is the covariance matrix when μ = 0

○ XXt ∈ ℝN×N is the dot matrix (similarity matrix)

Property 1. (At)t = A

Property 2. (A + B)t = At + Bt

Property 3. (rA)t = rAt

Property 4. The transpose of AB is (AB)t = BtAt

Property 5. If A and B are symmetric and AB = BA, then AB is symmetric

Property 6. The inverse of a symmetric matrix is also symmetric

⑸ Trace : denoted by trace or tr

Property 1. tr(E) = n, tr(O) = 0

Property 2. tr(A + B) = tr(A) + tr(B)

Property 3. tr(cA) = ctr(A), c ∈ ℝ

Property 4. tr(AT) = tr(A)

Property 5. tr(AB) = tr(BA)

Property 6. In general, tr(AB) ≠ tr(A) × tr(B)

Property 7. tr(ATB) = vec(A)·vec(B)

Property 8. tr(ABC) = tr(BCA) = tr(CAB) ( associativity)

Property 9. If A and B are positive semi-definite matrices, then tr(AB) ≤ tr(A)·tr(B)

Property 10. tr(An) = tr(P-1DnP) = tr(P-1PDn) = tr(Dn) = ∑i λin

○ However, the equation above appears to hold independently of diagonalizability

Property 11. If X ~ 𝒩(X̄, ∑) and S is a symmetric matrix, then 𝔼[XSX] = X̄SX̄ + tr(S∑)

Application 1. Hilbert-Schmidt test

○ Calculates trace of 2 kernels to determine whether they are dependent non-parameteric, non-linear pattern detection.

○ For two matrices M(1) and M(2) to which kernels have been applied, the cosine similarity is

T := (1/n) ∑i,k M(1)ik M(2)ik = (1/n) ∑i,k M(1)ik M(2)ki = (1/n) tr(M(1)×M(2))

⑹ Gaussian Elimination

① System of Linear Equations

② Augmented matrix : an m × (n + ℓ) matrix denoted by (A B)

③ Gauss-Jordan elimination : operations are performed by row

Operation 1. Swap the positions of two row vectors : changes the order of the equations

Operation 2. Multiply a specific row vector by c : multiplies both sides of one equation by c

Operation 3. Multiply one row vector by c and add it to another row vector : multiplies one equation by c and adds it to another equation

○ Final step : a reduced row echelon form (RREF) must be obtained

④ Advantage : it can be implemented through programming

2. Vector Spaces(vector space)

⑴ Conditions for a Vector Space : one type of algebraic structure, such as ℝn

① Closed under addition : if x, y ∈ V, then x + y ∈ V

② Closed under scalar multiplication : if ∀a ∈ ℝ and ∀x ∈ V, then ax ∈ V

③ Additive identity : for every element x of V, there exists 0 in V such that x + 0 = x

④ Additive inverse : if x ∈ V, then there exists -x in V such that x + (-x) = 0

⑤ Commutativity of addition : if x , y ∈ V, then x + y = y + x

⑥ Associativity of addition : if x , y , z ∈ V, then (x + y) + z = x + (y + z)

⑦ Distributive law : c(x + y) = cx + cy

⑧ Distributive law : (c1 + c2)x = c1x + c2x

⑨ Multiplicative identity : 1x = x

⑩ Associativity of scalar multiplication : c1(c2x) = (c1c2)x

⑵ Linear Transformations

① Linear transformation : a mapping T : U → V that satisfies the following conditions for x , y ∈ U and c ∈ F

② Linear combination

③ Linearly dependent

Linearly independent

Spanning set : a set S such that every vector in the vector space V can be expressed as a linear combination of elements of S

⑥ Basis : a linearly independent spanning set

⑶ If the inverse map T-1 : V → U exists, then T-1 is also a linear transformation

⑷ Inner Product Space

① Conditions for an inner product (dot product)

② Standard inner product : the standard inner product in V = ℂn is defined as follows

③ Properties of an inner product

④ Inner product space : a vector space on which an inner product is defined

○ Real inner product space : a case in which the scalar field is the set of real numbers

○ Complex inner product space : a case in which the scalar field is the set of complex numbers

⑤ Norm : a vector space equipped with a norm is called a norm vector space or normed linear space

Theorem 1. Cauchy-Schwartz inequality

Theorem 2. Triangle inequality

⑷ Metric Space

① Distance function, metric

Types of Distance Functions

② Relationship between a norm and distance

○ If a norm is defined, a distance d can be defined

○ Even if a distance is defined, a corresponding norm does not always exist

③ Kernel

○ A distance d may be defined differently using a nonlinear map φ(x) : X → H

○ Definition of a kernel : k(x i, y i) ≡ φ(x i)Tφ(x j)

○ Characteristics of kernels

Kernel Examples

○ φ(x, y) = (x, y, x2 + y2)T

○ k(v , u) = u Tv (in this case, φ(·) is the identity function)

○ Gaussian kernel : k(v , u) = exp(-   v - u   2 / 2σ2)

○ Polynomial kernel : k(v , u) = (u Tv + 1)d (where d is the polynomial degree)

○ Sigmoid kernel : k(v , u) = tanh(αu Tv + β)

○ RBF kernel (radial basis function kernel) : k(v , u) = exp(-γ   v - u   2)

○ Lorentz kernel

○ Heat kernel

○ Graph Matérn kernel

⑸ Subspace : a spanning set spans a subspace

Posted: 2020.04.07 21:36

Modified: 2024.10.10 08:12

results matching ""

    No results matching ""