Review of Linear Algebra

advertisement
Introduction to Pattern Recognition for Human ICT
Review of Linear Algebra
2014. 9. 12
Hyunki Hong
Contents
•
Vector and matrix notation
•
Vectors
•
Matrices
•
Vector spaces
•
Linear transformations
•
Eigenvalues and eigenvectors
Vector and matrix notation
• A ๐‘‘-dimensional (column) vector ๐‘ฅ and its transpose are
written as:
• An ๐‘›×๐‘‘ (rectangular) matrix and its transpose are written
as
• The product of two matrices is
cf. The matrix product is associative. If three matrices A, B, and C are respectively m×p, p×q, and q×r matrices, then there
are two ways of grouping them without changing their order, and ABC = A(BC) = (AB)C is an m × r matrix.
Vectors
also known as
• The inner product (a.k.a. dot product or scalar product) of two
vectors is defined by:
• The magnitude of a vector is
• The orthogonal projection of vector ๐‘ฆ onto vector ๐‘ฅ is <๐‘ฆT, ๐‘ข๐‘ฅ>
๐‘ข๐‘ฅ.
where vector ๐‘ข๐‘ฅ has unit magnitude and the same direction as ๐‘ฅ
• The angle between vectors ๐‘ฅ and ๐‘ฆ is
• Two vectors ๐‘ฅ and ๐‘ฆ are said to be
1. orthogonal if ๐‘ฅ๐‘‡๐‘ฆ = 0
๐‘‡
2. orthonormal
๏ƒฆ cos ๏ฑ -ifsin๐‘ฅ ๏ฑ๐‘ฆ๏ƒถ = 0 and |๐‘ฅ| = |๐‘ฆ| = 1
๏ƒง
๏ƒท
ex) ๏ƒง sin ๏ฑ cos ๏ฑ ๏ƒท
๏ƒจ
๏ƒธ
Vectors
• A set of vectors ๐‘ฅ1, ๐‘ฅ2, …, ๐‘ฅ๐‘› are said to be linearly dependent
if there exists a set of coefficients ๐‘Ž1, ๐‘Ž2, …, ๐‘Ž๐‘› (at least one
different than zero) such that
๐‘Ž 1 ๐‘ฅ 1 + ๐‘Ž 2 ๐‘ฅ 2 + … + ๐‘Ž๐‘›๐‘ฅ๐‘› = 0
• Alternatively, a set of vectors ๐‘ฅ1, ๐‘ฅ2, …, ๐‘ฅ๐‘› are said to be
linearly independent if
๐‘Ž 1 ๐‘ฅ 1 + ๐‘Ž 2 ๐‘ฅ 2 + … + ๐‘Ž๐‘›๐‘ฅ๐‘› = 0 ⇒ ๐‘Ž๐‘˜ = 0 ∀๐‘˜
Matrices
d
๏ƒฅ
• The determinant of a square matrix ๐ด๐‘‘×๐‘‘ is
k ๏€ฝ1
th
where ๐ด๐‘–๐‘˜ is the minor formed by removing the i row and the
kth column of ๐ด
NOTE) The determinant of a square matrix and its transpose
is the same: |๐ด|=|๐ด๐‘‡|
• The trace of a square matrix ๐ด๐‘‘×๐‘‘ is the sum of its diagonal
d
elements.
| A|๏€ฝ
tr ( A ) ๏€ฝ
๏ƒฅa
a ik | Aik | (-1)
kk
k ๏€ฝ1
• The rank of a matrix is the number of linearly independent
rows (or columns).
• A square matrix is said to be non-singular if and only if its
rank equals the number of rows (or columns).
1. A non-singular matrix has a non-zero determinant.
k ๏€ซi
Matrices
• A square matrix is said to be orthonormal if ๐ด๐ด๐‘‡ = ๐ด๐‘‡๐ด = ๐ผ
• For a square matrix ๐ด
1. if ๐‘ฅ๐‘‡๐ด๐‘ฅ > 0 ∀๐‘ฅ≠0, then ๐ด is said to be positive-definite (i.e., the
covariance matrix)
2. ๐‘ฅ๐‘‡๐ด๐‘ฅ ≥ 0 ∀๐‘ฅ≠0, then ๐ด is said to be positive-semi-definite
• The inverse of a square matrix ๐ด is denoted by ๐ด−1 and is such
that ๐ด๐ด−1 = ๐ด −1 ๐ด = ๐ผ
1. The inverse ๐ด−1 of a matrix ๐ด exists if and only if ๐ด is non-singular.
• The pseudo-inverse matrix ๐ด† is typically used whenever ๐ด−1
does not exist (because ๐ด is not square or ๐ด is singular).
1. One-sided inverse (left inverse or right inverse) If the matrix ๐ด has
dimensions and is full rank then use the left inverse if and the
right inverse if
cf. Formally, given a matrix
satisfies the condition
and a matrix
,
is a generalized inverse of if it
Vector spaces
• The n-dimensional space in which all the n-dimensional
vectors reside is called a vector space.
• A set of vectors {๐‘ข1, ๐‘ข2, …, ๐‘ข๐‘›} is said to form a basis for a
vector space if any arbitrary vector ๐‘ฅ can be represented by a
linear combination of the {๐‘ข๐‘–}
๐‘ฅ = ๐‘Ž1๐‘ข 1 + ๐‘Ž2๐‘ข 2 + โ‹ฏ + ๐‘Ž๐‘›๐‘ข๐‘›
1. The coefficients {๐‘Ž1, ๐‘Ž2, …, ๐‘Ž๐‘›} are called the components of
vector ๐‘ฅ with respect to the basis {๐‘ข๐‘–}.
2. In order to form a basis, it is necessary and sufficient that the {๐‘ข๐‘–}
vectors be linearly independent.
• A basis {๐‘ข๐‘–} is said to be
• A basis {๐‘ข๐‘–} is said to be
๏ƒฌ๏‚น 0 i ๏€ฝ j
T
u
u
orthogonal if i j ๏ƒญ
๏ƒฎ๏€ฝ 0 i ๏‚น j
๏ƒฌ๏€ฝ1 i ๏€ฝ j
T
orthonormal if u i u j ๏ƒญ ๏€ฝ 0 i ๏‚น j
๏ƒฎ
1. As an example, the Cartesian coordinate base is
an orthonormal base.
Vector spaces
• Given n linearly independent vectors {๐‘ฅ1, ๐‘ฅ2, …, ๐‘ฅ๐‘›}, we can
construct an orthonormal base {๐œ™ 1, ๐œ™ 2, …, ๐œ™๐‘›} for the vector
space spanned by {๐‘ฅ๐‘–} with the Gram-Schmidt
orthonormalization procedure.
• The distance between two points in a vector space is defined
as the magnitude of the vector difference between the points
1
2
๏ƒฉ
2๏ƒน
d E ( x, y ) ๏€ฝ | x - y | ๏€ฝ ๏ƒช๏ƒฅ ( xk - yk ) ๏ƒบ
๏ƒซ k ๏€ฝ1
๏ƒป
d
This is also called the Euclidean distance.
Linear transformations
• A linear transformation is a mapping from a vector space ๐‘‹๐‘
onto a vector space ๐‘Œ๐‘€, and is represented by a matrix.
1. Given vector ๐‘ฅ ๐œ– ๐‘‹๐‘, the corresponding vector y on ๐‘Œ๐‘€ is computed
as
2. Notice that the dimensionality of the two spaces does not need to
be the same.
3. For pattern recognition we typically have ๐‘€<๐‘ (project onto a
lower-dimensional space).
Linear transformations
• A linear transformation represented by a square matrix ๐ด is
said to be orthonormal when ๐ด๐ด๐‘‡=๐ด๐‘‡๐ด=๐ผ
1. This implies that ๐ด๐‘‡=๐ด−1
2. An orthonormal x form has the property of preserving the
magnitude of the vectors
| y|๏€ฝ
y y ๏€ฝ
T
( Ax ) Ax ๏€ฝ
T
x A Ax ๏€ฝ
T
T
x x ๏€ฝ|x|
T
3. An orthonormal matrix can be thought of as a rotation of the
reference frame.
ex) ๏ƒฆ cos ๏ฑ - sin ๏ฑ ๏ƒถ
๏ƒง๏ƒง
๏ƒจ sin ๏ฑ
๏ƒท๏ƒท
cos ๏ฑ ๏ƒธ
์„ ํ˜•๋ณ€ํ™˜ ์ถ”๊ฐ€์ž๋ฃŒ ์ฐธ์กฐ
The rotation takes the vector (1,0) to (cosθ, sinθ) and the
vector (0, 1) to (cosθ, -sinθ) . This is just what we need,
since in a matrix the first column is just the output when
you put in a unit vector along the x-axis; the second
column is the output for a unit vector along the y-axis, and
so on. So the 2D rotation matrix is.. (cf. ์‹œ๊ณ„๋ฐฉํ–ฅ์ด๋ฉด,..)
Eigenvectors and eigenvalues
• Given a matrix ๐ด๐‘×๐‘, we say that ๐‘ฃ is an eigenvector* if there
exists a scalar ๐œ† (the eigenvalue) such that
๐ด๐‘ฃ = ๐œ†๐‘ฃ
• Computing the eigenvalues
* The "eigen-" in "eigenvector" translates
as "characteristic“.
• The matrix formed by the column eigenvectors is called the
modal matrix M
1. Matrix Λ is the canonical form of A: a diagonal matrix with
eigenvalues on the main diagonal
• Properties
1. If Λ is non-singular, all eigenvalues are non-zero.
2. If Λ is real and symmetric, all eigenvalues are real.
The eigenvectors associated with distinct eigenvalues are
orthogonal.
3. If Λ is positive definite, all eigenvalues are positive
• If we view matrix ๐ด as a linear transformation, an eigenvector
represents an invariant direction in vector space.
1. When transformed by ๐ด, any point lying on the direction defined by
๐‘ฃ will remain on that direction, and its magnitude will be multiplied
by ๐œ†.
2. For example, the transform that rotates 3-d vectors about the ๐‘
axis has vector [0 0 1] as its only eigenvector and ๐œ† = 1 as its
eigenvalue.
๏ƒฉ cos ๏ข
๏ƒช
A ๏€ฝ sin ๏ข
๏ƒช
๏ƒช๏ƒซ 0
๏€ญ sin ๏ข
cos ๏ข
0
0๏ƒน
๏ƒบ
0
๏ƒบ
1 ๏ƒบ๏ƒป
• Given the covariance matrix Σ of a Gaussian distribution
1. The eigenvectors of Σ are the principal directions of the
distribution.
2. The eigenvalues are the variances of the corresponding principal
directions
• The linear transformation defined by the eigenvectors of Σ
leads to vectors that are uncorrelated regardless of the form
of the distribution.
1. If the distribution happens to be Gaussian, then the transformed
vectors will be statistically independent.
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๋ฒกํ„ฐ์˜ ํ‘œํ˜„
๏‚ง ๋ฒกํ„ฐ : ํฌ๊ธฐ์™€ ๋ฐฉํ–ฅ์„ ๊ฐ€์ง€๋Š” ์ž„์˜์˜ ๋ฌผ๋ฆฌ๋Ÿ‰
๏‚ง ํŒจํ„ด ์ธ์‹์—์„œ๋Š” ์ธ์‹ ๋Œ€์ƒ์ด ๋˜๋Š” ๊ฐ์ฒด๊ฐ€ ํŠน์ง•์œผ๋กœ ํ‘œํ˜„๋˜๊ณ , ํŠน์ง•์€ ์ฐจ์›์„ ๊ฐ€์ง„
๋ฒกํ„ฐ๋กœ ํ‘œํ˜„๋œ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฒกํ„ฐ๋ฅผ ํŠน์ง• ๋ฒกํ„ฐ(feature vector)๋ผ๊ณ  ํ•œ๋‹ค.
๏‚ง ํŠน์ง• ๋ฒกํ„ฐ์— ๋Œ€ํ•œ ๋Œ€์ˆ˜ํ•™์  ๊ณ„์‚ฐ์„ ์œ„ํ•ด์„œ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ํ–‰๋ ฌ๋กœ ํ‘œํ˜„ํ•˜์—ฌ d์ฐจ์›
๊ณต๊ฐ„์ƒ์˜ ํ•œ ์ ์˜ ๋ฐ์ดํ„ฐ๋กœ ํŠน์ง•์„ ๋‹ค๋ฃจ๊ฒŒ ๋œ๋‹ค.
16
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๋ฒกํ„ฐ์˜ ์ „์น˜ (transpose)
๏‚ง N×1ํ–‰๋ ฌ์„ 1×Nํ–‰๋ ฌ๋กœ, ํ˜น์€ 1×N ํ–‰๋ ฌ์„ N×1ํ–‰๋ ฌ๋กœ ํ–‰๊ณผ ์—ด์„ ๋ฐ”๊พผ ํ–‰๋ ฌ
๏ถ ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ
๏‚ง ์›์ ์—์„œ ๋ฒกํ„ฐ ๊ณต๊ฐ„์ƒ์˜ ํ•œ ์ ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ
๏ถ ๋‹จ์œ„ ๋ฒกํ„ฐ
๏‚ง ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ๊ฐ€ 1์ธ ๋ฒกํ„ฐ.
๏‚ง ๋งŒ์•ฝ, ๋ฒกํ„ฐ v๊ฐ€ 0์ด ์•„๋‹Œ ๋ฒกํ„ฐ๋ผ๋ฉด v๋ฐฉํ–ฅ์˜ ๋‹จ์œ„๋ฒกํ„ฐ u
๏‚ง ๋ฒกํ„ฐ v๋ฐฉํ–ฅ์˜ ๋‹จ์œ„ ๋ฒกํ„ฐ๊ณ„์‚ฐ: ์ •๊ทœํ™”
17
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๋ฒกํ„ฐ์˜ ๊ณฑ์…ˆ ๋‚ด์ , ์™ธ์ 
๏‚ง ์Šค์นผ๋ผ๊ณฑ
๏‚ง ์ž„์˜์˜ ๋ฒกํ„ฐ์— ์ž„์˜์˜ ์Šค์นผ๋ผ(์‹ค์ˆ˜)๋ฅผ ๊ณฑํ•˜๊ธฐ
๏‚ง ๋‚ด์  (dot product)
๏‚ง ์ฐจ์›์ด ๋™์ผํ•œ ๋‘ ๊ฐœ์˜ ๋ฒกํ„ฐ A,B์— ๋Œ€ํ•˜์—ฌ ๋Œ€์‘๋˜๋Š” ์„ฑ๋ถ„ ๋ณ„๋กœ ๊ณฑํ•˜์—ฌ ํ•ฉํ•˜๋Š” ๊ฒƒ์„
๋‘ ๋ฒกํ„ฐ์˜ '๋‚ด์ '์ด๋ผ๊ณ  ํ•จ
๏‚ง ๋ฒกํ„ฐ์˜ ๋‚ด์ ์˜ ๊ฒฐ๊ณผ๋Š” ์‹ค์ˆ˜ ์Šค์นผ๋ผ
= BTA
๏‚ง ๋‘ ๋ฒกํ„ฐ ์‚ฌ์ด์˜ ๊ฐ θ๊ฐ€ ์ฃผ์–ด์งˆ ๊ฒฝ์šฐ, ๋‚ด์  ์Šค์นผ๋ผ C
18
01_๋ฒกํ„ฐ ์ด๋ก 
๏‚ง ์™ธ์ 
๏‚ง A,B∈R3 (A,B๊ฐ€ 3์ฐจ์› ๋ฒกํ„ฐ ๊ณต๊ฐ„์ƒ์— ์†ํ•œ๋‹ค)์ธ ๋ฒกํ„ฐ A,B๊ฐ€ ๋‹ค์Œ๊ณผ ๊ฐ™์Œ
๏‚ง ๋ฒกํ„ฐ ์™ธ์ ์˜ ํฌ๊ธฐ๋Š” A์™€ B๋ฅผ ์ด์›ƒํ•˜๋Š” ๋‘ ๋ณ€์œผ๋กœ ํ•˜๋Š” ํ‰ํ–‰ ์‚ฌ๋ณ€ํ˜•์˜ ๋ฉด์ ๊ณผ ๊ฐ™์Œ
๏‚ง ์™ธ์ ์˜ ๊ฒฐ๊ณผ๋Š” A์™€ B์— ๋™์‹œ์— ์ˆ˜์ง์ด๋ฉฐ, ์˜ค๋ฅธ์†์˜ ์—„์ง€์™€ ์ธ์ง€์™€ ์ค‘์ง€๋ฅผ ์„œ๋กœ ์ˆ˜์ง์ด
๋˜๊ฒŒ ํŽด์„œ ์ธ์ง€๋ฅผ A๋ฐฉํ–ฅ, ์ค‘์ง€๋ฅผ B๋ฐฉํ–ฅ์œผ๋กœ ํ•  ๋•Œ ์—„์ง€์˜ ๋ฐฉํ–ฅ์„ ๊ฐ€๋ฅด์น˜๋Š” ๋ฒกํ„ฐ๊ฐ€ ๋จ.
19
01_๋ฒกํ„ฐ ์ด๋ก 
๏‚ง ๋‹จ์œ„๋ฒกํ„ฐ์˜ ๋‚ด์  ๋ฐ ์™ธ์ 
z
k
j
y
i
x
20
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ์ˆ˜์ง ์‚ฌ์˜ (vector projection)
๏‚ง ๋ฒกํ„ฐ x ์— ๋Œ€ํ•œ ๋ฒกํ„ฐ y์˜ ๋ฐฉํ–ฅ ์„ฑ๋ถ„
๏‚ง y ๋ฒกํ„ฐ๋ฅผ x ๋ฒกํ„ฐ๋กœ ์‚ฌ์˜ ๏ƒ  ๋ฒกํ„ฐ x์˜ ๋ฐฉํ–ฅ์œผ๋กœ์˜ ๋ฐฉํ–ฅ์„ฑ๋ถ„ ๊ณ„์‚ฐ
๏‚ง ์‚ฌ์˜ ๋ฒกํ„ฐ๋Š” ๋‚ด์ ์˜ ์ •์˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜
๏ƒ˜ux ์€ x์™€ ๊ฐ™์€ ๋ฐฉํ–ฅ์˜ ๋‹จ์œ„ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง€๋Š” ๋‹จ์œ„ ๋ฒกํ„ฐ
โ—€ ๋ฒกํ„ฐ์˜ ๋‚ด์ 
๏‚ง ๋‘ ๋ฒกํ„ฐ x์™€ y๊ฐ€ ๋งŒ์•ฝ, xTy = 0 ์ด๋ฉด ๏ƒ  ๋‘ ๋ฒกํ„ฐ x์™€ y๋Š” ์ˆ˜์ง(orthogonal)
๏‚ง xTy = 0 ์ด๊ณ  |x |= |y |= 1 ์ด๋ฉด, ๋‘ ๋ฒกํ„ฐ x์™€ y๋Š” ์ •๊ทœ ์ง๊ต(orthonormal)
๏ƒ˜ Ex: ๊ฐ ์ขŒํ‘œ์ถ• ๋ฐฉํ–ฅ์œผ๋กœ์˜ ๋ฐฉํ–ฅ๋ฒกํ„ฐ
21
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ์„ ํ˜• ๊ฒฐํ•ฉ
๏‚ง ๋ฒกํ„ฐ ์ง‘ํ•ฉ {x1,x2,…,xm} ๊ณผ ์Šค์นผ๋ผ ๊ณ„์ˆ˜ ์ง‘ํ•ฉ {α 1, α2,…, α m} ๊ณผ์˜ ๊ณฑ์˜ ํ•ฉ์œผ๋กœ ํ‘œํ˜„๋œ
๊ฒฐ๊ณผ๋ฅผ ๋ฒกํ„ฐ x์˜ ‘์„ ํ˜• ๊ฒฐํ•ฉ(linear combination)’ ํ˜น์€ ‘1์ฐจ ๊ฒฐํ•ฉ’ ์ด๋ผ๊ณ  ํ•จ
๏ถ ์„ ํ˜• ์ข…์†๊ณผ ์„ ํ˜• ๋…๋ฆฝ
๏‚ง ์„ ํ˜• ์ข…์† (linearly dependent)
๏‚ง ์ž„์˜์˜ ๋ฒกํ„ฐ ์ง‘ํ•ฉ์„ ๋‹ค๋ฅธ ๋ฒกํ„ฐ๋“ค์˜ ์„ ํ˜• ๊ฒฐํ•ฉ์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด, ์ด ๋ฒกํ„ฐ ์ง‘ํ•ฉ์€ '์„ ํ˜•
์ข…์†(linearly dependent)'์ด๋ผ๊ณ  ํ•จ
๏‚ง <{a1, a2, …, am}, {x1,x2,…,xm}> = 0
๏‚ง ์„ ํ˜• ๋…๋ฆฝ (linearly independent)
๏‚ง ๋งŒ์•ฝ, ์•„๋ž˜ ์‹์„ ๋งŒ์กฑํ•˜๋Š” ์œ ์ผํ•œ ํ•ด๊ฐ€ ๋ชจ๋“  i์— ๋Œ€ํ•˜์—ฌ αk = 0, {x1, x2, …, xm} ๋Š” '์„ ํ˜•
๋…๋ฆฝ(linearly independent)'์ด๋ผ๊ณ  ํ•จ
22
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๊ธฐ์ € ์ง‘ํ•ฉ
๏‚ง N ์ฐจ์›์˜ ๋ชจ๋“  ๋ฒกํ„ฐ๋ฅผ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ณธ ๋ฒกํ„ฐ์˜ ์ง‘ํ•ฉ
๏‚ง ์ž„์˜์˜ ๋ฒกํ„ฐ๋Š” ๊ธฐ์ € ๋ฒกํ„ฐ ์ง‘ํ•ฉ์„ ํ†ตํ•˜์—ฌ N×1 ๋ฒกํ„ฐ ๊ณต๊ฐ„์— ํŽผ์ณ์ง„๋‹ค๊ณ  ํ‘œํ˜„ (span)
๏‚ง ๋งŒ์•ฝ, {vi}1≤i≤N ์ด ๊ธฐ์ € ๋ฒกํ„ฐ ์ง‘ํ•ฉ์ด๋ผ๋ฉด, ์ž„์˜์˜ N×1 ๋ฒกํ„ฐ x๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ‘œํ˜„ํ•จ
๏‚ง ์ž„์˜์˜ ๋ฒกํ„ฐ x๊ฐ€ {ui} ์˜ ์„ ํ˜• ์กฐํ•ฉ์œผ๋กœ ํ‘œํ˜„๋œ๋‹ค๋ฉด ๋ฒกํ„ฐ ์ง‘ํ•ฉ {u1, u2, …, un}์„ N ์ฐจ์›
๊ณต๊ฐ„์˜ ๊ธฐ์ €(basis) ๋ผ๊ณ  ํ•จ.
๏‚ง ๋ฒกํ„ฐ์ง‘ํ•ฉ {ui}์ด ๊ธฐ์ €๋ฒกํ„ฐ๊ฐ€ ๋˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์„œ๋กœ ์„ ํ˜• ๋…๋ฆฝ์ด์–ด์•ผ ํ•จ.
๏ƒ˜๋‹ค์Œ ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋ฉด ๊ธฐ์ € {ui}๋Š” ์ง๊ต
๏ƒ˜๋‹ค์Œ ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋ฉด ๊ธฐ์ € {ui}๋Š” ์ •๊ทœ์ง๊ต
» ์ง๊ฐ์ขŒํ‘œ๊ณ„์˜ ๋‹จ์œ„๋ฒกํ„ฐ
23
01_๋ฒกํ„ฐ ์ด๋ก 
24
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๊ทธ๋žจ-์Šˆ๋ฏธํŠธ ์ •๊ทœ ์ง๊ตํ™” (Gram-schmidt Orthonormalization)
๏‚ง ์„œ๋กœ ์„ ํ˜• ๋…๋ฆฝ์ธ n๊ฐœ์˜ ๊ธฐ์ € ๋ฒกํ„ฐ {x1, x2, …, xm} ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ •๊ทœ์ง๊ต(orthonormal)
๊ธฐ์ € ๋ฒกํ„ฐ ์ง‘ํ•ฉ {p1, p2, …, pn} ์„ ๊ณ„์‚ฐํ•˜๋Š” ๊ณผ์ •
๏ƒ˜ ๊ธฐ์ € ๏ƒ  ์ง๊ต๊ธฐ์ €๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ณผ์ •
๏‚ง {v1, …, vn} ์„ ๋ฒกํ„ฐ๊ณต๊ฐ„ V์˜ ๊ธฐ์ €(basis)๋ผ๊ณ  ํ•˜๋ฉด, ์ง๊ต ๋ฒกํ„ฐ ์ง‘ํ•ฉ {u1, …, un}์€ ๋‹ค์Œ
๊ด€๊ณ„๋กœ๋ถ€ํ„ฐ ๊ณ„์‚ฐ ๊ฐ€๋Šฅ
๏ƒ˜ V์— ๋Œ€ํ•œ ์ •๊ทœ์ง๊ต๊ธฐ์ €(orthonormal basis)๋Š” ๊ฐ๊ฐ์˜ ๋ฒกํ„ฐ u1, …, un ์„ ์ •๊ทœํ™”ํ•˜๋ฉด ๋จ.
์ •๊ทœ ์ง๊ต ๋ฒกํ„ฐ
25
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๊ทธ๋žจ-์Šˆ๋ฏธํŠธ ์ •๊ทœ ์ง๊ตํ™” (Gram-schmidt Orthonormalization)
๏‚ง Example:
26
01_๋ฒกํ„ฐ ์ด๋ก 
๏‚ง ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ
๏ƒ˜ ๋ฒกํ„ฐ ๊ณต๊ฐ„์ƒ์—์„œ ๋‘ ์  ๊ฐ„์˜ ๊ฑฐ๋ฆฌ๋Š” ์  ์‚ฌ์ด ๋ฒกํ„ฐ ์ฐจ์˜ ํฌ๊ธฐ๋กœ ์ •์˜
27
01_๋ฒกํ„ฐ ์ด๋ก 
๏ถ ๋ฒกํ„ฐ ๊ณต๊ฐ„, ์œ ํด๋ฆฌ๋“œ ๊ณต๊ฐ„,ํ•จ์ˆ˜ ๊ณต๊ฐ„, ๋„ ๊ณต๊ฐ„ (null space)
๏‚ง ๋ชจ๋“  n์ฐจ์› ๋ฒกํ„ฐ๋“ค์ด ์กด์žฌํ•˜๋Š” n์ฐจ์› ๊ณต๊ฐ„
๏‚ง ์‹ค์ œ, ๋ฒกํ„ฐ ๊ณต๊ฐ„์€ ์‹ค์ˆ˜์— ์˜ํ•˜์—ฌ ๋ฒกํ„ฐ ๋ง์…ˆ๊ณผ ๊ณฑ์…ˆ์— ๋Œ€ํ•œ ๊ทœ์น™์— ๋‹ซํ˜€์žˆ๋Š” ๋ฒกํ„ฐ ์ง‘ํ•ฉ
๏‚ง ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ž„์˜์˜ ๋‘ ๋ฒกํ„ฐ์— ๋Œ€ํ•œ ๋ง์…ˆ๊ณผ ๊ณฑ์…ˆ์„ ํ†ตํ•˜์—ฌ ํ•ด๋‹น ๋ฒกํ„ฐ ๊ณต๊ฐ„ ๋‚ด์— ์žˆ๋Š” ์ƒˆ๋กœ์šด
๋ฒกํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Œ
๏‚ง ์ฆ‰, n์ฐจ์› ๊ณต๊ฐ„ Rn ์€ ๋ชจ๋‘ ์„ ํ˜•๋…๋ฆฝ์ธ n๊ฐœ์˜ n์ฐจ์› ๋ฒกํ„ฐ์— ์˜ํ•ด ์ƒ์„ฑ๋  ์ˆ˜ ์žˆ์Œ
๏‚ง ์ด ๋•Œ n์ฐจ์› ๊ณต๊ฐ„ Rn ์„ '์œ ํด๋ฆฌ๋“œ n์ฐจ์› ๊ณต๊ฐ„' ํ˜น์€ '์œ ํด๋ฆฌ๋“œ ๊ณต๊ฐ„'์ด๋ผ๊ณ  ํ•จ
๏‚ง ๋ฒกํ„ฐ์˜ ์ฐจ์›์ด ๋ฌดํ•œ๋Œ€์ผ ๊ฒฝ์šฐ, ๋ฒกํ„ฐ ๊ณต๊ฐ„์€ ‘ํ•จ์ˆ˜ ๊ณต๊ฐ„’์ด ๋จ
๏‚ง ํ–‰๋ ฌ A์˜ ๋„ ๊ณต๊ฐ„์€ Ax = 0๋ฅผ ๋งŒ์กฑํ•˜๋Š” ๋ชจ๋“  ๋ฒกํ„ฐ x๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋Š” ๊ณต๊ฐ„์„ ๋œปํ•œ๋‹ค.
28
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ์ „์น˜ํ–‰๋ ฌ (transpose)
๏ถ ์ •๋ฐฉํ–‰๋ ฌ (square matrix)
๏‚ง ํ–‰์˜ ์ˆ˜์™€ ์—ด์˜ ์ˆ˜๊ฐ€ ๋™์ผํ•œ ํ–‰๋ ฌ
29
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๋Œ€๊ฐ ํ–‰๋ ฌ (diagonal matrix)
๏‚ง ํ–‰๋ ฌ์˜ ๋Œ€๊ฐ ์„ฑ๋ถ„์„ ์ œ์™ธํ•˜๊ณ ๋Š” ๋ชจ๋‘ 0์ธ ํ–‰๋ ฌ
๏‚ง ์Šค์นผ๋ผ ํ–‰๋ ฌ : ๋Œ€๊ฐ ์„ฑ๋ถ„์ด ๋ชจ๋‘ ๊ฐ™๊ณ , ๋น„๋Œ€๊ฐ ์„ฑ๋ถ„์ด ๋ชจ๋‘ 0์ธ ์ •๋ฐฉํ–‰๋ ฌ
๏ถ ํ•ญ๋“ฑ ํ–‰๋ ฌ ํ˜น์€ ๋‹จ์œ„ ํ–‰๋ ฌ (identity matrix)
๏‚ง ๋Œ€๊ฐ ์„ฑ๋ถ„์ด ๋ชจ๋‘ 1์ด๊ณ  ๊ทธ๋ฐ–์˜ ์„ฑ๋ถ„์ด ๋ชจ๋‘ 0์ธ ์ •๋ฐฉํ–‰๋ ฌ
30
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๋Œ€์นญ ํ–‰๋ ฌ (symmetric matrix)
๋Œ€์นญํ–‰๋ ฌ ์˜ˆ: ๊ณต๋ถ„์‚ฐํ–‰๋ ฌ
๏‚ง ๋Œ€๊ฐ์„ ์„ ์ถ•์œผ๋กœ ๋ชจ๋“  ์„ฑ๋ถ„์ด ๋Œ€์นญ๋˜๋Š” ํ–‰๋ ฌ
๏ถ ์˜ ํ–‰๋ ฌ
๏‚ง ๋ชจ๋“  ๊ตฌ์„ฑ ์„ฑ๋ถ„์ด 0์ธ ํ–‰๋ ฌ
๏ถ ์ง๊ต ํ–‰๋ ฌ (orthogonal matrix)
๏‚ง ์ฃผ์–ด์ง„ ํ–‰๋ ฌ A๊ฐ€ ์ •๋ฐฉํ–‰๋ ฌ์ผ ๋•Œ,
๏‚ง ํ–‰๋ ฌ์˜ ๊ฐ ์—ด(column)์ด ์„œ๋กœ ์ง๊ต
๏‚ง ํšŒ์ „ ๋ณ€ํ™˜๊ณผ ๊ด€๊ณ„ ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Œ.
๋ฅผ ๋งŒ์กฑํ•˜๋Š” ํ–‰๋ ฌ
For an
orthogonal
matrix,
U U ๏€ฝI
T
Det = 1:
rotational transformation
Det = -1:
reflective transformation,
or axis permutation.
์ •๋ฐฉํ–‰๋ ฌ A์˜ ๊ฐ ํ–‰๋ฒกํ„ฐ(๋˜๋Š” ์—ด๋ฒกํ„ฐ)๋“ค์ด
์ƒํ˜ธ์ง๊ต์ธ ๋‹จ์œ„๋ฒกํ„ฐ(orthonormal vector)๋กœ ์ด๋ฃจ์–ด์ง.
31
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ํ–‰๋ ฌ์˜ ๊ณฑ์…ˆ
๏ถ ํ–‰๋ ฌ์˜ ํŠธ๋ ˆ์ด์Šค(trace) – ์ •๋ฐฉํ–‰๋ ฌ์—์„œ ๋Œ€๊ฐ ์„ฑ๋ถ„์˜ ํ•ฉ
๏ƒ˜ ํ–‰๋ ฌ์˜ ๊ณ ์œ ๊ฐ’ ๋ฌธ์ œ์—์„œ ๊ณ ์œ ๊ทผ์„ ๊ตฌํ•  ๋•Œ ๋งค์šฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•จ
32
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ํ–‰๋ ฌ์˜ ๊ณ„์ˆ˜(rank)
๏‚ง ํ–‰๋ ฌ์—์„œ ์„ ํ˜• ๋…๋ฆฝ์ธ ์—ด๋ฒกํ„ฐ(ํ˜น์€ ํ–‰๋ฒกํ„ฐ)์˜ ๊ฐœ์ˆ˜
๏‚ง ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ •๋ฐฉํ–‰๋ ฌ A๊ฐ€ ์ฃผ์–ด์งˆ ๊ฒฝ์šฐ,
๏‚ง ํ–‰๋ ฌ A ๋Š” ์„ธ ๊ฐœ์˜ ์—ด๋ฒกํ„ฐ e1, e2, e3 ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ A = (e1, e2, e3) ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Œ
๏‚ง ํ–‰๋ ฌ A ์˜ ๊ณ„์ˆ˜๋Š” ์ •์˜์— ์˜ํ•ด ์ด๋“ค ์—ด๋ฒกํ„ฐ ์ค‘์—์„œ ์„ ํ˜• ๋…๋ฆฝ์ธ ๋ฒกํ„ฐ์˜ ๊ฐœ์ˆ˜๋ฅผ ๋งํ•จ
๏ƒ˜ A ํ–‰๋ ฌ์€ ์„ธ ๋ฒกํ„ฐ๊ฐ€ ๋ชจ๋‘ ๋‹จ์œ„ ๋ฒกํ„ฐ์ด๊ณ  ๋ชจ๋‘ ์„ ํ˜• ๋…๋ฆฝ์ด๋ฏ€๋กœ rank(A) = 3
๏‚ง ํ–‰๋ ฌ์˜ ๊ณ„์ˆ˜๋Š” ์ฃผ์–ด์ง„ ํ–‰๋ ฌ์˜ ํ–‰์˜ ์ˆ˜๋‚˜ ์—ด์˜ ์ˆ˜๋ณด๋‹ค ํด ์ˆ˜ ์—†์Œ
๏ƒ˜ rank(An × n) = n ๏ƒจ ํ–‰๋ ฌ A๋Š” ๋น„ํŠน์ด(nonsingular)ํ–‰๋ ฌ ํ˜น์€ ์ •์น™ํ–‰๋ ฌ
๏ƒ˜ rank(An × n) < n ๏ƒจ ํ–‰๋ ฌ A๋Š” ํŠน์ด(singular)ํ–‰๋ ฌ
» ์—ญํ–‰๋ ฌ์ด ์กด์žฌํ•˜์ง€ ์•Š์Œ. (not invertible, rank deficient, degenerate, etc…)
33
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ํ–‰๋ ฌ์‹ (determinant)
๋˜๋Š” ํ–‰๋ ฌ๊ฐ’
๏‚ง ํ–‰๋ ฌ์‹์€ ํ–‰๋ ฌ์„ ์–ด๋– ํ•œ ํ•˜๋‚˜์˜ ์‹ค์ˆ˜ ๊ฐ’์œผ๋กœ ํ‘œํ˜„ํ•œ ๊ฒƒ์„ ๋งํ•จ
๏‚ง d×d ์ •๋ฐฉ ํ–‰๋ ฌ A์— ๋Œ€ํ•ด ํ–‰๋ ฌ์‹์€ |A| ํ˜น์€ det A ์œผ๋กœ ํ‘œํ˜„ํ•˜๋ฉฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์„ฑ์งˆ์„ ๊ฐ€์ง
๏ƒ˜ ํ–‰๋ ฌ์‹์€ ์˜ค์ง ์ •๋ฐฉ ํ–‰๋ ฌ์—์„œ๋งŒ ์ •์˜๋œ๋‹ค.
๏ƒ˜ ๊ตฌ์„ฑ ์„ฑ๋ถ„์ด ํ•˜๋‚˜์ธ ํ–‰๋ ฌ์˜ ํ–‰๋ ฌ์‹์€ ๊ทธ ์„ฑ๋ถ„ ์ž์ฒด์ด๋‹ค.
๏ƒ˜ ํ–‰๋ ฌ์‹์˜ ๊ฐ’์€ ํ•˜๋‚˜์˜ ์ƒ์ˆ˜ ์ฆ‰, ์ž„์˜์˜ ์‹ค์ˆ˜์ด๋‹ค.
๏ƒ˜ n์ฐจ์˜ ํ–‰๋ ฌ์‹ |An×n| ์€ n๊ฐœ์˜ ํ–‰๊ณผ ์—ด์˜ ์œ„์น˜๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ์„ฑ๋ถ„๋“ค์˜ ๊ณฑ์˜ ํ•ฉ์œผ๋กœ
ํ‘œํ˜„๋œ๋‹ค.
» 2x2 ํ–‰๋ ฌ์˜ ํ–‰๋ ฌ์‹
» 3x3 ํ–‰๋ ฌ์˜ ํ–‰๋ ฌ์‹
34
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏‚ง ์†Œํ–‰๋ ฌ์‹ (minor)
๏‚ง ํ–‰๋ ฌ์—์„œ i๋ฒˆ์งธ ์—ด๊ณผ j๋ฒˆ์งธ ํ–‰์„ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ ์–ป๋Š” ํ–‰๋ ฌ
๏‚ง ํ–‰๋ ฌ์‹ ๊ณ„์‚ฐ์˜ ์ผ๋ฐ˜ํ™”
๏‚ง ์—ฌ๊ธฐ์„œ |Mij| ๋ฅผ i๋ฒˆ์งธ ์—ด๊ณผ j๋ฒˆ์งธ ํ–‰์„ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ ์–ป์–ด์ง€๋Š” ์†Œํ–‰๋ ฌ์‹์ด๋ผ๊ณ  ํ•จ
๏‚ง ์ž„์˜์˜ ํ–‰์ด๋‚˜ ์—ด์„ ์ค‘์‹ฌ์œผ๋กœ ์ „๊ฐœํ•˜์—ฌ๋„ ๊ฒฐ๊ณผ๋Š” ๊ฐ™์Œ ๏ƒจ ๋ผํ”Œ๋ผ์Šค(Laplace) ์ „๊ฐœ
๏‚ง ๋ผํ”Œ๋ผ์Šค ์ „๊ฐœ ์‹œ ๋ถ€ํ˜ธ
๏ƒ˜ aij ์˜ ์•„๋ž˜์ฒจ์ž ํ˜น์€ ์†Œํ–‰๋ ฌ์‹ |Mij|์˜ ์•„๋ž˜์ฒจ์ž์˜ ํ•ฉ์ด ์ง์ˆ˜๋ฉด ๏ผ‹, ํ™€์ˆ˜๋ฉด ๏ผ๊ฐ€ ๋จ
๏‚ง ์†Œํ–‰๋ ฌ์‹ |Mij|์— ๋ถ€ํ˜ธ ๋ถ€๋ถ„ (-1)i+j ๊นŒ์ง€ ๊ณฑํ•œ ํ•ญ์„ ์—ฌ์ธ์ˆ˜ Aij๋ผ๊ณ  ํ•จ
35
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏‚ง Ai | j ๋ฅผ d x d ํ–‰๋ ฌ์ด๋ผ๊ณ  ํ•  ๋•Œ ์†Œํ–‰๋ ฌ์‹์œผ๋กœ A์˜ ํ–‰๋ ฌ์‹์€ ์ˆœํ™˜์ ์œผ๋กœ ๊ตฌํ•  ์ˆ˜ ์žˆ์Œ
๏ƒ˜ i๋ฒˆ์งธ ํ–‰์„ ์ค‘์‹ฌ์œผ๋กœ ์ „๊ฐœ
๏‚ง ํ–‰๋ ฌ์‹์˜ ์„ฑ์งˆ
๏‚ง ์‚ผ๊ฐํ–‰๋ ฌ์˜ ํ–‰๋ ฌ์‹์˜ ๊ฐ’์€ ๋Œ€๊ฐ ์„ฑ๋ถ„์˜ ๊ณฑ
๏‚ง ์ „์น˜ํ–‰๋ ฌ์˜ ํ–‰๋ ฌ์‹์€ ์›๋ž˜ ํ–‰๋ ฌ์˜ ํ–‰๋ ฌ์‹๊ณผ ๊ฐ™๋‹ค.
36
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ์—ญํ–‰๋ ฌ (inverse matrix)
๏‚ง ๋Œ€์ˆ˜ ์—ฐ์‚ฐ์—์„œ ์ž„์˜์˜ ์ˆ˜๋ฅผ ๊ณฑํ•˜์—ฌ 1์ด ๋  ๋•Œ, ์ด๋ฅผ '์—ญ์ˆ˜'๋ผ๊ณ  ํ•จ
๏‚ง ์—ญ์ˆ˜๋ฅผ ํ–‰๋ ฌ ๋Œ€์ˆ˜์— ์ ์šฉํ–ˆ์„ ๋•Œ, AX = I ๊ฐ€ ๋˜๋Š” X๊ฐ€ ์กด์žฌํ•  ๊ฒฝ์šฐ์— ์ด ํ–‰๋ ฌ X๋ฅผ A์˜
์—ญํ–‰๋ ฌ์ด๋ผ๊ณ  ํ•˜๋ฉฐ A-1 ๋กœ ํ‘œํ˜„ํ•œ๋‹ค.
๏ถ ์—ญํ–‰๋ ฌ์˜ ์„ฑ์งˆ
37
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๊ณ ์œ ๊ฐ’๊ณผ ๊ณ ์œ ๋ฒกํ„ฐ (eigenvalues and eigenvectors)
๏‚ง ํ–‰๋ ฌ A๊ฐ€ n×n์˜ ์ •๋ฐฉ ํ–‰๋ ฌ์ด๊ณ , x ≠ 0์ธ ๋ฒกํ„ฐ x ∈ Rn ๊ฐ€ ์กด์žฌํ•  ๋•Œ
๏‚ง ๋‹ค์Œ ๊ด€๊ณ„๋ฅผ ๋งŒ์กฑํ•˜๋Š” ์Šค์นผ๋ผ λ๋ฅผ ํ–‰๋ ฌ A์˜ ๊ณ ์œ ๊ฐ’์ด๋ผ๊ณ  ํ•จ
๏ƒ˜ ๋ฒกํ„ฐ x๋Š” λ์— ๋Œ€์‘ํ•˜๋Š” A์˜ ๊ณ ์œ  ๋ฒกํ„ฐ๋ผ๊ณ  ํ•จ
Ax ๏€ฝ ๏ฌ x
๏‚ง ๊ณ ์œ ๊ฐ’์˜ ๊ณ„์‚ฐ
A x ๏€ฝ ๏ฌ x ๏ƒž A x ๏€ญ ๏ฌ x ๏€ฝ 0 ๏ƒž ( A ๏€ญ ๏ฌ I ) x ๏€ฝ 0 ๏ƒž x ๏€ฝ 0 or ( A ๏€ญ ๏ฌ I ) x ๏€ฝ 0
๏‚ง ๋™์ฐจ์ผ์ฐจ ์—ฐ๋ฆฝ๋ฐฉ์ ์‹ Ax = 0์—์„œ x = 0์ด ์•„๋‹Œ ํ•ด๋ฅผ ์–ป๋Š” ์œ ์ผํ•œ ๊ฒฝ์šฐ๋Š” |A| = 0์ธ ๊ฒฝ์šฐ
๏‚ง ๋”ฐ๋ผ์„œ ์œ„ ์‹์„ ๋งŒ์กฑํ•˜๋ ค๋ฉด |A –λI| = 0์ผ ๋•Œ x ≠0 ์ธ ํ•ด๊ฐ€ ์กด์žฌํ•˜๊ฒŒ ๋œ๋‹ค.
๏‚ง ์ด๋•Œ |A –λI| = 0์ด๋ผ๋Š” ์‹์„ A์˜ 'ํŠน์„ฑ ๋ฐฉ์ •์‹'์ด๋ผ๊ณ  ํ•จ
(A ๏€ญ ๏ฌI) ๏€ฝ 0 ๏ƒž A ๏€ญ ๏ฌI ๏€ฝ 0 ๏ƒž ๏ฌ
N
๏€ซ a1 ๏ฌ
N ๏€ญ1
๏€ซ ... ๏€ซ a N ๏€ญ1 ๏ฌ ๏€ซ a 0 ๏€ฝ 0
๏ƒ˜ A๋ฅผ n×n ํ–‰๋ ฌ์ด๋ผ ํ•˜๊ณ , λ๋ฅผ A์˜ ๊ณ ์œ ๊ฐ’์ด๋ผ๊ณ  ํ•œ๋‹ค.
๏ƒ˜ N ๊ฐœ์˜ ๊ณ ์œ ๊ฐ’๊ณผ ๊ณ ์œ  ๋ฒกํ„ฐ๋ฅผ ๊ตฌํ•  ์ˆ˜ ์žˆ๋‹ค.
๏ƒ˜ ๊ณ ์œ ๋ฒกํ„ฐ๋กœ ์ •์˜๋˜๋Š” ๋ถ€๋ถ„ ๊ณต๊ฐ„์„ A์˜ ๊ณ ์œ  ๊ณต๊ฐ„์ด๋ผ๊ณ  ํ•œ๋‹ค.
38
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๊ณ ์œ ๊ฐ’๊ณผ ๊ณ ์œ ๋ฒกํ„ฐ (eigenvalues and eigenvectors)
๏‚ง ๊ธฐํ•˜ํ•™์  ์˜๋ฏธ
ํ–‰๋ ฌ(์„ ํ˜•๋ณ€ํ™˜) A์˜ ๊ณ ์œ ๋ฒกํ„ฐ๋Š” ์„ ํ˜•๋ณ€ํ™˜ A์— ์˜ํ•ด ๋ฐฉํ–ฅ์€ ๋ณด์กด๋˜๊ณ  ์Šค์ผ€
์ผ(scale)๋งŒ ๋ณ€ํ™”๋˜๋Š” ๋ฐฉํ–ฅ ๋ฒกํ„ฐ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , ๊ณ ์œ ๊ฐ’์€ ๊ทธ ๊ณ ์œ ๋ฒกํ„ฐ์˜ ๋ณ€ํ™”๋˜
๋Š” ์Šค์ผ€์ผ ์ •๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฐ’.
์˜ˆ) ์ง€๊ตฌ์˜ ์ž์ „์šด๋™๊ณผ ๊ฐ™์ด 3์ฐจ์› ํšŒ์ „๋ณ€ํ™˜์„ ์ƒ๊ฐํ–ˆ์„ ๋•Œ, ์ด ํšŒ์ „๋ณ€ํ™˜์—
์˜ํ•ด ๋ณ€ํ•˜์ง€ ์•Š๋Š” ๊ณ ์œ ๋ฒกํ„ฐ๋Š” ํšŒ์ „์ถ• ๋ฒกํ„ฐ์ด๊ณ  ๊ทธ ๊ณ ์œ ๊ฐ’์€ 1
39
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๊ณ ์œ ๊ฐ’๊ณผ ๊ณ ์œ ๋ฒกํ„ฐ (eigenvalues and eigenvectors)
๏‚ง Application example
In this shear mapping, the red arrow changes direction but the blue arrow does not.
Therefore the blue arrow is an eigenvector, with eigenvalue 1 as its length is
unchanged
40
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๊ณ ์œ ๊ฐ’๊ณผ ๊ณ ์œ ๋ฒกํ„ฐ์˜ ์„ฑ์งˆ
๏‚ง ๋Œ€๊ฐ ํ–‰๋ ฌ์˜ ๊ณ ์œ ๊ฐ’์€ ๋Œ€๊ฐ ์„ฑ๋ถ„ ๊ฐ’์ด๋‹ค
๏‚ง ์‚ผ๊ฐ ํ–‰๋ ฌ์˜ ๊ณ ์œ ๊ฐ’์€ ์ด ํ–‰๋ ฌ ๋Œ€๊ฐ ์„ฑ๋ถ„ ๊ฐ’์ด๋‹ค
๏‚ง ๋ฒกํ„ฐ x๊ฐ€ ํ–‰๋ ฌ A์˜ ๊ณ ์œ  ๋ฒกํ„ฐ์ด๋ฉด ๋ฒกํ„ฐ x์˜ ์Šค์นผ๋ผ ๊ณฑ์ธ kx๋„ ๊ณ ์œ  ๋ฒกํ„ฐ์ด๋‹ค
๏‚ง ์ „์น˜ํ•˜์—ฌ๋„ ๊ณ ์œ ๊ฐ’์€ ๋ณ€ํ•˜์ง€ ์•Š๋Š”๋‹ค.
๏ƒ˜ ํ–‰๋ ฌ A์˜ ๊ณ ์œ ๊ฐ’๊ณผ ์ „์น˜ ํ–‰๋ ฌ AT ์˜ ๊ณ ์œ ๊ฐ’์€ ๋™์ผํ•˜๋‹ค
๏‚ง ์—ญํ–‰๋ ฌ์˜ ๊ณ ์œ ๊ฐ’์€ ์›๋ž˜ ํ–‰๋ ฌ์˜ ๊ณ ์œ ๊ฐ’์˜ ์—ญ์ˆ˜๊ฐ€ ๋œ๋‹ค.
๏‚ง ํ–‰๋ ฌ A์˜ ๋ชจ๋“  ๊ณ ์œ ๊ฐ’์˜ ๊ณฑ์€ A์˜ ํ–‰๋ ฌ์‹๊ณผ ๊ฐ™๋‹ค
๏‚ง ์„œ๋กœ ๋‹ค๋ฅธ ๊ณ ์œ ๊ฐ’๊ณผ ๊ด€๋ จ๋œ ๊ณ ์œ  ๋ฒกํ„ฐ๋Š” ์„ ํ˜• ๋…๋ฆฝ์ด๋‹ค.
๏‚ง ์‹ค์ˆ˜ ๋Œ€์นญํ–‰๋ ฌ์˜ ๊ณ ์œ  ๋ฒกํ„ฐ๋Š” ์„œ๋กœ ์ง๊ตํ•œ๋‹ค.
๏‚ง ์‹ค์ˆ˜ ๋Œ€์นญ ํ–‰๋ ฌ์˜ ๊ณ ์œ ๊ฐ’ ๋˜ํ•œ ์‹ค์ˆ˜์ด๋‹ค.
๏‚ง ๋งŒ์•ฝ A๊ฐ€ ์–‘์˜ ์ •๋ถ€ํ˜ธ ํ–‰๋ ฌ์ด๋ผ๋ฉด ๋ชจ๋“  ๊ณ ์œ ๊ฐ’์€ ์–‘์ˆ˜์ด๋‹ค.
41
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ๊ณ ์œ ๊ฐ’๊ณผ ๊ณ ์œ ๋ฒกํ„ฐ (eigenvalues and eigenvectors)
42
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏‚ง ๋Œ€๊ฐํ™”์™€ ํŠน์ด๋ฒกํ„ฐ, ํŠน์ด๊ฐ’
๏‚ง A์˜ ๊ณ ์œ ๊ฐ’์ด λ1, …, λn ,์ด์— ๋Œ€์‘ํ•˜๋Š” 1์ฐจ ๋…๋ฆฝ์ธ ๊ณ ์œ ๋ฒกํ„ฐ๊ฐ€ v1, …, vn ์ด๋ผ๊ณ  ํ•  ๋•Œ, C๋ฅผ
๋‹ค์Œ๊ณผ ๊ฐ™์ด v1, …, vn ์„ ์—ด๋ฒกํ„ฐ๋กœ ํ•˜๋Š” ํ–‰๋ ฌ์ด๋ผ๊ณ  ํ•˜์ž.
๏‚ง Avn = λnvn ์ด๋ฏ€๋กœ, ํ–‰๋ ฌ ๊ณฑ์…ˆ์„ ์—ด๋กœ ํ‘œํ˜„ํ•˜๋ฉด ๋‹ค์Œ์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค.
AC = CΛ → A = CΛC-1
: ํ–‰๋ ฌ A๋Š” ์ž์‹ ์˜ ๊ณ ์œ ๋ฒกํ„ฐ๋“ค์„ ์—ด๋ฒกํ„ฐ๋กœ ํ•˜๋Š” ํ–‰๋ ฌ๊ณผ ๊ณ ์œ ๊ฐ’์„ ๋Œ€
๊ฐ์›์†Œ๋กœ ํ•˜๋Š” ํ–‰๋ ฌ์˜ ๊ณฑ์œผ๋กœ ๋Œ€๊ฐํ™” ๋ถ„ํ•ด ๊ฐ€๋Šฅ = eigen
decomposition
ํ–‰๋ ฌ A์˜ ๋Œ€๊ฐํ™”
๏‚ง n×nํ–‰๋ ฌ A๊ฐ€ n๊ฐœ์˜ 1์ฐจ ๋…๋ฆฝ์ธ ๊ณ ์œ ๋ฒกํ„ฐ๋ฅผ ๊ฐ€์ง„๋‹ค๋ฉด, ์ด ๊ณ ์œ ๋ฒกํ„ฐ๋“ค์€ A๋ฅผ ๋Œ€๊ฐํ™”ํ•˜๋Š” ํ–‰๋ ฌ
C์˜ ์—ด๋“ค๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋Œ€๊ฐํ–‰๋ ฌ์€ A์˜ ๊ณ ์œ ๊ฐ’์„ ๋Œ€๊ฐ์›์†Œ๋กœ ๊ฐ€์ง„๋‹ค.
43
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ 2์ฐจ ํ˜•์‹
๏‚ง 3๊ฐœ์˜ ๋ณ€์ˆ˜ x, y, z์˜ 2์ฐจ ํ˜•์‹(quadratic form)์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋™์ฐจ ํ•จ์ˆ˜์‹์„ ๋งํ•จ
๏‚ง 2์ฐจ ํ•จ์ˆ˜์ด๊ธฐ ๋•Œ๋ฌธ์— 2์ฐจ ํ˜•์‹์ด๋ผ๊ณ  ํ•œ๋‹ค.
F = ax2 + by2 + cz2 + 2fxy + 2gyz + 2hzx
๏‚ง ํ–‰๋ ฌ์„ ์ด์šฉํ•˜์—ฌ ํ‘œ์‹œํ•˜๋ฉด
๏‚ง ์ผ๋ฐ˜ํ™”ํ•˜๋ฉด
๏‚ง x=(x1, x2, …, xn), A = (aij)๋ผ๊ณ  ํ•  ๋•Œ
44
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ SVD : ํŠน์ด๊ฐ’ ์žฌ๊ตฌ์„ฑ (Singular Value Decomposition)
๏‚ง ์–ด๋–ค n×m ํ–‰๋ ฌ A๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ˜•ํƒœ์˜ ์„ธ ๊ฐ€์ง€ ํ–‰๋ ฌ์˜ ๊ณฑ์œผ๋กœ ์žฌ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค
๏‚ง U๋Š” ํŠน์ด ๋ฒกํ„ฐ๋ฅผ ์ด๋ฃจ๋Š” ์—ด๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, VT๋Š” ํŠน์ด ๋ฒกํ„ฐ๋ฅผ ์ด๋ฃจ๋Š” ํ–‰์œผ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค.
๏ƒ˜ m×mํ–‰๋ ฌ U์˜ ํ–‰์€ AAT ์˜ ๊ณ ์œ  ๋ฒกํ„ฐ.
๏ƒ˜ n×n ํ–‰๋ ฌ V์˜ ํ–‰์€ ATA ์˜ ๊ณ ์œ  ๋ฒกํ„ฐ.
๏‚ง Σ์€ n×m ์ธ ๋Œ€๊ฐํ–‰๋ ฌ๋กœ, ๋Œ€๊ฐ์„ฑ๋ถ„์€ 0์ด ์•„๋‹Œ ์–‘์ˆ˜๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์ด ๋Œ€๊ฐ์„ฑ๋ถ„์„
ํŠน์ด๊ฐ’์ด๋ผ๊ณ  ํ•œ๋‹ค.
๏‚ง Σ์˜ ํŠน์ด๊ฐ’๋“ค์€ AAT์™€ ATA ์˜ ๊ณ ์œ ๊ฐ’์˜ ์ž์Šน๊ทผ์— ํ•ด๋‹นํ•œ๋‹ค.
๏‚ง ์˜์ด ์•„๋‹Œ ํŠน์ด๊ฐ’์˜ ์ˆ˜๋Š” ํ–‰๋ ฌ A์˜ ํ–‰๋ ฌ์˜ ๊ณ„์ˆ˜(rank)์™€ ๊ฐ™๋‹ค.
45
† Singular
Value Decomposition
A matrix A can be factorized as the following form:
A m ๏‚ด n ๏€ฝ U m ๏‚ด m Σ m ๏‚ด n Vn๏‚ด n
T
๏ƒฉ๏ณ 1
๏ƒช
0
๏ƒช
๏ƒช ๏
Σ ๏€ฝ ๏ƒช
๏ƒช0
๏ƒช ๏
๏ƒช
๏ƒช๏ƒซ 0
0
๏Œ
๏ณ2
๏Œ
๏
๏
0
๏Œ
๏
0
๏Œ
0 ๏ƒน
๏ƒบ
0
๏ƒบ
๏ ๏ƒบ
๏ƒบ
๏ณn๏ƒบ
๏ ๏ƒบ
๏ƒบ
0 ๏ƒบ๏ƒป
(m ๏‚ณ n)
U and V are orthogonal, and Σ is
diagonal.
๏ณ1 ๏‚ณ ๏ณ 2 ๏‚ณ ๏Œ ๏‚ณ ๏ณ n ๏‚ณ 0
U U ๏€ฝI
T
det = 1:
rotational transformation
V V ๏€ฝI
T
det = -1:
(orthogonal) reflective transformation,
or axis permutation.
46
์ฐธ์กฐ: Singular Value Decomposition (SVD)
๏ถ A rectangular matrix A can be broken down into the product of three
matrices: an orthogonal matrix U, a diagonal matrix S, and the transpose of
an orthogonal matrix V.
Amn = UmmSmnVnnT
, where UTU = I, VTV = I ; the coloumns of U are orthonormal eigenvectors
of AAT, the columns of V are orthonormal eigenvectors of ATA. S is a
diagonal matrix containing the square roots of eigenvalues from U or V in
descending order.
๏ถ Example
๏ƒฉ 3
A๏€ฝ๏ƒช
๏ƒซ๏€ญ 1
1
3
1๏ƒน
๏ƒบ
1๏ƒป
- To find U, we have to start with AAT.
๏ƒฉ3
๏ƒช
T
A ๏€ฝ 1
๏ƒช
๏ƒช๏ƒซ 1
๏€ญ 1๏ƒน
๏ƒบ
3 ,
๏ƒบ
1 ๏ƒบ๏ƒป
AA
T
๏ƒฉ 3
๏€ฝ ๏ƒช
๏ƒซ๏€ญ 1
1
3
๏ƒฉ3
1๏ƒน ๏ƒช
๏ƒบ 1
1๏ƒป ๏ƒช
๏ƒช๏ƒซ 1
๏€ญ 1๏ƒน
๏ƒบ ๏ƒฉ11
3 ๏€ฝ ๏ƒช
๏ƒบ
๏ƒซ1
1 ๏ƒบ๏ƒป
1๏ƒน
๏ƒบ
11 ๏ƒป
47
์ฐธ์กฐ: Singular Value Decomposition (SVD)
- To find the eigenvalues & corresponding eigenvectors of AAT.
๏€จ11 ๏€ญ ๏ฌ ๏€ฉ x1 ๏€ซ x 2
11 x1 ๏€ซ x 2 ๏€ฝ ๏ฌ x1
๏ƒฉ x1 ๏ƒน
๏ƒฉ11 1 ๏ƒน ๏ƒฉ x1 ๏ƒน
๏ƒช
๏ƒบ๏ƒช ๏ƒบ ๏€ฝ ๏ฌ ๏ƒช ๏ƒบ
x1 ๏€ซ 11 x 2 ๏€ฝ ๏ฌ x 2
x1 ๏€ซ ๏€จ11 ๏€ญ ๏ฌ ๏€ฉ x 2
๏ƒซ 1 11 ๏ƒป ๏ƒซ x 2 ๏ƒป
๏ƒซ x2 ๏ƒป
๏€จ11 ๏€ญ ๏ฌ ๏€ฉ
1
1
๏€จ11 ๏€ญ ๏ฌ ๏€ฉ
๏€ฝ0
๏€จ11 ๏€ญ ๏ฌ ๏€ฉ๏€จ11 ๏€ญ ๏ฌ ๏€ฉ ๏€ญ 1 ๏ƒ— 1 ๏€ฝ 0
๏‚ฎ ๏€จ๏ฌ ๏€ญ 10 ๏€ฉ๏€จ๏ฌ ๏€ญ 12 ๏€ฉ ๏€ฝ 0
๏€ฝ0
- For λ= 10, ๏€จ11 ๏€ญ 10 ๏€ฉ x1 ๏€ซ x 2 ๏€ฝ 0
For λ= 12, ๏€จ11 ๏€ญ 12 ๏€ฉ x1 ๏€ซ x 2 ๏€ฝ 0
๏€ฝ0
๏‚ฎ
x1 ๏€ฝ ๏€ญ x 2
๏‚ฎ
x1 ๏€ฝ x 2
๏‚ฎ
๏‚ฎ
[1, ๏€ญ 1]
[1, 1]
- These eigenvectors become column vectors in a matrix
ordered by the size of the corresponding eigenvalue.
๏ƒฉ1
๏ƒช
๏ƒซ1
1 ๏ƒน
๏ƒบ
๏€ญ 1๏ƒป
- Convert this matrix into an orthogonal matrix which we do by
applying the Gram-Schmidt orthonormalization process to the
column vectors.
48
์ฐธ์กฐ: Singular Value Decomposition (SVD)
1) Begin by normalizing v1.
๏ƒฉ 1
w 2 ๏€ฝ v 2 ๏€ญ u 1 ๏ƒ— v 2 ๏€ช u 1 ๏€ฝ ๏›1, ๏€ญ 1๏ ๏€ญ ๏ƒช
,
๏ƒซ 2
2) normalize
u2๏€ฝ
- The calculation of V:
w2
w2
๏€ฝ
v1
๏ƒฉ 1
๏€ฝ๏ƒช
,
2
2
๏ƒซ 2
1 ๏€ซ1
[1, 1]
1 ๏ƒน
๏ƒฉ 1
๏›
๏
๏ƒ—
1
,
๏€ญ
1
๏€ช
,
๏ƒบ
๏ƒช
2๏ƒป
๏ƒซ 2
๏ƒฉ 1
๏€ฝ ๏ƒช
,
๏ƒซ 2
๏ƒฉ3
๏ƒช
T
A A๏€ฝ 1
๏ƒช
๏ƒช๏ƒซ 1
v1
u1 ๏€ฝ
๏€ญ1 ๏ƒน
๏ƒบ
2๏ƒป
๏€ญ 1๏ƒน
๏ƒบ๏ƒฉ 3
3 ๏ƒช
๏ƒบ ๏€ญ1
๏ƒซ
1 ๏ƒบ๏ƒป
1
3
๏ƒฉ10
1๏ƒน ๏ƒช
๏ƒบ๏€ฝ 0
1๏ƒป ๏ƒช
๏ƒช๏ƒซ 2
0
1 ๏ƒน
๏ƒบ
2๏ƒป
1 ๏ƒน
๏ƒบ ๏€ฝ ๏›1, ๏€ญ 1๏ ๏€ญ [ 0 , 0 ] ๏€ฝ ๏›1, ๏€ญ 1๏
2๏ƒป
1 ๏ƒน
๏ƒฉ 1
๏ƒช
๏ƒบ
2
2
U ๏€ฝ ๏ƒช
๏ƒบ
1
๏€ญ1
๏ƒช
๏ƒบ
๏ƒช๏ƒซ 2
2 ๏ƒบ๏ƒป
2๏ƒน
๏ƒบ
4
๏ƒบ
2 ๏ƒบ๏ƒป
10
4
1) find the eigenvalues of ATA by
๏ƒฉ10
๏ƒช
0
๏ƒช
๏ƒช๏ƒซ 2
0
10
4
2 ๏ƒน ๏ƒฉ x1 ๏ƒน
๏ƒฉ x1 ๏ƒน
๏ƒบ๏ƒช ๏ƒบ
๏ƒช ๏ƒบ
4 x2 ๏€ฝ ๏ฌ x2
๏ƒบ๏ƒช ๏ƒบ
๏ƒช ๏ƒบ
๏ƒช๏ƒซ x 3 ๏ƒบ๏ƒป
2 ๏ƒบ๏ƒป ๏ƒช๏ƒซ x 3 ๏ƒบ๏ƒป
๏€จ10 ๏€ญ ๏ฌ ๏€ฉ
0
2
0
๏€จ10 ๏€ญ ๏ฌ ๏€ฉ
4
2
4
๏€จ2 ๏€ญ ๏ฌ ๏€ฉ
๏€ฝ0
λ = 0, 10, 12
49
์ฐธ์กฐ: Singular Value Decomposition (SVD)
2) λ = 12์ผ ๋•Œ, v1 = [1, 2, 1]
๏ƒฉ1
๏ƒช
2
๏ƒช
๏ƒช๏ƒซ 1
λ = 10์ผ ๋•Œ, v2 = [2, -1, 0]
λ = 0์ผ ๋•Œ, v3 = [1, 2, -5]
3) an orthonormal matrix
w 2 ๏€ฝ v 2 ๏€ญ u 1 ๏ƒ— v 2 ๏€ช u 1 ๏€ฝ ๏›2 ,
u1 ๏€ฝ
๏€ญ 1,
0 ๏,
v1
v1
u2 ๏€ฝ
๏ƒฉ๏€ญ 2
w 3 ๏€ฝ v 3 ๏€ญ u1 ๏ƒ— v 3 ๏€ช u1 ๏€ญ u 2 ๏ƒ— v 3 ๏€ช u 2 ๏€ฝ ๏ƒช
,
3
๏ƒซ
4) Amn = UmmSmnVnn
๏ƒฉ 1
๏ƒช
๏€ฝ ๏ƒช 2
1
๏ƒช
๏ƒช๏ƒซ 2
1 ๏ƒน
๏ƒบ
2 ๏ƒฉ 12
๏ƒบ
๏€ญ1 ๏ƒช 0
๏ƒบ๏ƒซ
2 ๏ƒบ๏ƒป
0
10
๏ƒฉ
๏ƒช
๏ƒช
0๏ƒน ๏ƒช
๏ƒบ
0๏ƒป ๏ƒช
๏ƒช
๏ƒช
๏ƒซ
๏ƒฉ 1
๏€ฝ๏ƒช
,
๏ƒซ 6
w2
๏€ญ4
3
T
1
2
6
2
6
๏€ญ1
5
1
5
2
30
30
1 ๏ƒน
๏ƒบ
6
๏ƒบ
๏ƒฉ 3
0 ๏ƒบ๏€ฝ ๏ƒช
๏ƒบ ๏ƒซ๏€ญ 1
๏€ญ5 ๏ƒบ
๏ƒบ
30 ๏ƒป
1
3
,
1๏ƒน
๏ƒบ
1๏ƒป
๏ƒฉ
๏ƒช
๏ƒช
V ๏€ฝ ๏ƒช
๏ƒช
๏ƒช
๏ƒช
๏ƒซ
๏€ญ1
0
2
1 ๏ƒน
๏ƒบ
6๏ƒป
,
6
๏€ญ1
๏ƒฉ 2
๏€ฝ ๏ƒช
,
๏ƒซ 5
w2
1 ๏ƒน
๏ƒบ
2
๏ƒบ
๏€ญ 5 ๏ƒบ๏ƒป
2
,
5
๏ƒน
0๏ƒบ
๏ƒป
w3
10 ๏ƒน
๏ƒฉ 1
,
u
๏€ฝ
๏€ฝ
,
3
๏ƒช
3 ๏ƒบ๏ƒป
w3
๏ƒซ 30
1
2
6
2
5
๏€ญ1
6
1
5
6
0
1 ๏ƒน
๏ƒบ
30
๏ƒบ
2 ๏ƒบ
,
30 ๏ƒบ
๏€ญ5 ๏ƒบ
๏ƒบ
30 ๏ƒป
V
T
๏ƒฉ
๏ƒช
๏ƒช
๏€ฝ ๏ƒช
๏ƒช
๏ƒช
๏ƒช
๏ƒซ
2
,
30
1
2
6
2
6
๏€ญ1
5
1
5
2
30
30
๏€ญ5 ๏ƒน
๏ƒบ
30 ๏ƒป
1 ๏ƒน
๏ƒบ
6
๏ƒบ
0 ๏ƒบ
๏ƒบ
๏€ญ5 ๏ƒบ
๏ƒบ
30 ๏ƒป
50
† Singular
Value Decomposition
Given a matrix A, it can be factorized as the following form:
A m ๏‚ด n ๏€ฝ U m ๏‚ด m Σ m ๏‚ด n Vn๏‚ด n
T
Example:
(m ๏‚ณ n)
Ax ๏€ฝ x'
x
For a matrix A such that :
Rotation by V:
x
T
V x
And then, rotated again by U:
U Σ V x ๏€ฝ Ax
T
Scaling by Σ :
T
ΣV x
51
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏ถ ์„ ํ˜• ๋ณ€ํ™˜
๋ฐ์ดํ„ฐ์˜ ๋ณ€ํ™˜
๏‚ง ๋ฒกํ„ฐ ๊ณต๊ฐ„ XN ์œผ๋กœ๋ถ€ํ„ฐ ๋ฒกํ„ฐ ๊ณต๊ฐ„ YM ์ƒ์œผ๋กœ์˜ ์‚ฌ์ƒ (mapping)
๏‚ง ๋ฒกํ„ฐ x ∈XN ๊ฐ€ ์ฃผ์–ด์งˆ ๋•Œ YM ์ƒ์— ๋Œ€์‘ ๋˜๋Š” ๋ฒกํ„ฐ y๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐํ•œ๋‹ค.
๏‚ง ์„ ํ˜• ๋ณ€ํ™˜์ด ์ด๋ฃจ์–ด์ง€๋Š” ๋‘ ๋ฒกํ„ฐ ๊ณต๊ฐ„์˜ ์ฐจ์›์ด ๊ฐ™์„ ํ•„์š”๋Š” ์—†๋‹ค.
๏ƒ˜ ์„ ํ˜• ๋ณ€ํ™˜ ํ–‰๋ ฌ์ด ์ •๋ฐฉํ–‰๋ ฌ A์ด๊ณ  AAT = ATA = I ์ผ ๋•Œ, ์ •๊ทœ์ง๊ตํ•œ๋‹ค๊ณ  ๋งํ•œ๋‹ค.
๏ƒฉ cos ๏ฑ
๏ƒช
๏ƒซ sin ๏ฑ
- sin ๏ฑ ๏ƒน
๏ƒบ,
cos ๏ฑ ๏ƒป
๏ƒฉ- 1
๏ƒช
๏ƒซ0
0๏ƒน
๏ƒบ
- 1๏ƒป
52
02_ํ–‰๋ ฌ ๋Œ€์ˆ˜
๏‚ง ์ •๊ทœ ์ง๊ต์ด๋ฉด AT=A-1 ์ด๋‹ค.
๏‚ง ์ •๊ทœ์ง๊ต ๋ณ€ํ™˜ํ•˜๊ฒŒ ๋˜๋ฉด, ๋‹ค์Œ ์‹์—์„œ ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ๋ฅผ ๋ณด์กดํ•˜๋Š” ์„ฑ์งˆ์„ ๊ฐ€์ง์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.
๏‚ง ์ •๊ทœ์ง๊ต ๋ณ€ํ™˜์˜ ํ–‰๋ฒกํ„ฐ (a1,a2,…,aN) ๋Š” ์ •๊ทœ ์ง๊ต ๊ธฐ์ € ๋ฒกํ„ฐ์ง‘ํ•ฉ์„ ํ˜•์„ฑํ•œ๋‹ค.
53
Download