본문으로 건너뛰기

Solving Equations

선형방정식의 해를 체계적으로 분석하는 방법과 이를 바탕으로 한 vector space, eigenvalue, Newton's method, inverse function과 implicit function의 핵심 개념을 정리합니다.

2.1 Row Reduction

1) Linear System

선형방정식계는 matrix equation으로 표현할 수 있다.

Ax=b.A\mathbf{x}=\mathbf{b}.

여기서

  • AA: coefficient matrix
  • x\mathbf{x}: vector of unknowns
  • b\mathbf{b}: vector of constants

2) Augmented Matrix

선형방정식계는 augmented matrix로 나타낼 수 있다.

[Ab].[A\mid\mathbf{b}].

Meaning: 방정식의 coefficient와 constant를 하나의 matrix로 묶어 계산한다.

3) Row Operations

해를 바꾸지 않는 기본 row operation은 세 가지이다.

  1. 한 row에 nonzero scalar를 곱한다.
  2. 한 row의 scalar multiple을 다른 row에 더한다.
  3. 두 row를 교환한다.

Row operation을 적용해도

Ax=bA\mathbf{x}=\mathbf{b}

의 solution set은 변하지 않는다.

4) Echelon Form

책에서 사용하는 Echelon Form은 다음 조건을 만족한다.

  1. 각 nonzero row의 첫 nonzero entry는 11이다.
  2. 아래 row의 pivotal 11은 위 row의 pivotal 11보다 오른쪽에 있다.
  3. pivotal 11이 있는 column의 다른 entry는 모두 00이다.
  4. zero row는 모두 아래에 위치한다.

Meaning: 방정식의 해 구조를 바로 읽을 수 있도록 matrix를 정리한 형태이다.

5) Row Reduction

모든 matrix는 row operations를 통해 echelon form으로 변환할 수 있다.

이 과정을 Row Reduction 또는 Gaussian Elimination이라고 한다.

책에서 중요한 점은 주어진 matrix의 echelon form이 unique하다는 것이다.


2.2 Solving Linear Equations

1) Solution Classification

[Ab][A\mid\mathbf{b}]

를 echelon form

[A~b~][\widetilde A\mid\widetilde{\mathbf b}]

으로 row reduce한다.

해는 세 가지 경우로 나뉜다.

2) No Solution

b~\widetilde{\mathbf b} column에 pivotal 11이 있으면

Ax=bA\mathbf{x}=\mathbf{b}

는 solution을 가지지 않는다.

Meaning: 서로 모순되는 equation이 존재한다.

3) Unique Solution

b~\widetilde{\mathbf b}에 pivotal 11이 없고 A~\widetilde A의 모든 column에 pivotal 11이 있으면 unique solution을 가진다.

4) Infinitely Many Solutions

b~\widetilde{\mathbf b}에 pivotal 11이 없고 A~\widetilde A에 nonpivotal column이 있으면 infinitely many solutions을 가진다.

5) Pivotal Variable

Pivotal Variable: echelon form에서 pivotal 11을 포함하는 column에 대응하는 variable.

6) Nonpivotal Variable

Nonpivotal Variable: pivotal 11이 없는 column에 대응하는 variable.

Nonpivotal variables의 값을 자유롭게 선택하면 pivotal variables의 값이 결정된다.

Meaning: nonpivotal variable은 free variable이고 pivotal variable은 그것에 의존한다.

7) Invertibility Criterion

Square matrix AA에 대해

A is invertible    A row reduces to I.A\text{ is invertible} \iff A\text{ row reduces to }I.

따라서

Ax=bA\mathbf{x}=\mathbf{b}

가 모든 b\mathbf b에 대해 unique solution을 가지는 것과 AA가 invertible인 것은 같은 조건이다.


2.3 Matrix Inverses and Elementary Matrices

1) Solving with an Inverse

AA가 invertible이면

Ax=bA\mathbf{x}=\mathbf{b}

의 unique solution은

x=A1b\mathbf{x}=A^{-1}\mathbf{b}

이다.

2) Computing an Inverse

AA의 inverse는 augmented matrix

[AI][A\mid I]

를 row reduce하여 계산할 수 있다.

만약

[AI][IB][A\mid I] \longrightarrow [I\mid B]

이면

B=A1.B=A^{-1}.

반대로 AA 부분이 II로 row reduce되지 않으면 A1A^{-1}은 존재하지 않는다.

3) Elementary Matrix

Elementary Matrix: identity matrix에 하나의 row operation을 적용하여 얻은 matrix.

Matrix AA에 elementary matrix EE를 왼쪽에서 곱하는 것은 AA에 해당 row operation을 적용하는 것과 같다.

EA.EA.

Elementary matrix는 항상 invertible하며 그 inverse도 elementary matrix이다.


2.4 Linear Combinations, Span, and Linear Independence

1) Linear Combination

vectors

v1,,vk\mathbf v_1,\ldots,\mathbf v_k

Linear Combination

a1v1++akvka_1\mathbf v_1+\cdots+a_k\mathbf v_k

형태의 vector이다.

2) Linear Independence

vectors v1,,vk\mathbf v_1,\ldots,\mathbf v_k가 linearly independent라는 것은

a1v1++akvk=0a_1\mathbf v_1+\cdots+a_k\mathbf v_k=\mathbf 0

의 유일한 solution이

a1==ak=0a_1=\cdots=a_k=0

이라는 뜻이다.

Meaning: 어떤 vector도 나머지 vector들의 linear combination으로 만들어지지 않는다.

Linear independence는 solution의 uniqueness와 연결된다.

3) Span

Span: 주어진 vectors의 모든 linear combinations의 집합.

Span(v1,,vk)={i=1kaivi}.\operatorname{Span} (\mathbf v_1,\ldots,\mathbf v_k) = \left\{ \sum_{i=1}^{k}a_i\mathbf v_i \right\}.

Meaning: 주어진 vectors를 이용해 만들 수 있는 모든 vector의 집합이다.

Span은 solution의 existence와 연결된다.

4) Matrix Criterion

A=[v1  vk]A= [\mathbf v_1\ \cdots\ \mathbf v_k]

라고 하자.

v1,,vk\mathbf v_1,\ldots,\mathbf v_k가 linearly independent인 것은 row-reduced AA의 모든 column에 pivotal 11이 있는 것과 동치이다.

또한 v1,,vk\mathbf v_1,\ldots,\mathbf v_kRn\mathbb R^n을 span하는 것은 row-reduced AA의 모든 row에 pivotal 11이 있는 것과 동치이다.

5) Basis

Basis: vector space VV를 span하면서 linearly independent인 ordered set of vectors.

{v1,,vk}.\{\mathbf v_1,\ldots,\mathbf v_k\}.

Meaning: 공간의 모든 vector를 중복 없이 표현하기 위한 기본 좌표축이다.

basis를 이용하면 모든 vV\mathbf v\in V를 unique하게

v=a1v1++akvk\mathbf v = a_1\mathbf v_1+\cdots+a_k\mathbf v_k

로 표현할 수 있다.

6) Dimension

Dimension: vector space의 basis에 포함된 vector의 개수.

dimV.\dim V.

모든 basis는 동일한 수의 vectors를 가진다.

특히

dimRn=n.\dim\mathbb R^n=n.

7) Orthonormal Basis

Orthonormal Basis: basis vectors가 서로 orthogonal하고 모두 unit length인 basis.

vivj=0(ij)\mathbf v_i\cdot\mathbf v_j=0 \qquad (i\neq j) vi=1.|\mathbf v_i|=1.

orthonormal basis에서는

x=i=1n(xvi)vi.\mathbf x = \sum_{i=1}^{n} (\mathbf x\cdot\mathbf v_i)\mathbf v_i.

8) Orthogonal Matrix

Orthogonal Matrix: columns가 orthonormal basis를 이루는 square matrix.

ATA=AAT=I.A^TA=AA^T=I.

따라서

A1=AT.A^{-1}=A^T.

Orthogonal matrix는 dot product와 length를 보존한다.


2.5 Kernels, Images, and the Dimension Formula

1) Kernel

Linear transformation

T:RnRmT:\mathbb R^n\to\mathbb R^m

Kernel

kerT={xRnT(x)=0}.\ker T = \{ \mathbf x\in\mathbb R^n \mid T(\mathbf x)=\mathbf 0 \}.

Meaning: transformation에 의해 zero로 mapping되는 모든 vectors의 집합이다.

2) Image

Image:

imgT={T(x)xRn}.\operatorname{img}T = \{ T(\mathbf x) \mid \mathbf x\in\mathbb R^n \}.

Meaning: transformation이 실제로 만들어 낼 수 있는 모든 output의 집합이다.

Kernel과 image는 각각 domain과 codomain의 vector subspace이다.

3) Kernel and Uniqueness

kerT={0}\ker T=\{\mathbf 0\}

이면

T(x)=bT(\mathbf x)=\mathbf b

는 모든 b\mathbf b에 대해 solution을 최대 하나만 가진다.

즉,

kerT={0}    T is one-to-one.\ker T=\{\mathbf 0\} \iff T\text{ is one-to-one}.

4) Image and Existence

imgT=Rm\operatorname{img}T=\mathbb R^m

이면 모든 bRm\mathbf b\in\mathbb R^m에 대해

T(x)=bT(\mathbf x)=\mathbf b

의 solution이 존재한다.

즉,

imgT=Rm    T is onto.\operatorname{img}T=\mathbb R^m \iff T\text{ is onto}.

5) Basis for the Image

Matrix AAoriginal pivotal columns

imgA\operatorname{img}A

의 basis를 이룬다.

주의할 점은 row-reduced matrix의 columns가 아니라 original matrix의 pivotal columns를 사용한다는 것이다.

6) Dimension Formula

Linear transformation

T:RnRmT:\mathbb R^n\to\mathbb R^m

에 대해

dim(kerT)+dim(imgT)=n\boxed{ \dim(\ker T) + \dim(\operatorname{img}T) = n }

이다.

Meaning: domain의 dimension은 kernel과 image 사이에 나누어진다.

7) Rank

Rank: image의 dimension.

rankT=dim(imgT).\operatorname{rank}T = \dim(\operatorname{img}T).

따라서 dimension formula는

dim(kerT)+rankT=n\dim(\ker T)+\operatorname{rank}T=n

으로 쓸 수 있다.

8) Existence from Uniqueness

특히

T:RnRnT:\mathbb R^n\to\mathbb R^n

인 경우

kerT={0}\ker T=\{\mathbf 0\}

이면

imgT=Rn.\operatorname{img}T=\mathbb R^n.

따라서 square linear system에서는

one-to-one    onto    invertible.\text{one-to-one} \iff \text{onto} \iff \text{invertible}.

2.6 Abstract Vector Spaces

1) Vector Space

Vector Space: vectors를 더하고 scalar를 곱할 수 있으며 이러한 operations가 vector space axioms를 만족하는 집합.

대표적인 예:

Rn\mathbb R^n Mat(m,n)\operatorname{Mat}(m,n) PkP_k

그리고 function spaces도 vector space가 될 수 있다.

Meaning: vector는 반드시 숫자의 column일 필요가 없다.

2) Linear Transformation

Vector spaces V,WV,W 사이의 mapping

T:VWT:V\to W

가 linear transformation이라는 것은

T(av+bw)=aT(v)+bT(w)T(av+bw) = aT(v)+bT(w)

를 만족한다는 뜻이다.

3) Matrix with Respect to Bases

VV의 basis를

{v1,,vn}\{v_1,\ldots,v_n\}

WW의 basis를

{w1,,wm}\{w_1,\ldots,w_m\}

이라고 하자.

T(vk)T(v_k)

WW의 basis로 표현한 coefficients가 matrix [T][T]kk번째 column이 된다.

Meaning: abstract linear transformation도 basis를 선택하면 matrix로 표현할 수 있다.

4) Change of Basis Matrix

두 bases

{v},{v}\{v\}, \qquad \{v'\}

사이의 change of basis matrix를

PvvP_{v'\to v}

라고 하면

[x]v=Pvv[x]v[\mathbf x]_v = P_{v'\to v} [\mathbf x]_{v'}

이다.

Meaning: 같은 vector를 다른 coordinate system으로 변환한다.

5) Change of Basis Formula

Linear transformation TT의 matrix representation은 basis에 따라 달라진다.

[T]v,w=Pww1[T]v,wPvv.[T]_{v',w'} = P_{w'\to w}^{-1} [T]_{v,w} P_{v'\to v}.

특히 domain과 codomain이 같은 vector space이고 같은 basis를 바꾸는 경우

A=P1AP.A' = P^{-1}AP.

이 형태는 eigenbasis에서 매우 중요하다.


2.7 Eigenvectors and Eigenvalues

1) Eigenvector and Eigenvalue

Linear transformation

T:VVT:V\to V

에 대해 nonzero vector vv

Tv=λvTv=\lambda v

를 만족하면 vvEigenvector, λ\lambdaEigenvalue라고 한다.

Meaning: transformation을 적용해도 방향은 변하지 않고 크기만 λ\lambda배 되는 vector이다.

2) Eigenspace

Eigenvalue λ\lambda에 대응하는 eigenspace는

Eλ={vTv=λv}.E_\lambda = \{ v\mid Tv=\lambda v \}.

Eigenvalue의 multiplicity는 책에서 이 eigenspace의 dimension으로 정의한다.

3) Eigenbasis

Eigenbasis: 모든 basis vector가 eigenvector인 basis.

Eigenbasis를 사용하면 linear transformation이 매우 단순해진다.

4) Diagonalization

PP의 columns가 AA의 eigenbasis이고 corresponding eigenvalues가

λ1,,λn\lambda_1,\ldots,\lambda_n

이면

P1AP=[λ1000λ2000λn].P^{-1}AP = \begin{bmatrix} \lambda_1&0&\cdots&0\\ 0&\lambda_2&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\cdots&\lambda_n \end{bmatrix}.

따라서

A=PDP1.A = PDP^{-1}.

Meaning: 적절한 basis를 선택하면 transformation이 각 coordinate를 단순히 scaling하는 형태가 된다.

5) Powers of a Matrix

Diagonalization이 가능하면

Ak=PDkP1.A^k = PD^kP^{-1}.

그리고

Dk=[λ1k000λ2k000λnk].D^k = \begin{bmatrix} \lambda_1^k&0&\cdots&0\\ 0&\lambda_2^k&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\cdots&\lambda_n^k \end{bmatrix}.

Meaning: matrix의 높은 power를 계산하는 문제가 scalar powers 계산으로 분리된다.

6) Independent Eigenvectors

서로 다른 eigenvalues에 대응하는 eigenvectors는 linearly independent이다.

따라서 nn-dimensional vector space에서는 서로 다른 eigenvalue가 최대 nn개 존재한다.

7) Important Limitation

모든 matrix가 eigenbasis를 가지는 것은 아니다.

Meaning: 모든 square matrix가 diagonalizable한 것은 아니다.


2.8 Newton's Method

1) Nonlinear Equation

Nonlinear system

f(x)=0f(x)=0

을 직접 풀기 어려울 때 현재 guess 근처에서 ff를 linearize한다.

f(x)f(a)+Df(a)(xa).f(x) \approx f(a)+Df(a)(x-a).

2) Newton's Method

현재 approximation을 ana_n이라고 하면

f(an)+Df(an)(an+1an)=0.f(a_n)+Df(a_n)(a_{n+1}-a_n)=0.

따라서

an+1=an[Df(an)]1f(an)\boxed{ a_{n+1} = a_n - [Df(a_n)]^{-1}f(a_n) }

이다.

Meaning: nonlinear equation을 매 iteration마다 linear equation으로 바꾸어 푼다.

3) Practical Computation

실제 계산에서는

[Df(an)]1[Df(a_n)]^{-1}

을 직접 계산하기보다

Df(an)hn=f(an)Df(a_n)h_n = -f(a_n)

을 row reduction으로 풀고

an+1=an+hna_{n+1}=a_n+h_n

으로 계산한다.

4) Lipschitz Condition

Newton's method의 convergence를 보장하려면 derivative가 너무 빠르게 변하지 않아야 한다.

Df(x)Df(y)Mxy.|Df(x)-Df(y)| \leq M|x-y|.

여기서 MM은 Lipschitz ratio이다.

5) Kantorovich's Theorem

Kantorovich's theorem은 적절한 initial guess와 Lipschitz condition 아래에서 Newton sequence가 실제 root로 convergence한다는 것을 보장한다.

책의 sufficient condition 중 핵심 형태는

f(a0)Df(a0)12M12.|f(a_0)| \, |Df(a_0)^{-1}|^2 M \leq \frac12.

Meaning: initial residual, inverse derivative의 크기, derivative의 변화율이 충분히 작으면 Newton's method의 convergence를 보장할 수 있다.


2.9 Superconvergence

1) Newton's Method Convergence

Kantorovich condition에서

f(a0)Df(a0)12M<12|f(a_0)| \, |Df(a_0)^{-1}|^2 M < \frac12

처럼 strict inequality가 성립하면 Newton's method는 훨씬 빠르게 convergence할 수 있다.

책에서는 이를 Superconvergence라고 부른다.

2) Main Idea

successive increment를

xi=ai+1aix_i=|a_{i+1}-a_i|

라고 하면 superconvergence에서는 이 값이 iteration마다 매우 빠르게 감소한다.

Meaning: Newton's method가 root에 충분히 가까워지면 정확한 digits의 수가 매우 빠르게 증가한다.


2.10 Inverse and Implicit Function Theorems

1) Local Inverse

Function

f:RnRnf:\mathbb R^n\to\mathbb R^n

이 global하게 invertible하지 않더라도 특정 point 주변에서는 inverse를 가질 수 있다.

이를 locally invertible이라고 한다.

2) Inverse Function Theorem

ff가 continuously differentiable이고

Df(x0)Df(x_0)

가 invertible이면 ffx0x_0 근처에서 locally invertible하다.

즉, y0=f(x0)y_0=f(x_0) 근처에 inverse function

g=f1g=f^{-1}

가 존재한다.

Meaning: derivative가 invertible이면 nonlinear function도 충분히 작은 영역에서는 invertible linear transformation처럼 행동한다.

3) Derivative of the Inverse

Dg(y)=[Df(g(y))]1.Dg(y) = [Df(g(y))]^{-1}.

특히

D(f1)(f(x))=[Df(x)]1.D(f^{-1})(f(x)) = [Df(x)]^{-1}.

4) Local vs Global Invertibility

Df(x)Df(x)

가 모든 point에서 invertible이라고 해서 ff가 global하게 invertible한 것은 아니다.

Inverse function theorem은 local theorem이다.

5) Implicit Function

Equation

F(x,y)=0F(x,y)=0

이 어떤 variables를 다른 variables의 function으로 결정할 때 이를 implicit function이라고 한다.

예를 들어 locally

x=g(y)x=g(y)

의 형태로 나타낼 수 있다.

6) Implicit Function Theorem

F:RnRnkF:\mathbb R^n\to\mathbb R^{n-k}

이고

F(c)=0F(c)=0

이라고 하자.

만약

DF(c)DF(c)

가 onto라면 적절한 nkn-k개의 pivotal variables를 나머지 kk개의 nonpivotal variables의 function으로 locally 표현할 수 있다.

즉,

x=g(y).x=g(y).

Meaning: nonlinear equation에서도 linear equation의 pivotal / nonpivotal variable 구조가 local하게 유지된다.

7) Derivative of an Implicit Function

variables를

F(x,y)=0F(x,y)=0

로 나누고

DxFD_xF

가 invertible이면

Dg=(DxF)1DyF\boxed{ Dg = -(D_xF)^{-1}D_yF }

이다.


Essential Study Checklist

반드시 이해해야 할 내용:

  • Linear system
Ax=bA\mathbf{x}=\mathbf{b}
  • Augmented Matrix
[Ab][A\mid\mathbf b]
  • Row Operations
  • Echelon Form
  • Pivotal / Nonpivotal Variables
  • Unique / None / Infinite Solutions
  • Matrix Inverse
[AI][IA1][A\mid I] \to [I\mid A^{-1}]
  • Linear Combination
  • Linear Independence
  • Span
  • Basis
  • Dimension
  • Orthonormal Basis
  • Orthogonal Matrix
A1=ATA^{-1}=A^T
  • Kernel
kerT\ker T
  • Image
imgT\operatorname{img}T
  • Rank
  • Dimension Formula
dim(kerT)+dim(imgT)=dim(domain)\boxed{ \dim(\ker T) + \dim(\operatorname{img}T) = \dim(\text{domain}) }
  • Change of Basis
A=P1APA' = P^{-1}AP
  • Eigenvector / Eigenvalue
Av=λvAv=\lambda v
  • Eigenbasis
  • Diagonalization
A=PDP1A=PDP^{-1}
  • Newton's Method
an+1=an[Df(an)]1f(an)\boxed{ a_{n+1} = a_n-[Df(a_n)]^{-1}f(a_n) }
  • Lipschitz Condition
  • Kantorovich's Theorem의 의미
  • Superconvergence의 의미
  • Inverse Function Theorem
Df(x0) invertiblef locally invertibleDf(x_0)\text{ invertible} \Rightarrow f\text{ locally invertible}
  • Implicit Function Theorem
  • Implicit Function Derivative
Dg=(DxF)1DyFDg = -(D_xF)^{-1}D_yF