hizhaolei 2019-09-15 10:48 原文
- Chap 1: Linear Equations and Matrix
- Linear equations
- Gaussian elimination
- Pivot;
- Triangularize;
- Back substitution;
- Coefficient matrix, augmented matrix, row vector & column vector;
- the meaning of Ai*, A*j;
- 3 situations of solution existence (under the view of linear equations): 0,1 or infinite;
- Computational complexity: n^3/3+...;
- Gaussian-Jordan Method
- Computational complexity: n^3/2+...;
- Roundoff error
- Form of floating number: f=± .d1 d2 ... dt * b^n (d1≠0);
- Roundoff error: caused by the different magnitudes between the different columns;
- Partial pivoting: search the position BELOW the pivotal position for the coefficient in maximum magnitude;
- Complete pivoting: search the position BELOW and on the RIGHT of the pivotal position for the max coefficient;
- Partial & Complete pivoting: whether using elementary column operation. The partial one is used more frequently because the elementary column operation is not easy to use;
- the ill-conditioned system
- the solution of an ill-conditioned system is extremely sensitive to a small perturbation on the coefficients;
- Geometrical view: two linear systems are almost parallel so that their cross point will move sensitively when any one system moved;
- How to notice the ill-condition of a linear system: enumerating ( it's not easy to find whether a system is ill-conditioned);
- 2 way to solve the problem: bite the bullet and compute the accurate solution, or redesign the experiment setup to avoid producing the ill-conditioned system. The latter one is better empirically. Finding a system is an ill-conditioned one as early as possible will save much time;
- Row echelon form
- Notation: E;
- Cause: linear correlation between different column vectors and modified Gaussian elimination;
- The echelon form (namely the position of pivots) is uniquely determined by the entries in A. However, the entries in E is not uniquely determined by A.
- Basic column: the columns in A which contain the pivotal position;
- Rank: the number of pivots = the number of nonzero rows in E = the number of basic columns in A;
- Reduced row echelon form: produced by Gaussian-Jordan Method( [0 0 1 0]T ), notated by EA;
- Both form and entries of EA is uniquely determined by A;
- EA can show the hidden relationships among the different columns of A;
- Consistency of linear system
- A system is consistent if it has at least one solution. Otherwise, it is inconsistent.
- When n (the number of equations) is two or three, the consistency of the system can be shown geometrically, the common point.
- If n>3, we can judge through the following method:
- In the augmented matrix [A|b], 0=a≠0 does not exist;
- In [A|b], b is the nonbasic column;
- rank(A|b)=rank(A);
- b is the combination of the basic column in A.
- Homogeneous system
- Homogeneous and nonhomogeneous;
- Trivial solution;
- A homogeneous system must be a consistent system;
- General solution: basic variable, free variable;
- Nonhomogeneous system
- General solution;
- The system possesses a unique solution if and only if:
- rank(A) = the number of the unknowns;
- no free variable;
- the associated homogeneous system only has a trivial solution;
- Chap 2: Matrix Algebra
- Addition
- Addition and addition inversion;
- Addition properties;
- Scalar multiplication
- Transpose
- Transpose and conjugate transpose;
- Properties;
- Symmetry;
- Symmetric matrix, skew-symmetric matrix, hermitian matrix, skew-hermitian matrix;
- Multiplication
- Linear function: f(x1+x2)=f(x1)+f(x2), f(kx)=kf(x) <=> f(kx+y)=kf(x)+f(y);
- Affine function: translation of linear function;
- Matrix multiplication;
- Properties: distributive law(left one or tight one) and associative law, but no commutative law;
- Trace
- Definition: the sum of diagonal entries;
- Properties: trace(AB) = trace(BA), trace(ABC) = trace(BCA) = trace(CAB) ≠ trace(ACB);
- Meaning of rows and columns in a product
- [AB]i* = linear combination of row vectors in B based on i-th row vector in A;
- [AB]*j = linear combination of column vectors in A based on j-th column vector in B;
- column vector * row vector = a matrix whose rank is 1 ( outer product);
- row vector * column vector <=> inner product;
- Identity matrix；
- Power: nonnegative;
- Block matrix multiplication;
- Inversion
- Only square matrices have matrix inversion;
- AB=I and BA=I ( When only square matrix involved, any one of the two equations is sufficing);
- Nonsingular matrix and singular matrix;
- When an inversion exists, it is unique. That means:
- If A is nonsingular, the equation Ax=b has the unique solution x=A'b;
- If A is nonsingular, rank(A) =n (full rank);
- If A is nonsingular, the unknown x has no free variable;
- If A is nonsingular, the associated homogeneous system has a trivial solution only;
- Existence of matrix inversion: A' exists <=> rank(A)=n <=> A can be transformed to I via Gauss-Jordan Method <=> Ax=0 only has a trivial solution;
- Computing an inversion: transforming [A|I] to [I|A'] via Gauss-Jordan Method;
- Complexity(x=A'b) > Complexity(Gaussian Elimination):
- C(GE) ≈ n^3/3;
- C(x=A'b) = C(computing A') + C(A'b) ≈ 2*(n^3/2) + n*n*n = 2n^3;
- Properties:
- (A')' = A;
- A, B are nonsingular, AB is also nonsingular;
- (AB)' = B'A';
- (A')T = (AT)' as well as (A')* = (A*)';
- Inversion of sum and sensitivity:
- Directly discuss the relationship of (A+B)' and A', B' is meaningless;
- Sherman-Morrison Formula: a small perturbation;
- Neumann Series:
- If limn->infiniteAn=0 and (I-A) is nonsingular, (I-A)'=I + A + A2 + ... =ΣiAi;
- To solve (A+B)', the expression can be transformed into A(I-(-A'B))';
- (A+B)' ≈ A' + A'BA': A perturbation B on A, will make inversion change by A'BA'. When A' is large, a small perturbation will change the result a lot;
- Condition number;
- Elementary Matrices and Equivalence
- Elementary matrix: I-uv^T, u and v are column vectors;
- The inversion of an elementary matrix is also an elementary matrix;
- Elementary matrices associated with three types of elementary row (or column) operation;
- A is a nonsingular matrix <=> A is the product of elementary matrices of Type I, II and III row (or column) operation;
- Equivalence: A~B <=> PAQ=B for nonsingular P and Q;
- Row equivalence and column equivalence;
- Rank normal form: if A is an m*n matrix such that rank(A)=r, then A~Nr=[[Ir, 0]^T, [0, 0]^T], Nr is called rank normal form for A;
- A~B <=> rank(A)=rank(B);
- Corollary: Multiplication by nonsingular matrices cannot change rank;
- rank(A^T)=rank(A);
- rank(A*)=rank(A);
- LU factorization
- Origin: Gaussian Elimination;
- LU factorization: A=LU, L: lower triangular matrix, U: upper triangular matrix;
- Observation of LU: Advantages of LU factorization:
- L:
- a lower triangular matrix;
- 1's on the diagonal: means itself row plus other rows' multiplication with a scalar;
- the entries below the diagonal record the multipliers used to eliminate;
- U:
- an upper triangular matrix;
- the result of the elimination on A;
- *L and U are unique;
- proof: A=L1U1=L2U2 => L2'L1=U2U1', L2'L1 is a lower triangular matrix, U2U1' is an upper triangular matrix. They are equal to each other. So I=I => L2'L1=U2U1'=I.
- *If exchanging of two rows is emerging during LU factorizing, the consistency of triangular form will be destroyed;
- Advantages of LU factorization:
- If only one system Ax=b need to be solved, the Gaussian Elimination is enough;
- If more then one systems which coefficient matrices are the same need to be solved, the LU factorization is better;
- Once the LU factors of A are known, any other system Ax=b can be solved in n^2 multiplications and n^2-n additions;
- Existence of LU:
- No zero pivot emerges during row reduction to upper triangular form with type III operation;
- Another characterization method associated with principle submatrix: each leading principle submatrices is nonsingular;
- PLU factorization: PA=LU;
- LDU factorization: A=LDU, D=diag(u11, u22, ..., unn);
- Vector Spaces
- Spaces and subspaces
- Vector space;
- Scalar field F: R for real numbers and C for complex numbers;
- null
推荐阅读