Numerical linear algebra

In the field of numerical analysis, numerical linear algebra is an area to study methods to solve problems in linear algebra by numerical computation.^[1][2][3] The following problems will be considered in this area:

1. Numerically solving a system of linear equations.^[4]
2. Numerically solving an eigenvalue problem for a given matrix.^[5]
3. Computing approximate values of a matrix-valued function.^[6]

Numerical errors can occur in any kind of numerical computation including the area of numerical linear algebra. Errors in numerical linear algebra are considered in another area called "validated numerics".^[7]

Latest Studies

See also: Matrix analysis

Methods for numerical linear algebra has been created by numerical analysts from many generations.^[1][2][3] But today, some of them have been rejected due to their computation speed or accuracy.^[1][2][3] Currently, the following methods are widely investigated:

• QZ method^[8]
• dqds method (differential quotient difference with shift)^[9]
• oqds method (orthogonal quotient difference with shift)^[10][11][12]
• MRRR method (multiple relatively robust representations)^[13]
• MRTR method^[14]
• Sakurai-Sugiura method^[15]
• CIRR method (Rayleigh-Ritz type method with contour integral)^[16]

Krylov Subspace Methods

In the field of numerical linear algebra, numerical methods based on the theory of Krylov subspaces are known as Krylov subspaces methods. They are considered to be one of the most successful studies in numerical linear algebra.^[17][18] The next list is the examples of them:

• MINRES (minimal residual) method^[19]
• CR (conjugate residual) method^[20]
• QMR type methods
  • QMR (quasi minimal residual) method^[21][22]
  • QMR-SYM method^[23][24]
  • TFQMR (transpose free quasi minimal residual) method^[25]

Conjugate Gradient Methods

The conjugate gradient (CG) method is one of the best linear equation solving method. It was originally limited to specific linear systems.^[26] In order to overcome this difficulty, many kinds of CG variants have benn created:

• CGS (conjugate gradient squared method)^[27]
• PCG (preconditioned conjugate gradient method)
• SCG (scaled conjugate gradient)^[28]
• ICCG (incomplete Cholesky conjugate gradient method)
• COCG (conjugate orthogonal conjugate gradient method)^[29]
• GPBiCG^[30]
• Stabilized methods
  • BiCGSTAB (biconjugate gradient stabilized method)^[31]
  • BiCGSTAB2^[32]
  • QMRCGSTAB^[33]
  • GBi-CGSTAB^[34]
• Block versions (dividing a given matrix into block matrices is a frequently used technique in numerical linear algebra^[1][2][3])
  • Block CG^[35][36]
  • Block BiCGSTAB^[37]
  • Block BiCGGR^[38]
  • Block BiCGGR2^[39]
  • Block GWBiCGSTAB^[40]

Validated Numerics for Numerical Linear Algebra

While high accuracy and high speed methods in above have been cretaed, some experts have studied how to evaluate numerical errors in numerical linear algebra.^[7] The following are their results:

• Validating numerical solutions of a given system of linear equations^[41][42]
  • Validated numerics for ill-conditioned problems^[43][44][45] (Ill-conditioned problems are problems which are hard to compute accurately^[1][2][3])
  • Pre-conditioning^[43][44][45] (Pre-conditioning is a procedure to allow the given system of linear equations to be easily solved^[1][2][3])
• Validating numerically obtained eigenvalues^[46][47][48]
  • Validating numerical solutions of inverse eigenvalue problems^[49][50] (In inverse eigenvalue problems, you will compute and identify an unknown matrix by a given eigenvalue)
• Rigorously computing determinants^[51]
• Validating numerical solutions of matrix equations^{[52][53][54][55][56][57][58]}
• Computing matrix functions rigorously (Approximate computation has been studied by N. J. Higham and others^{[59][60][61][62]})
  • Matrix exponential^[63]
  • Matrix logarithm^[64]
  • Matrix root^[65]

Software

Today, there are many tools for numerical linear algebra. One of the most famous one is MATLAB (matrix laboratory).^[66][67][68] This was made by MathWorks.