Mathematical Optimisation, Chpt 2: Linear Equations and inequalities

Similar documents
Mathematical Optimisation, Chpt 2: Linear Equations and inequalities

Math Camp Lecture 4: Linear Algebra. Xiao Yu Wang. Aug 2010 MIT. Xiao Yu Wang (MIT) Math Camp /10 1 / 88

Math 102, Winter Final Exam Review. Chapter 1. Matrices and Gaussian Elimination

MATH 20F: LINEAR ALGEBRA LECTURE B00 (T. KEMP)

Preliminary/Qualifying Exam in Numerical Analysis (Math 502a) Spring 2012

Review problems for MA 54, Fall 2004.

Cheat Sheet for MATH461

This can be accomplished by left matrix multiplication as follows: I

MAT Linear Algebra Collection of sample exams

Problem Set (T) If A is an m n matrix, B is an n p matrix and D is a p s matrix, then show

Knowledge Discovery and Data Mining 1 (VO) ( )

Algebra C Numerical Linear Algebra Sample Exam Problems

Conceptual Questions for Review

MATH 240 Spring, Chapter 1: Linear Equations and Matrices

Linear Algebra in Actuarial Science: Slides to the lecture

Foundations of Matrix Analysis

Math 520 Exam 2 Topic Outline Sections 1 3 (Xiao/Dumas/Liaw) Spring 2008

MATH 1120 (LINEAR ALGEBRA 1), FINAL EXAM FALL 2011 SOLUTIONS TO PRACTICE VERSION

LINEAR ALGEBRA KNOWLEDGE SURVEY

A Review of Linear Algebra

Chapter 3 Transformations

1 9/5 Matrices, vectors, and their applications

Linear Algebra Massoud Malek

Applied Linear Algebra in Geoscience Using MATLAB

Linear Analysis Lecture 16

MTH 464: Computational Linear Algebra

A Brief Outline of Math 355

Recall the convention that, for us, all vectors are column vectors.

Equality: Two matrices A and B are equal, i.e., A = B if A and B have the same order and the entries of A and B are the same.

2. Every linear system with the same number of equations as unknowns has a unique solution.

Math 18, Linear Algebra, Lecture C00, Spring 2017 Review and Practice Problems for Final Exam

Lecture 4 Orthonormal vectors and QR factorization

ANALYTICAL MATHEMATICS FOR APPLICATIONS 2018 LECTURE NOTES 3

Ir O D = D = ( ) Section 2.6 Example 1. (Bottom of page 119) dim(v ) = dim(l(v, W )) = dim(v ) dim(f ) = dim(v )

Math Camp II. Basic Linear Algebra. Yiqing Xu. Aug 26, 2014 MIT

Glossary of Linear Algebra Terms. Prepared by Vince Zaccone For Campus Learning Assistance Services at UCSB

Total 170. Name. Final Examination M340L-CS

MATH 315 Linear Algebra Homework #1 Assigned: August 20, 2018

G1110 & 852G1 Numerical Linear Algebra

MATH 369 Linear Algebra

1 Last time: least-squares problems

MATH 304 Linear Algebra Lecture 20: The Gram-Schmidt process (continued). Eigenvalues and eigenvectors.

Linear Algebra Review. Vectors

Eigenvalues and Eigenvectors A =

Linear Algebra review Powers of a diagonalizable matrix Spectral decomposition

Math Linear Algebra Final Exam Review Sheet

Quantum Computing Lecture 2. Review of Linear Algebra

Review of Linear Algebra

LINEAR ALGEBRA 1, 2012-I PARTIAL EXAM 3 SOLUTIONS TO PRACTICE PROBLEMS

Linear Algebra March 16, 2019

This operation is - associative A + (B + C) = (A + B) + C; - commutative A + B = B + A; - has a neutral element O + A = A, here O is the null matrix

EE731 Lecture Notes: Matrix Computations for Signal Processing

IMPORTANT DEFINITIONS AND THEOREMS REFERENCE SHEET

LECTURE NOTES ELEMENTARY NUMERICAL METHODS. Eusebius Doedel

HOMEWORK PROBLEMS FROM STRANG S LINEAR ALGEBRA AND ITS APPLICATIONS (4TH EDITION)

Lecture Notes in Linear Algebra

Jim Lambers MAT 610 Summer Session Lecture 1 Notes

Study Guide for Linear Algebra Exam 2

Linear Algebra review Powers of a diagonalizable matrix Spectral decomposition

DS-GA 1002 Lecture notes 0 Fall Linear Algebra. These notes provide a review of basic concepts in linear algebra.

Math 407: Linear Optimization

Solution of Linear Equations

a s 1.3 Matrix Multiplication. Know how to multiply two matrices and be able to write down the formula

Linear Algebra: Matrix Eigenvalue Problems

5.6. PSEUDOINVERSES 101. A H w.

Scientific Computing: Dense Linear Systems

AMS526: Numerical Analysis I (Numerical Linear Algebra)

Orthonormal Transformations and Least Squares

ELE/MCE 503 Linear Algebra Facts Fall 2018

ORIE 6300 Mathematical Programming I August 25, Recitation 1

Math 265 Linear Algebra Sample Spring 2002., rref (A) =

LINEAR ALGEBRA SUMMARY SHEET.

orthogonal relations between vectors and subspaces Then we study some applications in vector spaces and linear systems, including Orthonormal Basis,

Linear Algebra. Matrices Operations. Consider, for example, a system of equations such as x + 2y z + 4w = 0, 3x 4y + 2z 6w = 0, x 3y 2z + w = 0.

Lecture notes: Applied linear algebra Part 1. Version 2

Elementary linear algebra

Linear Algebra Primer

Theorem A.1. If A is any nonzero m x n matrix, then A is equivalent to a partitioned matrix of the form. k k n-k. m-k k m-k n-k

235 Final exam review questions

MTH 2032 SemesterII

5. Orthogonal matrices

Applied Linear Algebra

Linear Algebra Highlights

MIT Final Exam Solutions, Spring 2017

Math 1553, Introduction to Linear Algebra

Lecture 7: Positive Semidefinite Matrices

SUMMARY OF MATH 1600

MAT 610: Numerical Linear Algebra. James V. Lambers

Elementary maths for GMT

CS 246 Review of Linear Algebra 01/17/19

Math Linear Algebra II. 1. Inner Products and Norms

ANSWERS. E k E 2 E 1 A = B

Review of Some Concepts from Linear Algebra: Part 2

I. Multiple Choice Questions (Answer any eight)

Practical Linear Algebra: A Geometry Toolbox

Math 240 Calculus III

AMS526: Numerical Analysis I (Numerical Linear Algebra)

Numerical Methods I Solving Square Linear Systems: GEM and LU factorization

1. General Vector Spaces

Mathematical foundations - linear algebra

Transcription:

Mathematical Optimisation, Chpt 2: Linear Equations and inequalities Peter J.C. Dickinson p.j.c.dickinson@utwente.nl http://dickinson.website version: 12/02/18 Monday 5th February 2018 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 1/40

Table of Contents 1 Introduction 2 Gauss-elimination [MO, 2.1] 3 Orthogonal projection, Least Squares, [MO, 2.2] 4 Linear Inequalities [MO, 2.4] 5 Integer Solutions of Linear Equations [MO, 2.3] Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 2/40

Organization and Material Lectures (hoor-colleges) and 3 exercise classes for motivation, geometric illustration, and proofs. Script Mathematical Optimization, cited as: [MO,?] Sheets of the course (on Blackboard) Mark based on written exam Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 3/40

Chapter 1: Real vector spaces A collection of facts from Analysis and Linear Algebra, self-instruction Chapter 2: Linear equations, linear inequalities Gauss elimination and application Gauss elimination provides constructive proofs of main theorems in matrix theory. Fourier-Motzkin elimination This method provides constructive proofs of the Farkas Lemma (which is strong LP duality in disguise). Least square approximation, Fourier approximation Integer solutions of linear equations (discrete optimization) Chapter 3: Linear programs and applications LP duality, sensitivity analysis, matrix games. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 4/40

Chapter 4: Convex analysis Properties of convex sets and convex functions. Applications in optimization Chapter 5: Unconstrained optimization optimality conditions algorithms: descent methods, Newton s method, Gauss-Newton method, Quasi-Newton methods, minimization of nondifferentiable functions. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 5/40

Chapter 2. Linear equations, inequalities We start with some definitions: Definitions in matrix theory M = (m ij ) is said to be lower triangular: if m ij = 0 for i < j, upper triangular: if m ij = 0 for i > j. P = (p ij ) R m m is a permutation matrix if p ij {0, 1} and each row and each column of P contains exactly one coefficient 1. Note that P T P = I, implying that P 1 = P T. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 6/40

The set of symmetric matrices in R n n is denoted by S n. Q S n, is called Positive Semidefinite (denoted Q O or Q PSD n ) if x T Qx 0 for all x R n, Positive Definite (denoted: Q O) if x T Qx > 0 for all x R n \ {0}, Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 7/40

Table of Contents 1 Introduction 2 Gauss-elimination [MO, 2.1] Gauss-elimination (for solving Ax = b) Explicit Gauss Algorithm Implications of the Gauss algorithm Gauss-Algorithm for symmmetric A 3 Orthogonal projection, Least Squares, [MO, 2.2] 4 Linear Inequalities [MO, 2.4] 5 Integer Solutions of Linear Equations [MO, 2.3] Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 8/40

Gauss-elimination (for solving Ax = b) In this Section 2.1 We shortly survey the Gauss algorithm and use the Gauss algorithm to give a constructive proof of important theorems in Matrix Theory Motivation: We give a simple example which shows that successive elimination is equivalent with the Gauss algorithm. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 9/40

General Idea: To eliminate x 1, x 2... is equivalent with transforming Ax = b or (A b) to triangular normal form (Ã b) (with same solution set). Then solve Ãx = b, recursively: a 11 a 12... a 1n b 1 a 21 a 22... a 2n b 2.. a m1 a m2... a mn b m Transformation into form (Ã b): ã 1 j1 ã 1 j2 ã 1 jr 1 ã 1 jr b1 ã 2 j2 ã 2 jr 1 ã 2 jr b2.... ã r 1 jr 1 ã r 1 jr b r 1 ã r jr b r. b m Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 10/40

This Gauss elimination uses 2 types of row operations: (G1) (i, j)-pivot: k > i add λ k = a kj a ij times row i to row k. (G2) interchange row i with row k The matrix form of these operations are: [MO, Ex.2.3] The matrix form of (G1), (A b) (Ã b), is given by (Ã b) = M (A b) with a nonsingular lower triangular M R m m. [MO, Ex.2.4] The matrix form of (G2), (A b) (Ã b), is given by (Ã b) = P (A b) with a permutation matrix P R m m. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 11/40

Explicit Gauss Algorithm 1 Set i = 1 and j = 1. 2 While i m and j n do 3 If k i such that a kj 0 then 4 Interchange row i and row k. (G2) 5 Apply (i, j)-pivot. (G1) 6 Update i i + 1 and j j + 1. 7 else 8 Update j j + 1. 9 end if 10 end while Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 12/40

Implications of the Gauss algorithm Lemma 2.1 The set of lower triangular matrices is closed under addition, multiplication and inversion (provided square and nonsingular). Theorem 2.2 ([MO, Thm. 2.1]) For every A R m n, there exists an (m m)-permutation matrix P and an invertible lower triangular matrix M R m m such that U = MPA is upper triangular. Corollary 2.3 (Cor. 2.1, LU-factorization) For A R m n, there exists an (m m)-permutation matrix P, an invertible lower triangular L R m m and an upper triangular U R m n such that LU = PA. Rem.: Solve Ax = b by using the decomposition PA = LU! (How?) Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 13/40

Corollary 2.4 ([MO, Cor. 2.2], Gale s Theorem) Exactly one of the following statements is true for A R m n : (a) The system Ax = b has a solution x R n. (b) There exists y R m such that: y T A = 0 T and y T b 0. Remark: In normal form A Ã, the number r gives dimension of the space spanned by the rows of A. This equals the dimension of the space spanned by columns of A. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 14/40

Modified Gauss-Algorithm for symmmetric A Perform same row and column operators: 1 Set i = 1. 2 While i n do 3 If a kk 0 for some k i 4 Apply (G2 ): Interchange row i and row k. Interchange col. i and col. k, A PAP T. 5 Apply (G1 ): For all j > i add λ j = a ij a ii times row i to row j, and λ j times col. i to col. j, A MAM T. 6 Update i i + 1. 7 Else if a ik 0 for some k > i then 8 Add row k to row i and add col. k to col. i, A BAB T. 9 Else 10 Update i i + 1. 11 End if 12 End while. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 15/40

Implications of the symmetric Gauss algorithm Note: By symmetric Gauss the solution set of Ax = b is destroyed!!! But it is useful to get the following results. Theorem 2.5 ([MO, Thm. 2.2]) A R n n symmetric. Then with some nonsingular Q R n n QAQ T = D = diag(d 1,..., d n ), d R n. Look out: The d i s are in general not the eigenvalues of A. Recall: Q PSD n is positive semidefinite (notation: Q O) if: x T Qx 0 for all x R n. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 16/40

Corollary 2.6 ([MO, Cor. 2.3]) Let A S n, Q R n n nonsingular s.t. QAQ T = diag(d 1,..., d n ). Then (a) A O d i 0, i = 1,..., n (b) A O d i > 0, i = 1,..., n Implication: Checking A O can be done by the Gauss-algorithm. Corollary 2.7 ([MO, Cor. 2.4]) Let A S n. Then (a) A O A = BB T for some B R n m (b) A O A = BB T for some nonsingular B R n n Complexity of Gauss algorithm The number of ±,, / flop s (floating point operations) needed to solve Ax = b, where A R n n, is less than or equal to n 3. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 17/40

Table of Contents 1 Introduction 2 Gauss-elimination [MO, 2.1] 3 Orthogonal projection, Least Squares, [MO, 2.2] Projection and equivalent condition Constructing a solution Gram Matrix Gram-Schmidt Algorithm Eigenvalues 4 Linear Inequalities [MO, 2.4] 5 Integer Solutions of Linear Equations [MO, 2.3] Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 18/40

Projection and equivalent condition In this section: We see that the orthogonal projection problem is solved by solving a system of linear equations; and present some more results from matrix theory. Assumption: V is a linear vector space over R with inner product x, y and (induced) norm x = x, x. Minimization Problem: Given x V, subspace W V find x W such that: x x = min x y (1) y W The vector x is called the projection of x onto W. Lemma 2.8 ([MO, Lm. 2.1]) x W is the (unique) solution to (1) x x, w = 0 w W. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 19/40

Constructing a solution via Lm 2.8 Let a 1,..., a m be a basis of W, i.e., a 1,..., a m are linearly independent and W = span{a 1,..., a m }. Write x := m i=1 z ia i Then x x, w = 0, w W is equivalent with x m z ia i a j = 0, j = 1,..., m i=1 or m a i, a j z i = x, a j, j = 1,..., m i=1 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 20/40

Gram Matrix Defining the Gram-matrix G = ( a i, a j ) i,j S m and considering b = ( x, a i ) i R m, this leads to the linear equation (for z): (2.16) Gz = b with solution ẑ = G 1 b Ex. 2.1 Prove that the Gram-matrix is positive definite, and thus non-singular (under our assumption). Summary Let W = span{a 1,..., a m }. Then the solution x of the minimization problem (1) is given by x := m i=1 ẑia i where ẑ is computed as the solution of G ẑ = b. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 21/40

Special case 1: V = R n, x, y = x T y and a 1,..., a m a basis of W. Then with A := [a 1,..., a m ] the projection of x onto W is given by ẑ = arg min{ x Az } = (A T A) 1 A T x, z x = arg min y { x y : y AR m } = Aẑ = A(A T A) 1 A T x. Special case 2: V = R n, x, y = x T y, a 1,..., a m R n lin. indep. and W = {w R n a T i w = 0, i = 1,..., m}. Then the projection of x onto W is given by x = arg min{ x y : a T y i y = 0 i} = x A(A T A) 1 A T x Special case 3: W = span{a 1,..., a m } with {a i }, an orthonormal basis, i.e., I = ( a i, a j ) i,j ). Then the projection of x onto W is given by ẑ = ( a j, x ) j R m, ( Fourier coefficients ) x = arg min{ x y : y W } = m ẑja j y j=1 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 22/40

Gram-Schmidt Algorithm Given subspace W with basis a 1,..., a m, construct an orthonormal basis c 1,... c m for W, i.e. ( c i, c j ) i,j = I. Start with b 1 := a 1 and c 1 = b 1 / b 1. For k = 2,..., m let b k = a k k 1 i=1 c i, a k c i and c i = b i / b i. Gram-Schmidt in matrix form: With W V := R n. Put A = [a 1,..., a m ] T, B = [b 1,..., b m ] T, C = [c 1,..., c m ] T. Then the Gram-Schmidt-steps are equivalent with: add multiple of row j < k to row k (for B) multiply row k by scalar (for C) Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 23/40

Matrix form of Gram-Schmidt : Given A R m n with full row rank, there is a decomposition C = LA with lower triangular nonsingular matrix L (l ii = 1) and the rows c j of C are orthogonal, i.e. c i, c j = 0, i j. A corollary of this fact: Lemma 2.9 ([MO, Prop. 2.1]) Prop. 2.1 (Hadamard s inequality) Let A R m n with rows a T i. Then m 0 det (AA T ) a T i a i i=1 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 24/40

Eigenvalues Definition. λ C is an eigenvalue of A R n n if there is an (eigenvector) 0 x C n with Ax = λx. The results above (together with the Theorem of Weierstrass) allow a proof of: Theorem 2.10 ([MO, Thm. 2.3], Spectral theorem for symmetric matrices) Let A S n. Then there exists an orthogonal matrix Q R n n (i.e. Q T Q = I) and eigenvalues λ 1,..., λ n R such that Q T AQ = D = diag (λ 1,..., λ n ) Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 25/40

Table of Contents 1 Introduction 2 Gauss-elimination [MO, 2.1] 3 Orthogonal projection, Least Squares, [MO, 2.2] 4 Linear Inequalities [MO, 2.4] Fourier-Motzkin Algorithm Solvability of linear systems Application: Markov chains 5 Integer Solutions of Linear Equations [MO, 2.3] Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 26/40

Linear Inequalities Ax b In this Section we learn: Finding a solution to Ax b. What the Gauss algorithm is for linear equations; is the Fourier-Motzkin algorithm for linear inequalities. The Fourier-Motzkin algorithm leads to a construcive proof of the Farkas Lemma (the basis for strong duality in linear programming). Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 27/40

Fourier-Motzkin algorithm for solving Ax b. Eliminate x 1 : a r1 x 1 + a s1 x 1 + n a rj x j b r j=2 n a sj x j b s j=2 n a tj x j b t j=2 r = 1,..., k s = k + 1,..., l t = l + 1,..., m with a r1 > 0, a s1 < 0. Divide by a r1, a s1, leading to (for r and s) x 1 + x 1 + n a rjx j b r j=2 n a sjx j b s j=2 r = 1,..., k s = k + 1,..., l Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 28/40

These two sets of inequalities have a solution x iff n a sjx j b s x 1 b r n j=2 j=2 a rjx j { r = 1,..., k s = k + 1,..., l or equivalently: n { (a sj + a rj)x j b s + b r = 1,..., k r s = k + 1,..., l j=2 Remark Explosion of number of inequalities. Before number = k + (l k) + (m l). Now the number = k (l k) + (m l) Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 29/40

So Ax b has a solution x = (x 1,.., x n ) if and only if there is a solution x = (x 2,.., x n ) of n (a sj + a rj)x j b r + b s r = 1,..., k; s = k + 1..., l j=2 n a tj x j b t t = l + 1,..., m. j=2 In matrix form: Ax b has a solution x = (x 1,.., x n ) if and only if there is a solution of the transformed system: A x b or ( 0 A )x b Remark: of (A b): Any row of (0 A b ) is a positive combination of rows any row is of the form y T (A b), y 0 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 30/40

By eliminating x 1, x 2,..., x n, in this way we finally obtain an equivalent system à (n) x b where à (n) = 0 which is solvable iff 0 b i, i. Theorem 2.11 (Projection Theorem, [MO, Thm. 2.5]) Let P = {x R n Ax b}. Then all for k = 1,..., n, the projection P (k) = {(x k+1,.., x n ) (x 1,.., x k, x k+1,.., x n ) P is the solution set of a linear system in n k variables x (k) = (x k+1,..., x n ). for suitable x 1,..., x k R} A (k) x (k) b (k) In principle: Linear inequalities can be solved by FM. However this might be inefficient! (Why?) Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 31/40

Solvability of linear systems Theorem 2.12 (Farkas Lemma, [MO, 2.6]) Exactly one of the following statements is true: (I) Ax b has a solution x R n. (II) There exists y R m such that y T A = 0 T, y T b < 0 and y 0 Ex. 2.2 Let A R m n, C R k n, b R m, c R k. Then precisely one of the alternatives is valid. (I) There is a solution x of: Ax b, Cx = c (II) There is a solution µ R m, µ 0, λ R k of : ( A T b T ) µ + ( C T c T ) λ = ( ) 0 1 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 32/40

Corollary 2.13 (Gordan, [MO, Cor. 2.5]) Given A R m n, exactly one of the following alternatives is true: (I) Ax = 0, x 0 has a solution x 0. (II) y T A < 0 T has a solution y. Remark: As we shall see in Chapter 3, the Farkas Lemma in the following form is the strong duality of LP in disguise. Corollary 2.14 (Farkas, implied inequalities, [MO, Cor. 2.6]) Let A R m n, b R m, c R n, z R. Assume that Ax b is feasible. Then the following are equivalent: (a) Ax b c T x z (b) y T A = c T, y T b z, y 0 has a solution y. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 33/40

Application: Markov chains (Existence of a steady state) Def. A vector π R n + with 1 T π := n i=1 π i = 1 is called a probability distribution on {1,.., n}. A matrix P = (p ij ) where each row P i is a probability distribution is called a stochastic matrix, i.e. P R+ n n and P1 = 1 In a stochastic process: (individuals in n possibly states) π i proportion of population is in state i p ij is probability of transition from state i j So the transition step k k + 1 is: π (k+1) = P T π (k) A probability distribution π is called steady state if As a corollary of Gordan s result: π = P T π Each stochastic matrix P has a steady state π. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 34/40

Table of Contents 1 Introduction 2 Gauss-elimination [MO, 2.1] 3 Orthogonal projection, Least Squares, [MO, 2.2] 4 Linear Inequalities [MO, 2.4] 5 Integer Solutions of Linear Equations [MO, 2.3] Two variables Lattices Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 35/40

Integer Solutions of Linear Equations Example Equation 3x 1 2x 2 = 1 has solution x = (1, 1) Z 2. But equation 6x 1 2x 2 = 1 does not allow an integer solution x. Key remark: Let a 1, a 2 Z and let a 1 x 1 + a 2 x 2 = b have a solution x 1, x 2 Z. Then b = λc with λ Z, c = gcd(a 1, a 2 ) Here: gcd(a 1, a 2 ) denotes the greatest common divisor of a 1, a 2. Lemma 2.15 (Euclid s Algorithm, [MO, Lm. 2.2]) Let c = gcd(a 1, a 2 ). Then L(a 1, a 2 ) := {a 1 λ 1 + a 2 λ 2 λ 1, λ 2 Z} = {cλ λ Z} =: L(c). (The proof of) this result allows to solve (if possible) a 1 x 1 + a 2 x 2 = b (in Z). Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 36/40

Algorithm to solve, a 1 x 1 + a 2 x 2 = b (in Z) Compute c = gcd(a 1, a 2 ). If λ := b/c / Z, no integer solution exists. If λ := b/c Z, compute solutions λ 1, λ 2 Z of λ 1 a 1 + λ 2 a 2 = c. Then (λ 1 λ)a 1 + (λ 2 λ)a 2 = b. General problem: Given a 1,..., a n, b Z m, find x = (x 1,, x n ) Z n such that ( ) a 1 x 1 + a 2 x 2 +... + a n x n = b or equivalently Ax = b where A := [a 1,..., a n ]. Def. We introduce the lattice generated by a 1,..., a n, { n } L = L(a 1,..., a n ) = a jλ j λ j Z R m. j=1 Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 37/40

Assumption 1: rank A = m (m n); w.l.o.g., a 1,..., a m are linearly independent. To solve the problem: Find C = [c 1... c m ] Z m m such that ( ) L(c 1,..., c m ) = L(a 1,..., a n ). Then ( ) has a solution x Z n iff λ := C 1 b Z m Bad news: As in the case of one equation: in general Lemma 2.16 ([MO, Lm. 2.3]) L(a 1,..., a m ) L(a 1,..., a n ). Let c 1,..., c m L(a 1,..., a n ). Then L(c 1,..., c m ) = L(a 1,..., a n ) if and only if for all j = 1,..., n, the system Cλ = a j has an integral solution. Last step: Find such c i s Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 38/40

Main Result: The algorithm Lattice Basis INIT: C = [c 1,..., c m ] = [a 1,..., a m ] ; ITER: Compute C 1 ; If C 1 a j Z m for j = 1,..., n, then stop; If λ = C 1 a j / Z m for some j, then Let a j = Cλ = m i=1 λ ic i and compute c = m i=1 (λ i [λ i ])c i = a j m i=1 [λ i]c i ; Let k be the largest index i such that λ i / Z ; Update C by replacing c k with c in column k; next iteration stops after at most K = log 2 (det[a 1,..., a m ]) steps with a matrix C satisfying ( ). Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 39/40

As an exercise: Explain how an integer solution x of ( ) can be constructed with the help of the results from the algorithm above. Complexity From the algorithm above we see: To solve integer systems of equations is polynomial. Under additional inequalities (such as x 0) the problem becomes NP-hard Theorem 2.17 ([MO, Thm. 2.4]) Let A Z m n and b Z m be given. Then exactly one of the following statements is true: (a) There exists some x Z n such that Ax = b. (b) There exists some y R m such that y T A Z n and y T b Z. Peter J.C. Dickinson http://dickinson.website MO18, Chpt 2: Linear Equations and inequalities 40/40