Basics of Linear Algebra Properties of matrix calculation (𝐴, 𝐵 are matrices, 𝑥⃑ is a vector/matrix with one column) (𝐴𝐵)𝑥⃑ = 𝐴(𝐵𝑥⃑) 𝐼𝐴 = 𝐴𝐼 = 𝐴 (𝐴𝐵)𝐶 = 𝐴(𝐵𝐶) 𝐴(𝐵 + 𝐶) = 𝐴𝐵 + 𝐴𝐶 𝐴𝐵 ≠ 𝐵𝐴 (𝐴𝐵)−1 = 𝐵 −1 𝐴−1 𝐼 −1 = 𝐼 Gaussian Elimination 𝑥 + 2𝑦 + 𝑧 = 5 {𝑥 + 𝑦 + 2𝑧 = 3 2𝑥 + 𝑦 + 𝑧 = 4 1 2 15 1 2 1 5 1 2 1 5 (1 1 2|3) ~ (0 −1 1 |−2) ~ (0 −1 1 |−2) 2 1 14 0 −3 −1 −6 0 0 −4 0 𝑅2 → 𝑅2 − 𝑅1 , 𝑅3 → 𝑅3 − 2𝑅1 , 𝑅3 → 𝑅3 + 3𝑅2 A leading entry in a row is the left-most non-zero entry in that row Gaussian elimination is the conversion of a matrix into echelon form (each leading entry is to the right of the leading entry in the row above it). Leading entries in echelon form are called pivots A system of linear equations is consistent if it has solutions A system has a unique solution iff an echelon form has a pivot in all columns Elementary row operations (can be applied in Gaussian elimination) Multiplying a row by a non-zero constant Exchanging rows Substituting a row by that row minus a multiple of another row 𝑅𝑖 → 𝑅𝑖 − 𝑐𝑅𝑗 Identity matrix 𝐼𝑛 1 0 0 (0 1 0 ) 0 0 1 Elementary matrix 𝐸𝑖𝑗 1 0 0 (−𝑙 1 0) 0 0 1 Diagonal matrix 𝐷 1 0 0 (0 2 0) 0 0 3 Symmetric matrix 𝑆 2 6 4 (6 1 9 ) 4 9 3 Upper triangular matrix 𝑈 1 7 6 (0 3 4) 0 0 1 Lower triangular matrix 𝐿 −1 0 0 ( 7 3 0) 4 2 1 Row exchange matrix 𝑃 is obtained by exchanging the rows of 𝐼, and 𝑃𝐴 is the rows of 𝐴 exchanged in the same way 𝐴 is invertible iff the inverse 𝐴−1 such that 𝐴𝐴−1 = 𝐴−1 𝐴 = 𝐼 exists LDU Factorization 𝐴 = 𝐿𝐷𝑈 or 𝑃𝐴 = 𝐿𝐷𝑈 Example: if the steps of Gaussian elimination for 𝐴3×3 are 𝑅2 → 𝑅2 − 𝑅1 , 𝑅3 → 𝑅3 − 2𝑅1 , 𝑅3 → 𝑅3 + 3𝑅2 then 1 0 0 𝐿 = (1 1 0) 2 −3 1 𝐷 is a matrix with pivots, and 𝑈 is the echelon form with rows divided by their pivots If 𝐴 is invertible, then the 𝐿𝐷𝑈 factorization is unique The transpose matrix 𝐴𝑇 is obtained by switching the rows and columns of 𝐴 𝐴 is symmetric if 𝐴𝑇 = 𝐴 Properties of transpose matrices (𝐴𝑇 )𝑇 = 𝐴 (𝐴𝐵)𝑇 = 𝐵 𝑇 𝐴𝑇 (𝐴𝑇 )−1 = (𝐴−1 )𝑇 if 𝐴 is invertible If 𝐴 is symmetric and invertible, then 𝐴−1 is symmetric as well 𝐴𝑇 𝐴 and 𝐴𝐴𝑇 are symmetric If 𝐴 = 𝐿𝐷𝑈 is symmetric, then 𝑈 = 𝐿𝑇 i.e. 𝐴 = 𝐿𝐷𝐿𝑇 𝑁(𝐴𝑇 𝐴) = 𝑁(𝐴) Vector spaces If 𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑘 are in ℂ𝑛 then 𝑐1 𝑣⃑1 , 𝑐2 𝑣⃑2 , … , 𝑐𝑘 𝑣⃑𝑘 is a linear combination of 𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑘 , where 𝑐1 , 𝑐2 , … , 𝑐𝑘 are constants A subspace of a vector space is a non-empty subset of that space that contains the origin (so it’s non-empty), and any sum of two of its elements, and any scalar multiplication of its elements The set of linear combinations of 𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑘 in a vector space 𝑉 is a subspace of 𝑉 The column space of 𝐴 or 𝐶(𝐴) is the set of all linear combinations of the columns of 𝐴 The nullspace of 𝐴 or 𝑁(𝐴) is the set of vectors 𝑥⃑ such that 𝐴𝑥⃑ = ⃑0⃑ Finding the nullspace of 𝐴 from its echelon form: 1 2 30 (0 1 2|0) 0 0 00 𝑥3 = 𝛼 (free variable) 𝑥2 = −2𝑥3 = −2𝛼 𝑥1 = −2𝑥2 − 3𝑥3 = 4𝛼 − 3𝛼 = 𝛼 1 𝑁(𝐴) = {𝛼 (2)} 1 𝐴𝑛×𝑛 is invertible ⟺ 𝐴 has 𝑛 pivots ⟺ 𝐴𝑥⃑ = 𝑏⃑⃑ has exactly one solution 𝑥⃑ = 𝐴−1 𝑏⃑⃑ The solutions to 𝐴𝑥⃑ = 𝑏⃑⃑ have the form 𝑥⃑ = 𝑥⃑𝑛 + 𝑥⃑𝑝 where 𝑥⃑𝑛 is the general solution of the homogeneous equation (= ⃑0⃑), 𝐴𝑥⃑ = ⃑0⃑ and 𝑥⃑𝑝 is a particular solution of the equation 𝐴𝑥⃑ = 𝑏⃑⃑ 𝐶(𝐴𝐵) ⊆ 𝐶(𝐴), 𝑁(𝐵) ⊆ 𝑁(𝐴𝐵) In a reduced echelon form, all the pivots are 1 and all the elements above the pivots are zero The rank of 𝐴 is the number of pivots in its echelon form Linear independence, basis and dimension If 𝑐1 𝑣⃑1 , 𝑐2 𝑣⃑2 , … , 𝑐𝑘 𝑣⃑𝑘 = ⃑⃑ 0 is true only for 𝑐1 , 𝑐2 , … , 𝑐𝑘 = 0, then 𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑘 are linearly independent The columns of 𝐴 are linearly independent ⟺ its echelon form has a pivot in ⃑⃑} every column ⟺ 𝑁(𝐴) = {0 If the vectors 𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑘 are linearly dependent, then one of these vectors is a linear combination of the others If all elements in 𝑉 are linear combinations of 𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑘 , then 𝑉 is a span of those vectors. A basis for a vector space 𝑉 is a set of vectors that are linearly independent and span 𝑉. The standard basis of ℝ𝑛 can be expressed as {𝑒⃑1 , 𝑒⃑2 , … , 𝑒⃑𝑛 } where 𝑒⃑𝑖 is the ith column of 𝐼𝑛 The dimension of 𝑉 is the cardinality of its basis (which is the same for all bases) Any linearly independent set of vectors in 𝑉 can be extended to a basis by adding more vectors (if necessary) 𝑑𝑖𝑚𝐶(𝐴) = 𝑟, 𝑑𝑖𝑚𝑁(𝐴) = 𝑛 − 𝑟 where n is the number of columns and r is the rank of 𝐴 The row space 𝐶(𝐴𝑇 ) is the span of the rows of 𝐴. 𝑑𝑖𝑚𝐶(𝐴𝑇 ) = 𝑟 The left nullspace 𝑁(𝐴𝑇 ) consists of all the row vectors 𝑦⃑ such that 𝑦⃑𝐴 = ⃑⃑ 0 𝑇 𝑇 i.e. 𝐴 𝑦⃑ = ⃑0⃑ Linear Transformations A function 𝑇: 𝑉 → 𝑊 is a linear transformation iff 𝑇(𝑐1 𝑣⃑1 + 𝑐2 𝑣⃑2 ) = 𝑐1 𝑇(𝑣⃑1 ) + 𝑐2 𝑇(𝑣⃑2 ) The set of polynomials with real coefficients of degree at most 𝑛 is denoted as 𝑃𝑛 . Its standard basis is {1, 𝑡, 𝑡 2 , … , 𝑡 𝑛 } Suppose ℬ1 = {𝑣⃑1 , 𝑣⃑2 , … , 𝑣⃑𝑛 } is a basis for 𝑉 and ℬ2 = {𝑤 ⃑⃑⃑1 , 𝑤 ⃑⃑⃑2 , … , 𝑤 ⃑⃑⃑𝑛 } is a basis for 𝑊. Each linear transformation from 𝑉 to 𝑊 is represented by the matrix 𝑎1,1 𝑎1,2 ⋯ 𝑎1,𝑛 𝑎2,1 𝑎2,2 ⋯ 𝑎2,𝑛 𝐴=( ⋮ ⋮ ⋱ ⋮ ) 𝑎𝑚,1 𝑎𝑚,2 ⋯ 𝑎𝑚,𝑛 Where 𝑇(𝑣⃑𝑗 ) = 𝑎1,𝑗 𝑤 ⃑⃑⃑1 + 𝑎2,𝑗 𝑤 ⃑⃑⃑2 + ⋯ + 𝑎𝑚,𝑗 𝑤 ⃑⃑⃑𝑚 The kernel for a linear transformation 𝑇: 𝑉 → 𝑊: 𝐾𝑒𝑟𝑇 = {𝑣⃑ ∈ 𝑉|𝑇(𝑣⃑) = ⃑0⃑}. 𝐾𝑒𝑟𝑇 is a subspace of 𝑉 The range for 𝑇: 𝑅𝑎𝑛𝑔𝑒𝑇 = {𝑇(𝑣⃑)|𝑣⃑ ∈ 𝑉}. 𝑅𝑎𝑛𝑔𝑒𝑇 is a subspace of 𝑊 𝑇 is surjective ⟺ 𝐶(𝐴) = 𝔽𝑚 ⟺ 𝑅𝑎𝑛𝑔𝑒𝑇 = 𝑊 ⃑⃑} ⟺ 𝑇 is injective ⟺ the columns of 𝐴 are linearly independent ⟺ 𝑁(𝐴) = {0 ⃑⃑} 𝐾𝑒𝑟𝑇 = {0 𝑇: 𝑉 → 𝑊 is invertible iff there exists 𝑇 −1 : 𝑊 → 𝑉 s.t. 𝑇(𝑇 −1 (𝑤 ⃑⃑⃑)) = 𝑤 ⃑⃑⃑, −1 𝑇 (𝑇(𝑣⃑)) = 𝑣⃑ If 𝑇: 𝑉 → 𝑊 is invertible and has matrix 𝐴 then the matrix of 𝑇 −1 is 𝐴−1 Let 𝑉, 𝑈, 𝑊 be vector spaces with bases ℬ𝑣 , ℬ𝑢 , ℬ𝑤 . Let 𝑇1 : 𝑉 → 𝑈 and 𝑇2 : 𝑈 → 𝑊 be linear transformations. Let 𝐴1 and 𝐴2 be the matrices for 𝑇1 and 𝑇2 with ℬ𝑣 , ℬ𝑢 and ℬ𝑢 , ℬ𝑤 , respectively. Then 𝑇2 𝑇1 : 𝑉 → 𝑊 is a linear transformation, and the linear transformation matrix 𝐴 = 𝐴2 𝐴1 with ℬ𝑣 , ℬ𝑤 Orthogonality in ℝ𝑛 For any 𝑥⃑ ∈ ℝ𝑛 , ||𝑥⃑|| = √(𝑥1 )2 + (𝑥2 )2 + ⋯ + (𝑥𝑛 )2 Vectors 𝑥⃑ and 𝑦⃑ are orthogonal (perpendicular) iff 𝑥⃑ 𝑇 𝑦⃑ = 0 (𝑥⃑ 𝑇 𝑦⃑ = 𝑦⃑ 𝑇 𝑥⃑) Let 𝑉 be a subspace of ℝ𝑛 . A basis for 𝑉 that consists of mutually orthogonal vectors is called an orthogonal basis Two subspaces 𝑉 and 𝑊 of ℝ𝑛 are orthogonal iff 𝑣⃑ 𝑇 𝑤 ⃑⃑⃑ = 0 for all 𝑣⃑ ∈ 𝑉, 𝑤 ⃑⃑⃑ ∈ 𝑊, denoted as 𝑉 ⊥ 𝑊 𝐶(𝐴𝑇 ) is orthogonal to 𝑁(𝐴) and 𝐶(𝐴) is orthogonal to 𝑁(𝐴𝑇 ) Given a subspace 𝑉 of ℝ𝑛 , its orthogonal complement 𝑉 ⊥ is the set of all the vectors orthogonal to 𝑉. 𝑉 ⊥ = 𝑈 ⟺ 𝑈 ⊥ = 𝑉 Projections The projection 𝑝⃑ of 𝑏⃑⃑ onto the line through 𝑎⃑ is 𝑎⃑𝑇 𝑏⃑⃑ 𝑝⃑ = 𝑇 𝑎⃑ 𝑎⃑ 𝑎⃑ 𝑎⃑𝑎⃑𝑇 = 𝑇 𝑏⃑⃑ 𝑎⃑ 𝑎⃑ = 𝑃𝑏⃑⃑ where 𝑃 is the n by n projection matrix. For any projection matrix 𝑃, 𝑃 = 𝑃𝑇 = 𝑃2 Consider 𝐴𝑥⃑=𝑏⃑⃑ where there is no solution for 𝑥⃑. Let 𝑝⃑ = 𝐴𝑥̂ is the projection of 𝑏⃑⃑ onto 𝐶(𝐴). Then the normal equation: 𝐴𝑇 𝐴𝑥̂ = 𝐴𝑇 𝑏⃑⃑ holds If 𝐴𝑇 𝐴 is invertible, then 𝑝⃑ = 𝑃𝑏⃑⃑ where 𝑃 = 𝐴(𝐴𝑇 𝐴)−1 𝐴𝑇 𝐴𝑇 𝐴 is symmetric and is invertible when the columns of 𝐴 are linearly independent ⊥ If 𝑏⃑⃑ is in (𝐶(𝐴)) = 𝑁(𝐴), then its projection onto 𝐶(𝐴) is 0 If 𝐴 is square, then 𝑝⃑ = 𝑏⃑⃑ since the columns of 𝐴 spans ℝ𝑛 Orthonormal bases and Gram-Schmidt The vectors 𝑞⃑1 , 𝑞⃑2 , … , 𝑞⃑𝑘 are orthonormal in ℝ𝑛 if 0, 𝑖 ≠ 𝑗 𝑞⃑𝑖𝑇 𝑞⃑𝑗 = { 1, 𝑖 = 𝑗 A real matrix 𝑄 has orthonormal columns ⟺ 𝑄 𝑇 𝑄 = 𝐼 A real matrix 𝑄 that is square and has orthonormal columns is called orthogonal. 𝑄 is orthogonal ⟺ 𝑄 𝑇 𝑄 = 𝐼 and 𝑄 is square ⟺ 𝑄 −1 = 𝑄 𝑇 Matrix multiplication by a real matrix 𝑄 preserves lengths and angles If 𝑄 is orthogonal then it has orthonormal rows If 𝑄 is real and has orthonormal columns then the least-squares solution of 𝑄𝑥⃑ =𝑏⃑⃑ is 𝑥̂ = 𝑄 𝑇 𝑏⃑⃑. The projection of 𝑏⃑⃑ onto 𝐶(𝑄) is 𝑝⃑ = 𝑃𝑏⃑⃑ where 𝑃 = 𝑄𝑄 𝑇 Gram-Schmidt Process takes linearly independent vectors 𝑎⃑1 , 𝑎⃑2 , … , 𝑎⃑𝑘 ∈ ℝ𝑛 that span 𝑉, and gives orthonormal vectors 𝑞⃑1 , 𝑞⃑2 , … , 𝑞⃑𝑘 that form a basis for 𝑉 𝐴⃑ 𝑇 𝑞⃑𝑖 = 𝑖 , where 𝐴⃑𝑖 = 𝑎⃑𝑖 − (𝑞⃑1𝑇 𝑎⃑𝑖 )𝑞⃑1 − (𝑞⃑2𝑇 𝑎⃑𝑖 )𝑞⃑2 − ⋯ − (𝑞⃑𝑖−1 𝑎⃑𝑖 )𝑞⃑𝑖−1 . Subtracting |𝐴⃑𝑖 | the projection onto a line through a vector gives a vector orthogonal to it. Every m by n matrix 𝐴 with real entries and linearly independent columns can be factored into 𝐴 = 𝑄𝑅, where 𝑄 has orthonormal columns and 𝑅 is n by n upper-triangular and invertible Determinants Determinant is a value such that 𝑑𝑒𝑡𝐼 = 1, −𝑑𝑒𝑡𝐵 = 𝑑𝑒𝑡𝐴 where 𝐵 is obtained from 𝐴 by exchanging two rows, and depends linearly on the rows of 𝐴: 𝑐𝑎1,1 + 𝑑𝑏1,1 𝑐𝑎1,1 + 𝑑𝑏1,1 ⋯ 𝑐𝑎1,1 + 𝑑𝑏1,1 𝑎2,1 𝑎2,2 ⋯ 𝑎2,𝑛 det ( ) ⋮ ⋮ ⋱ ⋮ 𝑎𝑛,1 𝑎𝑛,2 ⋯ 𝑎𝑛,𝑛 𝑎1,1 𝑎1,2 ⋯ 𝑎1,𝑛 𝑏1,1 𝑏1,2 ⋯ 𝑏1,𝑛 𝑎2,1 𝑎2,2 ⋯ 𝑎2,𝑛 𝑎2,1 𝑎2,2 ⋯ 𝑎2,𝑛 = 𝑐det ( ⋮ ) ⋮ ⋱ ⋮ ) + 𝑑det ( ⋮ ⋮ ⋱ ⋮ 𝑎𝑛,1 𝑎𝑛,2 ⋯ 𝑎𝑛,𝑛 𝑎𝑛,1 𝑎𝑛,2 ⋯ 𝑎𝑛,𝑛 The determinant of an upper/lower triangular matrix is the product of the entries on the main diagonal Matrix 𝐴 is invertible ⟺ 𝑑𝑒𝑡𝐴 ≠ 0 det(𝐴𝐵) = 𝑑𝑒𝑡𝐴 ∙ 𝑑𝑒𝑡𝐵 𝑑𝑒𝑡𝐴 = 𝑑𝑒𝑡𝐴𝑇 The cofactor 𝐶𝑖𝑗 = (−1)𝑖+𝑗 𝑑𝑒𝑡𝑀𝑖𝑗 where the minor 𝑀𝑖𝑗 is the determinant of the matrix formed by removing the ith row and jth column of 𝐴 The determinant of 𝐴 is a combination of any row times its cofactors: 𝑑𝑒𝑡𝐴 = 𝑎𝑖1 𝐶𝑖1 + 𝑎𝑖2 𝐶𝑖2 + ⋯ + 𝑎𝑖𝑛 𝐶𝑖𝑛 1 If 𝐴 is invertible then 𝐴−1 = 𝑑𝑒𝑡𝐴 𝐶 𝑇 where 𝐶 is the cofactor matrix of 𝐴 with 𝐶𝑖𝑗 on the ith row and jth column 𝑎1,1 𝑎1,2 ⋯ 𝑎1,𝑛 𝑎2,1 𝑎2,2 ⋯ 𝑎2,𝑛 Cramer’s Rule: suppose 𝐴 = ( ⋮ ⋮ ⋱ ⋮ ) is invertible. We can 𝑎𝑛,1 𝑎𝑛,2 ⋯ 𝑎𝑛,𝑛 𝑥1 𝑥2 𝑑𝑒𝑡𝐵 solve for 𝑥𝑖 in 𝐴 ( ⋮ ) = 𝑏⃑⃑ by 𝑥𝑖 = 𝑑𝑒𝑡𝐴𝑖 where 𝐵𝑖 is obtained by substituting 𝑥𝑛 the ith column of 𝐴 with 𝑏⃑⃑ Eigenvalues and eigenvectors Let 𝐴 be an n by n matrix and suppose that 𝑥⃑ ≠ 0, and 𝐴𝑥⃑ = 𝜆𝑥⃑. Then 𝜆 is called an eigenvalue of 𝐴, and 𝑥⃑ an eigenvector of 𝐴. The eigenvalues of 𝐴 are the solutions of det(𝐴 − 𝜆𝐼) = 0 The eigenvalues can be calculated by solving for the values of 𝜆 that makes det(𝐴 − 𝜆𝐼) = 0, then solve for the eigenvectors by substituting 𝜆 into (𝐴 − 𝜆𝐼)𝑥⃑ = ⃑⃑ 0 The eigenvalues of a projection matrix are 1 or 0 The eigenvalues of 𝐴 are the roots of the characteristic polynomial 𝑝(𝜆) = det(𝐴 − 𝜆𝐼) The trace of a square matrix 𝐴 is the sum of the elements on its main diagonal 𝑑𝑒𝑡𝐴 = 𝜆1 𝜆2 … 𝜆𝑛 𝑡𝑟𝑎𝑐𝑒𝐴 = 𝜆1 + 𝜆2 + ⋯ + 𝜆𝑛 If a square matrix 𝐴 is upper or lower triangular, then its eigenvalues are on its main diagonal If 𝑥⃑ is an eigenvector of 𝐴 with corresponding eigenvalue 𝜆, then inverse 𝑥⃑ is also an eigenvector with corresponding eigenvalue 𝜆̅ Diagonalization Suppose an n by n matrix 𝐴 has n linearly independent eigenvectors. Let 𝑆 be a matrix whose columns are the linearly independent eigenvectors of 𝐴. Then 𝐴 = 𝑆𝐷𝑆 −1 where 𝐷 is the diagonal matrix with the eigenvalues of 𝐴 on the main diagonal: 𝜆1 0 ⋯ 0 0 𝜆2 ⋯ 0 𝐴 = 𝑆𝐷𝑆 −1 = (𝑣⃑1 𝑣⃑2 ⋯ 𝑣⃑𝑛 ) ( ) (𝑣⃑1 𝑣⃑2 ⋯ 𝑣⃑𝑛 )−1 ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ 𝜆𝑛 If eigenvectors 𝑥⃑1 , 𝑥⃑2 , … , 𝑥⃑𝑘 correspond to distinct eigenvalues 𝜆1 , 𝜆2 , … , 𝜆𝑘 of 𝐴, then 𝑥⃑1 , 𝑥⃑2 , … , 𝑥⃑𝑘 are linearly independent Suppose 𝐴 is n by n and has n distinct eigenvalues. Then 𝐴 is diagonalizable. Not all square matrices are diagonalizable. Let 𝜆1 be an eigenvalue of 𝐴. The algebraic multiplicity of 𝜆1 is the greatest positive integer 𝑘 such that (𝜆 − 𝜆1 )𝑘 divides 𝑝(𝜆) = det(𝐴 − 𝜆𝐼) The geometric multiplicity of 𝜆1 is the number of linearly independent eigenvectors that correspond to 𝜆1 , i.e. dim(𝐴 − 𝜆𝐼). It is equal to the number of free variables in the echelon form for 𝐴 − 𝜆𝐼 To be diagonalizable, an n by n matrix 𝐴 has to have n linearly independent eigenvectors. To be invertible, none of eigenvalues can be 0. Let 𝐴 have eigenvectors 𝑥⃑1 , 𝑥⃑2 , … , 𝑥⃑𝑘 corresponding to eigenvalues 𝜆1 , 𝜆2 , … , 𝜆𝑘 . Then 𝑥⃑1 , 𝑥⃑2 , … , 𝑥⃑𝑘 are linearly independent 𝐴 is diagonalizable if for each eigenvalue of 𝐴, its algebraic multiplicity equals the geometric multiplicity. If 𝐴 has eigenvalues 𝜆1 , 𝜆2 , … , 𝜆𝑛 , then 𝐴𝑘 has eigenvalues 𝜆1𝑘 , 𝜆𝑘2 , … , 𝜆𝑘𝑛 . If 𝐴 is 1 1 1 invertible, then 𝐴−1 has eigenvalues 𝜆 , 𝜆 , … , 𝜆 1 2 𝑛 If 𝐴 is square 𝐴 and 𝐴𝑇 have same characteristic polynomial If 𝐴 = 𝑆𝐷𝑆 −1, then 𝐴𝑘 = 𝑆𝐷𝑘 𝑆 −1 for some natural number 𝑘 To solve a recurrence 𝑎𝑘+2 = 𝑏𝑘+1 𝑎𝑘+1 + 𝑏𝑘 𝑎𝑘 , find the eigenvalues of 𝐴 such 𝑎𝑘+2 𝑎𝑘+1 that (𝑎 ) = 𝐴 ( 𝑎 ), then solve 𝑎𝑘 = 𝑐1 𝜆𝑘1 + 𝑐2 𝜆𝑘2 for the constants by 𝑘+1 𝑘 substituting given values of 𝑎 A probability vector 𝑝⃑ has non-negative entries that add up to one A Markov matrix is a square matrix whose columns are probability vectors Any Markov matrix has 𝜆 = 1 as an eigenvalue A probability vector 𝑝⃑ is a steady-state vector for a Markov matrix 𝑀 if 𝑀𝑝⃑ = 𝑝⃑. It can be calculated by solving (𝑀 − 𝐼)𝑝⃑ = ⃑0⃑ If lim 𝑀𝑛 (= 𝑆𝐷𝑛 𝑆 −1 ) = 𝐿, then all the columns of 𝐿 are steady-state vectors 𝑛→∞ A Markov matrix 𝑀 is regular if there exists a positive integer 𝑝 such that 𝑀𝑝 has all positive entries. If a Markov matrix is regular, it has a unique steady-state vector If 𝐴 and 𝐵 are 𝑛x𝑛 matrices, then 𝐴 and 𝐵 −1 𝐴𝐵 are similar Similar matrices have the same eigenvalues 1 1 If the elements of each row of 𝐴 sum up to 𝑟, then 𝑥⃑ = ( ) is an eigenvector ⋮ 1 with 𝜆 = 𝑟 For any eigenvalue, its geometric multiplicity is less than or equal to its algebraic multiplicity Complex Space The conjugate transpose of 𝐴 is 𝐴𝐻 = ̅̅ 𝐴̅̅𝑇 The inner product of vectors 𝑥⃑, 𝑦⃑ in ℂ𝑛 is 𝑥⃑ 𝐻 𝑦⃑ The length of a complex vector 𝑥⃑ is ||𝑥|| = √𝑥⃑ 𝐻 𝑥⃑ Vectors 𝑥⃑, 𝑦⃑ are orthogonal if 𝑥⃑ 𝐻 𝑦⃑ = 0 (𝐴𝐵)𝐻 = 𝐵 𝐻 𝐴𝐻 Matrix 𝐴 is Hermitian if 𝐴 = 𝐴𝐻 If 𝐴 is Hermitian, the elements on the main diagonal are real A real symmetric matrix is Hermitian If 𝐴 is Hermitian and 𝑥⃑ is a complex vector, 𝑥⃑ 𝐻 𝐴𝑥⃑ is a real number The eigenvalues of a Hermitian matrix are real Real symmetric matrices have real eigenvalues Two eigenvectors of a Hermitian matrix are orthogonal to each other if they come from different eigenvalues The Spectral Theorem: A real symmetric matrix can be factored into 𝐴 = 𝑄𝐷𝑄 𝑇 where 𝐷 is diagonal and 𝑄 is an orthogonal matrix If 𝐴 is symmetric, it can be represented as a linear combination of projection matrices of rank 1: 𝐴 = 𝜆1 𝑞⃑1 𝑞⃑1𝑇 + 𝜆2 𝑞⃑2 𝑞⃑2𝑇 + ⋯ + 𝜆𝑛 𝑞⃑𝑛 𝑞⃑𝑛𝑇 The normal equation in complex space: 𝐴𝐻 𝐴𝑥̂ = 𝐴𝐻 𝑏⃑⃑ 𝐻 𝑎⃑⃑𝑎⃑⃑ The projection of 𝑏⃑⃑ onto 𝐶(𝐴) in complex space is 𝑝⃑ = 𝑎⃑⃑𝐻𝑎⃑⃑ 𝑏⃑⃑ 𝑁(𝐴𝐻 𝐴) = 𝑁(𝐴), 𝑁(𝐴)⊥ = 𝐶(𝐴𝐻 ), 𝐶(𝐴)⊥ = 𝑁(𝐴𝐻 )