03 Machine Learning Linear Algebra
12 likes6,674 views
The document provides an introduction to linear algebra concepts for machine learning. It defines vectors as ordered tuples of numbers that express magnitude and direction. Vector spaces are sets that contain all linear combinations of vectors. Linear independence and basis of vector spaces are discussed. Norms measure the magnitude of a vector, with examples given of the 1-norm and 2-norm. Inner products measure the correlation between vectors. Matrices can represent linear operators between vector spaces. Key linear algebra concepts such as trace, determinant, and matrix decompositions are outlined for machine learning applications.
03 Machine Learning Linear Algebra
- 1. Machine Learning for Data Mining Linear Algebra Review Andres Mendez-Vazquez May 14, 2015 1 / 50
- 2. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 2 / 50
- 3. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 3 / 50
- 4. What is a Vector? A ordered tuple of numbers x = x1 x2 ... xn Expressing a magnitude and a direction 4 / 50
- 5. What is a Vector? A ordered tuple of numbers x = x1 x2 ... xn Expressing a magnitude and a direction Magnitude Direction 4 / 50
- 6. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 5 / 50
- 7. Vector Spaces Definition A vector is an element of a vector space Vector Space V It is a set that contains all linear combinations of its elements: 1 If x, y ∈ V then x + y ∈ V . 2 If x ∈ V then αx ∈ V for any scalar α. 3 There exists 0 ∈ V then x + 0 = x for any x ∈ V . A subspace It is a subset of a vector space that is also a vector space 6 / 50
- 8. Vector Spaces Definition A vector is an element of a vector space Vector Space V It is a set that contains all linear combinations of its elements: 1 If x, y ∈ V then x + y ∈ V . 2 If x ∈ V then αx ∈ V for any scalar α. 3 There exists 0 ∈ V then x + 0 = x for any x ∈ V . A subspace It is a subset of a vector space that is also a vector space 6 / 50
- 9. Vector Spaces Definition A vector is an element of a vector space Vector Space V It is a set that contains all linear combinations of its elements: 1 If x, y ∈ V then x + y ∈ V . 2 If x ∈ V then αx ∈ V for any scalar α. 3 There exists 0 ∈ V then x + 0 = x for any x ∈ V . A subspace It is a subset of a vector space that is also a vector space 6 / 50
- 10. Vector Spaces Definition A vector is an element of a vector space Vector Space V It is a set that contains all linear combinations of its elements: 1 If x, y ∈ V then x + y ∈ V . 2 If x ∈ V then αx ∈ V for any scalar α. 3 There exists 0 ∈ V then x + 0 = x for any x ∈ V . A subspace It is a subset of a vector space that is also a vector space 6 / 50
- 11. Vector Spaces Definition A vector is an element of a vector space Vector Space V It is a set that contains all linear combinations of its elements: 1 If x, y ∈ V then x + y ∈ V . 2 If x ∈ V then αx ∈ V for any scalar α. 3 There exists 0 ∈ V then x + 0 = x for any x ∈ V . A subspace It is a subset of a vector space that is also a vector space 6 / 50
- 12. Classic Example Euclidean Space R3 7 / 50
- 13. Span Definition The span of any set of vectors {x1, x2, ..., xn} is defined as: span (x1, x2, ..., xn) = α1x1 + α2x2 + ... + αnxn What Examples can you Imagine? Give it a shot!!! 8 / 50
- 14. Span Definition The span of any set of vectors {x1, x2, ..., xn} is defined as: span (x1, x2, ..., xn) = α1x1 + α2x2 + ... + αnxn What Examples can you Imagine? Give it a shot!!! 8 / 50
- 15. Subspaces of Rn A line through the origin in Rn A plane in Rn 9 / 50
- 16. Subspaces of Rn A line through the origin in Rn A plane in Rn 9 / 50
- 17. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 10 / 50
- 18. Linear Independence and Basis of Vector Spaces Fact 1 A vector x is a linearly independent of a set of vectors {x1, x2, ..., xn} if it does not lie in their span. Fact 2 A set of vectors is linearly independent if every vector is linearly independent of the rest. The Rest 1 A basis of a vector space V is a linearly independent set of vectors whose span is equal to V 2 If the basis has d vectors then the vector space V has dimensionality d. 11 / 50
- 19. Linear Independence and Basis of Vector Spaces Fact 1 A vector x is a linearly independent of a set of vectors {x1, x2, ..., xn} if it does not lie in their span. Fact 2 A set of vectors is linearly independent if every vector is linearly independent of the rest. The Rest 1 A basis of a vector space V is a linearly independent set of vectors whose span is equal to V 2 If the basis has d vectors then the vector space V has dimensionality d. 11 / 50
- 20. Linear Independence and Basis of Vector Spaces Fact 1 A vector x is a linearly independent of a set of vectors {x1, x2, ..., xn} if it does not lie in their span. Fact 2 A set of vectors is linearly independent if every vector is linearly independent of the rest. The Rest 1 A basis of a vector space V is a linearly independent set of vectors whose span is equal to V 2 If the basis has d vectors then the vector space V has dimensionality d. 11 / 50
- 21. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 12 / 50
- 22. Norm of a Vector Definition A norm u measures the magnitud of the vector. Properties 1 Homogeneity: αx = α x . 2 Triangle inequality: x + y ≤ x + y . 3 Point Separation x = 0 if and only if x = 0. Examples 1 Manhattan or 1-norm : x 1 = d i=1 |xi|. 2 Euclidean or 2-norm : x 2 = d i=1 x2 i . 13 / 50
- 23. Norm of a Vector Definition A norm u measures the magnitud of the vector. Properties 1 Homogeneity: αx = α x . 2 Triangle inequality: x + y ≤ x + y . 3 Point Separation x = 0 if and only if x = 0. Examples 1 Manhattan or 1-norm : x 1 = d i=1 |xi|. 2 Euclidean or 2-norm : x 2 = d i=1 x2 i . 13 / 50
- 24. Norm of a Vector Definition A norm u measures the magnitud of the vector. Properties 1 Homogeneity: αx = α x . 2 Triangle inequality: x + y ≤ x + y . 3 Point Separation x = 0 if and only if x = 0. Examples 1 Manhattan or 1-norm : x 1 = d i=1 |xi|. 2 Euclidean or 2-norm : x 2 = d i=1 x2 i . 13 / 50
- 25. Norm of a Vector Definition A norm u measures the magnitud of the vector. Properties 1 Homogeneity: αx = α x . 2 Triangle inequality: x + y ≤ x + y . 3 Point Separation x = 0 if and only if x = 0. Examples 1 Manhattan or 1-norm : x 1 = d i=1 |xi|. 2 Euclidean or 2-norm : x 2 = d i=1 x2 i . 13 / 50
- 26. Norm of a Vector Definition A norm u measures the magnitud of the vector. Properties 1 Homogeneity: αx = α x . 2 Triangle inequality: x + y ≤ x + y . 3 Point Separation x = 0 if and only if x = 0. Examples 1 Manhattan or 1-norm : x 1 = d i=1 |xi|. 2 Euclidean or 2-norm : x 2 = d i=1 x2 i . 13 / 50
- 27. Norm of a Vector Definition A norm u measures the magnitud of the vector. Properties 1 Homogeneity: αx = α x . 2 Triangle inequality: x + y ≤ x + y . 3 Point Separation x = 0 if and only if x = 0. Examples 1 Manhattan or 1-norm : x 1 = d i=1 |xi|. 2 Euclidean or 2-norm : x 2 = d i=1 x2 i . 13 / 50
- 28. Examples Example 1-norm and 2-norm 14 / 50
- 29. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 15 / 50
- 30. Inner Product Definition The inner product between u and v u, v = n i=1 uivi. It is the projection of one vector onto the other one Remark: It is related to the Euclidean norm: u, u = u 2 2. 16 / 50
- 31. Inner Product Definition The inner product between u and v u, v = n i=1 uivi. It is the projection of one vector onto the other one 16 / 50
- 32. Properties Meaning The inner product is a measure of correlation between two vectors, scaled by the norms of the vectors if u · v > 0, u and v are aligned 17 / 50
- 33. Properties Meaning The inner product is a measure of correlation between two vectors, scaled by the norms of the vectors if u · v > 0, u and v are aligned 17 / 50
- 34. Properties The inner product is a measure of correlation between two vectors, scaled by the norms of the vectors 18 / 50
- 35. Properties The inner product is a measure of correlation between two vectors, scaled by the norms of the vectors 19 / 50
- 36. Properties The inner product is a measure of correlation between two vectors, scaled by the norms of the vectors 20 / 50
- 37. Definitions involving the norm Orthonormal The vectors in orthonormal basis have unit Euclidean norm and are orthonorgonal. To express a vector x in an orthonormal basis For example, given x = α1b1 + α2b2 x, b1 = α1b1 + α2b2, b1 = α1 b1, b1 + α2 b2, b1 = α1 + 0 Likewise, x, b2 = α2 21 / 50
- 38. Definitions involving the norm Orthonormal The vectors in orthonormal basis have unit Euclidean norm and are orthonorgonal. To express a vector x in an orthonormal basis For example, given x = α1b1 + α2b2 x, b1 = α1b1 + α2b2, b1 = α1 b1, b1 + α2 b2, b1 = α1 + 0 Likewise, x, b2 = α2 21 / 50
- 39. Definitions involving the norm Orthonormal The vectors in orthonormal basis have unit Euclidean norm and are orthonorgonal. To express a vector x in an orthonormal basis For example, given x = α1b1 + α2b2 x, b1 = α1b1 + α2b2, b1 = α1 b1, b1 + α2 b2, b1 = α1 + 0 Likewise, x, b2 = α2 21 / 50
- 40. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 22 / 50
- 41. Linear Operator Definition A linear operator L : U → V is a map from a vector space U to another vector space V satisfies: L (u1 + u2) = L (u1) + L (u2) Something Notable If the dimension n of U and m of V are finite, L can be represented by m × n matrix: A = a11 a12 · · · a1n a21 a22 · · · a2n · · · am1 am2 · · · amn 23 / 50
- 42. Linear Operator Definition A linear operator L : U → V is a map from a vector space U to another vector space V satisfies: L (u1 + u2) = L (u1) + L (u2) Something Notable If the dimension n of U and m of V are finite, L can be represented by m × n matrix: A = a11 a12 · · · a1n a21 a22 · · · a2n · · · am1 am2 · · · amn 23 / 50
- 43. Thus, product of The product of two linear operator can be seen as the multiplication of two matrices AB = a11 a12 · · · a1n a21 a22 · · · a2n · · · am1 am2 · · · amn b11 b12 · · · b1p b21 b22 · · · b2p · · · bn1 bn2 · · · bnp = n i=1 a1ibi1 n i=1 a1ibi2 · · · n i=1 a1ibip n i=1 a2ibi1 n i=1 a2ibi2 · · · n i=1 a2ibip · · · n i=1 amibi1 n i=1 amibi2 · · · n i=1 amibip Note: if A is m × n and B is n × p, then AB is m × p. 24 / 50
- 44. Thus, product of The product of two linear operator can be seen as the multiplication of two matrices AB = a11 a12 · · · a1n a21 a22 · · · a2n · · · am1 am2 · · · amn b11 b12 · · · b1p b21 b22 · · · b2p · · · bn1 bn2 · · · bnp = n i=1 a1ibi1 n i=1 a1ibi2 · · · n i=1 a1ibip n i=1 a2ibi1 n i=1 a2ibi2 · · · n i=1 a2ibip · · · n i=1 amibi1 n i=1 amibi2 · · · n i=1 amibip Note: if A is m × n and B is n × p, then AB is m × p. 24 / 50
- 45. Transpose of a Matrix The transpose of a matrix is obtained by flipping the rows and columns AT = a11 a21 · · · an1 a12 a22 · · · an2 · · · a1m a2m · · · anm Which the following properties AT T = A (A + B)T = AT + BT (AB)T = BT AT Not only that, we have the inner product u, v = uT v 25 / 50
- 46. Transpose of a Matrix The transpose of a matrix is obtained by flipping the rows and columns AT = a11 a21 · · · an1 a12 a22 · · · an2 · · · a1m a2m · · · anm Which the following properties AT T = A (A + B)T = AT + BT (AB)T = BT AT Not only that, we have the inner product u, v = uT v 25 / 50
- 47. Transpose of a Matrix The transpose of a matrix is obtained by flipping the rows and columns AT = a11 a21 · · · an1 a12 a22 · · · an2 · · · a1m a2m · · · anm Which the following properties AT T = A (A + B)T = AT + BT (AB)T = BT AT Not only that, we have the inner product u, v = uT v 25 / 50
- 48. Transpose of a Matrix The transpose of a matrix is obtained by flipping the rows and columns AT = a11 a21 · · · an1 a12 a22 · · · an2 · · · a1m a2m · · · anm Which the following properties AT T = A (A + B)T = AT + BT (AB)T = BT AT Not only that, we have the inner product u, v = uT v 25 / 50
- 49. Transpose of a Matrix The transpose of a matrix is obtained by flipping the rows and columns AT = a11 a21 · · · an1 a12 a22 · · · an2 · · · a1m a2m · · · anm Which the following properties AT T = A (A + B)T = AT + BT (AB)T = BT AT Not only that, we have the inner product u, v = uT v 25 / 50
- 50. As always, we have the identity operator The identity operator in matrix multiplication is defined as I = 1 0 · · · 0 0 1 · · · 0 · · · 0 0 · · · 1 With properties For any matrix A, AI = A. I is the identity operator for the matrix product. 26 / 50
- 51. As always, we have the identity operator The identity operator in matrix multiplication is defined as I = 1 0 · · · 0 0 1 · · · 0 · · · 0 0 · · · 1 With properties For any matrix A, AI = A. I is the identity operator for the matrix product. 26 / 50
- 52. As always, we have the identity operator The identity operator in matrix multiplication is defined as I = 1 0 · · · 0 0 1 · · · 0 · · · 0 0 · · · 1 With properties For any matrix A, AI = A. I is the identity operator for the matrix product. 26 / 50
- 53. Column Space, Row Space and Rank Let A be an m × n matrix We have the following spaces... Column space Span of the columns of A. Linear subspace of Rm. Row space Span of the rows of A. Linear subspace of Rn. 27 / 50
- 54. Column Space, Row Space and Rank Let A be an m × n matrix We have the following spaces... Column space Span of the columns of A. Linear subspace of Rm. Row space Span of the rows of A. Linear subspace of Rn. 27 / 50
- 55. Column Space, Row Space and Rank Let A be an m × n matrix We have the following spaces... Column space Span of the columns of A. Linear subspace of Rm. Row space Span of the rows of A. Linear subspace of Rn. 27 / 50
- 56. Column Space, Row Space and Rank Let A be an m × n matrix We have the following spaces... Column space Span of the columns of A. Linear subspace of Rm. Row space Span of the rows of A. Linear subspace of Rn. 27 / 50
- 57. Important facts Something Notable The column and row space of any matrix have the same dimension. The rank The dimension is the rank of the matrix. 28 / 50
- 58. Important facts Something Notable The column and row space of any matrix have the same dimension. The rank The dimension is the rank of the matrix. 28 / 50
- 59. Range and Null Space Range Set of vectors equal to Au for some u ∈ Rn. Range (A) = {x|x = Au for some u ∈ Rn } It is a linear subspace of Rm and also called the column space of A. Null Space We have the following definition Null Space (A) = {u|Au = 0} It is a linear subspace of Rm. 29 / 50
- 60. Range and Null Space Range Set of vectors equal to Au for some u ∈ Rn. Range (A) = {x|x = Au for some u ∈ Rn } It is a linear subspace of Rm and also called the column space of A. Null Space We have the following definition Null Space (A) = {u|Au = 0} It is a linear subspace of Rm. 29 / 50
- 61. Range and Null Space Range Set of vectors equal to Au for some u ∈ Rn. Range (A) = {x|x = Au for some u ∈ Rn } It is a linear subspace of Rm and also called the column space of A. Null Space We have the following definition Null Space (A) = {u|Au = 0} It is a linear subspace of Rm. 29 / 50
- 62. Range and Null Space Range Set of vectors equal to Au for some u ∈ Rn. Range (A) = {x|x = Au for some u ∈ Rn } It is a linear subspace of Rm and also called the column space of A. Null Space We have the following definition Null Space (A) = {u|Au = 0} It is a linear subspace of Rm. 29 / 50
- 63. Important fact Something Notable Every vector in the null space is orthogonal to the rows of A. The null space and row space of a matrix are orthogonal. 30 / 50
- 64. Important fact Something Notable Every vector in the null space is orthogonal to the rows of A. The null space and row space of a matrix are orthogonal. 30 / 50
- 65. Range and Column Space We have another interpretation of the matrix-vector product Au = (A1 A2 · · · An) u1 u2 ... un =u1A1 + u2A2 + · · · + unAn Thus The result is a linear combination of the columns of A. Actually, the range is the column space. 31 / 50
- 66. Range and Column Space We have another interpretation of the matrix-vector product Au = (A1 A2 · · · An) u1 u2 ... un =u1A1 + u2A2 + · · · + unAn Thus The result is a linear combination of the columns of A. Actually, the range is the column space. 31 / 50
- 67. Range and Column Space We have another interpretation of the matrix-vector product Au = (A1 A2 · · · An) u1 u2 ... un =u1A1 + u2A2 + · · · + unAn Thus The result is a linear combination of the columns of A. Actually, the range is the column space. 31 / 50
- 68. Matrix Inverse Something Notable For an n × n matrix A: rank + dim(null space) = n. if dim(null space)= 0 then A is full rank. In this case, the action of the matrix is invertible. The inversion is also linear and consequently can be represented by another matrix A−1. A−1 is the only matrix such that A−1A = AA−1 = I. 32 / 50
- 69. Matrix Inverse Something Notable For an n × n matrix A: rank + dim(null space) = n. if dim(null space)= 0 then A is full rank. In this case, the action of the matrix is invertible. The inversion is also linear and consequently can be represented by another matrix A−1. A−1 is the only matrix such that A−1A = AA−1 = I. 32 / 50
- 70. Matrix Inverse Something Notable For an n × n matrix A: rank + dim(null space) = n. if dim(null space)= 0 then A is full rank. In this case, the action of the matrix is invertible. The inversion is also linear and consequently can be represented by another matrix A−1. A−1 is the only matrix such that A−1A = AA−1 = I. 32 / 50
- 71. Matrix Inverse Something Notable For an n × n matrix A: rank + dim(null space) = n. if dim(null space)= 0 then A is full rank. In this case, the action of the matrix is invertible. The inversion is also linear and consequently can be represented by another matrix A−1. A−1 is the only matrix such that A−1A = AA−1 = I. 32 / 50
- 72. Matrix Inverse Something Notable For an n × n matrix A: rank + dim(null space) = n. if dim(null space)= 0 then A is full rank. In this case, the action of the matrix is invertible. The inversion is also linear and consequently can be represented by another matrix A−1. A−1 is the only matrix such that A−1A = AA−1 = I. 32 / 50
- 73. Orthogonal Matrices Definition An orthogonal matrix U satisfies UT U = I. Properties U has orthonormal columns. In addition Applying an orthogonal matrix to two vectors does not change their inner product: Uu, Uv = (Uu)T Uv =uT UT Uv =uT v = u, v 33 / 50
- 74. Orthogonal Matrices Definition An orthogonal matrix U satisfies UT U = I. Properties U has orthonormal columns. In addition Applying an orthogonal matrix to two vectors does not change their inner product: Uu, Uv = (Uu)T Uv =uT UT Uv =uT v = u, v 33 / 50
- 75. Orthogonal Matrices Definition An orthogonal matrix U satisfies UT U = I. Properties U has orthonormal columns. In addition Applying an orthogonal matrix to two vectors does not change their inner product: Uu, Uv = (Uu)T Uv =uT UT Uv =uT v = u, v 33 / 50
- 76. Example A classic one Matrices representing rotations are orthogonal. 34 / 50
- 77. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 35 / 50
- 78. Trace and Determinant Definition (Trace) The trace is the sum of the diagonal elements of a square matrix. Definition (Determinant) The determinant of a square matrix A, denoted by |A|, is defined as det (A) = n j=1 (−1)i+j aijMij where Mij is determinant of matrix A without the row i and column j. 36 / 50
- 79. Trace and Determinant Definition (Trace) The trace is the sum of the diagonal elements of a square matrix. Definition (Determinant) The determinant of a square matrix A, denoted by |A|, is defined as det (A) = n j=1 (−1)i+j aijMij where Mij is determinant of matrix A without the row i and column j. 36 / 50
- 80. Special Case For a 2 × 2 matrix A = a b c d |A| = ad − bc The absolute value of |A|is the area of the parallelogram given by the rows of A 37 / 50
- 81. Special Case For a 2 × 2 matrix A = a b c d |A| = ad − bc The absolute value of |A|is the area of the parallelogram given by the rows of A 37 / 50
- 82. Properties of the Determinant Basic Properties |A| = AT |AB| = |A| |B| |A| = 0 if and only if A is not invertible If A is invertible, then A−1 = 1 |A| . 38 / 50
- 83. Properties of the Determinant Basic Properties |A| = AT |AB| = |A| |B| |A| = 0 if and only if A is not invertible If A is invertible, then A−1 = 1 |A| . 38 / 50
- 84. Properties of the Determinant Basic Properties |A| = AT |AB| = |A| |B| |A| = 0 if and only if A is not invertible If A is invertible, then A−1 = 1 |A| . 38 / 50
- 85. Properties of the Determinant Basic Properties |A| = AT |AB| = |A| |B| |A| = 0 if and only if A is not invertible If A is invertible, then A−1 = 1 |A| . 38 / 50
- 86. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 39 / 50
- 87. Eigenvalues and Eigenvectors Eigenvalues An eigenvalue λ of a square matrix A satisfies: Au = λu for some vector , which we call an eigenvector. Properties Geometrically the operator A expands when (λ > 1) or contracts (λ < 1) eigenvectors, but does not rotate them. Null Space relation If u is an eigenvector of A, it is in the null space of A − λI, which is consequently not invertible. 40 / 50
- 88. Eigenvalues and Eigenvectors Eigenvalues An eigenvalue λ of a square matrix A satisfies: Au = λu for some vector , which we call an eigenvector. Properties Geometrically the operator A expands when (λ > 1) or contracts (λ < 1) eigenvectors, but does not rotate them. Null Space relation If u is an eigenvector of A, it is in the null space of A − λI, which is consequently not invertible. 40 / 50
- 89. Eigenvalues and Eigenvectors Eigenvalues An eigenvalue λ of a square matrix A satisfies: Au = λu for some vector , which we call an eigenvector. Properties Geometrically the operator A expands when (λ > 1) or contracts (λ < 1) eigenvectors, but does not rotate them. Null Space relation If u is an eigenvector of A, it is in the null space of A − λI, which is consequently not invertible. 40 / 50
- 90. More properties Given the previous relation The eigenvalues of A are the roots of the equation |A − λI| = 0 Remark: We do not calculate the eigenvalues this way Something Notable Eigenvalues and eigenvectors can be complex valued, even if all the entries of A are real. 41 / 50
- 91. More properties Given the previous relation The eigenvalues of A are the roots of the equation |A − λI| = 0 Remark: We do not calculate the eigenvalues this way Something Notable Eigenvalues and eigenvectors can be complex valued, even if all the entries of A are real. 41 / 50
- 92. Eigendecomposition of a Matrix Given Let A be an n × n square matrix with n linearly independent eigenvectors p1, p2, ..., pn and eigenvalues λ1, λ2, ..., λn We define the matrices P = (p1 p2 · · · pn) Λ = λ1 0 · · · 0 0 λ2 · · · 0 · · · 0 0 · · · λn 42 / 50
- 93. Eigendecomposition of a Matrix Given Let A be an n × n square matrix with n linearly independent eigenvectors p1, p2, ..., pn and eigenvalues λ1, λ2, ..., λn We define the matrices P = (p1 p2 · · · pn) Λ = λ1 0 · · · 0 0 λ2 · · · 0 · · · 0 0 · · · λn 42 / 50
- 94. Properties We have that A satisfies AP = PΛ In addition P is full rank. Thus, inverting it yields the eigendecomposition A = PΛP−1 43 / 50
- 95. Properties We have that A satisfies AP = PΛ In addition P is full rank. Thus, inverting it yields the eigendecomposition A = PΛP−1 43 / 50
- 96. Properties We have that A satisfies AP = PΛ In addition P is full rank. Thus, inverting it yields the eigendecomposition A = PΛP−1 43 / 50
- 97. Properties of the Eigendecomposition We have that Not all matrices are diagonalizable/eigendecomposition. Example 1 1 0 1 Trace (A) = Trace (Λ) = n i=1 λi |A| = |Λ| = Πn i=1λi The rank of A is equal to the number of nonzero eigenvalues. If λ is anonzero eigenvalue of A, 1 λ is an eigenvalue of A−1 with the same eigenvector. The eigendecompositon allows to compute matrix powers efficiently: Am = PΛP−1 m = PΛP−1PΛP−1PΛP−1 . . . PΛP−1 = PΛmP−1 44 / 50
- 98. Properties of the Eigendecomposition We have that Not all matrices are diagonalizable/eigendecomposition. Example 1 1 0 1 Trace (A) = Trace (Λ) = n i=1 λi |A| = |Λ| = Πn i=1λi The rank of A is equal to the number of nonzero eigenvalues. If λ is anonzero eigenvalue of A, 1 λ is an eigenvalue of A−1 with the same eigenvector. The eigendecompositon allows to compute matrix powers efficiently: Am = PΛP−1 m = PΛP−1PΛP−1PΛP−1 . . . PΛP−1 = PΛmP−1 44 / 50
- 99. Properties of the Eigendecomposition We have that Not all matrices are diagonalizable/eigendecomposition. Example 1 1 0 1 Trace (A) = Trace (Λ) = n i=1 λi |A| = |Λ| = Πn i=1λi The rank of A is equal to the number of nonzero eigenvalues. If λ is anonzero eigenvalue of A, 1 λ is an eigenvalue of A−1 with the same eigenvector. The eigendecompositon allows to compute matrix powers efficiently: Am = PΛP−1 m = PΛP−1PΛP−1PΛP−1 . . . PΛP−1 = PΛmP−1 44 / 50
- 100. Properties of the Eigendecomposition We have that Not all matrices are diagonalizable/eigendecomposition. Example 1 1 0 1 Trace (A) = Trace (Λ) = n i=1 λi |A| = |Λ| = Πn i=1λi The rank of A is equal to the number of nonzero eigenvalues. If λ is anonzero eigenvalue of A, 1 λ is an eigenvalue of A−1 with the same eigenvector. The eigendecompositon allows to compute matrix powers efficiently: Am = PΛP−1 m = PΛP−1PΛP−1PΛP−1 . . . PΛP−1 = PΛmP−1 44 / 50
- 101. Properties of the Eigendecomposition We have that Not all matrices are diagonalizable/eigendecomposition. Example 1 1 0 1 Trace (A) = Trace (Λ) = n i=1 λi |A| = |Λ| = Πn i=1λi The rank of A is equal to the number of nonzero eigenvalues. If λ is anonzero eigenvalue of A, 1 λ is an eigenvalue of A−1 with the same eigenvector. The eigendecompositon allows to compute matrix powers efficiently: Am = PΛP−1 m = PΛP−1PΛP−1PΛP−1 . . . PΛP−1 = PΛmP−1 44 / 50
- 102. Properties of the Eigendecomposition We have that Not all matrices are diagonalizable/eigendecomposition. Example 1 1 0 1 Trace (A) = Trace (Λ) = n i=1 λi |A| = |Λ| = Πn i=1λi The rank of A is equal to the number of nonzero eigenvalues. If λ is anonzero eigenvalue of A, 1 λ is an eigenvalue of A−1 with the same eigenvector. The eigendecompositon allows to compute matrix powers efficiently: Am = PΛP−1 m = PΛP−1PΛP−1PΛP−1 . . . PΛP−1 = PΛmP−1 44 / 50
- 103. Eigendecomposition of a Symmetric Matrix When A symmetric, we have If A = AT then A is symmetric. The eigenvalues of symmetric matrices are real. The eigenvectors of symmetric matrices are orthonormal. Consequently, the eigendecomposition becomes A = UΛUT for Λ real and U orthogonal. The eigenvectors of A are an orthonormal basis for the column space and row space. 45 / 50
- 104. Eigendecomposition of a Symmetric Matrix When A symmetric, we have If A = AT then A is symmetric. The eigenvalues of symmetric matrices are real. The eigenvectors of symmetric matrices are orthonormal. Consequently, the eigendecomposition becomes A = UΛUT for Λ real and U orthogonal. The eigenvectors of A are an orthonormal basis for the column space and row space. 45 / 50
- 105. Eigendecomposition of a Symmetric Matrix When A symmetric, we have If A = AT then A is symmetric. The eigenvalues of symmetric matrices are real. The eigenvectors of symmetric matrices are orthonormal. Consequently, the eigendecomposition becomes A = UΛUT for Λ real and U orthogonal. The eigenvectors of A are an orthonormal basis for the column space and row space. 45 / 50
- 106. Eigendecomposition of a Symmetric Matrix When A symmetric, we have If A = AT then A is symmetric. The eigenvalues of symmetric matrices are real. The eigenvectors of symmetric matrices are orthonormal. Consequently, the eigendecomposition becomes A = UΛUT for Λ real and U orthogonal. The eigenvectors of A are an orthonormal basis for the column space and row space. 45 / 50
- 107. Eigendecomposition of a Symmetric Matrix When A symmetric, we have If A = AT then A is symmetric. The eigenvalues of symmetric matrices are real. The eigenvectors of symmetric matrices are orthonormal. Consequently, the eigendecomposition becomes A = UΛUT for Λ real and U orthogonal. The eigenvectors of A are an orthonormal basis for the column space and row space. 45 / 50
- 108. We can see the action of a symmetric matrix on a vector u as... We can decompose the action Au = UΛUT u as Projection of u onto the column space of A (Multiplication by UT ). Scaling of each coefficient Ui, u by the corresponding eigenvalue (Multiplication by Λ). Linear combination of the eigenvectors scaled by the resulting coefficient (Multiplication by U). Final equation Au = n i=1 λi Ui, u Ui It would be great to generalize this to all matrices!!! 46 / 50
- 109. We can see the action of a symmetric matrix on a vector u as... We can decompose the action Au = UΛUT u as Projection of u onto the column space of A (Multiplication by UT ). Scaling of each coefficient Ui, u by the corresponding eigenvalue (Multiplication by Λ). Linear combination of the eigenvectors scaled by the resulting coefficient (Multiplication by U). Final equation Au = n i=1 λi Ui, u Ui It would be great to generalize this to all matrices!!! 46 / 50
- 110. We can see the action of a symmetric matrix on a vector u as... We can decompose the action Au = UΛUT u as Projection of u onto the column space of A (Multiplication by UT ). Scaling of each coefficient Ui, u by the corresponding eigenvalue (Multiplication by Λ). Linear combination of the eigenvectors scaled by the resulting coefficient (Multiplication by U). Final equation Au = n i=1 λi Ui, u Ui It would be great to generalize this to all matrices!!! 46 / 50
- 111. We can see the action of a symmetric matrix on a vector u as... We can decompose the action Au = UΛUT u as Projection of u onto the column space of A (Multiplication by UT ). Scaling of each coefficient Ui, u by the corresponding eigenvalue (Multiplication by Λ). Linear combination of the eigenvectors scaled by the resulting coefficient (Multiplication by U). Final equation Au = n i=1 λi Ui, u Ui It would be great to generalize this to all matrices!!! 46 / 50
- 112. We can see the action of a symmetric matrix on a vector u as... We can decompose the action Au = UΛUT u as Projection of u onto the column space of A (Multiplication by UT ). Scaling of each coefficient Ui, u by the corresponding eigenvalue (Multiplication by Λ). Linear combination of the eigenvectors scaled by the resulting coefficient (Multiplication by U). Final equation Au = n i=1 λi Ui, u Ui It would be great to generalize this to all matrices!!! 46 / 50
- 113. Outline 1 Introduction What is a Vector? 2 Vector Spaces Definition Linear Independence and Basis of Vector Spaces Norm of a Vector Inner Product Matrices Trace and Determinant Matrix Decomposition Singular Value Decomposition 47 / 50
- 114. Singular Value Decomposition Every Matrix has a singular value decomposition A = UΣV T Where The columns of U are an orthonormal basis for the column space. The columns of V are an orthonormal basis for the row space. The Σ is diagonal and the entries on its diagonal σi = Σii are positive real numbers, called the singular values of A. The action of Aon a vector u can be decomposed into Au = n i=1 σi Vi, u Ui 48 / 50
- 115. Singular Value Decomposition Every Matrix has a singular value decomposition A = UΣV T Where The columns of U are an orthonormal basis for the column space. The columns of V are an orthonormal basis for the row space. The Σ is diagonal and the entries on its diagonal σi = Σii are positive real numbers, called the singular values of A. The action of Aon a vector u can be decomposed into Au = n i=1 σi Vi, u Ui 48 / 50
- 116. Singular Value Decomposition Every Matrix has a singular value decomposition A = UΣV T Where The columns of U are an orthonormal basis for the column space. The columns of V are an orthonormal basis for the row space. The Σ is diagonal and the entries on its diagonal σi = Σii are positive real numbers, called the singular values of A. The action of Aon a vector u can be decomposed into Au = n i=1 σi Vi, u Ui 48 / 50
- 117. Singular Value Decomposition Every Matrix has a singular value decomposition A = UΣV T Where The columns of U are an orthonormal basis for the column space. The columns of V are an orthonormal basis for the row space. The Σ is diagonal and the entries on its diagonal σi = Σii are positive real numbers, called the singular values of A. The action of Aon a vector u can be decomposed into Au = n i=1 σi Vi, u Ui 48 / 50
- 118. Properties of the Singular Value Decomposition First The eigenvalues of the symmetric matrix AT A are equal to the square of the singular values of A: AT A = V ΣUT UT ΣV T = V Σ2V T Second The rank of a matrix is equal to the number of nonzero singular values. Third The largest singular value σ1 is the solution to the optimization problem: σ1 = max x=0 Ax 2 x 2 49 / 50
- 119. Properties of the Singular Value Decomposition First The eigenvalues of the symmetric matrix AT A are equal to the square of the singular values of A: AT A = V ΣUT UT ΣV T = V Σ2V T Second The rank of a matrix is equal to the number of nonzero singular values. Third The largest singular value σ1 is the solution to the optimization problem: σ1 = max x=0 Ax 2 x 2 49 / 50
- 120. Properties of the Singular Value Decomposition First The eigenvalues of the symmetric matrix AT A are equal to the square of the singular values of A: AT A = V ΣUT UT ΣV T = V Σ2V T Second The rank of a matrix is equal to the number of nonzero singular values. Third The largest singular value σ1 is the solution to the optimization problem: σ1 = max x=0 Ax 2 x 2 49 / 50
- 121. Properties of the Singular Value Decomposition Remark It can be verified that the largest singular value satisfies the properties of a norm, it is called the spectral norm of the matrix. Finally In statistics analyzing data with the singular value decomposition is called Principal Component Analysis. 50 / 50
- 122. Properties of the Singular Value Decomposition Remark It can be verified that the largest singular value satisfies the properties of a norm, it is called the spectral norm of the matrix. Finally In statistics analyzing data with the singular value decomposition is called Principal Component Analysis. 50 / 50