DOCSLIB.ORG

Matrix Norms and Singular Value Decomposition

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Matrix Norms and Singular Value Decomposition Lectures on Dynamic Systems and Control Mohammed Dahleh Munther A. Dahleh George Verghese Department of Electrical Engineering and Computer Science 1 Massachuasetts Institute of Technology 1 c � Chapter 4 Matrix Norms and Singular Value Decomp osition 4.1 Intro duction In this lecture, we intro duce the notion of a norm for matrices. The singular value decom- position or SVD of a matrix is then presented. The SVD exp oses the 2-norm of a matrix, value to us go es much further: it enables the solution of a class of matrix perturbation but its problems that form the basis for the stability robustness concepts intro duced later� it solves - least squares problem, which is a generalization of the least squares estima the so-called total tion problem considered earlier� and it allows us to clarify the notion of conditioning, in the context of matrix inversion. These applications of the SVD are presented at greater length in the next lecture. Example 4.1 To provide some immediate motivation for the study and applica- tion of matrix norms, we b egin with an example that clearly brings out the issue resp ect to inversion. The question of interest is how of matrix conditioning with sensitive the inverse of a matrix is to p erturbations of the matrix. Consider inverting the matrix � � 100 100 A � (4.1) 100:2 100 A quick calculation shows that � � ;5 5 ;1 A � (4.2) 5:01 ;5 Now supp ose we invert the p erturb ed matrix � � 100 100 A + �A � (4.3) 100:1 100 The result now is � � ;10 10 ;1 ;1 ;1 (A + �A) � A + �(A ) � (4.4) 10:01 ;10 ;1 Here �A denotes the p erturbation in A and �(A ) denotes the resulting p er- ;1 turbation in A . Evidently a 0.1% change in one entry of A has resulted in a ;1 . If we want to solve the problem Ax � b where change in the entries of A 100% T ;1 T , then x � A b � [;10 10:01] , while after p erturbation of A we 1] b � [1 ; ;1 T b � [;20 20:01] . Again, we see a 100% change in the + �x � (A + �A) get x entries of the solution with only a 0.1% change in the starting data. The situation seen in the ab ove example is much worse than what can ever arise in the ;1 ;1 )�(a ) � ;da�a, so the fractional change in the d(a scalar case. If a is a scalar, then inverse of a has the same maginitude as the fractional change in a itself. What is seen in the ab ove example, therefore, is a purely matrix phenomenon. It would seem to b e related to dep endent, its the fact that A is nearly singular | in the sense that its columns are nearly determinant is much smaller than its largest element, and so on. In what follows (see next we shall develop a sound way to measure nearness to singularity, and show how this lecture), sensitivity under inversion. measure relates to Before understanding such sensitivity to p erturbations in more detail, we need ways to vectors and matrices. We have already intro duced the notion measure the \magnitudes" of vector norms in Lecture 1, so we now turn to the de�nition of matrix norms. of 4.2 Matrix Norms An m � n complex matrix may b e viewed as an op erator on the (�nite dimensional) normed n vector space C : m�n n m A : ( C � k � k ) ;! ( C � k � k ) (4.5) 2 2 where the norm here is taken to b e the standard Euclidean norm. De�ne the induced 2-norm follows: of A as kAxk 4 2 kAk � sup (4.6) 2 kxk x6�0 2 : (4.7) � max kAxk 2 kxk �1 2 The term \induced" refers to the fact that the de�nition of a norm for vectors such as Ax and ab ove de�nition of a matrix norm. From this de�nition, it follows that x is what enables the amount of \ampli�cation" the matrix A provides to vectors the induced norm measures the n on the unit sphere in C , i.e. it measures the \gain" of the matrix. Rather than measuring the vectors x and Ax using the 2-norm, we could use any p-norm, interesting cases b eing p � 1� 2� 1. Our notation for this is the kAk � max kAxk : (4.8) p p kxk �1 p An imp ortant question to consider is whether or not the induced norm is actually a norm, vectors in Lecture 1. Recall the three conditions that de�ne a norm: in the sense de�ned for 1. kxk � 0, and kxk � 0 () x � 0� 2. k�xk � j�j kxk � 3. kx + y k � kxk + ky k . m�n is a norm on C | using the preceding de�nition: Now let us verify that kAk p 1. kAk � 0 since kAxk � 0 for any x. Futhermore, kAk � 0 () A � 0, since kAk is p p p p calculated from the maximum of kAxk evaluated on the unit sphere. p 2. k�Ak � j�j kAk follows from k�y k � j�j ky k (for any y ). p p p p 3. The triangle inequality holds since: kA + B k � max k(A + B )xk p p kxk �1 p � � � max kAxk + kB xk p p kxk �1 p � kAk + kB k : p p Induced norms have two additional prop erties that are very imp ortant: 1. kAxk � kAk kxk , which is a direct consequence of the de�nition of an induced norm� p p p m�n n�r 2. For A , B , kAB k � kAk kB k (4.9) p p p which is called the submultiplicative property. This also follows directly from the de�ni- tion: kAB xk � kAk kB xk p p p � kAk kB k kxk for any x: p p p Dividing by kxk : p kAB xk p � kAk kB k � p p kxk p from which the result follows. Before we turn to a more detailed study of ideas surrounding the induced 2-norm, which b e the fo cus of this lecture and the next, we make some remarks ab out the other induced will interest, namely the induced 1-norm and induced 1-norm. We shall also norms of practical say something ab out an imp ortant matrix norm that is not an induced norm, namely the Frobenius norm. It is a fairly simple exercise to prove that m X kAk � max ja j (max of absolute column sums of A) � (4.10) ij 1 1�j �n i�1 and n X j (max of absolute kAk � max ja row sums of A) : (4.11) ij 1 1�i�m j �1 (Note that these de�nitions reduce to the familiar ones for the 1-norm and 1-norm of column vectors in the case n � 1.) The pro of for the induced 1-norm involves two stages, namely: 1. Prove that the quantity in Equation (4.11) provides an upp er b ound � : kAxk � � kxk 8x � 1 1 2. Show that this b ound is achievable for some x � x^ : kAx^ k � � kx^ k for some x^ : 1 1 In order to show how these steps can b e implemented, we give the details for the 1-norm n and consider case. Let x 2 C n X kAxk � max j a x j ij j 1 1�i�m j �1 n X � max ja jjx j ij j 1�i�m j �1 0 1 n X @ A � max ja j max jx j ij j 1�i�m 1�j �n j �1 0 1 n X @ A � max ja j kxk ij 1 1�i�m j �1 The ab ove inequalities show that an upp er b ound � is given by n X max kAxk � � � max ja j: ij 1 1�i�m kxk �1 1 j �1 � Now in order to show that this upp er b ound is achieved by some vector x^ , let i b e an index P n at which the expression of � achieves a maximum, that is � � ja j. De�ne the vector x^ � j �1 ij as 2 3 sg n(a ) � i 1 6 7 ) sg n(a 6 7 � i 2 6 7 x^ � : . 6 7 . 4 5 ) sg n(a � in Clearly kx^ k � 1 and 1 n X kAx^ k � ja j � � : � 1 ij j �1 The pro of for the 1-norm pro ceeds in exactly the same way, and is left to the reader. There are matrix norms | i.e. functions that satisfy the three de�ning conditions stated imp ortant example of this for us is the earlier | that are not induced norms. The most Frob enius norm: 1 0 1 2 n m X X 4 2 @ A kAk � ja j (4.12) ij F j �1 i�1 1 ; � 0 2 � trace (A A) (verify) (4.13) In other words, the Frob enius norm is de�ned as the ro ot sum of squares of the entries, i.e. mn the usual Euclidean 2-norm of the matrix when it is regarded simply as a vector in C . b e shown that it is not an induced matrix norm, the Frob enius norm still has Although it can submultiplicative prop erty that was noted for induced norms. Yet other matrix norms the may b e de�ned (some of them without the submultiplicative prop erty), but the ones ab ove interest to us.
Load more

Recommended publications
  • Singular Value Decomposition (SVD)
    San José State University Math 253: Mathematical Methods for Data Visualization Lecture 5: Singular Value Decomposition (SVD) Dr. Guangliang Chen Outline • Matrix SVD Singular Value Decomposition (SVD) Introduction We have seen that symmetric matrices are always (orthogonally) diagonalizable. That is, for any symmetric matrix A ∈ Rn×n, there exist an orthogonal matrix Q = [q1 ... qn] and a diagonal matrix Λ = diag(λ1, . , λn), both real and square, such that A = QΛQT . We have pointed out that λi’s are the eigenvalues of A and qi’s the corresponding eigenvectors (which are orthogonal to each other and have unit norm). Thus, such a factorization is called the eigendecomposition of A, also called the spectral decomposition of A. What about general rectangular matrices? Dr. Guangliang Chen | Mathematics & Statistics, San José State University3/22 Singular Value Decomposition (SVD) Existence of the SVD for general matrices Theorem: For any matrix X ∈ Rn×d, there exist two orthogonal matrices U ∈ Rn×n, V ∈ Rd×d and a nonnegative, “diagonal” matrix Σ ∈ Rn×d (of the same size as X) such that T Xn×d = Un×nΣn×dVd×d. Remark. This is called the Singular Value Decomposition (SVD) of X: • The diagonals of Σ are called the singular values of X (often sorted in decreasing order). • The columns of U are called the left singular vectors of X. • The columns of V are called the right singular vectors of X. Dr. Guangliang Chen | Mathematics & Statistics, San José State University4/22 Singular Value Decomposition (SVD) * * b * b (n>d) b b b * b = * * = b b b * (n<d) * b * * b b Dr.
    [Show full text]
  • C H a P T E R § Methods Related to the Norm Al Equations
    ¡ ¢ £ ¤ ¥ ¦ § ¨ © © © © ¨ ©! "# $% &('*),+-)(./+-)0.21434576/),+98,:%;<)>=,'41/? @%3*)BAC:D8E+%=B891GF*),+H;I?H1KJ2.21K8L1GMNAPO45Q5R)S;S+T? =RU ?H1*)>./+EAVOKAPM ;<),5 ?H1G;P8W.XAVO4575R)S;S+Y? =Z8L1*)/[%\Z1*)]AB3*=,'Q;<)B=,'41/? @%3*)RAP8LU F*)BA(;S'*)Z)>@%3/? F*.4U ),1G;^U ?H1*)>./+ APOKAP;_),5a`Rbc`^dfeg`Rbih,jk=>.4U U )Blm;B'*)n1K84+o5R.EUp)B@%3*.K;q? 8L1KA>[r\0:D;_),1Ejp;B'/? A2.4s4s/+t8/.,=,' b ? A7.KFK84? l4)>lu?H1vs,+D.*=S;q? =>)m6/)>=>.43KA<)W;B'*)w=B8E)IxW=K? ),1G;W5R.K;S+Y? yn` `z? AW5Q3*=,'|{}84+-A_) =B8L1*l9? ;I? 891*)>lX;B'*.41Q`7[r~k8K{i)IF*),+NjL;B'*)71K8E+o5R.4U4)>@%3*.G;I? 891KA0.4s4s/+t8.*=,'w5R.BOw6/)Z.,l4)IM @%3*.K;_)7?H17A_8L5R)QAI? ;S3*.K;q? 8L1KA>[p1*l4)>)>lj9;B'*),+-)Q./+-)Q)SF*),1.Es4s4U ? =>.K;q? 8L1KA(?H1Q{R'/? =,'W? ;R? A s/+-)S:Y),+o+D)Blm;_8|;B'*)W3KAB3*.4Urc+HO4U 8,F|AS346KABs*.,=B)Q;_)>=,'41/? @%3*)BAB[&('/? A]=*'*.EsG;_),+}=B8*F*),+tA7? ;PM ),+-.K;q? F*)75R)S;S'K8ElAi{^'/? =,'Z./+-)R)K? ;B'*),+pl9?+D)B=S;BU OR84+k?H5Qs4U ? =K? ;BU O2+-),U .G;<)Bl7;_82;S'*)Q1K84+o5R.EU )>@%3*.G;I? 891KA>[ mWX(fQ - uK In order to solve the linear system 7 when is nonsymmetric, we can solve the equivalent system b b [¥K¦ ¡¢ £N¤ which is Symmetric Positive Definite. This system is known as the system of the normal equations associated with the least-squares problem, [ ­¦ £N¤ minimize §* R¨©7ª§*«9¬ Note that (8.1) is typically used to solve the least-squares problem (8.2) for over- ®°¯± ±²³® determined systems, i.e., when is a rectangular matrix of size , .
    [Show full text]
  • Eigen Values and Vectors Matrices and Eigen Vectors
    EIGEN VALUES AND VECTORS MATRICES AND EIGEN VECTORS 2 3 1 11 × = [2 1] [3] [ 5 ] 2 3 3 12 3 × = = 4 × [2 1] [2] [ 8 ] [2] • Scale 3 6 2 × = [2] [4] 2 3 6 24 6 × = = 4 × [2 1] [4] [16] [4] 2 EIGEN VECTOR - PROPERTIES • Eigen vectors can only be found for square matrices • Not every square matrix has eigen vectors. • Given an n x n matrix that does have eigenvectors, there are n of them for example, given a 3 x 3 matrix, there are 3 eigenvectors. • Even if we scale the vector by some amount, we still get the same multiple 3 EIGEN VECTOR - PROPERTIES • Even if we scale the vector by some amount, we still get the same multiple • Because all you’re doing is making it longer, not changing its direction. • All the eigenvectors of a matrix are perpendicular or orthogonal. • This means you can express the data in terms of these perpendicular eigenvectors. • Also, when we find eigenvectors we usually normalize them to length one. 4 EIGEN VALUES - PROPERTIES • Eigenvalues are closely related to eigenvectors. • These scale the eigenvectors • eigenvalues and eigenvectors always come in pairs. 2 3 6 24 6 × = = 4 × [2 1] [4] [16] [4] 5 SPECTRAL THEOREM Theorem: If A ∈ ℝm×n is symmetric matrix (meaning AT = A), then, there exist real numbers (the eigenvalues) λ1, …, λn and orthogonal, non-zero real vectors ϕ1, ϕ2, …, ϕn (the eigenvectors) such that for each i = 1,2,…, n : Aϕi = λiϕi 6 EXAMPLE 30 28 A = [28 30] From spectral theorem: Aϕ = λϕ 7 EXAMPLE 30 28 A = [28 30] From spectral theorem: Aϕ = λϕ ⟹ Aϕ − λIϕ = 0 (A − λI)ϕ = 0 30 − λ 28 = 0 ⟹ λ = 58 and
    [Show full text]
  • Singular Value Decomposition (SVD) 2 1.1 Singular Vectors
    Contents 1 Singular Value Decomposition (SVD) 2 1.1 Singular Vectors . .3 1.2 Singular Value Decomposition (SVD) . .7 1.3 Best Rank k Approximations . .8 1.4 Power Method for Computing the Singular Value Decomposition . 11 1.5 Applications of Singular Value Decomposition . 16 1.5.1 Principal Component Analysis . 16 1.5.2 Clustering a Mixture of Spherical Gaussians . 16 1.5.3 An Application of SVD to a Discrete Optimization Problem . 22 1.5.4 SVD as a Compression Algorithm . 24 1.5.5 Spectral Decomposition . 24 1.5.6 Singular Vectors and ranking documents . 25 1.6 Bibliographic Notes . 27 1.7 Exercises . 28 1 1 Singular Value Decomposition (SVD) The singular value decomposition of a matrix A is the factorization of A into the product of three matrices A = UDV T where the columns of U and V are orthonormal and the matrix D is diagonal with positive real entries. The SVD is useful in many tasks. Here we mention some examples. First, in many applications, the data matrix A is close to a matrix of low rank and it is useful to find a low rank matrix which is a good approximation to the data matrix . We will show that from the singular value decomposition of A, we can get the matrix B of rank k which best approximates A; in fact we can do this for every k. Also, singular value decomposition is defined for all matrices (rectangular or square) unlike the more commonly used spectral decomposition in Linear Algebra. The reader familiar with eigenvectors and eigenvalues (we do not assume familiarity here) will also realize that we need conditions on the matrix to ensure orthogonality of eigenvectors.
    [Show full text]
  • A Singularly Valuable Decomposition: the SVD of a Matrix Dan Kalman
    A Singularly Valuable Decomposition: The SVD of a Matrix Dan Kalman Dan Kalman is an assistant professor at American University in Washington, DC. From 1985 to 1993 he worked as an applied mathematician in the aerospace industry. It was during that period that he first learned about the SVD and its applications. He is very happy to be teaching again and is avidly following all the discussions and presentations about the teaching of linear algebra. Every teacher of linear algebra should be familiar with the matrix singular value deco~??positiolz(or SVD). It has interesting and attractive algebraic properties, and conveys important geometrical and theoretical insights about linear transformations. The close connection between the SVD and the well-known theo1-j~of diagonalization for sylnmetric matrices makes the topic immediately accessible to linear algebra teachers and, indeed, a natural extension of what these teachers already know. At the same time, the SVD has fundamental importance in several different applications of linear algebra. Gilbert Strang was aware of these facts when he introduced the SVD in his now classical text [22, p. 1421, obselving that "it is not nearly as famous as it should be." Golub and Van Loan ascribe a central significance to the SVD in their defini- tive explication of numerical matrix methods [8, p, xivl, stating that "perhaps the most recurring theme in the book is the practical and theoretical value" of the SVD. Additional evidence of the SVD's significance is its central role in a number of re- cent papers in :Matlgenzatics ivlagazine and the Atnericalz Mathematical ilironthly; for example, [2, 3, 17, 231.
    [Show full text]
  • The Column and Row Hilbert Operator Spaces
    The Column and Row Hilbert Operator Spaces Roy M. Araiza Department of Mathematics Purdue University Abstract Given a Hilbert space H we present the construction and some properties of the column and row Hilbert operator spaces Hc and Hr, respectively. The main proof of the survey is showing that CB(Hc; Kc) is completely isometric to B(H ; K ); in other words, the completely bounded maps be- tween the corresponding column Hilbert operator spaces is completely isometric to the bounded operators between the Hilbert spaces. A similar theorem for the row Hilbert operator spaces follows via duality of the operator spaces. In particular there exists a natural duality between the column and row Hilbert operator spaces which we also show. The presentation and proofs are due to Effros and Ruan [1]. 1 The Column Hilbert Operator Space Given a Hilbert space H , we define the column isometry C : H −! B(C; H ) given by ξ 7! C(ξ)α := αξ; α 2 C: Then given any n 2 N; we define the nth-matrix norm on Mn(H ); by (n) kξkc;n = C (ξ) ; ξ 2 Mn(H ): These are indeed operator space matrix norms by Ruan's Representation theorem and thus we define the column Hilbert operator space as n o Hc = H ; k·kc;n : n2N It follows that given ξ 2 Mn(H ) we have that the adjoint operator of C(ξ) is given by ∗ C(ξ) : H −! C; ζ 7! (ζj ξ) : Suppose C(ξ)∗(ζ) = d; c 2 C: Then we see that cd = (cj C(ξ)∗(ζ)) = (cξj ζ) = c(ζj ξ) = (cj (ζj ξ)) : We are also able to compute the matrix norms on rectangular matrices.
    [Show full text]
  • Limited Memory Block Krylov Subspace Optimization for Computing Dominant Singular Value Decompositions
    Limited Memory Block Krylov Subspace Optimization for Computing Dominant Singular Value Decompositions Xin Liu∗ Zaiwen Weny Yin Zhangz March 22, 2012 Abstract In many data-intensive applications, the use of principal component analysis (PCA) and other related techniques is ubiquitous for dimension reduction, data mining or other transformational purposes. Such transformations often require efficiently, reliably and accurately computing dominant singular value de- compositions (SVDs) of large unstructured matrices. In this paper, we propose and study a subspace optimization technique to significantly accelerate the classic simultaneous iteration method. We analyze the convergence of the proposed algorithm, and numerically compare it with several state-of-the-art SVD solvers under the MATLAB environment. Extensive computational results show that on a wide range of large unstructured matrices, the proposed algorithm can often provide improved efficiency or robustness over existing algorithms. Keywords. subspace optimization, dominant singular value decomposition, Krylov subspace, eigen- value decomposition 1 Introduction Singular value decomposition (SVD) is a fundamental and enormously useful tool in matrix computations, such as determining the pseudo-inverse, the range or null space, or the rank of a matrix, solving regular or total least squares data fitting problems, or computing low-rank approximations to a matrix, just to mention a few. The need for computing SVDs also frequently arises from diverse fields in statistics, signal processing, data mining or compression, and from various dimension-reduction models of large-scale dynamic systems. Usually, instead of acquiring all the singular values and vectors of a matrix, it suffices to compute a set of dominant (i.e., the largest) singular values and their corresponding singular vectors in order to obtain the most valuable and relevant information about the underlying dataset or system.
    [Show full text]
  • Matrices and Linear Algebra
    Chapter 26 Matrices and linear algebra ap,mat Contents 26.1 Matrix algebra........................................... 26.2 26.1.1 Determinant (s,mat,det)................................... 26.2 26.1.2 Eigenvalues and eigenvectors (s,mat,eig).......................... 26.2 26.1.3 Hermitian and symmetric matrices............................. 26.3 26.1.4 Matrix trace (s,mat,trace).................................. 26.3 26.1.5 Inversion formulas (s,mat,mil)............................... 26.3 26.1.6 Kronecker products (s,mat,kron).............................. 26.4 26.2 Positive-definite matrices (s,mat,pd) ................................ 26.4 26.2.1 Properties of positive-definite matrices........................... 26.5 26.2.2 Partial order......................................... 26.5 26.2.3 Diagonal dominance.................................... 26.5 26.2.4 Diagonal majorizers..................................... 26.5 26.2.5 Simultaneous diagonalization................................ 26.6 26.3 Vector norms (s,mat,vnorm) .................................... 26.6 26.3.1 Examples of vector norms.................................. 26.6 26.3.2 Inequalities......................................... 26.7 26.3.3 Properties.......................................... 26.7 26.4 Inner products (s,mat,inprod) ................................... 26.7 26.4.1 Examples.......................................... 26.7 26.4.2 Properties.......................................... 26.8 26.5 Matrix norms (s,mat,mnorm) .................................... 26.8 26.5.1
    [Show full text]
  • Lecture 5 Ch. 5, Norms for Vectors and Matrices
    KTH ROYAL INSTITUTE OF TECHNOLOGY Norms for vectors and matrices — Why? Lecture 5 Problem: Measure size of vector or matrix. What is “small” and what is “large”? Ch. 5, Norms for vectors and matrices Problem: Measure distance between vectors or matrices. When are they “close together” or “far apart”? Emil Björnson/Magnus Jansson/Mats Bengtsson April 27, 2016 Answers are given by norms. Also: Tool to analyze convergence and stability of algorithms. 2 / 21 Vector norm — axiomatic definition Inner product — axiomatic definition Definition: LetV be a vector space over afieldF(R orC). Definition: LetV be a vector space over afieldF(R orC). A function :V R is a vector norm if for all A function , :V V F is an inner product if for all || · || → �· ·� × → x,y V x,y,z V, ∈ ∈ (1) x 0 nonnegative (1) x,x 0 nonnegative || ||≥ � �≥ (1a) x =0iffx= 0 positive (1a) x,x =0iffx=0 positive || || � � (2) cx = c x for allc F homogeneous (2) x+y,z = x,z + y,z additive || || | | || || ∈ � � � � � � (3) x+y x + y triangle inequality (3) cx,y =c x,y for allc F homogeneous || ||≤ || || || || � � � � ∈ (4) x,y = y,x Hermitian property A function not satisfying (1a) is called a vector � � � � seminorm. Interpretation: “Angle” (distance) between vectors. Interpretation: Size/length of vector. 3 / 21 4 / 21 Connections between norm and inner products Examples / Corollary: If , is an inner product, then x =( x,x ) 1 2 � The Euclidean norm (l ) onC n: �· ·� || || � � 2 is a vector norm. 2 2 2 1/2 x 2 = ( x1 + x 2 + + x n ) . Called: Vector norm derived from an inner product.
    [Show full text]
  • Matrix Norms 30.4   
    Matrix Norms 30.4 Introduction A matrix norm is a number defined in terms of the entries of the matrix. The norm is a useful quantity which can give important information about a matrix. ' • be familiar with matrices and their use in $ writing systems of equations • revise material on matrix inverses, be able to find the inverse of a 2 × 2 matrix, and know Prerequisites when no inverse exists Before starting this Section you should ... • revise Gaussian elimination and partial pivoting • be aware of the discussion of ill-conditioned and well-conditioned problems earlier in Section 30.1 #& • calculate norms and condition numbers of % Learning Outcomes small matrices • adjust certain systems of equations with a On completion you should be able to ... view to better conditioning " ! 34 HELM (2008): Workbook 30: Introduction to Numerical Methods ® 1. Matrix norms The norm of a square matrix A is a non-negative real number denoted kAk. There are several different ways of defining a matrix norm, but they all share the following properties: 1. kAk ≥ 0 for any square matrix A. 2. kAk = 0 if and only if the matrix A = 0. 3. kkAk = |k| kAk, for any scalar k. 4. kA + Bk ≤ kAk + kBk. 5. kABk ≤ kAk kBk. The norm of a matrix is a measure of how large its elements are. It is a way of determining the “size” of a matrix that is not necessarily related to how many rows or columns the matrix has. Key Point 6 Matrix Norm The norm of a matrix is a real number which is a measure of the magnitude of the matrix.
    [Show full text]
  • 5 Norms on Linear Spaces
    5 Norms on Linear Spaces 5.1 Introduction In this chapter we study norms on real or complex linear linear spaces. Definition 5.1. Let F = (F, {0, 1, +, −, ·}) be the real or the complex field. A semi-norm on an F -linear space V is a mapping ν : V −→ R that satisfies the following conditions: (i) ν(x + y) ≤ ν(x)+ ν(y) (the subadditivity property of semi-norms), and (ii) ν(ax)= |a|ν(x), for x, y ∈ V and a ∈ F . By taking a = 0 in the second condition of the definition we have ν(0)=0 for every semi-norm on a real or complex space. A semi-norm can be defined on every linear space. Indeed, if B is a basis of V , B = {vi | i ∈ I}, J is a finite subset of I, and x = i∈I xivi, define νJ (x) as P 0 if x = 0, νJ (x)= ( j∈J |aj | for x ∈ V . We leave to the readerP the verification of the fact that νJ is indeed a seminorm. Theorem 5.2. If V is a real or complex linear space and ν : V −→ R is a semi-norm on V , then ν(x − y) ≥ |ν(x) − ν(y)|, for x, y ∈ V . Proof. We have ν(x) ≤ ν(x − y)+ ν(y), so ν(x) − ν(y) ≤ ν(x − y). (5.1) 160 5 Norms on Linear Spaces Since ν(x − y)= |− 1|ν(y − x) ≥ ν(y) − ν(x) we have −(ν(x) − ν(y)) ≤ ν(x) − ν(y). (5.2) The Inequalities 5.1 and 5.2 give the desired inequality.
    [Show full text]
  • A Degeneracy Framework for Scalable Graph Autoencoders
    Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19) A Degeneracy Framework for Scalable Graph Autoencoders Guillaume Salha1;2 , Romain Hennequin1 , Viet Anh Tran1 and Michalis Vazirgiannis2 1Deezer Research & Development, Paris, France 2Ecole´ Polytechnique, Palaiseau, France [email protected] Abstract In this paper, we focus on the graph extensions of autoen- coders and variational autoencoders. Introduced in the 1980’s In this paper, we present a general framework to [Rumelhart et al., 1986], autoencoders (AE) regained a sig- scale graph autoencoders (AE) and graph varia- nificant popularity in the last decade through neural network tional autoencoders (VAE). This framework lever- frameworks [Baldi, 2012] as efficient tools to learn reduced ages graph degeneracy concepts to train models encoding representations of input data in an unsupervised only from a dense subset of nodes instead of us- way. Furthermore, variational autoencoders (VAE) [Kingma ing the entire graph. Together with a simple yet ef- and Welling, 2013], described as extensions of AE but actu- fective propagation mechanism, our approach sig- ally based on quite different mathematical foundations, also nificantly improves scalability and training speed recently emerged as a successful approach for unsupervised while preserving performance. We evaluate and learning from complex distributions, assuming the input data discuss our method on several variants of existing is the observed part of a larger joint model involving low- graph AE and VAE, providing the first application dimensional latent variables, optimized via variational infer- of these models to large graphs with up to millions ence approximations. [Tschannen et al., 2018] review the of nodes and edges.
    [Show full text]