DOCSLIB.ORG

Norms on Complex Matrices Induced by Complete

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Norms on Complex Matrices Induced by Complete NORMS ON COMPLEX MATRICES INDUCED BY COMPLETE HOMOGENEOUS SYMMETRIC POLYNOMIALS KONRAD AGUILAR, ÁNGEL CHÁVEZ, STEPHAN RAMON GARCIA, AND JURIJ VOLCIˇ Cˇ Abstract. We introduce a remarkable new family of norms on the space of n × n complex matrices. These norms arise from the combinatorial properties of sym- metric functions, and their construction and validation involve probability theory, partition combinatorics, and trace polynomials in noncommuting variables. Our norms enjoy many desirable analytic and algebraic properties, such as an ele- gant determinantal interpretation and the ability to distinguish certain graphs that other matrix norms cannot. Furthermore, they give rise to new dimension- independent tracial inequalities. Their potential merits further investigation. 1. Introduction In this note we introduce a family of norms on complex matrices. These are ini- tially defined in terms of certain symmetric functions of eigenvalues of complex Hermitian matrices. The fact that we deal with eigenvalues, as opposed to their absolute values, is notable. First, it prevents standard machinery, such as the the- ory of symmetric gauge functions, from applying. Second, the techniques used to establish that we indeed have norms are more complicated than one might ex- pect. For example, combinatorics, probability theory, and Lewis’ framework for group invariance in convex matrix analysis each play key roles. These norms on the Hermitian matrices are of independent interest. They can be computed recursively or directly read from the characteristic polyno- mial. Moreover, our norms distinguish certain pairs of graphs which the standard norms (operator, Frobenius, Schatten-von Neumann, Ky Fan) cannot distinguish. Our norms extend in a natural and nontrivial manner to all complex matrices. These extensions of our original norms involve partition combinatorics and trace polynomials in noncommuting variables. A Schur convexity argument permits our norms to be bounded below in terms of the mean eigenvalue of a matrix. arXiv:2106.01976v2 [math.CO] 20 Jul 2021 These norms, their unusual construction, and their potential applications sug- gest a host of open problems. We pose several at the end of the paper. 1.1. Notation. Denote by N, R, and C, respectively, the set of natural numbers, real numbers, and complex numbers. Let Hn(C) denote the set of n × n complex Hermitian matrices and Mn(C) the set of n × n complex matrices. Denote the eigenvalues of A ∈ Hn(C) by λ1(A) ≥ λ2(A) ≥···≥ λn(A) and define n λ(A) = λ1(A), λ2(A),..., λn(A) ∈ R . Key words and phrases. norm, complete homogeneous symmetric polynomial, partition, trace, pos- itivity, convexity, expectation, complexification, trace polynomial, symmetric tensor power. Third named author supported by NSF grants DMS-1800123 and DMS-2054002. Fourth named author supported by NSF grant DMS-1954709. 1 2 K. AGUILAR, Á. CHÁVEZ, S.R. GARCIA, AND J. VOLCIˇ Cˇ We may use λ and λ1, λ2,..., λn if the matrix A is clear from context. Let diag(x1, x2,..., xn) ∈ Mn(C) denote the n × n diagonal matrix with diagonal en- tries x1, x2,..., xn, in that order. If x =(x1, x2,..., xn) is understood from context, we may write diag(x) for brevity. 1.2. Complete homogeneous symmetric polynomials. The complete homogeneous symmetric (CHS) polynomial of degree d in the n variables x1, x2,..., xn is hd(x1, x2,..., xn) = ∑ xi1 xi2 ··· xid , (1) 1≤i1≤···≤id≤n the sum of all degree d monomials in x1, x2,..., xn [28, Sec. 7.5]. For example, h0(x1, x2) = 1, h1(x1, x2) = x1 + x2, 2 2 h2(x1, x2) = x1 + x1x2 + x2, and 3 2 2 3 h3(x1, x2) = x1 + x1x2 + x1x2 + x2. n+d−1 Elementary combinatorics confirms that there are precisely ( d ) summands in Rn the definition (1). We often write hd(x), in which x =(x1, x2,..., xn) ∈ , when the number of variables is clear from context. Rn For d even and x ∈ , Hunter proved that hd(x) ≥ 0, with equality if and only if x = 0 [10]. This is not obvious because some of the summands that comprise hd(x) (for d even) may be negative. Hunter’s theorem has been reproved many times; see [2], [4, p. 69 & Thm. 3], [6, Cor. 17], [25, Thm. 2.3], and [29, Thm. 1]. 1.3. Partitions and traces. A partition of d ∈ N is an r-tuple π =(π1, π2,..., πr) ∈ r N such that π1 ≥ π2 ≥ ··· ≥ πr and π1 + π2 + ··· + πr = d; the number of terms r depends on the partition π. We write π ⊢ d if π is a partition of d. For π ⊢ d, define pπ(x1, x2,..., xn) = pπ1 pπ1 ··· pπr , k k k in which pk(x1, x2,..., xn) = x1 + x2 + ··· + xn. If the length of x =(x1, x2,..., xn) is clear from context, we often write pπ(x) and pk(x), respectively. Another expression for (1) is pπ(x1, x2,..., xn) hd(x1, x2,..., xn) = ∑ , (2) π ⊢ d zπ in which the sum runs over all partitions π =(π1, π2,..., πr) of d and mi zπ = ∏ i mi!, (3) i≥1 where mi is the multiplicity of i in π [28, Prop. 7.7.6]. For example, if π = 3 1 2 (4,4,2,1,1,1), then zπ =(1 3!)(2 1!)(4 2!) = 384 [28, (7.17)]. C If A ∈ Hn( ) has eigenvalues λ =(λ1, λ2,..., λn), then π1 π2 πr pπ(λ) = pπ1 (λ)pπ2(λ) ··· pπr (λ)=(tr A )(tr A ) ··· (tr A ). (4) This connects eigenvalues, traces, and partitions to symmetric polynomials. NORMS ON COMPLEX MATRICES 3 1.4. Main results. The following theorem provides a family of novel norms on the space Hn(C) of n × n Hermitian matrices. Some special properties of these norms are discussed in Section 4. Theorem 1. For even d ≥ 2, the following is a norm on Hn(C): 1/d |||A|||d = hd λ1(A), λ2(A),..., λn(A) . For example, equations (2) and (4) yield trace-polynomial representations 1 |||A|||2 = tr(A2) + (tr A)2 , 2 2 1 |||A|||4 = (tr A)4 + 6(tr A)2 tr(A2)+ 3(tr(A2))2 + 8(tr A) tr(A3)+ 6 tr(A4) . (5) 4 24 Theorem 1 is nontrivial for several reasons. (a) The sums (1) and (2) that characterize hd(λ(A)) may contain negative sum- mands. For example, (tr A) tr(A3) in (5) can be negative for Hermitian A. (b) The relationship between the spectra of (Hermitian) A, B, and A + B, conjec- tured by A. Horn in 1962 [9], was only established in 1998-9 by Klyachko [13] and Knutson–Tao [14]. Therefore, the triangle inequality is difficult to estab- lish. Even if A and B are diagonal, the result is not obvious; see (11). (c) The sums that define these norms do not involve the absolute values of the eigenvalues of A. Theorem 1 does not follow from standard considerations, but rather from delicate properties of multivariate symmetric polynomials. Example 2. Because CHS norms do not rely upon the absolute values of the eigenvalues of a Hermitian matrix (that is, its singular values), they can some- times distinguish singularly cospectral graphs (graphs with the same singular values) that are not cospectral. This feature is not enjoyed by many standard norms (e.g., operator, Frobenius, Schatten–von Neumann, Ky Fan). For example, 0 1 1 K = 1 0 1 , 1 1 0 which has eigenvalues 2, −1, −1, is the adjacency matrix for the complete graph on three vertices. The graphs with adjacency matrices K 0 0 K A = and B = 0 K K 0 are singularly cospectral but not cospectral: their eigenvalues are −1, −1, −1, −1, 6 6 2,2 and −2, −1, −1, 1, 1, 2, respectively. However, |||A|||6 = 120 6= 112 = |||B|||6. The norms of Theorem 1 extend in a natural and nontrivial fashion to the set Mn(C) of all n × n complex matrices. Theorem 3. Let d ≥ 2 be even and let π =(π1, π2,..., πr) be a partition of d. Define C R d d Tπ :Mn( ) → by setting Tπ(A) to be 1/(d/2) times the sum over the (d/2) possible locations to place d/2 adjoints ∗ among the d copies of A in (tr AA ··· A)(tr AA ··· A) ··· (tr AA ··· A). π1 π2 πr | {z } | {z } | {z } 4 K. AGUILAR, Á. CHÁVEZ, S.R. GARCIA, AND J. VOLCIˇ Cˇ Then T (A) 1/d |||A||| = ∑ π , (6) d z π π in which the sum runs over all partitions π of d and zπ is defined in (3), is a norm on Mn(C) that restricts to the norm on Hn(C) given by Theorem 1. Example 4. The two partitions of d = 2 satisfy z(2) = 2 and z(1,1) = 2. There are 2 ∗ (1) = 2 ways to place two adjoints in a string of two As. Therefore, 1 T (A) = tr(A∗ A) + tr(AA∗) = tr(A∗ A), and (2) 2 1 T (A) = (tr A∗)(tr A)+(tr A)(tr A∗) =(tr A∗)(tr A), (1,1) 2 so 1 1 |||A|||2 = tr(A∗ A) + (tr A∗)(tr A). (7) 2 2 2 Example 5. The five partitions of d = 4 satisfy z(4) = 4, z(3,1) = 3, z(2,2) = 8, 4 ∗ z(2,1,1) = 4, and z(1,1,1,1) = 24. There are (2) = 6 ways to place two adjoints in a string of four As. For example, ∗ ∗ ∗ ∗ ∗ ∗ 6T(3,1)(A)=(tr A A A)(tr A)+(tr A AA )(tr A)+(tr A AA)(tr A ) +(tr AA∗ A∗)(tr A)+(tr AA∗ A)(tr A∗)+(tr AAA∗)(tr A∗) = 3 tr(A∗2A)(tr A) + 3(tr A2 A∗)(tr A∗).
Load more

Recommended publications
  • A Direct Approach to Robustness Optimization
    A Direct Approach to Robustness Optimization Thesis by Seungil You In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy California Institute of Technology Pasadena, California 2016 (Defended August 12, 2015) ii c 2016 Seungil You All Rights Reserved iii To My Family iv Acknowledgments First and foremost, my gratitude goes to my advisor, John. He encouraged me to pursue my own ideas, and always gave me great feedback with a sense of humor. His unique research style made my academic journey at Caltech very special, and changed my way of looking at problems. I was simply lucky to have him as my advisor. I am also thankful to my thesis committee members: Richard Murray, Babak Hassibi, and Venkat Chandrasekaran. Richard has guided me at every milestone of my PhD studies, despite of his impossible schedule. Babak gladly agreed to be my committee member. I was in Venkat’s first class at Caltech on advanced topics in convex optimization. Of course, I was fascinated by the results he taught, and deeply influenced by his genuine point of view on solving mathematical problems. I learned convex optimization from Joel Tropp, and was immediately thrilled by the subject. His endless effort to establish a minimal, elegant theoretical framework for many challenging math problems had a profound impact on me. Lijun Chen taught me how to write a scientific paper, as well as how to formulate the problem in a concrete manner. Under his guidance, I was able to become an independent researcher. I met Ather Gattami in Florence, Italy, during CDC.
    [Show full text]
  • Front Matter
    Cambridge University Press 0521865654 - Combinatorial Matrix Classes Richard A. Brualdi Frontmatter More information ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS FOUNDED BY G.-C. ROTA Editorial Board P. Flajolet, M. Ismail, E. Lutwak Volume 108 Combinatorial Matrix Classes © Cambridge University Press www.cambridge.org Cambridge University Press 0521865654 - Combinatorial Matrix Classes Richard A. Brualdi Frontmatter More information ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS FOUNDED EDITOR G.-C. ROTA Editorial Board P. Flajolet, M. Ismail, E. Lutwak 40 N. White (ed.) Matroid Applications 41 S. Sakai Operator Algebras in Dynamical Systems 42 W. Hodges Basic Model Theory 43 H. Stahl and V. Totik General Orthogonal Polynomials 45 G. Da Prato and J. Zabczyk Stochastic Equations in Infinite Dimensions 46 A. Bj¨orner et al. Oriented Matroids 47 G. Edgar and L. Sucheston Stopping Times and Directed Processes 48 C. Sims Computation with Finitely Presented Groups 49 T. Palmer Banach Algebras and the General Theory of *-Algebras I 50 F. Borceux Handbook of Categorical Algebra I 51 F. Borceux Handbook of Categorical Algebra II 52 F. Borceux Handbook of Categorical Algebra III 53 V. F. Kolchin Random Graphs 54 A. Katok and B. Hasselblatt Introduction to the Modern Theory of Dynamical Systems 55 V. N. Sachkov Combinatorial Methods in Discrete Mathematics 56 V. N. Sachkov Probabilistic Methods in Discrete Mathematics 57 P. M. Cohn Skew Fields 58 R. Gardner Geometric Tomography 59 G. A. Baker, Jr., and P. Graves-Morris Pad´e Approximants, 2nd edn 60 J. Krajicek Bounded Arithmetic, Propositional Logic, and Complexity Theory 61 H. Groemer Geometric Applications of Fourier Series and Spherical Harmonics 62 H.
    [Show full text]
  • Gradients for the Loss!
    Gradients for the Loss! Towards Provable Gradient Algorithms for Non-Convex Problems Fotis Iliopoulos, Vrettos Moulous, Vaishaal Shankar, Max Simchowitz EE227 BT Department of EECSS Contents 1 Convergence of Stochastic Gradient for PCA3 1.1 SGD for Rank 1 Stochastic PCA........................3 1.1.1 A brief overview of Shamir's result...................4 1.1.2 Rank 1 Stochastic PCA Experiments..................7 1.2 The Alecton Algorithm for Higher Rank PCA.................7 1.2.1 Algorithm Intuition...........................8 1.2.2 Precise Setup...............................9 1.3 Sketching the Proof of Theorem 1.2.1...................... 10 1.3.1 Martingales................................ 10 1.3.2 Preliminaries............................... 10 1.3.3 Main Arguments and Completing the proof.............. 11 1.3.4 Experiments for Alecton......................... 14 1.3.5 Toy Convexity in Two Dimensions................... 14 1.3.6 Convexity Results in Higher Dimensions................ 15 1.3.7 Experimental Results.......................... 17 2 Dropping Convexity for Faster Semi-definite Optimization 19 2.1 Introduction: Back to Rank r-PCA....................... 19 2.2 From PCA to General Rank-Constrained Optimization............ 20 2.2.1 Notation and Assumptions....................... 21 2.2.2 Results.................................. 23 2.2.3 A Note on Step Size Selection...................... 24 2.3 Proof Sketch of Results............................. 24 2.4 Initialization................................... 27 2.4.1 Test case: Matrix sensing problem................... 27 3 Noisy Gradients in the Presence of Saddle Points 29 3.1 The Strict Saddle Property........................... 30 3.1.1 Sketch of the Proof............................ 31 3.1.2 Proof Sketch of Lemma 3.1.4...................... 32 3.2 Gradient Descent with Manifold Constraints.................. 34 3.2.1 The Tangent and Normal Space....................
    [Show full text]
  • Proscribed Normal Decompositions of Euclidean Jordan Algebras
    Proscribed normal decompositions of Euclidean Jordan algebras Michael Orlitzky June 30, 2020 Abstract Normal decomposition systems unify many results from convex ma- trix analysis regarding functions that are invariant with respect to a group of transformations—particularly those matrix functions that are unitarily-invariant and the affiliated permutation-invariant “spectral func- tions” that depend only on eigenvalues. Spectral functions extend in a nat- ural way to Euclidean Jordan algebras, and several authors have studied the problem of making a Euclidean Jordan algebra into a normal decom- position system. In particular it is known to be possible with respect to the “eigenvalues of” map when the algebra is essentially-simple. We show the converse, that essential-simplicity is essential to that process. 1 Notation All spaces we encounter will be finite-dimensional real inner-product spaces with the generic hx, yi denoting the inner product of x and y. If V is such a space, then B (V ) is the set of all linear operators on V and Aut (X) is the linear automorphism group of X ⊆ V . The inner product on V induces a norm, and Isom (V ) is the group of linear isometries on V under composition. Explicitly, −1 Aut (X) := L ∈ B (V ) L exists and L (X) = X , Isom (V ) := {L ∈ Aut (V ) | kL (x)k = kxk for all x in V } . The group Isom (V ) is endowed with the natural topology [16] so that Ln ap- proaches L in Isom (V ) if and only if Ln (x) approaches L (x) for all x ∈ V . The adjoint of any L ∈ B (V ) is denoted by L∗, and thus L∗ = L−1 when L ∈ Isom (V ).
    [Show full text]
  • Arxiv:1602.02719V1 [Math.CO] 8 Feb 2016 Monotonicity
    MONOTONE SUBSEQUENCES IN HIGH-DIMENSIONAL PERMUTATIONS NATHAN LINIAL AND MICHAEL SIMKIN Abstract. This paper is part of the ongoing effort to study high-dimensional permutations. We prove the analogue to the Erdős-Szekeres theorem: For every k ≥ 1, every order-n k-dimensional permutation contains a monotone p subsequence of length Ωk n , and this is tight. On the other hand, and unlike the classical case, the longest monotone subsequence in a random k- k dimensional permutation of order n is asymptotically almost surely Θk n k+1 . The study of monotone subsequences in permutations began with the famous Erdős-Szekeres theorem [5]. Since then numerous proofs and generalizations have emerged (see Steele’s survey [14]). We recall the theorem: Theoremp 1. Every permutation in Sn contains a monotone subsequence of length at least d ne, and this is tight: for every n there exists somep permutation in Sn in which all monotone subsequences are of length at most d ne. In order to derive a high-dimensional analogue of theorem 1 we need to define high-dimensional permutations and their monotone subsequences. If we view a permutation as a sequence of distinct real numbers, it is suggestive to consider se- quences of points in Rk, with coordinatewise monotonicity. The following argument is attributed by Kruskal [9] to de Bruijn: Repeatedly apply theorem 1 to conclude k that every sequence x1; x2; : : : ; xn 2 R must have a coordinatewise monotone sub- 1 sequence of length n 2k , and this is tight up to an additive constant. In [9] one considers projections of the points to a line and defines the length of the longest monotone subsequence according to the line with the longest such subsequence.
    [Show full text]
  • 2019 December .41 Contents Editorial
    CENTRO INTERNACIONAL DE MATEMÁTICA ISSN 2183-8062 (print version) ISSN 2183-8070 (electronic version) 2019 December .41 Contents Editorial 02 Editorial In this issue of the bulletin, we present an article about the topological by Jorge Milhazes de Freitas complexity of surfaces, which is a notion introduced by M. Farber in order 03 A Mathematical Celebration of the 150th to classify the complexity of the motion planning problem in robotics. Anniversary of the Periodic Table by José Francisco Rodrigues We also include a paper regarding the representation of commutative rings and respective modules by means of the prime spectrum, which 10 Nonlinear PDEs Workshop in Braga Report by Fernando Miranda and Lisa Santos is the set of prime ideals of a ring. We present an article about small 11 An interview with Graciano de Oliveira tubular neighbourhoods of the n-twisted Möbius strip described as affine by Ana Paula Santana varieties associated to explicit polynomials in three real variables, exhibiting 16 Topological complexity of surfaces illustrations produced with the software surfer. by Lucile Vandembroucq Inserted in the cycle of historical articles we feature a short biography 21 LxDS Spring School 2019 of Luis Woodhouse, which highlights, in particular, his efforts to pursue Report by João Lopes Dias, José Pedro Gavião, Pedro Miguel Duarte, and Telmo recognition for the work and research of Portuguese mathematicians such Peixe as José Maria Dantas Pereira and João Evangelista Torriani. 23 Luis Inacio Woodhouse: a Mathematics We feature an interview honouring Graciano de Oliveira for his dedication Professor to promoting mathematical research and dynamising the Portuguese by M.
    [Show full text]
  • Combinatorial Matrix Classes ENCYCLOPEDIA of MATHEMATICS and ITS APPLICATIONS
    ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS FOUNDED BY G.-C. ROTA Editorial Board P. Flajolet, M. Ismail, E. Lutwak Volume 108 Combinatorial Matrix Classes ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS FOUNDED EDITOR G.-C. ROTA Editorial Board P. Flajolet, M. Ismail, E. Lutwak 40 N. White (ed.) Matroid Applications 41 S. Sakai Operator Algebras in Dynamical Systems 42 W. Hodges Basic Model Theory 43 H. Stahl and V. Totik General Orthogonal Polynomials 45 G. Da Prato and J. Zabczyk Stochastic Equations in Infinite Dimensions 46 A. Bj¨orner et al. Oriented Matroids 47 G. Edgar and L. Sucheston Stopping Times and Directed Processes 48 C. Sims Computation with Finitely Presented Groups 49 T. Palmer Banach Algebras and the General Theory of *-Algebras I 50 F. Borceux Handbook of Categorical Algebra I 51 F. Borceux Handbook of Categorical Algebra II 52 F. Borceux Handbook of Categorical Algebra III 53 V. F. Kolchin Random Graphs 54 A. Katok and B. Hasselblatt Introduction to the Modern Theory of Dynamical Systems 55 V. N. Sachkov Combinatorial Methods in Discrete Mathematics 56 V. N. Sachkov Probabilistic Methods in Discrete Mathematics 57 P. M. Cohn Skew Fields 58 R. Gardner Geometric Tomography 59 G. A. Baker, Jr., and P. Graves-Morris Pad´e Approximants, 2nd edn 60 J. Krajicek Bounded Arithmetic, Propositional Logic, and Complexity Theory 61 H. Groemer Geometric Applications of Fourier Series and Spherical Harmonics 62 H. O. Fattorini Infinite Dimensional Optimization and Control Theory 63 A. C. Thompson Minkowski Geometry 64 R. B. Bapat and T. E. S. Raghavan Nonnegative Matrices with Applications 65 K.
    [Show full text]
  • Robust Mixing
    THE UNIVERSITY OF CHICAGO ROBUST MIXING A DISSERTATION SUBMITTED TO THE FACULTY OF THE DIVISION OF THE PHYSICAL SCIENCES IN CANDIDACY FOR THE DEGREE OF DOCTOR OF PHILOSOPHY DEPARTMENT OF COMPUTER SCIENCE BY MURALI KRISHNAN GANAPATHY CHICAGO, ILLINOIS AUGUST 2006 To my dear wife and parents Abstract How many times should a card shuffler shuffle to get the cards shuffled? Convergence rate questions like these are central to the theory of finite Markov Chains and arise in diverse fields including Physics, Computer Science as well as Biology. This thesis introduces two new approaches to estimating mixing times: robust mixing time of a Markov Chain and Markovian product of Markov Chains. The “robust mixing time” of a Markov Chain is the notion of mixing time which results when the steps of the Markov Chain are interleaved with that of an oblivious adversary under reasonable assumptions on the intervening steps. We develop the basic theory of robust mixing and use it to give a simpler proof of the limitations of reversible liftings of a Markov Chain due to Chen, Lov´asz, and Pak (1999). We also use this framework to improve the mixing time estimate of the random-to-cyclic transposition process (a non-Markovian process) given by Peres and Revelle (2004). The “Markovian product” of Markov Chains is like the direct product, except for a controlling Markov Chain which helps decide which component should be updated. Direct products as well as wreath products of Markov Chains are special cases. We show how a coupon collector type of analysis can be used to estimate the mixing times of these product chains under various distance measures.
    [Show full text]
  • And Leon Mirsky [4] Give Necessary and Sufficient Conditions for the Existence of an N-Square Matrix a That Has Prescribed Diagonal Elements and Prescribed Eigenvalues
    Electronic Journal of Linear Algebra ISSN 1081-3810 A publication of the International Linear Algebra Society Volume 18, pp. 438-441, July 2009 ELA http://math.technion.ac.il/iic/ela SHORT PROOFS OF THEOREMS OF MIRSKY AND HORN ON DIAGONALS AND EIGENVALUES OF MATRICES∗ ERIC A. CARLEN† AND ELLIOTT H. LIEB‡ Abstract. A theorem of Mirsky provides necessary and sufficient conditions for the existence of an N-square complex matrix with prescribed diagonal entries and prescribed eigenvalues. A simple inductive proof of this theorem is given. Key words. Majorization, Eigenvalues, Prescribed diagonals. AMS subject classifications. 15A42, 15A51. Theorems of Alfred Horn [3] and Leon Mirsky [4] give necessary and sufficient conditions for the existence of an N-square matrix A that has prescribed diagonal elements and prescribed eigenvalues. In the case of Horn’s Theorem, A is required to be Hermitian, and in the case of Mirsky’s theorem no constraint is imposed on A. While simple inductive proofs of Horn’s Theorem have been published, the only published proof of Mirsky’s Theorem that we are aware of is the original, which is neither simple nor transparent. In this note, we give a simple proof of Mirsky’s Theorem which we relate to Chan and Li’s proof [2] of Horn’s Theorem. We begin by recalling the results in [3, 4] in more detail. Let Λ = (λ1,...,λN )andD =(d1,...,dN ) be two sequences of N real numbers. Recall that Λ majorizes D (we write Λ D)ifforeachk =1,...,N, the sum of the k largest λ’s is at least as large as the sum of the k largest d’s, with equality for k = N.
    [Show full text]
  • “But You Have to Remember P. J. Daniell of Sheffield”
    “But you have to remember P. J. Daniell of Sheffield” JOHN ALDRICH1 Abstract: P. J. Daniell is a mysterious figure who appears at several turns in the history of mathematics in the 20th century, in the fields of integration, stochastic processes, statistics, control engineering and even in the history of English mathematical education. The main focus of this paper is on Daniell’s work in relation to the development of probability in the twentieth century. But as it seems that no survey of his work and life has been attempted for 60 years I try to consider all his contributions and place them in an appropriate historical context. R´esum´e: P. J. Daniell est un personnage myst´erieuxqui apparaˆıt`aplusieurs moments clefs de l’histoire des math´ematiquesdu 20`emesi`ecle, dans le do- maine de l’int´egration,des processus stochastiques, des statistiques, de la commande optimale et mˆemedans l’histoire de l’´educationmath´ematique en Angleterre. Ce papier se concentre sur le travail de Daniell en relation avec le d´eveloppement des probabilit´esau vingti`emesi`ecle.Comme aucune description de sa vie et de son œuvre n’a semble-t-il ´et´er´ealis´eedepuis 60 ans, nous essayons de dresser un tableau de l’ensemble de ses contributions et de les placer dans un contexte historique appropri´e. 1Economics Division. School of Social Sciences, University of Southampton. Southampton SO17 1BJ. UK. [email protected] P.J.Daniell (Courtesy of Woodson Research Center, Fondren Library, Rice University, Houston, USA c Rice University 2007) 1 Introduction In a conversation with David Kendall, the doyen of British probabilists, Bingham (1996, p.
    [Show full text]
  • Thesis-Part1.Pdf
    Spectral Methods for Natural Language Processing (Part I of the Dissertation) Jang Sun Lee (Karl Stratos) Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Graduate School of Arts and Sciences COLUMBIA UNIVERSITY 2016 c 2016 Jang Sun Lee (Karl Stratos) All Rights Reserved Table of Contents 1 A Review of Linear Algebra 1 1.1 Basic Concepts . .1 1.1.1 Vector Spaces and Euclidean Space . .1 1.1.2 Subspaces and Dimensions . .2 1.1.3 Matrices . .3 1.1.4 Orthogonal Matrices . .6 1.1.5 Orthogonal Projection onto a Subspace . .6 1.1.6 Gram-Schmidt Process and QR Decomposition . .8 1.2 Eigendecomposition . .9 1.2.1 Square Matrices . 10 1.2.2 Symmetric Matrices . 12 1.2.3 Variational Characterization . 14 1.2.4 Semidefinite Matrices . 15 1.2.5 Numerical Computation . 17 1.3 Singular Value Decomposition (SVD) . 22 1.3.1 Derivation from Eigendecomposition . 23 1.3.2 Variational Characterization . 26 1.3.3 Numerical Computation . 27 1.4 Perturbation Theory . 28 1.4.1 Perturbation Bounds on Singular Values . 28 1.4.2 Canonical Angles Between Subspaces . 29 1.4.3 Perturbation Bounds on Singular Vectors . 30 i 2 Examples of Spectral Techniques 36 2.1 The Moore{Penrose Pseudoinverse . 36 2.2 Low-Rank Matrix Approximation . 37 2.3 Finding the Best-Fit Subspace . 39 2.4 Principal Component Analysis (PCA) . 39 2.4.1 Best-Fit Subspace Interpretation . 40 2.5 Canonical Correlation Analysis (CCA) . 41 2.5.1 Dimensionality Reduction with CCA .
    [Show full text]
  • Spectral Methods for Natural Language Processing Jang Sun
    Spectral Methods for Natural Language Processing Jang Sun Lee (Karl Stratos) Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Graduate School of Arts and Sciences COLUMBIA UNIVERSITY 2016 c 2016 Jang Sun Lee (Karl Stratos) All Rights Reserved ABSTRACT Spectral Methods for Natural Language Processing Jang Sun Lee (Karl Stratos) Many state-of-the-art results in natural language processing (NLP) are achieved with sta- tistical models involving latent variables. Unfortunately, computational problems associ- ated with such models (for instance, finding the optimal parameter values) are typically intractable, forcing practitioners to rely on heuristic methods without strong guarantees. While heuristics are often sufficient for empirical purposes, their de-emphasis on theoretical aspects has certain negative ramifications. First, it can impede the development of rigorous theoretical understanding which can generate new ideas and algorithms. Second, it can lead to black art solutions that are unreliable and difficult to reproduce. In this thesis, we argue that spectral methods|that is, methods that use singular value decomposition or other similar matrix or tensor factorization|can effectively remedy these negative ramifications. To this end, we develop spectral methods for two unsupervised language processing tasks. The first task is learning lexical representations from unanno- tated text (e.g., hierarchical clustering of a vocabulary). The second task is estimating parameters of latent-variable models used in NLP applications (e.g., for unsupervised part- of-speech tagging). We show that our spectral algorithms have the following advantages over previous methods: 1. The algorithms provide a new theoretical framework that is amenable to rigorous analysis.
    [Show full text]