Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Gaussian elimination wikipedia , lookup
Symmetric cone wikipedia , lookup
Determinant wikipedia , lookup
Matrix (mathematics) wikipedia , lookup
Non-negative matrix factorization wikipedia , lookup
Singular-value decomposition wikipedia , lookup
Jordan normal form wikipedia , lookup
Eigenvalues and eigenvectors wikipedia , lookup
Four-vector wikipedia , lookup
Matrix multiplication wikipedia , lookup
Orthogonal matrix wikipedia , lookup
Matrix calculus wikipedia , lookup
BEST UPPER BOUNDS BASED ON THE ARITHMETIC-GEOMETRIC MEAN INEQUALITY Luc Knockaert1 Abstract In this paper we obtain a best upper bound for the ratio of the extreme values of positive numbers in terms of the arithmetic-geometric means ratio. This has immediate consequences for condition numbers of matrices and the standard deviation of equiprobable events. It also allows for a refinement of Schwarz’s vector inequality. Keywords Arithmetic mean, geometric mean, condition number, standard deviation, Schwarz inequality AMS 1991 Subject classification 15A42, 60E15 1 INTRODUCTION Bounds for the extreme eigenvalues of positive-definite matrices [1], [2] allow to localize their spectrum and to obtain useful estimates for their spectral condition number [3]. The arithmeticgeometric mean inequality is a classical subject [4] with developments and applications in [5]–[7], where the last paper deals with a trace-determinant based spectral condition number bound. We follow the approach of [7] to obtain a best upper bound for the spectral condition number of a positive definite matrix in terms of the arithmetic-geometric means ratio of the eigenvalues. This has consequences for the spherical condition number of general matrices and also, rather surprisingly, for the standard deviation of equiprobable events and Schwarz’s vector inequality. 2 MAIN RESULT Given a sequence of n positive real numbers 0 < λ1 ≤ λ2 ≤ . . . ≤ λn 1 (1) IMEC-INTEC, St. Pietersnieuwstraat 41, B-9000 Gent, Belgium. Tel: +32 9 264 33 43, Fax: +32 9 264 35 93. E-mail: [email protected] 2 the arithmetic-geometric mean inequality states that S(λ) = nn λi n ≥1 λi (2) It is clear that S(λ) is a homogeneous expression, i.e. S(αλ) = S(λ) for all α > 0. Another important simple homogeneous form is T (λ) = λn λ1 (3) When the numbers λi represent the eigenvalues of a positive definite matrix, T (λ) is known as the spectral condition number [3]. Our purpose is to obtain the best upper bound of the form T (λ) ≤ Φ (S(λ)) (4) where Φ(x) is a strictly increasing continuous function defined over the interval [1, ∞). We have the following Theorem : The best bound of the form (4) is obtained when Φ(x) = τ (x) ≡ (2x − 1) + (2x − 1)2 − 1 (5) Proof : Clearly τ (x) is a strictly increasing continuous function defined over [1, ∞). For n = 2 it is easy to show that the bound is satisfied with equality. For n > 2 we consider the minimization problem 1 Φ (S(λ)) J = min λ T (λ) (6) To account for homogeneity we can take λ1 = 1 leading to 1 ≤ λ2 ≤ . . . ≤ λ n (7) Hence problem (6) can be written as J = min λn 1 λn min λ2 ,···,λn−1 Φ (S(λ)) (8) Since Φ(x) is supposed to be strictly increasing, we can consider minimizing S(λ) for λn fixed. From symmetry considerations it is clear that the minimum is obtained for λi = 1 + λn 2 i = 2, . . . , n − 1 (9) 3 Putting β = λn it is seen that 1 (1 + β)2 J = min Φ β≥1 β 4β (10) Now since τ (u) is a homeomorphic transformation over [1, ∞), expression (10) can be written as J = min ξ(u) (11) Φ(u) τ (u) (12) u≥1 where ξ(u) = From (6) we obtain that T (λ) ≤ K τ (S(λ)) (13) where ξ (S(λ)) ≥1 minu ξ(u) K= (14) which completes the proof, since K = 1 if and only if ξ(u) is a constant function. Corollary 1 : Let A be a positive definite Hermitian matrix of dimension n and let κs (A) be its spectral condition number. Then κs (A) ≤ (2s − 1) + (2s − 1)2 − 1 (15) where s = n−n (trA)n / det(A) (16) Proof : Straightforward, since κs (A) = λn /λ1 , the ratio of the largest and smallest eigenvalues of A. Corollary 2 : Let A be a complex square matrix of dimension n and let κ2 (A) be its spherical condition number and AF its Frobenius norm. Then κ2 (A) ≤ (2s − 1) + (2s − 1)2 − 1 (17) where 2 s = n−n A2n F /| det(A)| (18) Proof : Straightforward, since κ2 (A) = σn /σ1 , the ratio of the largest and smallest singular value of A, and exploiting the fact that A2F = i j |Aij |2 = i σi2 (19) 4 Corollary 3 : Let X be a discrete random variable with a finite set of equiprobable outcomes µ1 ≤ µ2 ≤ . . . ≤ µn . Then the standard deviation σ satisfies the inequality µn − µ1 √ 2n σ≥ (20) Proof : We start with the generalization λn ≤ (2s(p) − 1) + λ1 2 1 (2s(p) − 1) − 1 where s(p) = nn λpi p p>0 (21) n p (22) λi Putting µi = ln λi , it is straightforward to show that we have the power expansion 1 s(p) = 1 + nσ 2 p2 + O p3 2 in the vicinity of p = 0, where of course σ2 = µ2i n After some algebra we obtain µn − µ1 ≤ − √ µi n (23) 2 2n σ + O(p) (24) (25) leading in the limit for p → 0+ to the required result. Corollary 4 : Refinement of Schwarz’s vector inequality. Let w, y be two column vectors in Rn with unit Euclidean norm. Then (wT y)2 ≤ 1 − 1 (∆v)2 2 (26) Here ∆v ≡ (maxi vi − mini vi ) stands for the variation of v, with v itself defined as v = w − (en − y) (en − y)T w 1 − y T en (27) where en is the constant unit norm vector √ √ √ en = (1/ n, 1/ n, · · · , 1/ n)T (28) Proof : The vector v can be written as v = Ωw, where Ω is the orthogonal Householder [3] reflection Ω = ΩT = I − 2uuT with u= en − y 2 − 2y T en (29) 5 since we have uT u = 1. The result of the previous Corollary may be cast for our purposes into the form (v T en )2 ≤ v T v − 1 1 (∆v)2 = 1 − (∆v)2 2 2 (30) since orthogonal transformations preserve the Euclidian norm. Hence (wT Ωen )2 ≤ 1 − 1 (∆v)2 2 (31) and the proof is complete, since it is easy to show that Ωen = y. For the degenerate case y = en , we simply take u = 0 and Ω = I. 3 AN EXAMPLE From Corollary 3 we obtain easily µk ≤ As an example, if we take µk = √ n µ2k 1 − (µn − µ1 )2 2 (32) k we obtain the inequality n √ k=1 k≤ √ n 2 n +2 n−1 2 (33) Of course, inequality (33) is satisfied with equality for n = 1, 2. For larger values of n it remains relatively tight since it is easy to show that the ratio n √ 2k R(n) = k=1 √ n n2 + 2 n − 1 (34) √ satisfies L ≤ R(n) ≤ 1 with L = 2 2/3 ≈ 0.942809. Figure 1 presents a plot of R(n) for 1 ≤ n ≤ 100. References [1] A. Dembo, Bounds on the extreme eigenvalues of positive-definite Toeplitz matrices, IEEE Trans. Inform. Theory, 2, (1988), 352-355. [2] E. M. Ma, C. J. Zarowski, On lower bounds for the smallest eigenvalue of a Hermitian positive-definite matrix, IEEE Trans. Inform. Theory, 2, (1995), 539-540. [3] A. S. Householder, The Theory of Matrices in Numerical Analysis, Dover, New York, 1975. 6 [4] P. S. Bullen, D. S. Mitrinovic, P. M. Vasic, Means and Their Inequalities, Reidel Publishing, Dordrecht, 1988. [5] K. Dzhaparidze, R. H. P. Janssen, A stochastic approach to an interpolation problem with applications to Hellinger integrals and arithmetic-geometric mean relationship, CWI Quarterly, 3, (1994), 245-258. [6] B. Ortner, A. R. Kräuter, Lower bounds for the determinant and the trace of a class of Hermitian matrices, Lin. Alg. Appl., 236, (1996), 147-180. [7] J. E. Dennis Jr., H. Wolkowicz, Sizing and least-change secant methods, SIAM J. Numer. Anal., 30, (1993), 1291-1314. Figure Captions Fig. 1: The ratio R(n) in the range 1 ≤ n ≤ 100. R(n) 1 0.99 0.98 0.97 0.96 0.95 0.94 0 10 20 30 40 50 n 60 70 80 90 100