You are on page 1of 98

Institute of Numerical Mathematics Russian Academy of Sciences

II INTERNATIONAL CONFERENCE ON MATRIX METHODS AND OPERATOR EQUATIONS


Web-site: bach.inm.ras.ru E-mail: "E. E. Tyrtyshnikov" <tee@inm.ras.ru>

PRINCIPAL ORGANIZERS: Institute of Numerical Mathematics, Russian Academy of Sciences Lomonosov Moscow State University

CO-ORGANIZERS & SPONSORS: Russian Foundation for Basic Research International Foundation for Technology and Investments (IFTI) Neurok Techsoft University of Insubria (Como, Italy)

ORGANIZING COMMITTEE: Dario Bini Gene Golub Alexander Guterman Yuri Nesterenko Vadim Olshevsky Stefano Serra-Capizzano Gilbert Strang Eugene Tyrtyshnikov University of Pisa, Italy Stanford University, USA Moscow State University Moscow State University University of Connecticut, USA University of Insubria, Italy M.I.T., USA INM RAS; Moscow State University

APPLICATION SECTION COMMITTEE: O. Diyankov (Neurok), S. Karabashev (IFTI), M. Olshansky (MSU), Yu. Vassilevski (INM RAS)

ABSTRACTS
Moscow, July 2327, 2007

On the spectra of the Laplacian matrices of a certain class Rag Agaev Institute of Control Sciences of the Russian Academy of Science, Moscow arpoye@rambler.ru A Laplacian matrix L = (ij ) Rnn has nonpositive o-diagonal entries and zero row sums. Let L = (lij ) be a Laplacian matrix dened as 1 2 ... i i + 1 ... n 1 n 1

In this paper the spectra of the matrices L1 and L0 are studied by means of the Chebyshev polynomials Pn (x) of the second kind. More particularly, we consider the polynomials Zn (x), n N, that satisfy the recurrence relation Zn (x) = (x 2)Zn1 (x) Zn2 (x) with the initial conditions Z0 (x) 1 and Z1 (x) x 1, in which case Zn (x) and Pn (x) are related by Zn (x) = P2n ( x). Theorem 1. The characteristic polynomial of L1 is L1 () = Zn ()(1)n and the eigenvalues k of L1 are 4 cos2 ( 2n+1(1)k+n ), k = 1, . . . , n. Theorem 2. The characteristic polynomial of L0 is L0 () = Zi ()Zni () (1)n . For an even n, all eigenvalues of L0 are real if and only if i = n , in which case they are 4 cos2 ( k ), 2 n k 4 cos2 ( n+2 ), k = 1, . . . , n . For an odd n, all eigenvalues of L0 are real if and only if i = n1 or 2 2 k i = n+1 ; in either case they are 0 and 4 cos2 ( n+1 ), k = 1, . . . , n1 , of multiplicity two. 2 2 In this work we also show that the theorems 1 and 2 can be applied to calculation of the number of spanning trees of digraphs of a certain class. A similar result for undirected graphs is due to Boesch and Prodinger [1] and Zhang, Yong and Golin [2]. References 1. Boesch F.T., Prodinger H. Spanning tree formulas and Chebyshev polynomials, Graphs and Combinatorics. 1986. V. 2. P. 191200. 2. Zhang Y., Yong X., Golin M. Chebyshev polynomials and spanning tree formulas for circulant and related graphs Discrete Mathematics. 2005. V. 298. P. 334364.

2 1 1 2 ... 1 + L = 1 2 ...

2 1 1 1

1 2 ... i i+1 ... n1 n

Spectral analysis of the anti-reective transform A. Aric` , M. Donatelli and S. Serra-Capizzano o Universit` degli studi di Cagliari, Italy a Universit` degli Studi dellInsubria, Via Ravasi, 2, 21100 Varese, Italy a 1

arico@unica.it, marco.donatelli@uninsubria.it, s.serracapizzano@uninsubria.it

Anti-reective boundary conditions have been studied in connection with fast deblurring algorithms, in the case of d-dimensional objects (signals for d = 1, images for d = 2). Here we show how, under the assumption of strong symmetry of the point spread functions and under mild degree conditions, the associated matrices depend on a symbol and dene an algebra homomorphism. Furthermore, the eigenvalues can be exhaustively described in terms of samplings of the symbol and other related functions, and appropriate O(nd log(n)) arithmetic operations algorithms can be derived for the related computations. These results, in connection with the use of the anti-reective transform, are of interest when employing ltering type procedures for the reconstruction of noisy and blurred objects. References 1. S. Serra-Capizzano, A note on anti-reective boundary conditions and fast deblurring models, SIAM J. Sci. Comput., 25 (3) (2003) 13071325.

Splitting algorithm for solving variational inequalities with inversely strongly monotone operators I. B. Badriev , O. A. Zadvornov Kazan State University, Russia The work was supported by RFBR, grant 06-01-00633 Ildar.Badriev@ksu.ru Let R n , n 1 be a bounded domain with a Lipschitz continuous boundary . We consider the following boundary valued problem relatively function u = (u1 , u2 , . . . , un ):
n

(
j=1

(i) v (x) + dij (x) uj (x) ) = fi (x), xj j vj (x)

x ,

i = 1, 2, . . . , n,

u(x) = 0,

x ,

(1)

gj : [0, +) [0, +) are the continuous functions satises the following conditions: gj (0) = 0, gj () > gj () > 0, j > 0 : | gj () gj () |

where f = (f1 , f2 , . . . , fn ) is given function, D = {dij } is a non symmetric matrix such that (D, ) 0 (D, D) for all R n , 0 > 0. We assume that the multi-valued functions gj can be represented in the form gj () = g0j () + j h( j ), where j , j are the given non negative constants, h is the multi-valued and g0j are the single-valued functions given by the formulas < 0, 0, 0, j , h() = [0, 1], = 0, g0j () = ( ), , gj j j 1, > 0, 1 | | , 0, j

gj (|u(x)/xj |) u(x) , |u(x)/xj | xj

x ,

j = 1, 2, . . . , n,

(2)

kj > 0, j 0 : gj (j ) kj j ,

We introduce the next notations: V = W 2 () , H = [L2 ()]n , Bj = /xj : V H, j = 1, 2, . . . , n. A solution of the problem (1), (2) is dened as the function u V satisfying for all V the variational inequality
n n

(1)

gj () gj () kj ( ) j . n

(A0 u, u)V +

j=1

(Aj Bj (u), Bj ( u))H + F0 () F0 (u) +

j=1

[Fj (Bj ) Fj (Bj u) ] 0, (3)

where Bj : H V , j = 1, 2, . . . , n, are conjugated to Bj operators, the operators A0 : V V and Aj : H H, j = 1, 2, . . . , n, are generated by the forms

(A0 u, )V =

(Qu, )dx, u, V,

(Aj y, z)H =

(Gj (y), z)dx, y, z H,

the operators Gj : R n R n and the functionals F0 : V R 1 , Fj : H R 1 , j = 1, 2, . . . , n, are dened by the formulas: Gj (y) = g0j (|y|) |y|1 y , y = 0, Gj (0) = 0, Fj (z) = j

F0 () =

(f, ) dx, V, 0, < 0, , 0.

(|z| j ) dx, z H,

() =

To nd a solution of the variational inequality (3), one can use the following splitting algorithm. (0) (0) We x j > 0, j = 0, 1, 2, . . . , n, and r > 0. Let u(0) V , yj H, j H, j = 0, 1, . . . , n, be arbitrary elements. For k = 0, 1, 2, . . ., and for known elements yj , j , j = 1, 2, . . . , n, we dene u(k+1) as a solution of the variational inequality 1 0 u(k+1) u(k) , u(k+1)
n V (k) (k)

+ F0 () F0 u(k+1) + A0 u(k) , u(k+1)


(k)

+ (4)

+(
j=1

Bj j + r j=1

(k)

Bj (Bj u(k) yj ) , u(k+1) )V 0

V.

Then we nd yj

(k+1)

, j = 1, 2, . . . , n, from the variational inequalities

1 (k+1) (k) (k+1) (k+1) (k) (k) (k+1) (yj yj , z yj )H + Fj (z) Fj (yj ) + (Aj yj j z yj )H + j +r (yj Finally we set j
(k+1) (k)

Bj u(k+1) , z yj
(k)

(k+1)

)H 0 z H,
(k+1)

j = 1, 2, . . . , n.

(5) (6)

= j + r (Bj u(k+1) yj

),

j = 1, 2, . . . , n.

To analyze the convergence of the method (4)-(6) we formulate it via the transition operator T : V H n H n V H n H n that takes each vector q = ( q0 , q1 , . . . , q2n ) = (u, Y, ), Y H n , H n to the element T q = ( T 0 q, T 1 q, . . . , T 2n q ) as follows
n n Bj q n+j j=1

T 0 q = Prox 0 F0 ( q0 0 [ A0 q0 +

+r
j=1

Bj (Bj q0 qj ) ] ) ,

T j q = Prox j Fj (qj j [Aj qj q n+j + r (qj Bj T 0 q ) ] ) , j = 1, 2, . . . , n, T n+j q = q n+j + r (Bj T 0 q T j q ) ,


(k) (k) (k)

j = 1, 2, . . . , n,

where Prox G is the proximal mapping. Then the method (4)-(6) can be represented in the form
(k) q (k+1) = T q (k) , q (k) = (u(k) , y1 , y2 , . . . , yn , 1 , 2 , . . . , (k) ), k = 0, 1, 2, . . . , n (k)

(7)

where q (0) is an arbitrary element. Theorem 1. The point q = (u, y1 , y2 , . . . , yn , 1 , 2 , . . . , n ) is a xed point of the T if and only
n

if yj = Bj u, j Fj (yj ) + Aj yj , j = 1, 2, . . . , n;

j=1

Bj j F0 (u) + A0 u. Moreover, the rst

component u of each xed point of the operator T is a solution of the problem (3). Theorem 2. Let the following conditions j < (2 j )/(2 j r + 1), j = 0, 1, 2, . . . , n, are hold and + the operator T has at least one xed point. Then the iterative q (k) k=0 constructed by the (7) converges weakly to q = T q and yj Bj u (k) H = 0, j = 1, 2, . . . , n as k +. It would be note that for the problem (1), (2) the splitting algorithm (4) (6) is reduced to solving of n boundary valued problems with Laplace operators.
(k)

Elementary equivalence of unitary linear groups Balmassov E. S. Moscow State University, Russia helenbunina@yandex.ru Elementary properties of linear groups were considered in 1961 by A.I.Maltsev. in the paper [1] for the groups GLn (K), SLn (K), P Gn (K), P SLn (K), it was proved that they are elementary equivalent if and only if their dimensions coincide, and the initial elds are elementary equivalent. This theory was continued in 1992, when K. Beidar and A.V. Mikhalev in [2] found the general approach to the problems of this type. With the help of logical constructions and model theory they generalized Maltsev theorem for the case of the group GL over prime rings with 1/2, over skewelds, and also for some other algebraic systems. Then in 1998 E.I. Bunina in the paper [3] studied elementary properties of unitary linear groups over elds. There were proved structural theorems on unitary linear groups over elds with involutions with forms of maximal rank, and there were used only methods from [1]. Besides, with the help of Beidar and Mikhalev theorems [2] and isomorphism theorem of I.Z. Golubchik and A.V. Mikhalev [4] in 1998 in the paper [5] E.I. Bunina proved theorems on elementary equivalence of unitary linear groups over associative rings and skewelds with 1/2 and 1/3. Now we generalize results of the paper [3] for unitary linear groups over elds, where the form is not necessarily of maximal rank. Let K be innite eld of characteristics = 2 with involution j. Let Q2p+q be the form of rank (2p, q). By U2p+q (K, j, Q) we denote the unitary group of matrices A GL2p+q (K) such that AQ2p+q A = Q2p+q , where A = (Aj )T . We prove the following theorem: Theorem. The groups U2p1 +q1 (K1 , j1 , Q2p1 +q1 ) and U2p2 +q2 (K2 , j2 , Q2p2 +q2 ), where K1 and K2 are innite elds of characteristics = 2, with involutions j1 and j2 , respectively, p1 , p2 2, q1 , q2 4, are elementary equivalent if and only if : 4

1) p1 = p2 , q1 = q2 and the elds K1 and K2 are elementary equivalent as elds with involutions; 2) the involution j1 is not identical, 2p1 = q2 , 2p2 = q1 and the elds K1 and K2 are elementary equivalent as the elds with involutions. References 1. Maltsev A.I., On elementary equivalence of linear groups, Problems of mathematics and mechanics, Novosibirsk, 1961, pp. 110-132 (in Russian). 2. Beidar C.I., Michalev A.V., On Malcevs theorem on elementary equivalence of linear groups, Contemporary mathematics, 131 (1992), pp. 2935. 3. Bunina E.I. Elementary equivalence of unitary linear groups over elds, Fundamentalnaya i prikladnaya matematika, 4 (1998), pp. 114 (in Russian). 4. Golubchik I.Z., Mikhalev A.V., Isomorphisms of unitary groups over associative rings, Zapiski nauchnyh seminarov LOMI AN USSR, 132 (1983), pp. 97109 (in Russian). 5. Bunina E.I., Elementary equivalence of unitary linear groups over rings and skewelds, Russian Mathematical Surveys, 53(2): 137138 (1998).

Constructing a Sparse Linear System Hardware Accelerator J. Vic Batson , Oleg V. Diyankov Corporation, 236 West Portal Avenue #221, San Francisco, CA 94127, USA Sparsix Corporation, Oktyabrsky pr., 19a, 120, Troitsk, Russia diyankov@aconts.com

Sparsix

Handling massive amounts of data and complex computations are common aspects of highperformance computing. Sparse linear systems, however, introduce an added level of complexity that greatly reduces the eciency of general-purpose computers: non-sequential data access. Modern computer architectures are tuned to access data in sequential blocks. To solve sparse linear systems, data must often be accessed in a non-sequential manner, leading to massive cache misses and dramatically reducing overall system performance. When analyzing large sparse linear systems, the combination of non-sequential data access and numerous, complex computations can easily cause simulation times to extend into the tens or hundreds of hours, even on powerful general-purpose computers. To address this problem, Sparsix Corporation has developed its Sparse System Accelerator (SSA), a eld programmable gate array (FPGA)-based application accelerator designed to signicantly improve the performance of general-purpose computers when solving large sparse linear systems. The SSA features two stages of performance improvement: a custom-designed memory subsystem based on Sparsixs patent-pending Intelligent Memory Controller (IMC) technology, and a recongurable coprocessor that provides hardware acceleration of certain mathematical functions used in sparse linear systems analysis. This balanced design is unique in that it addresses the issue of non-sequential data access as well as the computational challenge. Sparsix Corporation is an early-stage US-Russia technology rm developing solutions that dramatically accelerate the speed at which general-purpose computers process sparse linear systems. For more information, email info@sparsix.com.

Averaging operations on matrices Rajendra Bhatia Indian Statistical Institute New Delhi 110016, India rbh@isid.ac.in Various notions of means of positive numbers can be extended to positive denite matrices. Of these the geometric mean presents interesting diculties because of noncommutativity of matrix multiplication and subtleties of matrix order. We will survey dierent approaches to the problem advanced by physicists, engineers, and mathematicians.

On ranks of spatial matrices Maxim Bershadsky Ben-Gurion University of the Negev, Sapir Academic College maximb@bgu.ac.il, maximb@mail.sapir.ac.il The study of spatial matrices dates back to the 19-th century. In particular, the determinants and ranks of spatial matrices were dened and the classication of normal forms of 222 matrices over C and R was obtained. G. Belitskii, V. Sergrichuck and I have recently presented the classication of m n 2 - matrices and the list of normal forms for m 2 2 matrices over an arbitrary eld K , based on the Kronecker-Weierstrass theory. In this presentation two- and three-dimensional ranks of spatial matrices are considered and a new invariant with respect to equivalence of spatial matrices over a eld K is introduced. The invariant, rA (x) , is an integer-valued function on Kq for a given m n q - spatial matrix A . The function image r(A) : rA (Kq ) is particularly connected to the classical three-dimensional ranks and even provides more information than said ranks. As an application of the invariant, the classication of some normal forms of spatial 3 3 3 matrices are presented.

The Cyclic Reduction Algorithm: From Poissons Equation to Stochastic Processes and Beyond Dario A. Bini Dipartimento di Matematica, Universit` di Pisa a bini@dm.unipi.it Cyclic reduction (CR) is a well known algorithm for people working in the area of numerical linear algebra. It was introduced by Gene H. Golub in 1970 for the ecient solution of the Poisson equation in a rectangle and has been adapted and modied to cope with dierent computational needs. 6

In this talk we revisit this celebrated algorithm and describe recent applications in dierent computational frameworks like solving wide classes of queuing models, more general stochastic processes, and matrix equations of dierent kind like unilateral polynomial equations and the Riccati equation. We discuss the functional interpretation of CR which allows one to relate this algorithm with the Dandelin-Graee-Lobachevsky iteration, to analyze its convergence properties and to extend it to solving (generalized) block Hessenberg block Toeplitz systems. Finally, we report the results of a recent research which aims to relate CR with the Structured Doubling Algorithm (SDA) introduced to solve algebraic Riccati equations.

Multi-Level Partitioning Algorithms N. Bochkarev , O. Diyankov , V. Pravilnikov , B.Beckner , A. Usadi , I. Mishev , S. Maliassov Neurok TechSoft LLC, Troitsk, Moscow reg., Russia ExxonMobil Upstream Research Company, Houston, TX, USA diyankov@aconts.com The presentation describes a multi-level approach on the construction of eective partitioning of unstructured graphs used in parallel computations. The quality of partitioning is estimated by its using in parallel iterative solution of large sparse linear systems arising in discretization of PDE on unstructured grids. Various algorithms of coarsening, and renement, and balancing under certain constrains are considered and numerically compared on a set of test matrices.

Analytic method for computation of magnetic eld in complex domain A. Bogatyrv e Institute of Numerical Mathematics of RAS gourmet@inm.ras.ru The minimization of hard magnetic discs requires the highest density of recorded data. This in turn brings about the necessity of high accuracy (7-8 true digits) computation of magnetic eld. The problem in 2D domain bounded by the soft underlayer, two shields and the reading magnetoresistive head is solved analytically in terms of elliptic theta functions. The proposed method allows to solve boundary value problems for harmonic functions with extreme precision in other domains as well.

An approach to solution of nonlinear Helmholtz eigenvalue problems stemming from integrated optics with Jacobi-Davidson solvers 7

Mikhail Botchev University of Twente, the Netherlands m.a.botchev@math.utwente.nl We consider nonlinear, nonpolynomial eigenvalue problems stemming from nite element discretization of the Helmholtz equation and arising in simulation of integrated optical devices. The Helmholtz equation is originally posed in an innite domain but solved numerically in a bounded domain with articial boundary conditions. These so-called nonreecting or transparent boundary conditions should guarantee transparency of the domain boundary for outgoing waves. The nonreecting boundary conditions used in this work are the recently developed transparent-inux boundary conditions [Nicolau & Van Groesen, 2005]. These conditions are obtained by solving the problem in the exterior of the computational domain analytically and have a number of advantages as compared to other known formulations of transparent boundary conditions. A nite element discretization of the Helmholtz equation leads to an eigenvalue problem where the transparent-inux boundary conditions cause a nonlinear, nonpolynomial dependence of the matrix on the eigenvalue. Moreover, it is not trivial to express this nonlinearity in an explicit way. However, since the nonlinearity results from the boundary conditions, the nonlinear contributions to the matrix of the eigenvalue problem can be seen a smaller dimensional discrete operator. This allows for a relatively cheap low-rank parametrization of the nonlinear dependence so that it can be approximated by a low-degree matrix polynomial. Thus, we reduce the nonlinear nonpolynomial eigenvalue problem to a nonlinear polynomial one. Once this reduction is done, the Jacobi-Davidson method can readily applied. Depending on the accuracy requirements of the eigenvalue problem, the polynomial approximation can be rened during the Jacobi-Davidson iterations. This research is based on a joint work in progress with Ardhasena Sopaheluwakan, Gerard Sleijpen, Brenny van Groesen, Jaap van der Vegt and Manfred Hammer.

Automorphisms of adjoint Chevalley groups over local rings E. I. Bunina Department of Mechanics and Mathematics, Moscow State University, 119992, Moscow, Russia helenbunina@yandex.ru Suppose that Gad is a Chevalley-Demazure groups scheme associated with irreducible root system ; Gad (R, ) is the group of points Gad with values in R; Ead (R, ) is an elementary subgroup in Gad (R, ), where R is a local commutative ring with 1. We describe automorphisms of the groups Gad (R, ) and Ead (R, ) in the case when R is a local ring with 1/2, with such an invertible element , that 2 1 is also invertible, and if the root system has the type G2 , then 1/3 R, and we suppose that the root system has the rank > 1. Then every automorphism of Gad (R, ) (Ead (R, )) is standard (see below). These results for Chevalley groups over elds were obtained by R. Steinberg [1] for the nite case and by J. Humphreys [2] for the innite one. E. Abe [3] proved this result for Noetherian rings, but the class of local rings is not contained completely in the class of Noetherian rings, and proofs of the paper [3] can not be extended for arbitrary local rings. We use some methods from [4]. Let us dene 4 types of automorphisms of a Chevalley group G (R, ), that are called standard. Suppose that CG (R) is the center of G (R, ) and : G (R, ) CG (R) is a group homomorphism. 8

Then the mapping x (x)x from G (R, ) onto itself is an automorphism of G (R, ), that is called a central automorphism of G (R, ). Let : R R be an automorphism of the ring R. The mapping x x from G (R, ) onto itself is an automorphism of G (R, ), that is called a ring automorphism of G (R, ). Let V be the representation space of the group G (R, ), and suppose that g GL(V ) is a matrix such that gG (R, )g1 = G (R, ). Then the mapping x gxg1 from G (R, ) onto itself is an automorphism of G(R), that is called the inner automorphism of G(R), induced by the element g from GL(V ). Suppose that is a basis of (that is xed), and is an automorphism of such that = . Then there exists a unique automorphism G (R, ) such that for every and t R the element x (t) is mapped to x() (()t), where () = 1 for all and () = 1 for all . It is called a diagram automorphism of G (R, ). Similarly we can dene 4 types of automorphisms of the elementary subgroup E(R). An automorphism of the group G (R, ) (or E (R, )) is called standard, if it is a composition of automorphisms of these 4 types. Theorem. Suppose that Ead (R, ) is an elementary Chevalley group with irreducible root system of rank > 1, R is a commutative local ring with 1/2. Suppose that R contains an invertible element such that 2 1 is invertible in R, and if the root system has the type G2 , then 1/3 R. Then every automorphism of Ead (R, ) is standard. There is the similar theorem for Chevalley groups Gad (R, ). References 1. Steinberg R., Automorphisms of nite linear groups, Canad. J. Math., 121, 1960, 606615. 2. Humphreys J. F., On the automorphisms of innite Chevalley groups, Canad. J. Math., 21, 1969, 908-911. 3. Abe E. Automorphisms of Chevalley groups over commutative rings, Algebra and Analysis, 5(2), 1993, 7490. 4. Petechuk V. M. Automorphisms of groups SLn , GLn over some local rings, Mathematical Notes, 28(2), 1980, 187206 (in Russian).

Mirsky-based tools for the eigenvalue distribution in a non-Hermitian setting Stefano Serra-Capizzano Universit` degli Studi dellInsubria, Via Ravasi, 2, 21100 Varese, Italy a The work of the author is supported in part by MIUR grant no. 2006017542. s.serracapizzano@uninsubria.it Under mild trace norm assumptions on the perturbing sequence, we extend a recent Mirski-based perturbation result concerning the eigenvalues of a generic (non Hermitian) complex perturbation of a bounded Hermitian sequence of matrices. Some examples of application are considered, ranging from the algebra generated by Toeplitz sequences to the approximation of PDEs with various boundary conditions. A nal discussion on open questions and further lines of research ends the note. References 1. R. Bhatia, Matrix Analysis, Springer Verlag, New York, 1997. 2. R. Bhatia, Pinching, trimming, truncating, and averaging of matrices, Amer. Math. Monthly, 107 (2000), 602608. 9

3. D. Fasino and S. Serra-Capizzano, From Toeplitz matrix sequences to zero distribution of orthogonal polynomials, Contemp. Math. 323 (2003), 329340. 4. L. Golinskii and S. Serra-Capizzano, The asymptotic properties of the spectrum of non symmetrically perturbed Jacobi matrix sequences, J. Approx. Theory, 144-1 (2007), pp. 84102. 5. S. Holmgren, S. Serra-Capizzano, and P. Sundqvist, Can one hear the composition of a drum?, Mediterr. J. Math., 3-2 (2006), pp. 227249. 6. A. B. J. Kuijlaars and S. Serra-Capizzano, Asymptotic zero distribution of orthogonal polynomials with discontinuously varying recurrence coecients, J. Approx. Theory 113 (2001), 142155. 7. S. Serra-Capizzano, Generalized Locally Toeplitz sequences: spectral analysis and applications to discretized Partial Dierential Equations, Linear Algebra Appl. 366-1 (2003), 371402. 8. S. Serra-Capizzano, The GLT class as a Generalized Fourier Analysis and applications, Linear Algebra Appl., 419-1 (2006), pp. 180233. 9. S. Serra-Capizzano, D. Bertaccini and G. H. Golub, How to deduce a proper eigenvalue cluster from a proper singular value cluster in the non normal case, SIAM J. Matrix Anal. Appl., 27-1 (2005), pp. 8286. 10. S. Serra-Capizzano and P. Sundqvist, Stability of the notion of approximating class of sequences and applications, J. Comput. Appl. Math., in press. 11. P. Tilli, A note on the spectral distribution of Toeplitz matrices, Linear Multilin. Algebra, 45 (1998), 147159. 12. P. Tilli, Some results on complex Toeplitz eigenvalues, J. Math. Anal. Appl. 239-2 (1999), 390401. 13. E. Tyrtyshnikov and N. Zamarashkin, Spectra of multilevel Toeplitz matrices: advanced theory via simple matrix relationships, Linear Algebra Appl., 270 (1998), pp. 1527.

p-Sparse BEM for weakly singular integral equation on random boundary Alexey Chernov and Christoph Schwab Seminar for Applied Mathematics, ETH Z rich, 8092 Z rich, Switzerland u u achernov@math.ethz.ch, schwab@math.ethz.ch We consider the weakly singular boundary integral equation Vu() = g on a randomly perturbed smooth closed surface () (cf. [2] for stochastic g). The aim is the computation of the moments Mk u := E[k u], k 1, if the corresponding moments of the perturbation are known. The problem i=1 on the stochastic surface is reduced to a problem on the nominal deterministic surface with the random perturbation parameter (). Note, that u() depends nonlinearly on (). Resulting formulation for the kth moment is posed in the tensor product Sobolev spaces and involve the k-fold tensor product operators V (k) = k V. The complexity of the standard full i=1 tensor product Galerkin BEM is O(N k ), where N is the number of degrees of freedom needed to discretize the nominal surface . Based on [3], we develop the p-sparse grid Galerkin BEM to reduce the problem complexity to O(N (log N )k1 ). References

10

1. H. Harbrecht, R. Schneider, Ch. Schwab, Sparse second moment analysis for elliptic problems in stochastic domains, in press. 2. T. von Petersdorff, Ch. Schwab, Sparse nite element methods for operator equations with stochastic data, Appl. Math., 51 (2006), no. 2, 145180. 3. V. N. Temlyakov, Approximation of periodic functions, Nova Science Publ., New York, 1994.

Continuation method to solve symmetric indenite Toeplitz systems Andrey A. Chesnokov , Marc Van Barel Comp. Math. and Cyb., Moscow State University; Dept. Comp. Sci., Katholieke Universiteit Leuven {andrey.chesnokov, marc.vanbarel}@cs.kuleuven.be

Dept.

The method being presented starts with the identity matrix and gradually transforms it to the inverse of an indenite Toeplitz matrix. If the Toeplitz matrix is ill-conditioned, the method performs several Newton iterations at the very end. Overall memory requirement is of the order O(n). The computational complexity per iteration step is of the order O(n log n). The validity of the approach is illustrated by numerical experiments.

On a general approach to deriving the known classes of normal Hankel matrices Chugunov V. N. , Ikramov Kh. D. of Numerical Mathematics of RAS Moscow State University, Mechanics and Mathematics Dept. vadim@bach.inm.ras.ru, ikramov@cs.msu.su
Institute

The problem of characterizing normal Hankel matrices is still far from being completely solved. The available partial results are mainly the descriptions of specic subsets of this matrix class. These subsets were found by considerations that are specic for each subset. In this talk, we present a general approach that allows us to obtain all the known subsets as special cases of a unied scheme. Normal Hankel matrices outside of these subsets correspond to the main (and the most dicult) case of our scheme. This case, which at the moment does not yield to a complete analysis, gives origin to a new and very interesting matrix class that extends (and contains) the class of -circulants.

Decompositions of a Higher-Order Tensor in Block Terms 11

Lieven De Lathauwer Centre National de la Recherche Scientique Lab. ETIS (ENSEA, UCP, CNRS UMR 8051) 6, avenue du Ponceau, BP 44, F 95014, Cergy-Pontoise Cedex, France delathau@ensea.fr In this talk we introduce a new class of tensor decompositions. Intuitively, we decompose a given tensor block into blocks of smaller size, where the size is characterized by a set of mode-n ranks. We study dierent types of such decompositions. For each type we present conditions under which essential uniqueness is guaranteed. The Parallel Factor decomposition and Tuckers decomposition can be considered as special cases in the new framework. The results shed new light on fundamental aspects of tensor algebra.

Development of parallel iterative algorithms for the solution of 2D elasto-plastic problems Elena N. Akimova , Irina P. Demeshko , Anatolii V. Konovalov of Mathematics and Mechanics of UrB RAS, S. Kovalevskaya str.,16, Ekaterinburg, Institute of Engineering Science of UrB RAS, Komsomolskaya str.,34, Ekaterinburg aen@imm.uran.ru, avk@imach.uran.ru

Institute

In this work we investigate the applicability of parallel computations for the solution of elastoplastic problems with small elastic and large plastic strains using nite element method. As an example, we consider the problem of cylinder compression from elasto-plastic isotropic and isotropic-strengthened material by at plates. The solution is based on a principle of virtual capacity of the speed form ( + t) h dV +
V

(P + tP ) hd = 0.

(1)

Here is Cauchys strain tensor; P surface force density; t period for an increment interval of load; h and h kinematically admissible velocity elds variation and its inverted delta; V, solid volume and surface; dV, d volume and cylinder surface area elements. These dening equations for the strain tensor velocity are taken from the work [1]. On contact with plates the Coulombs law was applied. A lateral surface of the cylinder is free from loadings. By means of nite-element approximation the equation (1) is reduced to the linear algebraic equations system (LAES) Az = b, (2)

A, b, z initial matrix, right part vector and decision vector of the system, respectively. Here A is the asymmetrical banded matrix. Solving the problem (1) by means of Finite Element Analysis consist of 3 basic stages: making of the matrix A, solving of the LAES (2), calculating of deected mode at the end of loading iteration. Parallel algorithms for solving the rst and third stages were developed. This stages are easy for parallelization, but solving time of its are about half past of the problem time. The rest of time is spent for solving of the linear algebraic equations system. Time results of that algorithms showed that 12

with increasing number of processor a paralleling eectiveness is increase. However, with increasing matrix dimension a paralleling eectiveness reduces for every examined iterative method because of a send data time is increase. The comparison of computational eciency of parallel algorithms at the common memory multiprocessor computer with the time at the distributed memory multiprocessor computer are performed. The eectiveness of the solving at the common memory multiprocessor computer higher that at the distributed memory multiprocessor computer because of send data time higher at memory multiprocessor computer For the solving of the linear algebraic equations system a direct method and a iteration methods are reviewed. As a direct method the Gauss method was applied, but its parallelization [2] didnt show the reduction of the calculation time due to the matrix being banded. The parallel algorithms were developed for the following iterative methods [3]: the Prime Iteration method (MPI), the Minimal residual method (MMR), the Steepest descent method (MSD) and the Method of conjugate gradients (MCG). Parallelization algorithms is based on transformation of a banded matrix into a vertical strip and banding it on m blocks, and the A parent matrix, right part vector and decision vector of the system, accordingly vector of solving and the right part vector of the linear algebraic into m parts so, that n = m L, where n dimension of the equations system, m number of processors. Each processor calculates a part of the solving vector on each iteration. Numerical realization and parallelization of the iterative methods for solving the linear algebraic equations has been implemented on the Massively Parallel Computing System MVS-1000/32 with 32 processors with help of the MPI C++ program library. Solving results of the problem (2) by the parallel iterative algorithms for the banded matrix the width of which is 171 and the number of variables is 3362 are shown in the picture. It is evident that the Prime Iteration method, the Minimal residual method and the Steepest descent method show the same results. The method of conjugate gradients requires to do more iteration than the other examined methods and needs more time for solving the linear algebraic equations.

Fig 1. The LAES executing times by the parallel iterative algorithms With increasing matrix dimension a paralleling eectiveness rises for every examined iterative method. This happens because the increase of matrix dimensions leads to the increase of the laboriousness of each matrix parts, located on each processor. At the same time, data message level increases slightly. As a result, computation time: time of transmission increases rises, when matrix dimension increase. So, with the increasing matrix dimension we can obtain larger computational speedup at the larger processor number, if we use a parallel iterative algorithms. Also, comparison 13

of computational speedup and eciency of parallel algorithms and comparison the computing time with the computing time of the Gauss method were performed. The obtained results show, that the computing time of linear algebraic equations by the Gauss method on one processor is less than the computing time of this system by the iterative methods on one processor for moderate matrix. But, in comparison with the Gauss method, the iterative methods have higher eciency of parallelization. The use of parallel algorithms reduces the computing time, that allows to recommend them for the solving of greater elasto-plastic problems. References 1. Konovalov A.V. Constitutive equations for elasto-plastic medium with large plastic deformations, Proceedings of the RAS. Mechanics of rigid body. 1997. V. 5. P. 139149. 2. Korneev V.D. Parallel programming by MPI, Novosibirsk, The Russian Academy of Sciences. 2000. 213 p. 3. Fadeev V.K., Fadeeva V.N. Calculating methods of the linear algebra, Moscov: State Publishing House of the physicotechnical literature. 1963. 734 p.

Simple wavelets for irregular grid Yu. K. Demyanovich Saint-Petersburg State University, Russia Research is supported in part by RFFI grants 04-01-00269 and 04-01-00451 Yuri.Demjanovich@JD16531.spb.edu Excelent results of wavelet theory and their applications are obtained with powerful Fouriers apparatus (for example, see [1], [2]). The proposed work deals with an algebraic approach which leads to simple spline-wavelet decompositons. Let {X r }, r = 0, 1, 2, . . ., be a system of irregular grids in interval (, ) R1 , X r : . . . < x1 < x0 < x1 < . . . ;
(r) (r) (r) (r) lim x j j

= ,

(r) lim x j+ j

= ,

with embedding property X 0 X 1 X 2 . . .. We discuss a vector-function def t (, ), with measurable linear independent (t) = (0 (t), 1 (t), . . . , m (t))T , components k (t), k = 0, 1, . . . , m. A system of embedded (general speaking, non-polynomial) spline spaces S(Ar ,) S(Ar+1,) , r = 0, 1, 2, . . . , is constructed, and decomposition of the system into direct sum of wavelet spaces . S(Ar+1,) = S(Ar ,) + W r is realized. Sequential usage of the decomposition leads to represantation def of space S = + S(Ar ,) with direct sum, r=0 . . . . S = S(A0 ,) + W 0 + W 1 + W 2 + . . . . (1)

As a result, simple formulars of decomposition and reconstruction are obtained. The space S(A,) def def is dened with vector-function and with complete chain A = {mk }kZ of points mk = (xk , pk ) of def direct product M = (, ) P m of the interval (, ) and m-dimensional projective space P m . An involution operator (here it is signed by symbol ), which gives local orthogonal chain, is introduced 14

in the set of complete chains; then embedding of mentioned spaces can be written in the form S(A,) S(B ,) , where B is a complete extension of the chain A (notice that set of complete extensions has the power of the continuum). The wavelet decomposition (1) is derived with caliber relations (generalizing the multi-scaled equation) and with a functional system which is biorthogonal to the system of basic functions in the space S(Ar ,) for each r = 0, 1, 2, . . .. Basic wavelets of the proposed decompositions possess compact supports (of length of m + 1 grid intervals), and they have simple analytical expression. Under dierent assumptions as to vector-function (t) the necessary s and sucient conditions of belonging of the wavelets to spaces Wq and C s are formulated. The order of approximation with discussed wavelet spaces are asymptotically optimal with respect to N -width of standard compact sets. If the grid X and the vector-function C m1 (, ) are xed then the space S(A,) which belongs to C m1 (, ) is unique among discussed spline spaces; this space is called the space of B -splines of order m (see [3]). Some examples are discussed. 1. 2. 3. Ross. References Charles K. Chui, An Introduction to Wavelets, Academic Press, 1992. S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, 1999. Yu. K. Demyanovich, Smoothness of spline spaces and wavelet decompositions, Dokl. Akad. Nauk 401 (2005), no. 4, pp. 439442.

Solving Systems with Hierarchical Structure Patrick Dewilde TU Delft, the Netherlands p.dewilde@ewi.tudelft.nl The holy grail of solving large systems of equations is to do it (in a stable way) with a complexity equal to the number of independent parameters determining the system (the algebraic degree of freedom). Although in most cases it appears to be next to impossible to achieve this feat, we do dispose of techniques that are capable of handling some systems with structure eciently, such as some types of sparse systems, semi-separable systems, Toeplitz or Hankel systems and even combinations of those. A new challenge is to handle systems with hierarchical structure. We shall concentrate on systems consisting of a tree of subsystems which again may have structure at a lower level of the hierarchy. We shall connect with some classical results on the Toeplitz block Toeplitz case, consider the so called HSS (Hierarchically Semi Separable) systems pioneered by Chandrasekharan and Gu, look at the invariant multi-resolution case of Alpay and Volok, and derive a methodology and conditions under which solutions can generally be constructed.

Fuzzy Grid Method for Numerical Solution of PDE Oleg V. Diyankov Neurok Techsoft LLC, oce 120, 19a Oktyabrskiy ave., Troitsk, Moscow region, Russia, 142190 15

diyankov@aconts.com The Fuzzy Grid Method (FGM) for PDE solution is presented. The method is a natural generalization of meshless methods, based on Free-points method and SPH methods ideas. The method is illustrated with numerical samples of 2D Gas Dynamics equations and Poisson equation solution. There are dierent approaches to construction of numerical methods for PDE solution. The most popular are: Finite Dierence (FD), Finite Element (FE), Particle in Cell (PIC), Free Lagrange (FL). All of these methods have their own advantages and drawbacks. The paper is devoted to the introduction of new approach, developing the two mentioned above meshless approaches (PIC and FL), which inherits the advantages of meshless approaches allowing modeling of large deformations and overcomes drawbacks, rst of all linked with nonmonotonicity of the received numerical solutions. The main idea of the method is based on the approximation of the surface and volume integrals of the unknown functions in a particle as sums of these functions values in the points (particles), which are in the nearest vicinity to the given one.
V V

f dV |i f dS|i

ji

Ai,j fj Bi,j fi,j

(1)

ji

Here i is a set of numbers of neighbor particles for the particle number i, fi,j approximation of function f at the fuzzy face between i-th and j-th particles. The coecients in (1) should satisfy some relations. So, for each particle we have something like a cloud of particles around it, which is used for the approximation construction. The name particle is used as a close to reality one, but really we have a volume, which doesnt have strict bounds, but about which we know values, which are necessary for modeling of a specic applied problem. That is why, this method was called Fuzzy Grid.

Monotone transformations with respect to

cn

and

-partial orders

Mikhail A. Emov [ poster presentation ] Lomonosov Moscow State University mix2904@rambler.ru Let Mn (F) be an algebra of n n-matrices with entries from an arbitrary eld F. A generalized inverse matrix A for a xed matrix A Mn (F) is dened to be any solution of the matrix equation AA A = A. A generalized inverse matrix A for the matrix A, which in addition satises r the condition A AA = A , is called a reexive generalized inverse. A group generalized inverse r r r matrix A (or simply group inverse ) for the matrix A is dened to be a reexive generalized inverse matrix which commutes with the matrix A. A matrix A is said to be of index 1 if rk A2 = rk A. In this talk we deal with partial orders which can be dened in terms of group inverse. Their studying is important today in connection with dierent aspects of general algebra and mathematical statistics. Denition 1. [4] Let A Mn (F) be a matrix of index 1 and B Mn (F) be an arbitrary matrix. We say that A

B i AA = BA = A B.

16

We note, that is rather poor, namely, all matrices of index bigger than one are maximal. R. Hartwig and S.-K. Mitra in [3] introduced the order relation on the whole matrix algebra Mn (F) which extends -order. This relation is called -order. To dene this order we need to introduce the following decomposition for an arbitrary matrix A Mn (F). Let F be an arbitrary eld. The corenilpotent decomposition of a square matrix A Mn (F ) is the following decomposition: A = CA +NA , where NA is nilpotent matrix and CA is a matrix of index 1, moreover CA NA = NA CA = 0. This decomposition is unique and exists for any square matrix A, see [1, Chapter 4.8]. The other notion, we need, is the following well-known notion of minus-order : A B i rk (B A) = rk B rk A. Denition 2. [3] Let F be an arbitrary eld. For a pair of matrices A, B Mn (F) we have: A
cn cn

B, i

CA NA

CB NB
cn

We investigate transformations on matrix algebra, which preserve and -partial orders. In [2] linear bijective transformations preserving each of these orders are characterized. Here we remove the assumptions of bijectivity and show that linearity can be substituted by additivity. Namely, we prove the following: Theorem 1. Let F be a eld with |F | 3, n be a positive integer, n
cn

2. Let, moreover, additive

mapping T : Mn (F) Mn (F) is monotone with respect to or -partial orders. Then there exist F, P GLn (F) and Aut F, such that T has the form T (X) = P 1 (X)P for all X Mn (F) or T (X) = P 1 (X)t P for all X Mn (F), where X t denotes transposed matrix. I would like to thank my scientic advisor Dr. Alexander E. Guterman for posing this problem and fruitful discussions. References 1. Ben-Israel A., Greville T., Generalized Inverses: Theory and Applications, New York, John Wiley and Sons, 1974. 2. Bogdanov I. I., Guterman A. E. Monotone matrix transformations dened by the group inverse and simultaneous diagonalizability, Matematicheskii Sbornik, 198(1), 2007, 320. 3. Hartwig R. E., Mitra S. K. Partial orders based on outer inverses, Linear Algebra Appl., 176 (1982), 320. 4. Mitra S. K. On group inverses and the sharp order, Linear Algebra Appl. 92 (1987), 1737.

Ecient generators of quasiseparable matrices Yuli Eidelman and Israel Gohberg Tel Aviv University, Israel eideyu@post.tau.ac.il

17

Here pi , qj , ak are matrices of sizes mi ri1 , rj mj , rk rk1 respectively; these elements are . The elements g , h , b are matrices of said to be lower generators of the matrix R with orders rk i j k sizes mi ri , rj1 mj , rk1 rk respectively; these elements are said to be upper generators of the matrix R with orders rk . The diagonal entries dk are matrices of sizes mk mk . Let m = max1kN mk be the maximal size of blocks of R. We assume that m is a small number. Let r be the maximal order of generators of the matrix R:

We study a class of block structured matrices R follows: pi ai1 aj+1 qj , Rij = d, i gi bi+1 bj1 hj ,

= {Rij }N i,j=1 whose entries are specied as 1 j < i N, 1 i = j N, 1 i < j N.

r = max{ max

1kN 1

rk ,

1kN 1

max

rk }.

In terms of lower and upper generators and diagonal entries of a quasiseparable matrix one can obtain O(r 2 N ) algorithms to compute the product y = Rx and O(r 3 N ) algorithms to solve the system of linear algebraic equations Rx = y. In this talk we discuss a special type of generators for which the multiplication of a quasiseparable matrix by a vector and solution of a quasiseparable system may be done in O(rN ) and O(r 2 N ) operations respectively. We show that for any matrix such ecient generators exist. Various algorithms to compute these generators are presented.

A use of matrix methods for solution and stability analysis in convective problems Michael Ermakov Institute for Problems in Mechanics RAS, Moscow, Russia ermakov@ipmnet.ru An experience of a use of matrix methods for a solution of axisymmetric convective problems on the basis of Navier-Stokes equations for a viscous heat-conductive non-compressible uid is described. The governing equations consist of continuity, momentum and energy conservation equations. Approximation of governing equations in natural variables (velocity, pressure) is eected by nite volume method on staggered mesh. The boundary conditions and relations for ctious points are included in the joint system. The system of non-linear equations is solved by Newton-Raphson method. The solution of sparse linear system for Jacobi matrix is eected by stabilized bi-conjugate gradient (BiCGStab) method with ILU-preconditionning [1]. An advantage of the generalized residual method (GMRes) against BiCGStab is not recovered in our problems. The dimension of solved problems are up to 200000 and the calculation time is about 10 minutes. The eective solution of higher dimension problem is restricted by huge sizes of prenditionners. For a linear stability analysis a representation of perturbations at the same discrete mesh as a basic solution is used, in an azimuthal direction the solution is presented as a superposition of integer-valued normal modes. The solution of generalized eigen-value problem is eected by inverse iteration method. For solution of a linear system the version of BiCGStab and ILU factorization for sparse complex matrices. The time of obtaining of a pair eigen vector eigen value for 200000 dimension vector is about 20 miutes. 18

The advantages of matrix methods are in a relative simplicity of a numerical code development and a high numerical stability of created algorithms. A use of linear stability analysis allows to accelerate essentially, in compare to direct numerical modeling , a search of critical conditions. The proposed procedure of a matrix solution of Navier-Stokes equations and a linear stability analysis are carefully tested [2-4]. The applications of this research are a simplied models of crystal growth: half-zone model and Czochraski model [2-4]. An eciency of existing ILU-preconditionners, a development of optimal preconditionners and a perspectives of 3-D uid modeling are also discussed in the paper. References 1. Y. Saad, Iterative methods for sparse linear systems, 2nd edition, SIAM, 2003. 2. M. Ermakov, Linear Stability Analysis of Axisymmetric Convective Flows, AIAA 2006-3793. 3. M.K. Ermakov, M.S. Ermakova, Linear-stability analysis of thermocapillary convection in liquid bridges with highly deformed free surface, J. Crystal Growth 266 (2004) 160166. 4. V. Shevtsova, Thermal convection in liquid bridge with curved free surfaces: benchmark on numerical solutions, J. Crystal Growth 280 (2005) 632651. 5. W. Shi, M.K. Ermakov, N. Imaishi, Eect of pool rotation on thermocapillary convection in shallow annular pool of silicone oil, J. Crystal Growth 294 (2006) 474485.

Superoptimal approximation for ill-conditioned linear systems with unbounded symbols C. Estatico and S. Serra Capizzano Dipartimento di Matematica e Informatica, Via Ospedale 72, 09124 Cagliari (ITALY), Dipartimento di Fisica e Matematica, Via Valleggio 11, 22100 Como (ITALY) estatico@dima.unige.it, stefano.serrac@uninsubria.it, serra@mail.dm.unipi.it In this talk, we discuss some properties of the superoptimal Frobenius approximation for the solution of ill-conditioned linear systems with Toeplitz matrices generated by unbounded symbols. In particular, we use the superoptimal approximation as preconditioner for the CG method when a Fisher-Hartwig singularity is present in the symbol, with special regard to systems coming from times series and nancial applications. The idea behind the superoptimal approximation is to approximate the system matrix by considering an optimization procedure concerning a kind of relative error in the matrix sense, instead of the absolute error considered in the classical optimal preconditioning. On these grounds, we compare the very dierent behavior between (i) the relative approximation of the generating function and (ii) the relative approximation of the generated Toeplitz matrix. A theoretical discussion concerning classical circulant preconditioners and a numerical comparison with the Strang and with the optimal approximations are also presented particularly with reference to the presence of noise. References 1. F. Di Benedetto, C. Estatico, S. Serra-Capizzano, Superoptimal preconditioned conjugate gradient iteration for image deblurring, SIAM J. Sci. Comput., 26-3 (2005), pp. 10121035.

19

2. F. Di Benedetto, S. Serra-Capizzano, A note on the superoptimal matrix algebra operators, Linear Multilin. Algebra, 50 (2002), pp. 343372. 3. C. Hurvich, Y. Lu, On the complexity of the Preconditioned Conjugate Gradient algorithm for solving Toeplitz systems with a Fisher-Hartwig singularity, SIAM J. Matrix Anal. Appl., 27 (2005), pp. 638653. 4. E. Tyrtyshnikov, Optimal and superoptimal circulant preconditioners, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 459473.

Tensor product approximation in quantum chemistry Heinz-J rgen Flad u Institut fr Informatik, Christian-Albrechts-Universitt zu Kiel u a flad@mis.mpg.de Tensor products provide a versatile tool to approximate various quantities encountered in electronic structure calculations. The talk provides an overview on some recent results concerning the best N -term approximation of these quantities in anisotropic wavelet tensor product bases. Furthermore, so-called best separable approximations have been studied which do not rely on a xed multivariate basis anymore. In combination with stable quadrature schemes for the Coulomb interaction, these tensor products enable an ecient evaluation of two-electron integrals. It will be shown that substantial further compressions with respect to the tensor product rank can be achieved compared to traditional Gaussian-type basis sets.

Non-linear commutativity preserving maps on symmetric matrices Ajda Foner s University of Maribor, Faculty of Natural Sciences and Mathematics, Koroka cesta 160, s 2000 Maribor, Slovenia ajda.fosner@uni-mb.si Let Sn (C) be the set of all symmetric matrices in Mn (C), n 3. We will study bijective maps on the set Sn (C) that preserve commutativity in both directions. We will show that if : Sn (C) Sn (C), n 3, is a continuous bijective map preserving commutativity in both directions, then there exist an orthogonal matrix Q Mn (C) and for every A Sn (C) a complex polynomial pA such that either (A) = QpA (A)Qt , A Sn (C), or (A) = QpA (A)Qt , A Sn (C). We will also study commutativity preserving maps on Sn (C) without the continuity assumption.

20

On superderivations and local superderivations Ajda Foner and Maja Foner s s University of Maribor, Faculty of Natural Sciences and Mathematics, Koroka cesta 160, s 2000 Maribor, Slovenia ajda.fosner@uni-mb.si We describe superderivations in certain superalgebras by their actions on elements satisfying some special conditions. One of the main results is applied to local superderivations on some certain superalgebras.

Computing Zeros of Polynomials Walter Gander ETH Zrich u gander@inf.ethz.ch Jim Wilkinson discovered that the computation of zeros of polynomials is ill conditioned when the polynomial is given by its coecients. We review and demonstrate this fact. For many problems we need to compute zeros of polynomials, but we show that we do not necessarily need the coecients of the polynomials to solve the problem. We develop algorithms that avoid the coecients. They turn out to be stable, however, the drawback is often heavily increased computational eort. On the other hand modern processors are mostly idle and wait for crunching numbers so it may pay to accept more computations in order to increase stability. We show and demonstrate examples for the quadratic eigenvalue problem and for orthogonal polynomials.

Linear projectors on tropical spaces G. Cohen , S. Gaubert , J.-P. Quadrat 6-8, avenue Blaise Pascal Cit Descartes, Champs-sur-Marne 77455 Marne-La-Valle e e Cedex 2, France INRIA-Rocquencourt, B.P. 105, 78153 Le Chesnay Cedex, France Guy.Cohen@enpc.fr, Stephane.Gaubert@inria.fr, Jean-Pierre.Quadrat@inria.fr

ENPC

The max-plus semiring consists of the set R{} equipped with the addition (a, b) max(a, b) and the multiplication (a, b) a + b. The term tropical refers to such algebraic structures, in which the addition is idempotent. Many classical notions have tropical analogues. In particular, tropical linear spaces can be dened as modules, by replacing the ring of scalars by the max-plus semiring [6,2]. Such spaces have some features in common with classical convex polyhedra, see [4,5]. In this talk, we shall review the properties of projections on tropical spaces. Since these spaces are naturally ordered, we may dene the projection of a vector onto a tropical space to be the greatest vector of this tropical space that the former vector dominates. Such 21

projectors appear to be reminiscent of orthogonal projectors: they minimize a metric -the Hilberts projective metric-, and they can be used to prove separation theorems [2]. However, they turn out to be non-linear. Hence, a natural question is to characterize those tropical linear spaces which are the images of linear projectors. We shall consider here specially the case of nitely generated subspaces of the nfold Cartesian product of the max-plus semiring. Such spaces can be identied to the row or column spaces of tropical matrices. It is known [3] that the row (or column) space of a tropical matrix A is the image of a linear projector if and only if A is regular in the sense of von Neumann, meaning that there exists a matrix X (a generalized inverse) such that A = AXA. This is equivalent to the condition that the row (or column) space of A is a projective space. The main result of this talk is the following characterization, which extends a theorem proved by Zaretski [7] in the case of Boolean matrices: a tropical matrix is von Neumann regular if and only if its row (or column) space is a completely distributive lattice. When A has full column (or row) rank, the latter condition is equivalent to requiring the row (or column) space of A to be stable by the inmum operation. Moreover, there is always a linear projector on the complete lattice closure of the column (or row) space of A. We shall nally indicate an application of these ideas. Projectors on tropical spaces have been used in [1] to design an analogue of the nite element method, adapted to rst order HamiltonJacobi equations in which the Hamilonian is convex in the adjoint variable. Such equations govern the evolution of the value function of deterministic optimal control problems. In the classical nite element method, the solution is approximated by its projection on a space generated by nite elements. Essentially the same is true in the tropical case. References 1. M. Akian, S. Gaubert, and A. Lakhoua, The max-plus nite element method for solving deterministic optimal control problems: basic properties and convergence analysis, SIAM J. Control and Opt., arXiv:math.OC/0603619, to appear. 2. G. Cohen, S. Gaubert, and J.-P. Quadrat, Duality and separation theorems in idempotent semimodules, Linear Algebra and Appl., 379: 395422, 2004, arXiv:math.FA/0212294. 3. G. Cohen, S. Gaubert, and J.P. Quadrat, Linear projectors in the max-plus algebra, In Proceedings of the IEEE Mediterranean Conference, Cyprus, 1997. IEEE. 4. M. Develin and B. Sturmfels, Tropical convexity, Doc. Math., 9: 127, 2004. 5. S. Gaubert and R. Katz, Max-plus convex geometry, In R. A. Schmidt, editor, Proceedings of the 9th International Conference on Relational Methods in Computer Science and 4th International Workshop on Applications of Kleene Algebra (RelMiCS/AKA 2006), volume 4136 of Lecture Notes in Comput. Sci., pages 192206. Springer, 2006. 6. G.L. Litvinov, V.P. Maslov, and G.B. Shpiz, Idempotent functional analysis: an algebraic approach, Math. Notes, 69(5): 696729, 2001. 7. K.A. Zaretski, Regular elements in the semigroup of binary relations, Uspekhi Mat. Nauk, 17(3): 105108, 1962.

Cyclic projectors in semimodules over the max-plus semiring Stphane Gaubert , Serge Sergeev e 22

Department

B.P. 105, 78105 Le Chesnay Cdex, France e of Physics, Sub-Department of Quantum Statistics and Field Theory, Moscow State University, Moscow, 119992 Leninskie Gory, Russia The research of the rst author was supported by the joint RFBR/CNRS grant 05-01-02807. The research of the second author was supported by the RFBR grant 05-01-00824 and the joint RFBR/CNRS grant 05-01-02807. Stephane.Gaubert@inria.fr, sergiej@gmail.com

INRIA-Rocquencourt,

We study the properties of cyclic projectors in semimodules over the max-plus or tropical semiring Rmax = (R {}, = max, = +). A cyclic projector is an operator of the form Pk . . . P1 , where for i = 1, . . . , k, Pi is the projector onto a closed subsemimodule Vi Rn , the action of max which is dened by Pi (y) = max{x Vi : x y}. The study of these cyclic projectors is motivated, on the one hand, by the observation, due to Cuninghame-Green and Butkovi [3], that they give rise to an eective pseudo-polynomial algorithm c for solving two-sided systems of equations of the form A x = B y over the max-plus semiring. On the other hand, projectors Pi have been used to establish separation theorems in idempotent convex geometry, see works by Cohen, Quadrat, Singer, and the rst author [1], [2], and by Litvinov, Maslov, and Shpiz [4]. First we show that the orbit of a cyclic projector maximizes a certain objective function. We call this function Hilberts value of semimodules, as it is a natural extension of Hilberts projective metric. Then we characterize the (non-linear) spectrum of cyclic projectors in terms of Hilberts values. The observation that cyclic projectors are closely related to idempotent separation theorems leads us to a new separation result: that several closed semimodules with a trivial intersection can be separated from each other. This means that for each of these semimodules, we can construct an idempotent halfspace containing it, in such a way that the intersection of these halfspaces is trivial. As a corollary, we get an idempotent analogue of Hellys theorem (another proof of which has been obtained by Meunier and the rst author in an independent work). Some of our results still hold if the max-plus semiring is replaced by a conditionnally complete, or b-complete [4], idempotent semield. References 1. G. Cohen, S. Gaubert, and J.P. Quadrat, Duality and separation theorems in idempotent semimodules, Linear Algebra Appl., 379: 395422, 2004. E-print arXiv:math.FA/0212294. 2. G. Cohen, S. Gaubert, J.P. Quadrat, and I. Singer, Max-plus convex sets and functions, In G. Litvinov and V. Maslov, editors. Idempotent Mathematics and Mathematical Physics, volume 377 of Contemporary Mathematics , pages 105129. AMS, Providence, 2005. E-print arXiv:math.FA/0308166. 3. R. A. Cuninghame-Green and P. Butkovic, The equation A x = B y over (max,+), Theoretical Computer Science, 293:312, 2003. 4. G.L. Litvinov, V.P. Maslov, and G.B. Shpiz, Idempotent functional analysis. An algebraical approach, Math. Notes, 69(5):696729, 2001. E-print arXiv:math.FA/0009128.

The direct projection method and its modication for solving systems of linear equations with sparse unstructured matrices 23

Gogoleva S. Y. The Samara State Aerospace University named after academician S.P. Korolev, Applied mathematics department, Russia, 443086, Samara, Moscow Shosse, 34. Phone (846) 332-56-07 gogoleva s@mail.ru The mathematical models of many practical problems result in systems of linear algebraic equations (SLAE) with large and sparse matrices of coecients. When the large part of matrix coecients consists of zero, it is quite obvious, that we try to store only nonzero elements. The serious problem at a storage and processing of sparse matrices is represented by ll-in, i.e. occurrence of new nonzero elements. Reduction of ll-in accompany reduction of the requirements to memory volume and work acceleration of method. In the given work the direct projection method (DPM) of solving SLAE is considered [1]. In this method if to carry out analogy to methods based on decomposition of matrices (LU-decomposition), the matrix of equations system A is factored into two matrices LA and R. [1] For realization DPM probably to store in operative memory only matrix R, that requires accordingly n(n + 1)/2 machine words. Thus, DPM gives a prize in required operative memory volume twice, in comparison with the computation scheme based on LU - decomposition, that it is favourable to use at the solving of large sparse SLAE. The total number of arithmetic operations required for DPM is estimated by size 2n3 /3 + O(n2 ). It almost as much, how many in a method based on LU - decomposition. At the solving SLAE with sparse matrices by DPM the ll-in occurs only in a matrix R, whereas in LU - decomposition the ll-in can occur, both in a matrix L, and in a matrix U . In a case for SLAE with the sparse unstructured matrices the numerical researches of DPM were carried out on examples considered in [2]. The results of numerical researches have shown, that at the solving of SLAE with the given matrices DPM surpasses LU - decomposition in arithmetic operations number and operative memory. To increase numerical stability DPM apply pivot strategy. At work with sparse matrices, using a choice of a pivot , it is necessary to keep matrix sparsity. If to consider LU decomposition, in this case use the Markowitz strategy. The Markowitz strategy has two lacks. It is necessary to look through all elements active submatrix, we can encounter numerical instability. The second lack is connected to that circumstance, that, if as pivot the very small elements will be chosen, in result we shall receive the large errors in a vector of the decisions. Oered DPM with a pivoting strategy allows to nd the compromise between matrix sparsity preservation and stability. References 1. Zhdanov A.I., A direct consecutive method of the solving systems of linear algebraic equations, Report RAN, 1997. Vol. 356. No 4. P. 442-444. 2. sterby O., Zlatev Z., Direct methods for sparse matrices, Springer Verlag, 1983. 127pp.

Matrix Computations and the Secular Equation Gene H. Golub 24

Stanford University golub@stanford.edu The secular equation is a special way of expressing eigenvalue problems in a variety of applications. We describe the secular equation for several problems, viz. eigenvector problems with a linear constraint on the eigenvector and the solution of eigenvalue problems where the given matrix has been modied by a rank one matrix. Next we show how the secular equation can be approximated by use of the Lanczos algorithm. Finally, we discuss numerical methods for solving the approximate secular equation.

Application of Radon transform for fast solution of boundary value problems for elliptic PDE in domains with complicated geometry Alexander Grebennikov Benemrita Universidad Autnoma de Puebla, e o Facultad de ciencias F sico Matemticas, a Av. San Claudio y Ri verde, Ciudad Universitaria, CP 72570, Puebla, Mexico o agrebe@fcfm.buap.mx

Simulation of two-phase incompressible ow: discretization and preconditioning Sven Gross IGPM, RWTH Aachen University, Germany gross@igpm.rwth-aachen.de Two-phase systems play an important role in chemical engineering. Two examples are extraction columns where mass transport takes place between bubbles and a surrounding liquid (uid-uid system), or falling lms which are e.g. used for cooling by heat transfer from a thin liquid layer to the gaseous phase (liquid-gas system). In this talk we consider 3D ow simulations of such two-phase systems applying a nite element method on adaptive tetrahedral grids. For tracking the interface we use a level set approach. The eect of surface tension is modeled by an interfacial force term (CSF model). Numerical results for a single bubble obtained by our software package DROPS will be presented at the end of the talk. We will focus on several numerical challenges related to the solution of incompressible twophase ow problems. Special care has to be taken for the numerical treatment of surface tension, as otherwise articial spurious velocities are induced at the interface. We therefore analyze the Laplace-Beltrami discretization of the interfacial force term and introduce an extended nite element space for the representation of the discontinuous pressure (XFEM). The preconditioning of the Schur 25

complement has to be adapted to the jumping coecients in the PDE, which are induced by dierent material properties of the two phases. First results for the preconditioning of the discontinuous pressure Schur complement are given, too.

Primitivity Preservers for Matrix Tuples Alexander E. Guterman Moscow State University, Russia guterman@list.ru This talk is based on the joint work with LeRoy B. Beasley, Utah State University. Let S be a semiring. A matrix A Mn (S) is primitive if there is an integer k > 0 such that all entries of Ak are non-zero. The investigation of certain multi-dimensional dynamical systems leads to the following notion, which is the generalization of primitivity. Let A, B Mn (S) and h, k be some non-negative integers. The (h, k)-Hurwitz product is the sum of all matrices which are products of h copies of A and k copies of B. We denote it by (A, B)(h,k) . A pair (A, B) M2 (S) is n called primitive if there exist non-negative integers h, k such that the matrix (A, B)(h,k) is positive. We say that an operator, T : Mn (S) Mn (S) Mn (S) Mn (S), preserves primitivity for pairs if for primitive pair (A, B) we have that T (A, B) is also primitive. Theorem. Let B be a binary Boolean algebra, T : Mn (B) Mn (B) Mn (B) Mn (B) be a surjective additive operator which preserves primitive pairs. Then there is a permutation matrix P and permutations and of {1, , n} such that one of the following holds: T (X, Y ) = (P D (X)P t , P D (Y )P t ) for all (X, Y ) Mn (B) Mn (B); T (X, Y ) = (P D (Y )P t , P D (X)P t ) for all (X, Y ) Mn (B) Mn (B); T (X, Y ) = (P D (X t )P t , P D (Y t )P t ) for all (X, Y ) Mn (B) Mn (B); T (X, Y ) = (P D (Y t )P t , P D (X t )P t ) for all (X, Y ) Mn (B) Mn (B),

here D : Mn (S) Mn (S) is a diagonal replacement operator corresponding to a permutation Sn , i.e. there exists Sn such that D (Ei,j ) = Ei,j whenever i = j, and D (Ei,i ) = E(i),(i) for all i. We also provide the generalization of this theorem for k-tuples of matrices. We obtain analogs of these results in the case of arbitrary antinegative semirings without zero divisors, in particular, this includes non-negative matrices and matrices over max-algebras.

26

Fast and stable linear algebra Olga Holtz University of Berkeley, USA oholtz@eecs.berkeley.edu I will discuss my recent joint work with J. Demmel, I. Dumitiru and R. Kleinberg. The rst part of the talk will be devoted to the complexity and stability of matrix multiplications, the second to implications of our results for other problems or numerical linear algebra. In the rst part, we perform forward error analysis for a large class of recursive matrix multiplication algorithms in the spirit of [D. Bini and G. Lotti, Stability of fast algorithms for matrix multiplication, Numer. Math. 36 (1980), 6372]. As a consequence of our analysis, we show that the exponent of matrix multiplication (the optimal running time) can be achieved by numerically stable algorithms. We also show that new group-theoretic algorithms proposed in [H. Cohn, and C. Umans, A group-theoretic approach to fast matrix multiplication, FOCS 2003, 438449] and [H. Cohn, R. Kleinberg, B. Szegedy and C. Umans, Group-theoretic algorithms for matrix multiplication, FOCS 2005, 379388] are all included in the class of algorithms to which our analysis applies, and are therefore numerically stable. We perform detailed error analysis for some specic fast group-theoretic algorithms. In the second part, we extend our results to show that essentially all standard linear algebra operations, including LU decomposition, QR decomposition, linear equation solving, matrix inversion, solving least squares problems, (generalized) eigenvalue problems and the singular value decomposition can also be done stably (in a normwise sense) in O(n+ ) operations if matrix multiplication can be done in O(n+ ) operations for > 0 arbitrarily small.

Rational iterations for matrix functions Bruno Iannazzo Dipartimento di Fisica e Matematica, Universit` dellInsubria, Via Valleggio, 11, 22100 Como, a Italy iannazzo@mail.dm.unipi.it Rational iterations are commonly encountered in the study and computation of roots of polynomial equations. The generalization to the matrix case appears, for instance, in the computation of matrix functions and in the numerical solution of certain matrix equations. It raises problems somehow new: there can be innite xed points, the lack of commutativity of the matrix product can have eects on the convergence in nite arithmetic, it is not even straightforward how to dene a rational matrix iteration. We discuss these problems, providing general convergence results for a class of rational matrix iterations. Then, we prove some properties of specic classes of rational iterations. In particular it is shown that the principal Pad family of iterations for the matrix sign function e and the matrix square root is a special case of rational iterations due to Schrder. This characterio zation provides a family of iterations for the matrix pth root which preserve the structure of group of automorphisms associated with a scalar product. The rst iteration in that family is the Halley method for which is proved a convergence result. Finally, it is shown that the Schur-Newton method 27

for the matrix pth root applied to the inverse Newton iteration can be applied also to the direct case and to the Halley method.

On regularity in the sense of von Neumann of complete matrix semirings over additively regular semirings S. N. Ilyin, D. R. Timurshin Kazan State University, Russia Sergey.Ilyin@ksu.ru A semiring is an algebraic system (S, +, , 0) such that (S, +, 0) is a commutative monoid, (S, ) is a semigroup, (a + b)c = ac + bc, a(b + c) = ab + ac and a0 = 0a = 0 for all a, b, c S. A semiring S is regular (additively regular) i an equation axa = a (resp., a + x + a = a) is solvable for any a S. A semiring S is additively idempotent i a + a = a for all a S. Clearly, all rings and additively idempotent semirings are additively regular. A semiring S is regular in the sense of von Neumann i for each a S there exist x, y S such that a + axa = aya. Obviously, any regular semiring is regular in the sense of von Neumann. Conversely, any ring is regular if it is regular in the sense of von Neumann but this is not true for an arbitrary semiring. Let S be a semiring and Mn (S) be the semiring of all nn-matrices over S. It is well-known that if R is a ring then Mn (R) is regular i R is regular. The regularity criterion for matrix semirings is given in [1], in particular, for n 3 Mn (S) is not regular if S is not a regular ring. On the other hand, by [2] Mn (S) is regular in the sense of von Neumann for all n if S is a commutative semiring with identity 1 and for any a, b S (a = b) there exist x, y S such that 1 + ax + by = ay + bx. Also, it is easy to prove the following result: Proposition Let S be an additively idempotent semiring. Then Mn (S) is regular in the sense of von Neumann i so is S. So, the equivalence of regularity in the sense of von Neumann for S and Mn (S) holds for any ring and additively idempotent semiring. The theorem below gives an analogous result for additively regular semirings which satisfy some additional conditions. We denote by I(S) and R(S) the ideal of all additively idempotent elements in S and the ideal of all elements in S which have additive inverses, respectively. Theorem Let S be an additively regular semiring which is regular in the sense of von Neumann. If one of the following conditions holds 1) there exists u I(S) such that x + u = u for all x I(S), 2) R(S) or I(S) are nite, then for any n Mn (S) is regular in the sense of von Neumann. References 1. S.N.Ilyin, Regularity criterion for complete matrix semirings, Mat. Zamet. 70 (2001), 366374. 2. Shamik Ghosh, A note on regularity in matrix semirings, Kyungpook Math. J. 44 (2004), 14.

28

On the conjugate and semi-conjugate direction methods with preconditioning projectors V.P.Ilin Institute of Computational Mathematics and Mathematical Geophysics, SBRAS, Novosibirsk ilin@sscc.ru The goal of presented work consists in the improvement of an iterative convergence rate and corresponding estimates for multiplicative and additive type algorithms with preconditioning projectors, applied to domain decomposition approach and investigated by J. Bramble with his coauthors (1991). For the multiplicative method, the symmetrized alternating direction algorithm version is proposed. It can be accelerated by means of the conjugate gradient or conjugate residual algorithms in Krylov subspaces. The similar acceleration is suggested for the additive type iterative procedure. The acceleration of the multiplicative method without symmetrization is done by semi-conjugate residual algorithm which is stable modication of GCG method by H. Elman and others (1983). It is shown that the multiplicative and additive projector algorithms include, as the particular cases, the known Kaczmarzs (1937) and Cimminos (1938) methods respectively. Various generalizations and acceleration in Krylov subspaces of these algorithms are proposed. The estimations of the iterative convergence rate are obtained. The results of the numerical experiments for the model grid diusion-convection boundary value problems, with dierent coecients and mesh steps, are presented.

On the quaternionic equation AX + BXC + XD = E Drahoslava Janovsk , Gerhard Opfer a Institute of Chemical Technology, Department of Mathematics, Prague, Czech Republic University Hamburg Drahoslava.Janovska@vscht.cz We study the quaternionic equation AX + BXC + XD = E. For BC = 0 the resulting equation is called Sylvesters equation. For this case a complete solution will be given. For the general case we show that the solution can be found by solving a corresponding matrix equation.

On singular values of parameter dependent matrices Drahoslava Janovsk , Vladim Janovsk a r y of Chemical Technology, Department of Mathematics, Prague, Czech Republic Faculty of Mathematics and Physics, Charles University, Prague, Czech Republic The research of both authors was partially supported by the Grant Agency of the Czech Republic (grant No. 201/06/0356). Drahoslava.Janovska@vscht.cz, janovsky@karlin.mff.cuni.cz
Institute

29

We consider the Analytic Singular Value Decomposition, ASVD, of matrix valued functions. ASVD is smooth up to isolated parameter values at which either a multiple singular value or a zero singular value turns up on the path. These exceptional points are called non-generic, see [1]. They were classied in [2]. Note that ASVD-computations, see e.g. [1,2], require information on all singular values on the path and hence the algorithms were not able to cope with large sparse input data. In [3], we investigated a pathfollowing of just one simple singular value and the corresponding left/right singular vector. A breakdown of the continuation is related to non-generic points on the path. We apply Singularity Theory to analyze and classify these non-generic points. Our analysis will include the questions concerning structural stability. The classication will result in precise localization technique of these points. We compare our classication list with [2]. References 1. A. Bunse-Gerstner, R. Byers, V. Mehrmann, and N. K. Nichols, Numerical Computation of an Analytic Singular Value Decomposition of a Matrix Valued Function, Numer. Math., 60 (1991), pp. 139. 2. K. Wright, Dierential equations for the analytic singular value decomposition of a matrix, Numer. Math. 63 (1992), pp. 283295. 3. D. Janovska, V. Janovsky, The analytic SVD: On the non-generic points on the path, Proceedings of Joint GAMM-SIAM Conference on Applied Linear Algebra, Dsseldorf 2006. Submitted u to ETNA.

Multilevel Optimal Preconditioners Xiao-Qing Jin Department of Mathematics, University of Macau, Macao, China XQJin@umac.mo

Incomplete LU factorization of general sparse matrices Igor Kaporin Computing Center of Russian Academy of Sciences, Vavilova 40, Moscow 119991, Russia kaporin@ccas.ru We discuss a convergence estimate for Preconditioned GMRES-type iterative linear equation solvers. Formulation of the result is specically adjusted to the case when the preconditioning is based on an Incomplete Triangular factorization by value.

30

Consider the linear algebraic system Ax = b with general unsymmetric nonsingular sparse n n matrix A. The ILU preconditioned GMRES-type iterative methods use the preconditioner matrix C A of the form C = PL LU PR as is obtained from the ILU equation A = PL LU PR + E, where L and U are nonsingular lower and upper triangular matrices, respectively, while PL and PR are permutation matrices. The additive term E is the ILU error matrix, for which the magnitude of its entries are limited by a prescribed threshold value . Under a proper choice of permutation matrices PL and PR , one can also observe that not only the stability of the triangular factors is improved (indeed, it seems natural to require at least that cond(C) cond(A)), but also E 2 F trace(E T E) = O(n 2 ). The latter means that only relatively few nonzero entries of E attain the upper bound by their magnitude. Furthermore, as was noted by many authors, see, e.g. [2,4] and references therein, the ILU factorization by value applied to a properly two-side scaled coecient matrix AS = DL ADR may yield a much better preconditioning, especially in several hard-to-solve cases, see also [8]. Main Result. Based on the latter paper, we present the following superlinear residual norm bound for GMRES iterations: n rk K(AS ) 4e sin2 [CS , AS ] r0 k where K(B) = n1/2 B
F k/2

e = exp(1),
n

= V

V 1 ,

(1)

| det B|

(2)

is the unsymmetric K-condition number of a nonsingular matrix B, rk = b Axk is the kth residual generated by Preconditioned GMRES(), CS is the preconditioner for AS and V is the n n matrix of eigenvectors of AC 1 (i.e., we assume the diagonalizability of the preconditioned matrix). Note that the earlier estimate of similar type developed in [7] was formulated in terms of sin[In , AC 1 ] and min (AC 1 ), which quantities are very problematic to measure even a posteriori. The correctness of the above convergence estimate was also tested numerically using many hard test matrices taken from the University of Florida Sparse Matrix Collection [1]. Relation to Scaling. In view of (1), it is natural to require that the scaling should minimize the functional (2) with B = AS . It appears that the minimizer satises exactly the requirement that AS = DL ADR has Euclidean norms of each row and column equal to the same number, exactly as was recommended in [2,4]. Relation to ILU Error. Let the matrix be scaled in such a way that AS 2 = n. Then F trace(AT CS ) S sin [CS , AS ] 1 CS 2 AS 2 F F
2 2

min AS CS AS 2 F
2 F

2 F 2 F.

AS CS

2 F

AS

2 F

= ES

AS

2 F

= n1 ES

Hence, the above convergence estimate takes the form rk 0 r | det AS | 3.3 ES k
k/2 F

cf. [8]. It remains to note that if the threshold parameter is chosen suciently small, e.g. = 0.001, and the ILU factors are stable enough, then the typical values of ES F are not big (often one can observe ES F < 1 even for rather big problems). 31

Conclusions. First, a theoretical justication is found for the standard pre-scaling technique related to the ILU factorization, with implications for the practical implementation. (Namely, a more accurate evaluation of DL and DR may be useful, and faster algorithms should be developed for this purpose.) Second, an even better estimate for the reduction of the original (unscaled) residial can be obtained in terms of the scaled ILU error. References 1. University of Florida Sparse Matrix Collection. http://www.cise.u.edu/research/sparse/ matrices/ 2. V.F. de Almeida, A.M. Chapman, and J.J. Derby, On Equilibration and Sparse Factorization of Matrices Arising in Finite Element Solutions of Partial Dierential Equations, Numer. Methods Partial Dierent. Equ. 16, 1129, 2000. 3. M. Bollhofer and Y. Saad, Multilevel preconditioners constructed from inverse-based ILUs, SIAM J. Sci. Comput., 27, 16271650, 2006. 4. O. E. Livne and G. H. Golub, Scaling by Binormalization, Numer. Alg., 35, 97120, 2004. 5. M. H. Schneider and S. A. Zenios, A comparative study of algorithms for matrix balancing, Operations Research, 38, 439455, 1990. 6. N. Li, Y. Saad, and E. Chow, Crout versions of ILU for general sparse matrices, SIAM J. Sci. Comput., 25, 716728, 2003. 7. I. Kaporin, Superlinear convergence in minimum residual iterations, Numerical Linear Algebra with Applications 12, pp. 453470, 2005. 8. I. Kaporin, Scaling, Reordering, and Diagonal Pivoting in ILU Preconditionings, (29 pp.) To appear in Russian Journal of Numerical Analysis and Mathematical Modelling 12, no.4, 2007.

Recursive scaling, permutation and 22-block splitting in ILU preconditionings Igor Kaporin and Igor Konshin Computing Center of Russian Academy of Sciences, Vavilova 40, Moscow 119991, Russia kaporin@ccas.ru We discuss a technique for the Incomplete LU preconditioning of general unsymmetric nonsingular matrix aimed at the construction of robust linear equation solvers. The idea of such preconditioning (multilevel ILU) was proposed in [3]. The key component of the technique is a recursive construction of a 2 2 block splitting with diagonally dominant leading block. In the present work, we propose a new algorithm for the construction of such splittings. Consider the linear algebraic system Ax = b with a nonsingular general sparse n n matrix A. In the construction of preconditioned GMRES-type iterative methods, the key problem is to construct the preconditioner C A. By the substitution x = C 1 y the original problem is then reduced to the preconditioned system M y = b where M = AC 1 is the right preconditioned matrix. An iterative scheme is then applied to the latter problem. This normally yields a faster convergence of the iterations.

32

Let n1 and n2 be the block sizes of the splitting, n1 + n2 = n. The preconditioner C is obtained by making the rst n1 steps of an ILU factorization by value of the Crout type [6]: CSR = ASR E = = A11 A12 (SR) (SR) A21 A22 L11 0 L21 I2
(SR) (SR)

E U11 U12 0 I2 ,

(1)

I1 0 0 A

where E is the ILU error matrix (with the magnitudes of its entries limited by a prescribed threshold), the n2 n2 -matrix (SR) A = A22 L21 U12 is the approximate active submatrix (or Schur complement) and ASR = PL AS PR , AS = DL ADR

are the scaled-and-reordered and two-side scaled coecient matrices, respectively. Here PL , PR are permutation matrices and DL , DR are nonsingular diagonal matrices. The preconditioner C for the original matrix A can be found from the relation CSR = PL DL CDR PR . If the size of A in this decomposition is still large, then this construction is applied recursively, is replaced by C in the same way. When the size of A is reduced to an acceptable level, it i.e., A can be processed using the direct sparse LU factorization. As is shown in [3], in order to guarantee a good preconditioning quality in (1) one should impose the following two contradictive requirements: (i) the size n1 of the leading block in splitting (1) is as large as possible and (ii) the same leading block in (1) is well-conditioned. Typically, A11 is constructed as a diagonally dominant matrix, cf. [3]. Note that in our treatment we consider the diagonal dominance with respect to the scaled matrix AS . Two-side scaling stage. Similar to [2,4], we determine the scaling by (nearly) equalizing both row and column 1-norms of the scaled matrix AS = DL ADR , i.e. we approximately solve the nonlinear system j |(AS )ij | = 1, i |(AS )ij | = 1. The diagonal matrices DL and DR were obtained by the RAS (Row-Alternating Scaling) iterations (see, e.g. [5] and references therein). Pre-ordering stage. As the algorithm for two-side unsymmetric reordering of the matrix AS , we used the simplest possible greedy strategy, which constructs the pre-ordered matrix AS R = PL AS PR such that |(AS R )kk | |(AS R )ij |, i > k, j > k. If the array of nonzero elements of H is organized as binary heap, the corresponding algorithm requires O(nz(A) log nz(A)) operations. Splitting stage. Let us denote the entries of the scaled and pre-ordered matrix as aij = (AS R )ij . We sequentially construct the set D of n1 diagonal positions (to be included in the required leading block) by checking the diagonal dominance condition: D = ; for k = 1, 2, ..., n if a a jD{k}\{i} |i,j | < |i,i |, 1 i k then D := D {k} end for Using a data structure which provides a fast access both to AS R and AT R , one can implement the S above procedure in O(nz(A)) operations. In order to gather consecutively the positions dened by the index set D, one applies the corresponding symmetrical reordering to AS R thus obtaining the nal 2 2 split form: ASR = P T AS R P , i.e., PL = P T PL , PR = PR P . Note that the value n1 = |D| determined as above, typically grows with the number of scaling iterations used. In practice, we performed the RAS iterations until the relative spread of the row or column norms be reduced below 1 + , with in the range [0.1, 0.5]. 33

The eciency of the proposed procedure was tested numerically using many hard test matrices taken from the University of Florida Sparse Matrix Collection [1]. References 1. University of Florida Sparse Matrix Collection. http://www.cise.u.edu/research/sparse/ matrices/ 2. V.F. de Almeida, A.M. Chapman, and J.J. Derby, On Equilibration and Sparse Factorization of Matrices Arising in Finite Element Solutions of Partial Dierential Equations, Numer. Methods Partial Dierent. Equ. 16, 1129, 2000. 3. M. Bollhofer and Y. Saad, Multilevel preconditioners constructed from inverse-based ILUs, SIAM J. Sci. Comput., 27, 16271650, 2006. 4. O. E. Livne and G. H. Golub, Scaling by Binormalization, Numer. Alg., 35, 97120, 2004. 5. M. H. Schneider and S. A. Zenios, A comparative study of algorithms for matrix balancing, Operations Research, 38, 439455, 1990. 6. N. Li, Y. Saad, and E. Chow, Crout versions of ILU for general sparse matrices, SIAM J. Sci. Comput., 25, 716728, 2003.

Monotone matrices and nite volume schemes for diusion problems preserving non-negativity of solution I. Kapyrin Institute of Numerical Mathematics of RAS ivan.kapyrin@gmail.com A new nite volume scheme for 3D diusion problems with heterogeneous full diusion tensor is considered. The discretization uses nonlinear two-point ux approximation on unstructured tetrahedral grids. Monotonicity of the linearized operator allows us to guarantee non-negativity of the discrete solution.

Spectral model order reduction for control systems modelling passive integration circuits Yu. M. Nechepurenko , A. S. Potyagalova , I. A. Karaseva of Numerical Mathematics, Russian Academy of Sciences, ul. Gubkina 8, Moscow, 119333 Russia, Cadence Design Systems LLC, ul. B.Ordynka 44, Bldg.4, Moscow, 119017 Russia, Moscow Institute of Physics and Technology, ul. Kerchenskaya 1a, Bldg.1, Moscow, 117303. This work was supported by the Russian Foundation for Basic Research (project 07-01-00658) and Russian Academy of Sciences (project Optimization of numerical algorithms for solving the problems of mathematical physics) yumn@inm.ras.ru, alena@cadence.com, irina kar@bk.ru
Institute

34

Complexity reduction for generating compact models of interconnect RCLM networks have been an intensive research area in the past decade due to increasing signal integrity eects and rising couplings modelled with parasitic capacitors and inductors. Since the original systems were passive one of the main requirements to the reduced systems was the preservation of passivity. Many algebraic model order reduction methods preserving passivity were proposed based on implicit moment matching, congruence transformations and truncated balance realizations. In this report we present a novel algebraic spectral model order reduction algorithm equipped with ecient tools for preserving the passivity. For RC networks our approach is similar to the wellknown spectral reduction technique PACT (pole analysis via congruence transformations) and can be construed as its generalization. Up to the present any such a generalization preserving passivity has seemed impossible and other model order reduction methods were applied for the reduction of RCL and RCLM networks. The accuracy and reduction ratio of resulting reduced-order models are demonstrated with several industrial examples.

Toeplitz and Toeplitz-block-Toeplitz matrices and their correlation with syzygies of polynomials Houssam Khalil , Bernard Mourrain , Michelle Schatzman Institut Camille Jordan, Universit Claude Bernard Lyon1 e 43 boulevard du 11 novembre 1918 69622 Villeurbanne cedex France INRIA, GALAAD team 2004 route des Lucioles BP 93, 06902 Sophia Antipolis Cedex, France khalil@math.univ-lyon1.fr, mourrain@sophia.inria.fr, schatz@math.univ-lyon1.fr Structured matrices appear in various domains, such as scientic computing, signal processing, . . . They usually express, in a linearized way, a problem which depends on less parameters than the number of entries of the corresponding matrix. An important area of research is devoted to the development of methods for the treatment of such matrices, which depend on the actual parameters involved in these matrices. Among well-known structured matrices, Toeplitz and Hankel structures have been intensively studied. Nearly optimal algorithms are known for the multiplication or the resolution of linear systems, for such structure. Namely, if A is a such matrix of size n, multiplying it by a vector or solving a linear system with A requires O(n) arithmetic operations (where O(n) = O(n logc (n)) for some c > 0) arithmetic operations for some c > 0. Such algorithms are called super-fast, in opposition with fast algorithms requiring O(n2 ) arithmetic operations. The correlation with other types of structured matrices has also been well developed in the literature, allowing to treat so eciently other structures such as Vandermonde or Cauchy-like structures. Such problems are strongly connected to polynomial problems. For instance, the product of a Toeplitz matrix by a vector can be deduced from the product of two univariate polynomials, and thus can be computed eciently by evaluation-interpolation techniques, based on FFT. The inverse of a Hankel or Toeplitz matrix is connected to the Bezoutian of the polynomials associated to their generators. 35

However, most of these methods involve univariate polynomials. So far, few investigations have been pursued for the treatment of multilevel structured matrices, related to multivariate problems. Such linear systems appear for instance in resultant or in residue constructions, in normal form computations, or more generally in multivariate polynomial algebra. A main challenge here is to devise super-fast algorithms of complexity O(n) for the resolution of multi-structured systems of size n. In this paper, we consider block-Toeplitz matrices, where each block is a Toeplitz matrix. Such a structure, which is the rst step to multi-level structures, is involved in many bivariate problems, or in numerical linear problems. We re-investigate rst the resolution of Toeplitz systems T u = g, from a new point of view, by correlating the solution of such problems with syzygies of polynomials or moving lines. We show an explicit connection between the generators of a Toeplitz matrix and the generators of the corresponding module of syzygies. We show that this module is generated by two elements of degree n and the solution of T u = g can be reinterpreted as the remainder of an explicit vector depending on g, by these two generators. This approach extends naturally to multivariate problems and we describe for Toeplitz-blockToeplitz matrices, the structure of the corresponding generators. In particular, we show the known result that the module of syzygies of k non-zero bivariate polynomials is free of rank k 1, by a new elementary proof. Exploiting the properties of moving lines associated to Toeplitz matrices, we give a new way which, may be, we can use it to compute explicitly the solution of a Toeplitz-block-Toeplitz matrices with the complexity of such computation is bounded by O(m n) where n is the number of Toeplitz blocks and m their size.

Sparse Approximation of FEM Matrix for Sheet Current Integro-Dierential Equation M. M. Khapaev , M. Yu. Kupiyanov Comp. Sci. Dept. of Moscow State University, Russia vmhap@cs.msu.su In the report we consider the problem of eective numerical solution of 2D sheet current integrodierential equations. Typically the equations for sheet current are derived from Maxwell equations for magnetic potential. The equation contain weak singularity in the kernel and for normal conductors has the form Rs J(r) + j 4
S

J(r ) ds = (r), |r r |

(r) = 0,

r S.

(1)

Here S is 2D domain, r = (x, y). J(r), (r) are phasors for sheet current and voltage. is frequency, Re(Rs ) 0, Im(Rs ) 0 thin plate resistance [1]. For (1), well developed numerical techniques exist [1]. In many cases, the numerical solution of (1) is very time consuming even if GMRES and Fast Multipoles Method for kernel approximation are used [1]. We are interested in the solution of very similar to (1) task for superconductors [3]. In our case equations are real and does not depend on frequency. The equations we solve are singular and follows 36

from Biot-Savart formula J (r)


z

1 4

J (r) (r

1 ) |r r0 |

dxdy = 0,
z

(2)

where r = (x, y), r0 = (x0 , y0 ), r, r0 S, is real coecient. Equation (2) must be fullled by the boundary conditions for J (r). On the part of the boundary of S boundary conditions reduces to Jn = 0, on the rest of the boundary the non-zero value of Jn is specied. Equation (1) can be reduced to form similar to (2) using 2D rotor dierentiation. To solve (2) we rst rewrite it using stream function (r). Then we obtain the problem very similar to rst boundary problem for Laplace equation: (r0 ) + 1 4 (r), r 1 |r r0 | ds = 0, (r) = F (r), r S. (3)

The function F (r) is completely dened by boundary condition for current. Equation (3) for (r) is hypersingular with kernel 1/|r r0 |3 . It is known that operator in (3) is positive and self-adjoint [2] (if (r) = 0 on S). To solve (3) numerically we use nite element method with linear elements on triangular grids [3]. The resulting FEM matrix is dense die to integral operator. Let uh , uh be FEM basis functions, i j then matrix elements are a(uh , uh ) i j = (uh , uh )ds i j 1 + 4 ds
Si Sj

Si Sj

(uh (r ), uh (r)) j i ds , | |r r

(4)

The method of calculation of quadruple integrals is given in [3]. Due to integral term in (3) FEM results in fully populated symmetric matrix. The dimension of this matrix for practical problems is large (103 104 ). This fact is the key limitation of the method. We suggest a simple method for sparsifying the FEM matrix. We simply drop small elements. Symmetry of matrix is preserved and positive denetness also can be preserved if dropped elements are small enough. Consider matrix element for integral operator (4). Diagonal matrrix elements are large and positive but matrix is not diagonally dominant. Sub-diagonal elements can be of any sign. If Si Sj = then elements are negative and bij =
Si

ds
Sj

(uh (r ), uh (r)) j i ds = |r r |

ds
Si Sj

uh (r )uh (r) j i ds |r r |3

(5)

Let h be the estimation of diameters of triangles in the mesh. Then mij h |ri rj |, mij > 1 and from (5) follows h (6) |bij | M 3 . mij Thus matrix elements quickly decay if distance between nite elements grows. For equation with weak singularity this decaying is of rst order only. In the report, we present set of results of numerical experiments with sparse FEM matrix approximation and direct linear system solver. The conclusion is that the simple method can solve large problems and overcome memory and CPU time limitations of dense matrix technique. The paper is supported by ISTC project 3174. References 37

1. M. Kamon, M. J. Tsuk and J. K. White, FastHenry: a multipole-accelerated 3-D inductance extraction program, IEEE Trans. Microwave Theory and Techn. vol. 42, pp. 17501758, September 1994. 2. V. J. Ervin, E. P. Stephan, A Boundary Element Galerkin Method for a Hypersingular Integral Equation on Open Surfaces, Mathematical Methods in the Applied Sciences, 1990, v. 13, issue 4, pp. 281289. 3. M. Khapaev, Inductance Extraction of Multilayer Finite Thickness Superconductor Circuits, IEEE Trans. Microwave Theory and Techn., 2001, v.49, pp. 217220.

Some approaches to solving inverse eigenvalue problems V. N. Kublanovskaya , V. B. Khazanov Branch of the V. A. Steklov Mathematical Institute RAS St.-Petersburg State Marine Technical University The work was supported by the RFBR (project No. 05-01-00945a). verakub@pdmi.ras.ru, khazanovvb@mail.ru

St.-Petersburg

The methods are proposed for solution of inverse eigenvalue problems for regular polynomial matrices (both the one-parameter ones, including the pencils A I and A B, and the multiparameter ones) which are formulated in two ways: to replace known eigenvalues of the polynomial matrix by given numbers with the preservation of the right (or left) eigenvectors associated with the invariable eigenvalues, to calculate values of parameters of a parameter matrix for which the matrix has the prescribed spectrum points (eigenvalues). For the rst statement the polynomial matrix F () and its eigenvalues [F ], i = 1, . . . , p i are given. It is required to compute the polynomial matrix F () with spectrum [F ] = [F ] \ { , i = 1, . . . , p} { , i = 1, . . . , p}. The right (or left) eigenvectors associated with the invariable i i eigenvalues must be preserve. The following approaches to solution of the problem are proposed. 1. The transformations of the matrix, based on the rank factorizations of polynomial matrix by using of unimodular matrices [4]. The factorization F () = F1 ()F2 () is fullled, where [F1 ] = { , i = 1, . . . , p}. For each the respective eigenspace of the matrix F1 () is computed. Afterwards i i of the corresponding multiplicity is replaced by . The procedure is repeated if the geometric i i multiplicity is smaller then the geometric multiplicity. i 2. The transformations of the matrix, based on the generalizations of the exhaustation processes (Vilandt and Hotelling methods) of an eigenvalue of a constant matrix [3]. For each the right (or i left) eigenvectors of the matrix F () are computed and the transformations of the matrix are fullled until all are replaced by . If the block spectral characteristics are given the transformation is i i fullled the one step. For the second statement the multiparameter polynomial n n matrix F (, ) is given and it is required to compute the values of parameters = (1 , . . . , n ) for which the matrix has a prescribed set of the spectrum points , . . . , . In the case of the one-parameter matrix F (, ) n 1 the prescribed set of the eigenvalues , . . . , . The following approaches to solution of this problem n 1 are proposed. 38

1. Reduction to the system of nonlinear algebraic equations. By used of the Faddeev trace method the polynomial (, ) := detF (, ) is computed. Once the polynomials i () := (, ) i are computed, the initial problems is reduced to searching a solution of the system of equations i () = 0, i = 1, . . . , n. The corresponding methods are discussed in [1,2]. 2. Reduction to the system of nonlinear equations. The equivalent nonlinear system is considered instead of the system of determinant equations. The following characteristics of singularity of matrix can be used: the equality of the least singular value or the least diagonal elements of the left triangular matrix obtained by use of the normalized decomposition of constant matrix to zero. The method based on this approach is considered, for example, in [5]. 3. Reduction to the multiparameter spectral problem. The normalized eigenvectors xi associated with the spectrum points , i = 1, . . . , n are introduced and the initial problem is re-formulated as i a multiparameter spectral one: Fi ()xi := F (, )xi = 0 , i xT xi = 1 , i i = 1, . . . , n

In particular, the additive and multiplicative inverse problems are reduced to the problems for the matrix pencils Fi () = Fi n Bij . The obtained nonlinear system (which is of the order (n+1)n) j=1 can be solved using iterative methods (i.e. Newton, Tchebychev, Halley, inverse iterations, gradient methods) [6]. On the each step of these methods (with the exception of the gradient methods) the linear systems of the order n are solved. The linearization of polynomial matrix F () [7] of the degree s > 1 is realized by means to transfer to the accompanied pencil. Moreover the order is increased at sn or s+n times. n References 1. Kublanovskaya V. N. Solution of systems of nonlinear algebraic eguations. Methods and algorithms IV, Zap. nauchn. semin. POMI, 248 (1998), 124146. 2. Kublanovskaya V. N. An approach to solving inverse eigenvalue problems for matrices, Zap. nauchn. semin. POMI, 268 (2000), 950-114. 3. Kublanovskaya V. N., Khazanov V.B. Exhaustion in the spectral problems for matrix pencils, In: Computing processes and systems, 5, (ed. G. I. Marchuk), Moscow, Nauka, 1987, pp. 138147. 4. Kublanovskaya V. N., Khazanov V.B. Numerical methods for solving parametric problems of algebra. Part 1. One-parameter problems, St. Petersburg, Nauka, 2004. 5. Khazanov V. B. Methods for solving two-parameter eigenvalue problem, Proc. of LKI: The mathematical software for automatized shipbuilding systems, Leningrad, LKI, 1987, pp. 118125. 6. Khazanov V. B. Methods for solving spectral problems for multiparameter matrix pencils, Zap. nauchn. semin. POMI, 296 (2003), 139168. 7. Khazanov V. B. To solving spectral problems for multiparameter polynomial matrices, Zap. nauchn. semin. POMI, 334 (2006), 212231.

On solving spectral problems for nonlinear multiparameter matrices V. B. Khazanov St.-Petersburg State Marine Technical University 39

The work was supported by the RFBR (project No. 05-01-00945a). khazanovvb@mail.ru The approach proposed by V. N. Kublanovskaya to solving nonlinear one-parameter spectral problem (based on normalized decomposition of constant matrix [1]), is generalized for the case of nonlinear multiparameter matrices. The joint spectrum of the multiparameter regular matrices Fi () of the order ni , , i = 1, . . . , q, whose elements are the nonlinear dierential functions of the parameters = (1 , . . . , q ) is determined as zeros of the system of nonlinear equations det Fi () = 0, i = 1, . . . , q.

The following equivalent equations are used instead of determinant ones:


(i) fi () = lni (), (i)

i = 1, . . . , q,

where lni () is the last diagonal element of the left triangular matrix Li (), obtained using the normalized decomposition of constant matrix Fi () with the xed value of multiparameter : i ()Fi () = Li ()QT (), i i = 1, . . . , q.

Here i (), Qi () and Li () are the permutation, orthogonal and left triangular matrices of the order of ni correspondingly. The assumption that the value of the multiparameter belongs to a neighborhood of the spectrum point F1 , . . . , Fq , where the permutation matrix i () is stabilized, allows one to obtain the relationships for computing the partial derivatives of the left hand sides of the considered equations. In its turn it enables one to apply such methods as the Newton method and the method of steepest descent. The modications of the methods are considered which avoid solution of ill-conditioned systems of linear equations. References 1. Kublanovskaya V. N., On applying the Newton method to computing eigenvalues of matrices, Dokl. AS USSR, 188(5) (1968), 10041005. 2. Khazanov V. B., Methods for solving two-parameter eigenvalue problem, Proc. of LKI: The mathematical software for automatized shipbuilding systems, Leningrad, LKI, 1987, pp. 118125. 3. Yakovlev M. I. On some methods for solving nonlinear equations, Proc. MI AS USSR, 84 (1965), 840.

Multi-linear approximation of collocated convolution transform in Rd Boris N. Khoromskij Max-Planck-Institute for Mathematics in the Sciences, Inselstr. 22-26, D-04103 Leipzig, Germany bokh@mis.mpg.de 40

Modern methods of tensor-product decomposition allow an ecient data-sparse approximation to integral and more general nonlocal operators in higher dimensions (cf. [18]). Examples of such nonlocal mappings are classical Greens functions, solution operators of elliptic/parabolic BVPs, integral operators arising from the Hartree-Fock equation in electronic structure calculations, collision integrals from the deterministic Boltzmann equation as well as the convolution integrals from the Ornstein-Zernike equation in molecular dynamics. We discuss the approximation theory and numerical methods of structured low tensor-rank approximation to convolution transform in Rd . In some cases, the asymptotic complexity of such approximations can be estimated by O(dn log n + logq n), where N = nd is the discrete problem size (hence, avoiding the curse of dimensionality). In particular, we focus on applications arising from 1 the Hartree-Fock equation, which includes the convolution with the classical Newton potential |xy| , x, y R3 . References 1. S. R. Chinnamsetty, H.-J. Flad, W. Hackbusch, V. Khoromskaia and B. N. Khoromskij, Tensor approximation in fully discrete electronic structure calculations, MPI MIS, Leipzig 2007 (in progress). 2. I. P. Gavrilyuk, W. Hackbusch and B. N. Khoromskij, Tensor-Product Approximation to Elliptic and Parabolic Solution Operators in Higher Dimensions. Computing 74 (2005), 131-157. 3. W. Hackbusch, B.N. Khoromskij, and E. Tyrtyshnikov, Hierarchical Kronecker tensorproduct approximation, J. Numer. Math. 13 (2005), 119-156. 4. W. Hackbusch and B.N. Khoromskij, Low-rank Kronecker product approximation to multi-dimensional nonlocal operators. Parts I/II. Computing 76 (2006), 177-202/203-225. 5. B.N. Khoromskij, Structured data-sparse approximation to high order tensors arising from the deterministic Boltzmann equation, Preprint 4, MPI MIS, Leipzig 2005 (Math. Comp., to appear). 6. B. N. Khoromskij, Structured Rank-(r1 , ..., rd ) Decomposition of Function-related Tensors in Rd , Comp. Meth. in Applied Math., 6 (2006), 2, 194220. 7. B. N. Khoromskij and V. Khoromskaia, On the Best Rank-(r1 , ..., rd ) Tucker Approximation to the Classical Potentials, Preprint 105, MPI MIS, Leipzig 2006 (CEJoM 2007, to appear). 8. E. Tyrtyshnikov, Tensor approximation of matrices generated by asymptotically smooth functions, Mat. Sb. 194 (6), 2003, 147160.

Stabilizing and destabilizing eect of breaking the Hamiltonian and reversible symmetry Oleg N. Kirillov Moscow State Lomonosov University, Institute of Mechanics, Michurinskii pr. 1, 119192 Moscow, Russia kirillov@imec.msu.ru The eigenvalue problem is studied for the matrix polynomial L() = I2 + (D + G) + P + N determining the stability of a linear autonomous non-conservative system in presence of dissipative, gyroscopic, potential, and non-conservative positional forces. It is assumed that the matrices involved are real, I is the unit matrix, D and P are symmetric, and G and N are skew-symmetric. In 41

practically important cases of potential ( = = = 0), gyroscopic ( = = 0), and circulatory system ( = = 0) possessing the Hamiltonian or reversible symmetry, the necessary and sucient condition for marginal stability is that the eigenvalues are purely imaginary and semi-simple. In modern problems of wave propagation, acoustics of friction, and uid-structure interaction it is important to know how the marginal stability is destroyed or improved up to asymptotic stability due to action of non-Hamiltonian and non-reversible perturbations [1,2,3,4,5,6]. The present contribution shows that in both cases the boundary of the asymptotic stability domain of the perturbed system possesses numerous singularities such as Whitney umbrella and the Break of an edge that govern stabilization and destabilization. With the use of the perturbation theory of multiple eigenvalues approximations of the stability boundary near the singularities and estimates of the critical gyroscopic and circulatory parameters are found in analytic form. The role of dissipative forces with the indenite matrix D is investigated and new explicit condition is obtained for such forces to be stabilizing. The typical motion of eigenvalues is described. Applications to the stability problems of gyroscopic systems with stationary and rotating damping, such as the Tippe Top and Jellets egg, as well as to the problem of suppression of the squeal of the automotive disc brake are discussed. References J. Maddocks, M. L. Overton, Comm. Pure and Applied Math. 48, (1995), 583610. P. Freitas, M. Grinfeld, P.A. Knight, Adv. Math. Sci. Appl. 7(1), (1997), 437448. A. Akay, J. Acoust. Soc. Am. 111(4), (2002), 15251548. N.M. Kinkaid, O.M. OReilly, P. Papadopoulos, J. Sound Vib. 267, (2003), 105166. R. Krechetnikov, J. Marsden, Rev. Mod. Phys., 79(2), (2007), 519553. O. N. Kirillov, Int. J. Non-Lin. Mech. 42(1), (2007), 7187.

1. 2. 3. 4. 5. 6.

How to economically solve a multi-frequency problem with a nonnegative denite operator Vladimir Druskin , Leonid Knizhnerman , Mikhail Zaslavsky Research, 1 Hampshire St., Cambridge, MA 02139, USA Central Geophysical Expedition, Russia, 123298, Moscow, Narodnogo Opolcheniya St., house 38, building 3 druskin1@boston.oilfield.slb.com, mmd@cge.ru, mzaslavsky@slb.com
Schlumberger-Doll

Let us have a parametric family of equations (A + iI)u = , 0 < min max ,

where A is a self-adjoint non-negative denite operator in a Hilbert space and is a nonzero vector. Given n, we wish to nd basic frequencies j , 1 j n, such that u can be well approximated by linear combinations of uj . Some potential theory considerations invoked us to extract n/2 basic frequencies from Zolotaryovs approximation to 1/2 on a bounded positive interval and to add negative frequencies with the same modules. The coecients xj of linear combinations can be obtained via solution of Galerkin type

42

systems in the weighted spectral functional space:


n

( + i)
j=1

xj

1 1 1 , + ij + ik

1 k n.

Here 0 represents the spectrum of A and the weight function is 1/2 . We have experimentally found out that our choice of basic frequencies is good enough. A hypothesis on the approximation error has been formulated. It is illustrated with the results of numerical experiments.

The method of magnetic eld computation in the presence of an ideal conductive multilinked surface using the integro-dierential equation of the rst kind T. V. Kochubey Southern Scientic Centre of Russian Academy of Sciences tako84@yandex.ru The problem of stationary plane magnetic eld computation in the presence of the ideal conductive surface with the piecewise-smooth Lipschitz border has been considered. The surface is in a homogeneous medium with magnetic conductivity 0 = const. The problem of magnetic eld computation has been reduced to the following scalar boundary problem: = 0 outside of ,
0 n = Bn on , 0 Bn d, n d = k k (M ) 0, M

k = 1, N ,

where B0 is an induction of non-perturbed magnetic eld, is a scalar potential of magnetic eld of reaction. The balance of magnetic eld sources is also taken into account. Here =
N N k=0

is a closure of to closed surface = , where N is a number of internal apertures, can be represented in the form of potential of double layer. The boundary problem has been reduced in this case to the integral-dierential equation of the rst kind for scalar density: K = f on where K = k 1 2 n
k=0 The

, is an adjacent face for and . k k k

1 0 d, f = 2Bn . n r

It is possible to continue the equation on if the operator which average values of functions over is introduced, and following conditions are carried out: 43

where ck = const and is an external border of the surface. 0 It was shown that the operator of the equation is linear, self-adjoint and positive in L2 () that . It allowed to prove the is Hilbert space of square-integrable functions with constant value on k availability, uniqueness and stability of the solution of the equation in the operators energetic space [1] by applying the variational principle and Rieszs theorem [2]. The numerical solution of the equation is carried out by constructing the minimizing Ritzs sequence on the base of system of continuous piece-polynomial coordinate functions. This approach reduces the equation to the SLAE with the real positive-denite matrix. The software package for the numerical realization of the developed theory has been created. The comparing of the results of the software package with the results of similar packages has shown that the developed software package has higher accuracy and shorter time of computation. References 1. Astakhov V. I., Surface potentials and operators of potentials theory in Dirichlets spaces, Electromechanics, 2000, 2, pp. 118. (in Russian language) 2. Mikhlin S. G., The linear equations in partial derivatives, Moscow, Higher sch., 1977. 431 p. (in Russian language)

k k 0, M , 0 0 (M ) , M , (M ) = ck , M , k = 1, N , k k

f (M ) =

f (M ) , M , 2 0 Bn d, k = 0, N, mes( )

On the connection between the regularity of matrices over idempotent semirings and the solutions to the assignment problem Vassili N. Kolokoltsov Moscow Institute of Economics, Russia, and Department of Statistics, University of Warwick, Coventry CV4 7AL, UK vkolok@fsmail.net A start to (max, +)-algebra was seemingly rst given in [6], where it was noted that discrete equations of dynamic programming become linear in a exotic algebra, where the real numbers are considered to be equipped with two distributing binary operations ab = max(a, b) and ab = a+b, some non-trivial results for this (max, +)-algebra were obtained and the program of its systematic study was put forward. As the generalized addition is idempotent in this algebra meaning that a a = a, this algebra became the basic example of what was later called idempotent algebra (or idempotent dioid). A couple of monographs on this algebra were published to the time when V.P. Maslov observed that this exotic linearity can be fruitfully pursued further from discrete dynamic programming to the dierential Bellman equation of continuous time optimization theory, and the Idempotent analysis was initiated. It was baptized in [2] and its rst 10 years of development were summarized in monographs [3,4]. Apart from optimization theory a spectacular application of idempotency appeared in the theory of nite automata and formal languages. In honor of the 44

contribution to this eld made by the Brazilian mathematician Imre Simon, the idempotent algebras were started to be called tropical by some authors. The appearance and rapid development of tropical geometry (see e.g. reviews in [5]) gave a new twist to the development. Let B = (bij ), i, j = 1, ..., n, be a square matrix. Assignment problem consists in nding a bijection F : X = {1, ..., n} X maximizing the sum n biF (i) among all such bijections. On the i=1 other hand, to the matrix B there corresponds a mapping B : Rn Rn (denoted again by B with some abuse of notation) by the formula (Bf )i = supj {bij fj }. In the theory of discrete optimal control and discrete event systems one often encounters the problem of solving the equation Bf = g for a given vector g Rn . One says that the matrix B is strongly regular if there exists g such that the equation Bf = g has a unique solution. The long development of the theory of such equations lead eventually to the following beautiful result (obtained in [1]): B is strongly regular if an only if the corresponding assignment problem has a unique solution. In the aim of the present contribution a more general point of view on this topic will be presented aimed at the extension of this result to an innite dimensional setting, namely to the case of the state space X being a countable set. Theses results are based on the joint work of the author with M. Akian and St. Gaubert from INRIA France. References 1. P. Butkovic, F. Hevery, A condition for the strong regularity of matrices in the minimax algebra, Discrete Appl. Math. 11 (1985), 209-222. 2. V.N. Kolokoltsov, V.P. Maslov, Idempotent Analysis as a tool in control theory. Part 1, Funkts. Anal. i Prilosh. 23:1 (1989), 114. Engl. Transl. Funct. Anal. Appl. 3. V.N. Kolokoltsov, V.P. Maslov, Idempotent Analysis and its Application to Optimal Control, Moscow, Nauka, 1994 (in Russian). 4. V.N. Kolokoltsov, V.P. Maslov, Idempotent Analysis and its Applications, Kluwer, 1997. 5. G.L. Litvinov, V.P. Maslov (Eds.) Idempotent Mathematics and Mathematical Physics, Contemporary Mathematics v.377, AMS, Providence, 2005. 6. N.N. Vorobyev, Extremal algebra of positive matrices, Elektron. Informationsverarbeitung und Kybernetik 3 (1967), 39-71 (in Russian).

Dessins denfants, generalized Chebyshev polynomials, and anti-Vandermonde systems of equations Elena Kreines Moscow State University This work is done under the partial nancial support from the grant MK-2687.2007.1 kreines@itep.ru (Bicolored) dessin denfant is a compact connected smooth oriented surface M together with a (bipartite) graph on it such that the complement M \ is homeomorphic to a disjoint union of open discs. The theory of dessins denfants was initiated by A. Grothendieck in [1] and actively developed thereafter. Dessins denfants became rather popular within the last decades; they provide a possibility to describe in the easy and visually eective combinatorial language of graphs on surfaces many dicult and deep concepts and results of Group theory, Teichmller and moduli spaces, Maps u

45

and hypermaps, Matrix models, Quantum gravity, String theory, etc. Numerous applications can be found in [2,3] and references therein. Bicolored dessins denfant are canonically related with so-called Belyi pairs, a Belyi pair is an algebraic curve X together with non-constant rational function on this curve, : X CP 1 with at most 3 critical values. This function is usually called the Belyi function. In the case, it is a polynomial on C, it is called a generalized Chebyshev polynomial or a Shabat polynomial. The aforesaid correspondence is realized in the following way: if the critical values are {0, 1, }, then 1 ([0, 1]) is a dessin denfant, whose edges are { 1 ((0, 1))}, black vertices are { 1 (0)}, and white vertices are { 1 (1)}. Also for a given list of valencies (degrees) of a dessin it is possible to write down a system of algebraic equations determining the Belyi pair of this dessin. In the talk dierent properties of such systems will be discussed. In particular, it appears that among the solutions there are solutions which do not have the original list of valencies, but only a certain its sub-list. We call them parasitic solutions and say that a parasitic solution is geometric if it is a Belyi function and non-geometric otherwise. We classify all parasitic solutions. As corollaries the following results can be obtained: Theorem 1. The systems for dual dessins have dual parasitic solutions. Theorem 2. Each list of valencies has at most nite number of geometric parasitic solutions and either zero or innitely many non-geometric parasitic solutions. Theorem 3. Systems of equations for generalized Chebyshev polynomials have no parasitic solutions. In particular, it is proved that the following anti-Vandermonde system: 1 z1 + 2 z2 + . . . + n zn = 0 2 2 2 1 z1 + 2 z2 + . . . + n zn = 0 ... . . . .n1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . n1 n1 1 z1 + 2 z2 + . . . + n zn = 0

here i are given positive integers and zi are unknowns, has no parasitic solutions and has exactly (n 1)! solutions, if i = j for i = j. Several other combinatorial properties of this system will be discussed. References 1. A. Grothendieck, Esquisse dun programme, London Math. Soc. Lecture Notes Series, Cambridge Univ. Press. 242 (1997) 548. 2. S. K. Lando, A. K. Zvonkin, Graphs on surfaces and their applications, with an appendix by D. Zagier, Encycl. of Math. Sciences 141, Springer, 2004. 3. G. Malle, B. H. Matzat, Inverse Galois Theory, Springer-Verlag, Berlin, Heidelberg, New York, 1999.

Preserving zeros of Jordan triple products B. Kuzma Institute of Mathematics, Physics, and Mechanics, Jadranska 19, 1000 Ljubljana, Slovenia. bojan.kuzma@pef.upr.si 46

A Jordan triple product, A B := ABA is a natural product to consider on the space of Hermitian/Symmetric/Alternate matrices, as well as on Hilbert space eects i.e., self adjoint operators with 0 X I. Let F be a eld and n 3. Suppose S1 , S2 Mn (F) contain all rank-one idempotents. We will discuss the structure of surjections : S1 S2 satisfying ABA = 0 (A)(B)(A) = 0. Similar results are also obtained for (a) subsets of bounded operators acting on a complex or real Banach space X, (b) the space of Hermitian matrices acting on n-dimensional vectors over a skeweld D, (c) subsets of self-adjoint bounded linear operators acting on an innite dimensional complex Hilbert space. The obtained results can be applied to characterize mappings on matrices or operators such that F (ABA) = F ((A)(B)(A)) for all A, B, for functions F such as the spectral norm, Schatten p-norm, numerical radius and numerical range, etc. This is a joint work with M. Doboviek, G. Lenjak, C.-K. Li, and T. Petek. s s

Separation of variables in nonlinear Fermi equation Yu. I. Kuznetsov ICM and MG SB RAS, Novosibirsk sur@ommfao1.sscc.ru Such characteristics of seismic activity as clustering in space, in time, in seismic energy and so on, which are observed in last decades, provide evidence of the wave nature of seismic processes. Problem of modelling the elastic chain of rotating spherical blocks of equal size is reduced to solving 1D nonlinear sine-Gordon equation: 2 2 = sin(). 2 2 Variables = k0 x and = c0 k0 t are dimensionless coordinates, x is a coordinate in the chain, t is time, c0 and k0 are constants (in essence, characteristic velocity and wave number of the solution). The solution has the character of a soliton. Computations based on the sine-Gordon equation have a rather good agreement with real data corresponding to a seismic foci belt. In this work we consider another nonlinear equation [1] with a soliton solution: zk = (zk+1 2zk + zk1 ) 1 + (zk+1 zk1 ) , (1)

z0 = zN +1 = 0, k = 1(1)N , describing N uniformly spaced equal masses xed on a thread. Here zk = zk (t) is a deviation of k-th mass from the equilibrium state. In this system, spatial and temporal variables can be separated. It is known [1] that the oscillatory system described by (1) has a non-linear behavior: the initial momentum energy does not dissipate in the phase space with time, but is focused in a group of harmonics. In this paper, by analogy to [1], we use the basis of the linear problem (1) for = 0. However a special transformation is used for the coecients of decomposition. As a result, we get an equation for 47

the Fourier coecients which is independent from the considered harmonics. We get new conclusions arming that the dynamics of discrete systems is more rich than that of continuous systems [2]. Let Z(t) = (z1 (t), . . . , zN (t))T and
N

Z(t) =
i=1 (i) (i) (i) (i)

ci (t)Y (i) ,

(2)

where Y (i) = (y1 , . . . , yN )T , y0 = 0, yN +1 = 0 is a solution of the homogeneous equation yk+1 2yk + yk1 = i yk ,
(i) (i) (i) (i)

i = 1(1)N.

(3)

Here i is the squared frequency, Y (i) is the amplitude vector of the linear oscillator described by the equation (1) for = 0. Representation (2) allows to make a transformation which leads to the following statement. T Theorem.The vector C = c1 (t), . . . , cN (t) is bound by the following relation: C = (I + (BT BH )C, (4)

where I is an identity matrix, = diag(1 , . . . , N ), j are eigenvalues appearing in equation (3), BT is a symmetric Toeplitz matrix and BH is a persymmetric Hankel matrix: 0 t1 t2 tN 2 tN 1 t1 0 t1 tN 2 . .. .. . t2 . . t1 . . .. BT = . , . . .. . t1 t2 tN 2 t1 0 t1 tN 1 tN 2 t2 t1 0 t2 t3 tN 1 tN 0 t3 0 tN .. . tN 1 . . . .. . BH = . . . . . . tN 1 .. .. tN 0 t3 0 tN tN 1 t3 t2 References 1. E. Fermi, Collected papers (Note e memorie), University of Chicago Press, 1965. V. 2. 2. V. K. Mezentsev, S. L. Musher, I. V. Ryzhenkova, S. K. Turitsyn, Two-dimensional solitons in discrete systems, JETP Letters, 60 (11) (1994), 815821.

48

On representation of the polynomials of best approximation with the weight and their evaluation Lebedev V. I. Kurchatov Institute; Institute of Numerical Mathematics of RAS gourmet@inm.ras.ru Let w(x), f (x) C[1, 1] be respectively the positive weight, and the function to be approximated. We consider the classical minimax problem. To nd en , the minimum of the uniform norm ||(f (x) Qn (x))w(x)||C[1,1] , among all degree n polynomials Qn (x), as well as the polynomial Pn (x) where this minimum is attained. The latter polynomial is known as the best polynomial approximation of f (x) with the weight w(x) or, briey, PBA. Let rn (x) = (f (x) Pn (x))w(x), x = (x1 , . . . , xn+1 ), xi [1, 1], xi = xk , n+1 (x, x) = n+1 = i=1 (x xi ), and Ln (x, x) be the n-th degree Lagrange polynomial for f (x) with interpolation points x. Then Pn (x) = Ln (x, y ) = arg min max |n+1 (x, x)Ln (x, x)w(x)|.
x x[1,1]

Let x = cos , then the following theorem is true. Theorem. The polynomial Pn (x) is PBA if and only if there exist real E, integer m n and a function () C[0, ], 0 (0) > , 0 () < such that, rn (x) = E cos((m + 1) + ()). In this setting en = E. Zeros of rn (x) are the points yi = cos i , i = ((i 1/2) (i ))/(m + 1), i = 1, m + 1.

The representation of the error of approximation in trigonometric form in terms of the phase function () is the basis of the iterative method. The interpolation quadrature formula for the singular integrals with Hilbert kernel is used for the corrections of the phase function. This formula together with the formula for the zeros of the error determines the algorithm of collective dynamics of zeros and e-points when the weight function w(x) is disturbed. The iterative methods for evaluation of the PBA parameters based on the methods of inverse analysis, perturbation theory and asymptotic formulae are formulated. Numerical experiments show the high eciency of the proposed approach. References 1. Lebedev V. I., Proc. Lobachevskii Math. Inst., v. 20, Kazan, 2003, 3786. 2. Lebedev V. I., Sb. Math. v. 195, 10, 2004.

49

Fast numerical solution methods for two-dimensional Fredholm integral equations of the second kind Fu-Rong Lin , Wen-Jing Xie , Feng Liang Department of Mathematics, Shantou University, Shantou Guangdong 515063, China frlin@stu.edu.cn In this paper we consider numerical solution methods for two dimensional Fredholm integral equation of the second kind

f (x, y)

a(x, y, u, v) f (u, v) dudv = g(x, y),


(x, y) [, ] [, ],

where a(x, y, u, v) is smooth and g(x, y) is in L2 [, ]2 . We discuss polynomial interpolation methods for four-variable functions and then use the interpolating polynomial to approximate the kernel function a(x, y, u, v). Based on the approximation we deduce fast matrix-vector multiplication algorithms and ecient preconditioners for the above two dimensional integral equations. The residual correction scheme is used to solve the discretization linear system. Numerical results are given to illustrate the eciency of our algorithms.

Dequantization of linear operators G. L. Litvinov , G. B. Shpiz Independent University of Moscow glitvinov@mail.ru For functions dened on Cn (or Rn or Rn ) the well-known Maslov dequantization generates the + following dequantization transform: f f (x) = lim h log(| f (exp(x/h) |) = lim (1/s) log(| f (exp(sx) |),
h0 s+

where x Rn , and h, s = 1/h are real parameters. If f is a polynomial, then the subdierential of f at the origin coincides with the Newton polytope of f ; it is possible to generalize this result to a wide class of functions and convex sets (see, e.g., [1] for details). The Maslov dequantization generates a passage from the traditional mathematics to the so-called tropical mathematics. Our aim is to apply the dequantization transform to matrix elements of operator semigroups generated by linear operators. Suppose that S is a semigroup and s s is a linear representation of S in a complete (or quasicomplete) barreled locally convex space (by continuous operators). Denote by V the dual space to V and by v , v the value of a functional v V on an element v V . If s v ,v (s) = v , s v is a matrix element of , then its dequantization v ,v is dened by the formula: v ,v = lims (1/s) log(| v , s v |). We discuss the cases S = R+ or S = Z+ . If v and v are xed, then v ,v S {}. 50

Proposition 1. Let A be a linear operator in V , s = exp(sA), and dim V < . Then the set of all dequantizations {v ,v } coincides with the set of real parts of all eigenvalues of A. Suppose that for every v V there exists a number r > 0 such that the set {r s (| v , s v | ), s S} is bounded for every v V and the number r does not depend on v V . Then the representation is called exponential. Note, that if V is a Banach space and is weakly continuous, then is exponential. In the general case the spectral radius of is dened by the formula: = inf{r | r s s v 0 weakly for every v V as s +} Proposition 2. If A is a bounded linear operator in a Banach space V , S = Z+ , = As , then = (A) = lims As 1/s , i.e. is the traditional spectral radius of A. Theorem 1. If is exponential, then log = sup{v ,v | v V , v V }. Theorem 2. Suppose that A is a compact operator and s = As , where s S = Z+ . Then the set {v ,v } of all dequantizations of coincides with the set of all numbers of the form log(| |), where runs the spectrum of A. The work is supported by the RFBR-CNRS grant 05-01-02807. References 1. G. L. Litvinov, Maslov dequantization, idempotent and tropical mathematics: A brief introduction. J. Math. Sci., 2007, vol. 140, 3, pp. 426444. There are generalizations of this result for the case dim V = .

Linear methods and error autocorrection in rational approximation G. L. Litvinov Independent University of Moscow glitvinov@mail.ru Let {0 , 1 , . . . , n } and {0 , 1 , . . . , m } be two collections consisting of linearly independent functions of an argument x belonging to some (possibly multidimensional) set X. Consider the problem of constructing an approximant of the form R(x) = P (x) a0 0 + a1 1 + + an n = b0 0 + b1 1 + + bm m Q(x) (1)

to a given function f (x) dened on X. Let an abstract construction method for the approximant of the form (1) be linear in the sense that the coecients of (1) can be determined from a homogeneous system of linear algebraic equations Hy = 0, (2) where H is a matrix of dimension (m + n + 2) (m + n + 1) and y = {a0 , a1 , . . . , an , b0 , b1 , . . . , bm }. Of course, an additional condition (e.g., of the type b0 = 1, or bm = 1, or a0 = 1) can be added. 51

Pad approximations, linear PadChebyshev approximations etc. (and even best approximations) e e could be treated by this way (see, e.g., [1]). The system of equations can be very ill-conditioned and it can lead to very signicant errors in the approximant coecients. However, these errors do not aect the accuracy of the approximant (the error autocorrection eect, see, e.g., [1]). Let the coecients ai , bj give an exact or an approximate solution of (2) and let ai , bj give another approximate solution obtained in the same way. Set ai = ai ai , bj = bj bj ; these errors arise due to perturbations of the approximated function f (x) or due to calculation n m errors. Set P (x) = i=0 ai i , Q(x) = j=0 bj j . It is easy to see that the vector P y = {a0 , a1 , . . . , an , b0 , . . . , bm } is an approximate solution of (2). So the function Q can be treated as an approximant to the same function f (x). The following uncertainty relation is valid: ( P P P P ) = Q( ) Q ( f ) , Q Q Q Q

Q P where Q = Q+Q is the relative error of Q, the dierence Q f is the absolute error of the approximant P to the function f , and is the absolute theoretical error of our method; the Q argument x can be treated as xed.

The error autocorrection eect means that in a calculation all the intermediate calculation errors compensate each other, so the nal result is much more accurate than the intermediate results. In this case standard interval estimates (e.g., in the framework of Yu. Matijasevichs a posteriori interval analysis) are not realistic, they are too pessimistic. The error autocorrection eect appears in some popular numerical methods, e.g., in the least squares method. References 1. G. L. Litvinov, Error autocorrection in rational approximation and interval estimates. Central European Journal of Mathematics, vol. 1, 1, 2003, p. 3660. See also arXiv:math.NA/0207183 (http://arXiv.org).

Overcoming principal uncertainties in calibration of matrix population models Dmitrii O. Logofet Institute of atmospheric physics RAS, Moscow State University, Mechanics and Mathematics Dept. danilal@postman.ru Matrix models of stage-structured population dynamics take on a general form of dierence equations, x(t + 1) = Ax(t), where x(t) denotes the n-vector of population structure, i.e., the vector of abundances of n stagespecic groups of individuals within a population, and the projection matrix A consists of stagespecic vital rates, the matrix pattern reecting a variety of life cycles for individuals of a species under study. When the projection matrix of a model for perennial plants, like, e.g., raspberry Rubus idaeus colonizing forest clear-cut areas, is to be calibrated on data, the very sense of the vital rates 52

suggests monitoring of marked individual plants during successive years, thereafter calculating the rates directly from the data. This can be eectively done for stage-specic survival and ontogenetic transition rates. But the rates of reproduction bear principal uncertainty as far as the parent plants can hardly be determined for the (not yet marked!) recruitment. To overcome this insurmountable uncertainty in data and to conclude the model calibration, we apply a general principle that suggests substitution an adaptation conjecture (empirical generalization) for the lack of data/information. This principle of adaptive optimality implies that a set of the stage-specic reproduction rates can be found such that maximizes the overall population growth rate (the dominant eigenvalue of matrix A) under certain constraints on the rates which ensue from expert knowledge of population biology of the species in hand and from the data. The original constrained optimization problem is of a nonlinear kind, with another principal uncertainty in the local-vs.-global character of a solution to be found by a numerical routine. To overcome this kind of uncertainty, we substitute maximization of the potential growth indicator function for the original problem. The latter reduces to a linear programming problem, whose global solution can always be found.

Calculation of the Schein rank for Boolean matrices E. E. Marenich Murmansk State Pedagogic University marenich1@yandex.ru Let P = {0, 1} be a two-element Boolean lattice, P mn a set of all m n matrices over P , and A P mn . According to K.X. Kim [1], the denition of the Schein rank over a Boolean lattice P = {0, 1} was given by Schein. For denition of the Schein rank over semirings we refer to K.X. Kim and F.W. Roush [2]. Calculation methods for the Schein rank over chains was considered in [3], [4] by Di Nola and S.Sesa. Let = (V1 V2 , E) be a bipartite graph, and let U be a nite set. A function f : V1 V2 2U is called an U -subsets coding function for , or more simply, coding function, if for any vertices v1 V1 , v2 V2 the conditions {v1 , v2 } E and f (v1 ) f (v2 ) = are equivalent. We call f (v) the code of v V . Note that there exist coding functions for any bipartite graph . Denote by S the family of all nite U such that there exist U -subsets coding functions for . The intersection number of the bipartite graph is nint () = min |U |.
U S

A clique in a bipartite graph is any maximal complete bipartite subgraph containing at least one edge. Let and {1 , . . . , k } be a bipartite graph and a family of its subgraphs. We say that subgraphs 1 , . . . , k cover if each vertex and each edge of are contained in at least one subgraph. In the following theorem we show that intersection number calculations for any bipartite graph can be reduced to counting the number of cliques covering the graph. Theorem 1. Let = (V1 V2 , E) be a bipartite graph without isolated vertices. Then the intersection number of is equal to the least number of cliques covering . 53

Let be obtained from by deleting all isolated vertices. Then nint () = nint ( ). For any matrix A there exists an associated bipartite graph (A) = (V1 V2 , E) such that V2 = {1 , 2 , . . . , n }, {i, j } E i aij = 1. The following theorem reduces the Schein rank problem for any matrix A to the intersection number calculation for the associated bipartite graph (A). Theorem 2. Let P = {0, 1} and A P mn be a two-element Boolean lattice and a matrix. Then the Schein rank of A is equal to the intersection number of the associated bipartite graph (A). Corollary 1. Let P = {0, 1} and A P mn be a two-element Boolean lattice and a matrix. Then the Schein rank of A is equal to the least number of cliques covering , where is obtained from the associated bipartite graph (A) by deleting all isolated vertices. From Corollary 1 it obviously follows that the Schein rank of the identity matrix E = Enn nn equals n. The following corollary gives the criterion for A P mn to be a CR-matrix. P Corollary 2. Let P = {0, 1} be a two-element Boolean lattice, A P mn a nonzero matrix, and (A) the associated bipartite graph for A. The following conditions are equivalent: i) The Schein rank of A is equal to 1. ii) is a complete bipartite graph, where is obtained from (A) by deleting all isolated vertices. Theorem 2 reduces the Schein rank calculation for any matrix A to the calculation of the least number of cliques covering the graph, i.e., to the problem of minimum covering. One more application of Theorem 1 follows. Theorem 3. Let P = {0, 1} and E = Enn P nn be a two-element Boolean lattice and the identity matrix. Let E denote the complement of E. Then the Schein rank of E is equal to the minimal k such that k n . [k/2] This research is conducted in accordance with the Thematic plan of Russian Federal Educational Agency, theme 1.03.07.

References 1. Kim K. X., Boolean matrix theory and applications, Marcel Dekker, New York, 1982. 2.Kim K. X., Roush F. W., Generalized fuzzy matrices, Fuzzy Sets Systems 4 (1980), 293315. 3.Di Nola, Sessa S., On the Schein rank of matrices over linear lattice, Linear Algebra Appl. 118(1989), 155158. 4.Di Nola, Sessa S., Determining the Schein rank of matrices over linear lattice and nite relational equations, The Journal of Fuzzy Mathematics, vol. 1, No. 1, 1993, 3338.

Lattices of matrix rows and matrix columns V. E. Marenich Murmansk State Pedagogic University VMarenich@yandex.ru The following notations will be used throughout. Denote by (P, , , ) a lattice. Let P mn be a set of all m n matrices over P and A P mn . 54

If for given elements a, b P the greatest solution of the inequality a x b exists, then it is b b b denoted by a P = a , and is called the relative pseudocomplement of a in b. If a exists for all a, b P then (P, , , ) is called a Brouwerian lattice. The partially ordered set (P m1 , ) is a lattice with meet and join operations denoted by and . Any partially ordered set is called, more simply, a poset. Let A = ||aij || P mn . We call A(i) = (ai1 , . . . , ain ) the i-th row vector and A(j) = (a1j , . . . , amj )t the j-th column vector. The linear span of row vectors is called a row subspace for A and is denoted by Row(A). The linear span of column vectors is called a column subspace for A and is denoted by Column(A). Dene posets: (Column(A), ) with respect to the partial order induced by the lattice (P m1 , ) and (Row(A), ) with respect to the partial order induced by the lattice (P 1n , ). The poset (Column(A), ) is an upper semilattice, in which join coincides with the join operation in the lattice (P m1 , ), uv = u + v = u v, u, v Column(A). Suppose (P, , , ) is a complete distributive lattice. If for any subset S, S Column(A), there exists the least upper bound S Column(A), then the poset (Column(A), ) is a complete lattice, in which join coincides with the join operation in the lattice (P m1 , ). If (P, , , ) is a nite distributive lattice, then (Column(A), ) is a lattice. In some cases we can express the meet operation in the lattice (Column(A), ) by the following formulas. Theorem 1. Let (P, , , ) be a lattice, A P mn . i) If (P, , , ) is a complete Brouwerian lattice, then (Column(A), ) is a complete lattice with meet and join operations, and , where S = A S , S Column(A). A ii) If (P, , , ) is a Brouwerian lattice, then the poset (Column(A), ) is a lattice with meet and join operations, and , where v u uv = A uv = A A A , u, v Column(A). A iii) If (P, , , ) is a Boolean lattice, then the poset (Column(A), ) is a lattice with meet and join operations, and , where uv = A t A (u + v), u, v Column(A). Sometimes we can express the meet operation more simply. Theorem 2. Let (P, , , ) be a Brouwerian lattice and matrix A P nn an idempotent. Then i) the meet operation in the lattice (Column(A), ) is given by uv = A (u v), u, v Column(A); ii) (Column(A), ) is a distributive lattice. Note that a similar theorem holds for the lattice (Row(A), ). A matrix A is called a regular matrix if there exists B P nn such that ABA = A. According [1], A is a regular matrix whenever there exists an idempotent C such that Column(A) = Column(C). For the following result we refer to K.A. Zaretski, [2]. Theorem (Zaretski). Let P = {0, 1} and A P nn be a two-element Boolean lattice and a matrix. Then (Column(A), ) is a distributive lattice whenever A is a regular matrix. Zaretskis theorem can be extended to Brouwerian lattices. Theorem 3. Let (P, , , ) be a Brouwerian lattice, A P nn a regular matrix, and C P nn an idempotent such that Column(A) = Column(C). Then i) the meet operation in the lattice (Column(A), ) is given by uv = C (u v), u, v Column(A), ii) (Column(A), ) is a distributive lattice.

55

For semigroups of binary relations, the statement ii) of Theorem 3 was proved by K.A. Zaretski, see [2]. The similar result for fuzzy lattices was proved by K.X. Kim and F.W. Roush, see [3]. This research is conducted in accordance with the Thematic plan of Russian Federal Educational Agency, theme 1.03.07.

References 1. K. H. Kim, Boolean matrix theory and applications, Marcel Dekker, New York, 1982. 2. K. A. Zaretski, Regular elements in the semigroup of binary relations, Uspekhi mat. nauk, 17(3): 105108, 1962. 3. K. X. Kim, F. W. Roush, Generalized fuzzy matrices, Fuzzy Sets Systems 4 (1980), 293315.

Matrix algebras and their length O. V. Markova Moscow State University, Mathematics and Mechanics Dept. ov markova@mail.ru Let F be a eld and let A be a nite-dimensional F-algebra. A number l(S) is called a length of a nite generating set S of this algebra provided it equals the smallest number k, such that words of the length not greater than k generate A as a vector space. The length of the algebra A, denoted by l(A), is the maximum of lengths of all its generating sets. The problem of evaluating the length of the full matrix algebra in terms of its order was posed by A. Paz in [3] and has not been solved yet. Some known upper bounds (see [3,2]) are not linear. Using our results [1] on the length of upper-triangular matrix algebra and on the upper and lower bounds for the lengths of direct sums of algebras, we have evaluated lengths of certain uppertriangular matrix subalgebras. The linear upper bound for the length of commutative subalgebras of Mn (C) was obtained in [3]. We have generalized this result on commutative matrix subalgebras over an arbitrary eld. Also we have obtained the following characterization of commutative matrix subalgebras of the maximal length: Theorem 1. Let F be an algebraically closed eld. Commutative subalgebra R in Mn (F) is of the maximal length if and only if it is generated by a non-derogatory matrix. A positive answer has been obtained to the question whether the length of an algebra can be smaller than the length of its subalgebra. That is, for any natural number k and for an arbitrary eld F there exist such F-algebra A in upper triangular matrix algebra and its subalgebra A , that l(A ) l(A) = k. I would also like to thank my supervisor Dr. A. E. Guterman for the statement of the problem and for useful discussions. References 1. O. V. Markova, On the length of upper-triangular matrix algebra. Uspekhi Matem. Nauk, 60(2005), no. 5, 177178. [in Russian]; English translation: Russian Mathematical Surveys, 60(2005), no. 5, 984985. 56

2. C. J. Pappacena, An Upper Bound for the Length of a Finite-Dimensional Algebra, Journal of Algebra, 197(1997), 535545. 3. A. Paz, An Application of the CayleyHamilton Theorem to Matrix Polynomials in Several Variables, Linear and Multilinear Algebra, 15(1984), 161170.

A fast algorithm for computing the nullspace of some structured matrices Nicola Mastronardi , Marc Van Barel , Raf Vandebril per le Applicazioni del Calcolo M. Picone Bari, Italy Katholieke Universiteit Leuven, Department of Computer Science, Belgium n.mastronardi@ba.iac.cnr.it, Marc.VanBarel@cs.kuleuven.ac.be, raf.vandebril@cs.kuleuven.be
Istituto

The problem of computing the nullspace of (block) Toeplitz/Hankel matrices arises in dierent applications, such as system theory, linear prediction and parameter estimation. Many algorithms have already been available in the literature to compute the nullspace of the latter structured matrices. In this talk an algorithm based on the generalized Schur algorithm will be presented. Similarities and dierences with the existing algorithms will be also emphasized.

Algebraic Multigrid Methods and Block Preconditioning for Mixed Elliptic Hyperbolic Linear Systems, Applications to Stratigraphic and Reservoir Simulations Roland Masson , Y. Achdou , P. Bonneau , P. Quandalle Institut Franais du Ptrole, c e University Paris VII roland.masson@ifp.fr In multiphase ow in porous media, and stratigraphic modeling, the linear systems obtained after nite volume discretization in space, fully implicit time integration and Newton type linearization couple basically an elliptic/parabolic variable (say the pressure) with hyperbolic variables (say the compositions or saturations). For such systems, unknown based preconditioners [1,2,3] allow to adapt eciently the preconditioner to each type of variable using algebraic multigrid [4] for the pressure equations and e.g. incomplete factorizations for the remaining variables. The coupling between the variables is obtained by multiplicative combination of both the multigrid and incomplete factorization preconditioners. In this talk, such methods will be detailed and results will be presented on reservoir simulation, basin simulation, and stratigraphic modeling examples. In reservoir simulation, implicit well equations and high compressibility lead to strongly non diagonal dominant additional equations for the pressure block which usually have a bad eect on

57

the performance of algebraic multigrid methods. We will discuss several solutions to recover the eciency of the combinative method in such cases. References 1. S. Lacroix, Y.V. Vassilevski, and M.F. Wheeler, Decoupling preconditioners in the Implicit Parallel Accurate Reservoir Simulator (IPARS), Numerical Linear Algebra with Applications 8 (2001) pp. 537549. 2. R. Scheichl, R. Masson, J. Wendebourg, Decoupling and block preconditioning for sedimentary basin simulations, Computational Geosciences 7, pp. 295318, 2003. 3. Y. Achdou, P. Bonneau, R. Masson, P. Quandalle, lgebraic Multigrid Based Preconditioners for Oil Reservoir Simulation, ECMOR X Proceedings, Amsterdam, 2006. 4. J.W. Ruge and K. Stuben, Algebraic Multigrid (AMG), In Multigrid Methods (S.F. McCormick, ed.), Frontiers in Applied Mathematics, vol. 5, SIAM, Philadelphia, 1986.

On the sum of elements of a nonnegative two-by-two matrix Jorma K. Merikoski Department of Mathematics, Statistics and Philosophy, FI-33014 University of Tampere, Finland jorma.merikoski@uta.fi Let A be a nonnegative two-by-two matrix with row sums r1 and r2 , and column sums c1 and c2 . Let m be a nonnegative integer, and let su B denote the sum of elements of a matrix B. Drury, Merikoski, Laakso and Tossavainen [Electron. J. Linear Algebra 10 (2003), 280290] proved that if m m m is even, then su Am (r1 c1 ) 2 + (r2 c2 ) 2 . The proof is surprisingly hard. Some further questions, motivated by this inequality, are discussed.

Application of the multigrid method to solving diusion-type equations M. E. Ladonkina, O. Yu. Milyukova, and V. F. Tishkin Institute for Mathematical Modeling, Russian Academy of Sciences Miusskaya sq. 4a, Moscow, 125047 Russia miliukova@imamod.ru Numerical modeling of many problems of mathematical physics, such as gas dynamic or hydrodynamic problems, is impossible without taking into account diusion processes. Describing diusion processes involves solving parabolic equations. The application of explicit schemes to approximate parabolic-type equations imposes strong constraints on the time step. The application of implicit schemes removes these constraints, but it requires solving large systems of linear algebraic equations, which renders using such schemes unexpedient. We suggest a dierence algorithm for solving the initial-boundary value problem for the heat equation T = div k grad T (1) cv t 58

on the basis of the multigrid method [1]. In (1), T is the temperature, cv is the coecient of heat capacity, and k is the heat conduction. The value of the grid function at the (n + 1)th time layer is determined as follows [2]. First, one or several simple smoothing iterations for the equations representing the purely implicit scheme on a ne grid with step h Ah uh = fh . (2) are performed. The discrepancy thus obtained is projected onto a coarse grid with step H = 2h, on which the equation for corrections AH H = RH (3) is solved; here, RH is the discrepancy on the coarse grid. The correction H is then interpolated onto the ne grid and used to evaluate the grid function at the (n + 1)th time layer by the formula un+1 = us h , where h is the correction on the ne grid and us is the solution obtained at the h smoothing stage. Thus, only one iteration of the two-grid cycle is performed. Since solving equation (3) on the coarse grid requires much fewer operations than solving equation (2) and, moreover, only one or several smoothing iterations are performed, the suggested method for solving the multidimensional heat equation requires less computational time than the implicit scheme on the ne grid (that is, solving equation (2)). The stability and accuracy of the suggested algorithm were theoretically studied for model problems by using Fourier analysis. For a one-dimensional model problem, the absolute stability of the method with respect to the initial data and the right-hand side was proved [2]. For a two-dimensional model problem, it was shown that if = h and 0 < < 2, then uh un+1 L2 = O((h2 )h2(13.5) ), where is the time step and 0 < < 1/3.5 is an arbitrary given real number. Thus, the error uh un+1 L2 has higher order than uh uT L2 , where uT is the projection of the exact solution to equation (1) onto the ne grid, and the suggested algorithm makes it possible to obtain a very good approximation to the exact solution by using the implicit scheme on the ne grid. The theoretical results were conrmed by computations of model problems, including two-dimensional ones. Computations of model problems with discontinuous heat conduction also demonstrated the stability and good accuracy of the method. References 1. Fedorenko R. P., A relaxation method for solving dierence elliptic equations, Zh. Vychisl. Mat. Mat. Fiz. 1961. Vol.1. 5. P.922927. 2. Ladonkina M. E., Milyukova O. Yu., Tishkin V. F., A numerical algorithm for diusiontype equations based on the multigrid methods, Mat. Model. 2007. Vol.19. 4. P.7189.

New Smoothers in Multigrid Methods for Strongly Nonsymmetric Linear Systems Evgenia M. Andreeva, Lev A. Krukier, Galina V. Muratova South Federal University Rostov on Don, Russia This research is supported by the Russian Foundation for Basic Research (RFBR) under the Grants No. 06-01-39002, 06-01-00038. muratova@rsu.ru 59

Multi-grid methods (MGMs) are fast iterative solvers based on the multilevel or the multi-scale approach. A typical application of the multigrid method is to numerical solution of partial dierential equations in two or higher dimension. Alternatively, the MGM can be applied in combination with any of the common discretization techniques; in such case, it is among the fastest solution techniques known today. Multigrid method does not mean only a xed multigrid algorithm applying to all boundary value problems in the general setting, and it may usually refer to an algorithmic framework of the multigrid techniques. The eciency of the MGM heavily depends on the adjustment of the involved components, e.g., the basic linear solver or the smoothing procedure. We consider new smoothers resulted from a special class of triangular skew-symmetric splitting iteration methods for the multigrid methods used to solve the systems of linear equations with strongly nonsymmetric coecient matrices, which may be produced by the central-dierence approximation of a convection-diusion equation with dominant convection. In specic, we consider the system of linear equations Ay = f, (1)

where A is a large sparse nonsymmetric matrix. Everywhere we assume that the matrix A is positive real, i.e., its symmetric part A0 = 1 (A + AT ) is positive denite. Here, we denote by A = A0 + A1 , 2 with A0 = 1 (A + AT ) = AT > 0 and A1 = 1 (A AT ) = AT . Note that A1 is the skew-symmetric 0 1 2 2 part of the matrix A. Besides, we assume that the matrix A is strongly nonsymmetric, i.e., it holds that A0 A1 in the sense of certain matrix norm. We solve the linear system (1) by the multigrid methods with the smoothers resulted from the triangular skew-symmetric splitting iteration methods TKM1 and TKM2 described as follows: B TKM1: yn+1 yn + Ayn = f, (2) (3) (4)

B = E + 2K or B = E + 2Ku ,

TKM2: B = i E + 2K or B = i E + 2Ku ,

where is an iterative parameter, and K and Ku are, respectively, the strictly lower- and uppertriangular parts of the skew-symmetric matrix A1 . We remark that TKM1 is the triangular skew-symmetric splitting iteration method without parameter, and TKM2 is a suitable modication about it. These two splitting iteration methods are applicable to linear systems whose coecient matrices may be not diagonal dominant. We suggest the following formulas for choosing the parameters involved in both TKM1 and TKM2: = M , i = n |mij |, i = 0, n, j=0 where M = (mij ) := A0 + Ku K is the symmetric matrix. The convergence of the multigrid methods with the TKM1 and the TKM2 smoothers are proved [2], and some theoretical and numerical results about the convergence of the MGM with dierent smoothers are also presented. Acknowledgments: The authors are very much indebted to Prof. Zhong-Zhi Bai for his useful comments and suggestions, in particular, for his kind invitation of visiting State Key Laboratory of Scientic/Engineering Computing, Academy of Mathematics and Systems Science, of Chinese Academy of Sciences, P.R. China, during August-September in 2007, under the support of ChinaRussia (NSFC-RFBR) International Cooperative Research Project. 60

References 1. L. A. Krukier, L. G. Chikina, T. V. Belokon, Triangular skew-symmetric iterative solvers for strongly nonsymmetric positive real linear system of equations, Appl. Numer. Math., 41(2002), 89105. 2. G. V. Muratova, E. M. Andreeva, Multigrid solver with dierent smoothers for convectiondiusion problem, in Proceedings of International Summer School Iterative Methods and Matrix Computations, RSU publisher, Rostov on Don, (2002), pages 437443.

Stabilization of nite-dierence scheme for Stokes problem M. Olshanskii, E. Muravleva Dept. Mechanics and Mathematics, Moscow State University, Leninskije gory 1, Moscow 119899, Russia catmurav@gmail.com We present a new stabilized method for the Stokes problem. Staggered grids (components of the velocity eld are computed in grids crosspoints and pressure in the cells centers) are considered. We deal with second order discrete gradient operator which leads to high precision results. It has bidimensional kernel in 2D case and mesh-dependent kernel in 3D case (number of spurious pressure modes linearly depends on the quantity of discretization points). Due this fact LBB stability condition violates. Our method is independent of the space dimension and does not require choosing any mesh-dependent parameters. It always leads to symmetric linear systems. The new method is unconditionally stable and has simple and straightforward implementations. The stabilization strategy relies on projection operator whose action can be evaluated locally at the mesh level using only standart nodal data structures. As a result, an existing code can be easily modied to handle the stabilization procedures. So, we solve the system A BT B C u p = f g

where C is a sparse matrix composed of number of blocks, and each block is associated with any node. Numerical experiments were held not only for traditional viscous uid, but also for more complicated non-newtonian models (Bingham uids).

On a class of Singular Structured Matrices with Integer Spectra Tatjana Nahtman , Dietrich von Rosen of Mathematical Statistics, University of Tartu, Estonia Department of Biometry and Egineering, Swedish University of Agricultural Sciences dietrich.von.rosen@bt.slu.se
Institute

The objective of this paper is to consider a class of structured singular nonsymmetric matrices with integer spectrum which has applications in statistical sampling theory. The spectrum of the 61

matrices, and right and left eigenvectors are presented. The matrices of eigenvectors are connected to Vandermonde matrices. We shall also discuss several interesting properties of the class of matrices, for example, several triangular factorizations.

Operator equations for eddy currents on singular carriers J. A. Naumenko Southern Scientic Centre of Russian Academy of Science, Rostov-on-Don, Russia jan@novoch.ru There is a number of various practical problems that needs to compute numerically magnetic elds in the medium with the multiconnected crack. Lets consider the space L2 (S) that consists of two-component complex square-integrable on S vector functions. We suppose that the multiconnected surface S and its boundary satisfy the (C) Lipschitzs conditions. The space L2 (S) can be decomposed to the sum: L2 (S) = L2 (S) () () () (C) L2 (S) L2 (S). Here L2 (S) consists of potential elds generalized by Weyl [1], L2 (S) consists () of generalized solenoidal elds and L2 (S) consists of generalized harmonic elds. Lets designate (C) () L = L2 (S) L2 (S). We use the orthoprojector P = PL PS below where PS vanishes normal to S eld component and PL is orthoprojector L2 (S) L. We consider the following operator equations on S: K = f1 = K + f2 . = Here K = P, =
1 4 rNM

(1) (2) (3)

K + f3 (t). t

dSN , is some imaginary parameter, f1 , f2 , f3 (t) L.

We proof the existence, uniquiness and numerical stability of the equation (1) solution in the 1/2 1/2 energy space LK of operator K in the most useful case f1 B2 (S) [2,3]. Here B2 (S) is the Besov space [4]. The existence, uniquiness and numerical stability of the equation (2) solution in L are proofed by the theory of Fredholm. f2 L is enough. The similar properties of the equation (3) solution is proofed by the spectral method. The operator K is linear, sel-ajoint and positive. This propeties open broad possibilities for eective numerical solving of equations (1)(3). The software package was built on the basis of the described theory. Some practical problems of electrodynamics were computed by it. References 1. Weyl H., The method of orthogonal projection in potential theory, Duke Math. J. 1940. V. 7. P. 411444. 2. Naumenko J. A., Astakhov V. I., Mathematical modeling of magnetic eld in presence of ideal conducting surface with boundary, Electromechanics, 2003. 5. P. 1116.

62

3. Naumenko J. A., Local attribute of correctness and numerical computing of integral equation of rst kind for density of eddy currents, The bulletin of the Southern Science Centre, 2005. 4. P. 38. 4. Besov O.V., Ilin V.P., Nikolsky S. M., Integral representations of functions and embedding theorems, Moscow, Nauka, 1996. 480 p.

Augmented Lagrangian approach for saddle-point problems with indenite (1, 1)-block Maxim A. Olshanskii Lomonosov Moscow State University maxim.olshanskii@mtu-net.ru We discuss block preconditioners for a system of linear algebraic equations of saddle-point type having indenite and not necessarily symmetric (1, 1) block. The approach is based on considering an augmented system and a special multilevel method as an inner solver for a preconditioned Krylov subspace method. To explain convergence properties of the method we rst prove bounds on the eigenvalues of the augmented (1, 1) block. Next we consider the application of the method to solve a problem arising in the linear instability analysis for the Navier-Stokes equations. Several new results are proved for the Schur complement of the linearized Navier-Stokes equations. This provides further insight into the convergence properties of the approach.

Quasiseparable matrices/polynomials and Lipschitz stability of Jordan bases Vadim Olshevsky University of Connecticut, USA olshevsky@math.uconn.edu This talk consists of two parts that are not related to each other. (i) In the rst part several recent algorithms that exploit properties of matrices having special quasiseparable structure will be described. In particular, several generalizations of the BjorckPereyra and Traub algorithms for quasiseparable-Vandermonde matrices will be presented. An interpretation via signal ow graphs will be described to fully characterize the classes of quasiseparable, semiseparable and well-free matrices. Joint work with Tom Bella, Yuli Eidelman, Israel Gohberg and Eugene Tyrtyshnikov. (ii) In the second part we study rst partial stability of the Jordan chains matrices under general perturbations. Under certain conditions, Lipschitz-type bounds are obtained. Secondly, we study Jordan-structure-preserving perturbations of matrices selfadjoint in the indenite inner product. The main result is Lipschitz stability, under small perturbations, of canonical Jordan bases (i.e., eigenvectors and generalized eigenvectors enjoying a certain ipped orthonormality 63

relation) of matrices selfadjoint in the indenite inner product. The proof relies upon the analysis of small perturbations of invariant subspaces, where the size of a perturbation of an invariant subspace is measured using the concepts of a gap and of a semigap. Joint work with Tom Bella and Upendra Prasad.

A new class of low-tensor rank matrices closed under inversion Ivan Oseledets , Eugene Tyrtyshnikov , Nikolay Zamarashkin Institute of Numerical Mathematics of RAS, Moscow, Russia ivan.oseledets@gmail.com, tee@inm.ras.ru, kolya@bach.inm.ras.ru Low tensor-rank matrices play increasingly important role in design of fast algorithms for dealing with huge dense unstructured matrices. It is a common belief that such matrices can be the main tool in breaking the curse of dimensionality. The approximation techniques by low-tensor rank matrices are now well-developed. However, the inversion of such matrices is treated only in the iterative way, by using matrix multiplication and recompression. Explicit formulaes for the inversion of low-tensor rank matrices are available only in the simplest case, when the matrix has tensor rank 1: (A B)1 = A1 B 1 . Recently we have discovered the new class of low-tensor rank matrices that is closed under inversion, i.e. the tensor rank of the inverse matrices is bounded from above independent from the dimension. One illuminating example is the matrix of the following form: A = I + D R + R D, where D is a diagonal matrix and R is a matrix of rank 1. Then the tensor rank of A1 is not higher than 5. The generalization of the result is available. The theorem comes along with a constructive algorithm for the inversion with the required number of operations of order O( N ), yielding a sublinear algorithm. We will also show that almost every practically important two-level matrices (i.e., obtained from two-dimensional equation) can be reduced to the proposed form, thus leading to a theoretically justied sublinear algorithm for the inversion of such matrices.

Fast multiplication of matrix polynomials and some tricks for the computation of the trilinear decomposition in GF (2) Ivan Oseledets Institute of Numerical Mathematics of RAS, Moscow, Russia ivan.oseledets@gmail.com

64

Suppose we want to compute the product of two matrix polynomials


n1 n1

C(x) = A(x)B(x), A(x) =


i=0

Ai xi , B(x) =
i=0

Bi xi ,

where Ai and Bi are matrices over the eld with two elements 0, 1 of appropriate sizes. Of course, we can compute such product as a product of two matrices with polynomial coecients and using well-established algorithms for the multiplication of large integers. If Ai are, for example, 6464 it is convenient to store them as 64 words. Thus we are interested in some other algorithm. One possible way is to use Karatsuba-like algorithms, but they do not achieve linear complexity. However, their important property is that they use only addition and multiplication of constants, which in our case, are matrices (thus multiplication is much more computationally expensive than addition). It was shown in the seminal work of Kaltofen and Cantor in 1991 that polynomials can be multiplied in O(n log n log log n) time for an arbitrary algebra (however, we are not aware about its practical implementations). So, our main goal is to construct the algorithm that is optimal in some sense. All of the algorithms for computing bilinear form c = c(a, b) can be written in the following form: c = V ((U a) (U b)) , where, for the multiplication of polynomials, a, b are vectors of polynomial coecients, c is a vector of coecients for the product. The matrices U and V are independent from a, b and are (2n 1) r and n r respectively. Such representation is valid not only for the computation of polynomial product but, for example, for the computation of a periodic convolution. We will show that r is a tensor rank of a certain 2n n n tensor and computing U and V is equivalent to the computation of trilinear decomposition of a certain tensor. The computation of such decomposition in GF (2) is not well studied. We will present several new (sometimes heuristic sometimes not) greedy-type approaches that allow us to construct trilinear decomposition for such tensors, obtaining some known as well as some new values for r. For example a multiplication of a 7 7 circulant by a vector in GF (2) requires only 13 active multiplications.

Additive Preconditioning and Aggregation in Matrix Computations Victor Pan Department of Mathematics and Computer Science, Lehman College of the City University of New York, Bronx, NY 10468 USA v y pan@yahoo.com Multiplicative preconditioning is a popular SVD-based technique for the solution of linear systems of equations, but our SVD-free additive preconditioners are more readily available and better preserve matrix structure. We combine additive preconditioning with aggregation and other relevant techniques to facilitate the solution of linear systems of equations and some other fundamental matrix computations. Our analysis and experiments show the power of our approach, guide us in selecting most eective policies of preconditioning and aggregation, and provide some new insights into these and related subjects of matrix computations.

65

Mappings preserving the idempotency of products of operators Tatjana Petek University of Maribor, FERI, Smetanova 17, SI-2000 Maribor, Slovenia tatjana.petek@uni-mb.si We will discuss mappings on the algebra of all bounded linear operators on a nite or innite dimensional Banach space X, which preserve the idempotency of products or triple Jordan products () of operators. The motivation for this study was the recently published paper on mappings preserving the nilpotency of products of operators by Li, Sze and Semrl. When studying preserving idempotency instead it turns out that these mappings if unital are basically mappings on idempotents. About the mappings on idempotents several results are known. We use some of them in order to obtain the characterization of all surjective unital mappings on B (X) having the property AB is a nonzero idempotent if and only if (A) (B) is a nonzero idempotent (A,B in B(X)). Note that there is no additivity assumption.

Laguerre-Pseudospectral Discretization vs. Finite-Dierences for Discrete Green Function Calculation Podgornova O.V., Sofronov I.L. Keldysh Institute of Applied Mathematics RAS, Moscow olgapodgornova@gmail.com The construction of non-reecting boundary conditions for hyperbolic equations in anisotropic media is discussed here. We develop the idea of discrete transparent boundary conditions (DTBC) [1]. According to this approach, generation of the boundary operator is made using Greens function of an exterior discrete initial-boundary value problem correspondent to a discrete counterpart of the original governing equations (e.g. nite-dierence equations as in [1]). We consider the circular boundary for 2D wave equation in polar coordinates for two applications: moving media and two-layer media [2]. A second-order nite-dierence scheme is used for discretization here. This leads to some issues because of oscillatory nature of the discrete Greens function. In particular, accuracy of our DTBC can be insucient while restricting admissible computational costs of the boundary operator. Trying to move ahead we looked at the other way to solve exterior problems with the requirements of high resolution and good performance. We apply Laguerre functions expansion method for time dependence of the solution [3,4]. It results in an elliptic equation for each coecient at Laguerre functions, spectrum of which has negative real part. An additional advantage is that the resolution matrix for elliptic equations does not depend on the order of Laguerre functions. Further improvement of the Laguerre approach is a combination of it with pseudospectral spatial discretization. At this stage we consider sequence of 1D wave problems in (t, r) variables correspondent to 2D isotropic wave problem after azimuth Fourier decomposition. We compare eciency of such hybrid Laguerre-pseudospectral method with previous nite-dierence approach using several representative test problems. References 66

1. Ryabenkii V.S., Exact Transfer of Dierence Boundary Conditions, Functional Anal. Appl. 24, (1990), 251253. 2. Sofronov I.L., Podgornova O.V., A spectral approach for generating non-local boundary conditions for external wave problems in anisotropic media, Journal of Scientic Computing, V. 27, 3, 2006, pp. 419430. 3. Konyukh G.V., Mikhailenko B.G., Mikhailov A.A., Application of the integral Laguerre transforms for forward seismic modeling, J. of Computational Acoustic. Vol. 9, 4, 2001, pp. 1523 1541. 4. Young-Seek Chung, Tapan K. Sarkar, Baek Ho Jung, and Magdalena SalazarPalma, An Unconditionally Stable Scheme for the Finite-Dierence Time-Domain Method, IEEE Transactions on Microwave Theory and Techniques, Vol. 51, 3, March 2003.

Exploiting rank structure in the solution of a class of nonsymmetric algebraic Riccati equations F. Poloni (joint work with D. A. Bini, B. Iannazzo, B. Meini) Scuola Normale Superiore, Pisa f.poloni@sns.it Nonsymmetric algebraic Riccati equations (NAREs) XCX + B AX XD = 0 with structured n n block coecients A, B, C, D, appear in many applications. In particular, in certain problems of transport theory the matrices A, B, C and D are such that M = D C is a diagonal plus rank-1 B A M -matrix. We show how to develop structure-preserving versions of some classical algorithms for solving NAREs in order to exploit the rank-structure of this problem. Three methods are analyzed: Newtons iteration, the structured doubling algorithm, and cyclic reduction. New algorithms are provided for implementing the three methods above with a cost of O(n2 ) arithmetic operations per step instead of O(n3 ), thus substantially accelerating the solution of the equation. The shift acceleration technique for speeding up the convergence in the critical case is implemented as well.

Reduction of matrices over a domain of principal ideals to the Smith normal forms by means of the same one-sided transformations V. M. Prokip Institute for Applied Problems of Mechanics and Mathematics; 3B Naukova Str. Lviv-60, 79060, Ukraine vprokip@mail.ru Let R be a domain of principal ideals (a commutative ring with identity e = 0 without nonzero divisors of zero and such that every ideal is principal) and Rn the ring of (n n)-matrices over R. 67

It is known that any matrix A Rn , rankA = r, is equivalent to the diagonal matrix SA = U AV = diag(a1 , . . . , ar1 , ar , 0, . . . , 0), where U, V GL(n, R) and aj for 1 j r are the nonzero elements of R such that a1 | a2 | . . . | ar (divides). The diagonal matrix SA is known as the Smith normal form of the matrix A (see:M. Newman. Integral Matrices. Academic Press: New York, 1972.). Theorem. Let A1 , A2 , . . . , Ak Rn be the nonsingular matrices. If (det Ai , det Aj ) = e for all i = j , i, j = 1, 2, . . . , k, then there exist matrices U GL(n, R) and V1 , V2 , . . . , Vk GL(n, R) such that U Ak Vk = SAk , k = 1, 2, . . . , k. Note: The Theorem presented above is true for an adequate ring (see: O. Helmer. The elementary divisors for certain rings without chain condition. Bull. Amer. Math. Soc., 49, 225236 (1943)). Furthermore, the obtained Theorem can be easily extended to the elementary divisor domains R, where, for any three elements a = 0, b = 0 and c from R , there exists an element r R such that the greatest common divisor of (a, b, c) coincides with the the greatest common divisor of (a+ rc, rb).

Bounded semigroups of matrices and the graph tractability problem Vladimir Protasov Department of Mechanics and Mathematics, Moscow State University, 119992, Moscow, Russia v-protassov@yandex.ru This is a joint work with Vincent Blondel and Raphael Jungers (University Catolique de Louvain, Belgium). For a given family of n n - matrices A = {A1 , . . . , Am } with nonnegative integer entries we consider the asymptotic growth of the value Fk = max{ A , A Sk } as k , where Sk = Ad1 Adk , dj {1, . . . , m} and is an arbitrary matrix norm. Thus, Fk is the largest norm of products of these matrices of length k. The problem is to determine whether Fk is uniformly bounded over k N, or it grows polynomially in k, or exponentially. For general sets of matrices this problem is well-known and has found numerous applications in functional analysis, optimal control, approximation theory, probability etc. The rate of growth of Fk can be expressed in terms of the joint spectral radius (JSR) of the family A. However, the computing or estimating JSR is known to be hard even for small dimensions n. For example, for matrices with rational coecients this problem is NP-hard, and the problem to determine whether JSR is less than 1 is, moreover, algorithmically undecidable. We show that for nonnegative integer matrices there is a relatively simple solution and ecient polynomial time algorithms to decide the type of growth of Fk . Theorem 1. For any family A of nonnegative integer matrices there is the following alternative: 1) A is mortal, i.e., Fk = 0 for all suciently large k; 2) A is not mortal, but product bounded, i.e. Fk C for all k; 3) A has polynomial growth, i.e. Fk k r for some natural r n 1; 68

4) A has exponential growth, i.e. Fk Ck for some > 1. Each of the conditions 1 4 can be checked by the following criteria. We denote by S = kN Sk the semigroup of matrices generated by A1 , . . . , Am . Theorem 2. For a given family A condition (1) can be checked within O(mn2 ) arithmetic operations. Condition (4) is satised i there is A S such that Aii 2 for some i. This can be checked within O(mn2 ) operations. Condition (3) is satised i there is A S such that Aii , Aij , Ajj 1 for some i, j. This can be checked within O(n3 (m + n2 )) operations. Thus, all the conditions of Theorem 1 can be decided in polynomial time. Moreover, the rate of polynomial growth r (condition 3) can also be found within O(n3 (m + n2 )) operations. Corollary 1. For a given family A of nonnegative integer matrices the niteness of the generated semigroup S can be decided in polynomial time. These results are applied to the graph tractability problem in large sensor networks [2]. A directed graph, whose edges are colored with m colors is called tractable if the total number of paths corresponding to any sequence of colors is uniformly bounded by some constant. Corollary 2. The tractability of any colored graph is polynomially decidable. References 1. R. Jungers, V. Protasov and V. Blondel, Ecient algorithms for deciding the type of growth of products of integer matrices, to appear in Linear Algebra and Applications. 2. V. Grespi, G.V. Cybenko and G. Jiang, The theory of tractability with applications to sensor networks, Technical report TR2005-555, Dartmouth College, Computer Science, Hanover NH (2005).

Singular values for the nite sections of non-Fredholm Toeplitz operators A. Rogozhin IBM (Moscow, Russia) a.rogozhin@mail.ru Let L (T) denote the usual Lebesgue space on the unit circle T = {t C : |t| = 1}. A function a L (T) is called piecewise continuous if it has one-sided limits a(t 0) for all t T. The following criteria for the Fredholmness of Toeplitz operators with piecewise continuous symbols is well-known. Theorem 1. Let a L (T) be a piecewise continuous function. The Toeplitz operator T (a) L(2 ) is Fredholm if and only if a (t, ) = 0, (t, ) T [0, 1], where the function a (t, ) is dened on the cylinder T [0, 1] by a (t, ) = a(t 0) + (a(t + 0) a(t 0)). In this talk we consider the behaviour of the singular values for nite sections of non-Fredholm Toeplitz operators with piecewise continuous symbols in the case where the symbol of the Toeplitz operator does not vanish on the unit circle. That is we analyse the case where a (t, ) = 0 for some (t, ) T (0, 1) ((t, ) the interior of the cylinder). 69

A special kind of this case was considered in the papers [2,3]. In these papers the behaviour of the singular values for the Cauchy-Toeplitz matrices An = 1 ij +
n

,
i,j=1

R,

is investigated. Note that An = Tn ( ) with the piecewice continuous symbol ei = ei ei , sin 0 < 2.

It was shown that the singular values of Tn ( ) are clustered at /| sin |, only o(n) singular values lie outside a given -neighborhood of the clustering point. Our consideration is based on the fact that the singular values are not only the nonnegative square roots of the eigenvalues of Tn (a)Tn (a), but also they are the approximation numbers. In view of this fact we use the results of the paper [1]. More precisely, our investigation is leaned on the following theorem (see [1], Theorem 1). Theorem 2. Let a be a piecewise continuous function on the unit circle, and let 1 < p < . If the two Toeplitz operators T (a) and T (), a(t) = a(1/t), are Fredholm in the Banach space L(p ), a then there is an integer k 0 such that
n

lim sk (Tn (a)) = 0 and

lim inf sk+1 (Tn (a)) > 0,


n

where sm (Tn (a)), m = 0, 1, . . . , n, is the mth approximation number of Tn (a), i.e. the distance of Tn (a) to the n n matrices of rank at most n m. The following theorem is the main result of the talk (we order the singular values so that 0 s1 (Tn (a)) s2 (Tn (a)) . . . sn (Tn (a)) = Tn (a) L(2 ) ). n Theorem 3. Let a be a piecewise continuous function on the unit circle such that a(t 0) = 0 for all t T. If the Toeplitz operator T (a) L(2 ) is not Fredholm, then for all m N sm (Tn (a)) 0 as n , n N, but there is an integer k 0 such that for any > 0 sk (Tn (a)) c()n , where the constant c() > 0 does not depend on n. References 1. A. Rogozhin, B. Silbermann, Approximation numbers for the nite sections of Toeplitz operators with piecewise continuous symbols, Journal of Functional Analysis, 237(1), 2006, pp. 135 149. 2. E. Tyrtyshnikov, Cauchy-Toeplitz matrices and some applications, LAA, 149, 1991, pp. 1 18. 3. E. Tyrtyshnikov, Singular values of Cauchy-Toeplitz matrices, LAA, 161, 1992, pp. 99116.

70

Matrix tricks in tensor approximation: 3D-cross method and maximum-volume principle Dmitry V. Savostyanov Institute of Numerical Mathematics of RAS drraug@gmail.com We consider Tucker-like approximations with an rrr core tensor for three-dimensional nnn arrays in the case of r n and possibly very large n (up to 104 106 ). As the approximation contains only O(rn + r 3 ) parameters, it is natural to ask if it can be computed using only a small amount of entries of the given array. A similar question for matrices (two-dimensional tensors) was asked and positively answered 10 years ago in [1]. We extend the positive answer to the case of three-dimensional tensors. More specically, it is shown that if the tensor admits a good Tucker approximation for some (small) rank r, then this approximation can be computed using only O(nr) entries with O(nr 3 ) complexity. But the existence is good, but the algorithm is better. We also propose an algorithm to compute such approximation, based on the same ideas as incomplete cross approximation method, proposed for matrices in [2]. It successively computes rows, columns and bers of 3D array A = [aijk ] (and hence is named 3D-cross method, or simply Cross-3D) using a neat adaptive technique. However, the good implementation of the method (which is even better than algorithm) requires to solve a number of matrix problems. We have found a bunch of them, rather interesting and new. For example, given a factors low-rank matrix A = U V T (A is n n and U, V are n r) how to nd maximum modulus element of A in O(nr a ) ops? (and how to provide the least a?) This, and some other questions are answered using the maximum-volume principle [3]. We are happy to notice, that the implementation of maximum-volume principle in 3D-cross method boosts its speed up to hundreds times. Finally, we demonstrate an application of our methods to the solution of 3D volume integral equation. For grid with 512 512 512 elements, the collocation matrix requires 128PB of memory to be stored and more than a year to be computed. 3D-Cross computes the approximation with relative accuracy 105 at 30 minutes and stores it in 114MB of memory. References 1. S. A. Goreinov, E. E. Tyrtyshnikov, N. L. Zamarashkin, A theory of pseudoskeleton approximations, Linear Algebra Appl. 261 (1997), 121. 2. E. E. Tyrtyshnikov, Incomplete cross approximation in the mosaic-skeleton method, Computing, 4 (2000), 367380. 3. S. A. Goreinov, E. E. Tyrtyshnikov, The maximal-volume concept in approximation by low-rank matrices, Contemporary Mathematics, 208 (2001), 4751.

Spectral theory of reducible nonnegative matrices in classical and max linear algebra: A comparison Hans Schneider Mathematics Department, University of Wisconsin hans@math.wisc.edu 71

We consider two forms of nonnegative linear algebra: classical and max. We compare the spectral theories of reducible nonnegative matrices in the two forms of linear algebra, stressing the similarities and identifying the dierences.

Exponential parametrization and linear scaling Density Functional Theory in BigDFT Reinhold Schneider University of Kiel as@informatik.uni-kiel.de Electronic structure calculation plays an increasing role in the numerical simulation of molecular phenomena. Density functional theory povides an eective one particle model for the computation of the groundstate energy of a quantum mechanical system of N electrons moving in a Coulombic eld given by xed nuclei. Instead of the high dimensional Schrdinger equation o one has to solve o nonquadratic optimization problem in three space dimensions. Since the exchange correlation potential is not yet known there remains a modelling error. The groundstate energy is the minimum of the Kohn-Sham engergy functional depending on N pairwise orthogonal orbital functions. This is a minimization problem on a prescribed Stiefel resp. Grassman manifold. This problem can be treated either as a constraint optimization problem, or by local parameterization of the manifold. The analysis of gradient type minimization procedure in both cases will be considered, and the connection of both approaches will be highlighted. For insulating materials it is known, that the subspace generated by the orbital functions can be spanned, approximatively, by local functions, e.g. Wannier orbital. Linear scaling with respect to the number of atoms of the computational costs can be achieved by enforcing the locality of the iterates in each iteration step. For this purpose local, or better multi-scale high order basis functions in conjunction with smooth, nonlocal, pseudo-potentials can provide an ecent alternative to either plane wave basis sets or Gaussian functions. This approach has been integrated in the BigDFT electronic structure program.

The matrix of bordering functions and the problems of the restoration of signals on discrete samples A. V. Sedov The Southern Scientic Centre of Russian Academy of Sciences, 132, Prosveshchenia Str., Novocherkassk, Rostov Region, 346428, Russia sedov a.v@mail.ru The classical theorem of sampling (Kotelnikov-Shannons theorem) for an innite signal x(t) , < t < in time, with the nal energy and the spectrum x() limited by frequency h , denes 72

the condition T < /h of formation of discrete samples x [kT ] = x (kT ) , k Z and the formula of the exact restoration of a signal on these samples:

x(t) =
k=

t t/T x[kT ] sinc( T k) = S

p X ,

(1)

p where X = [..., x[2T ], x[T ], x[0], x[T ], ...] T is the innite-dimensional vector of samples x[kT ] t/T ; S = . . . , S
t/T 2

, S

t/T 1

, S , S

t/T

t/T +1

,...

is the innite-dimensional vector consisting of

equidistant samples of a function S = sinc (t/T k). In case of a nal signal xp (t) , t [0, TS ) in time which can be considered as the period of the signal x(t) with the nal energy and with the spectrum x() limited by frequency h the other formula of restoration [1]
N 1

t/T k

x(t) =
k=0

xp [kT ]

1 N

2 N

<N/2> n=1

cos 2n (t/T k) N

N 1

=
k=0

xp [kT ]SN

t/T k

(2)

is a function of samples for a nal signal xp (t) , that is satised. The bordering function SN is it is the same for each of N samples xp [kT ], but it is displaced on k steps on an axis of time. Besides it is the periodic one with the period N . The sign N /2 is the integer part of number N /2. t/T k t The interrelation of formulas (1) and (2) and so the functions of samples sinc( T k) and SN is dened by the limit lim SN
N t/T k

on discrete samples xp [kT ], k = 0, N 1 is oered, where T = TS /N and the condition T < /h


t/T k

In practical tasks of interpolation the formula (2) is used for the reception on samples xp [kT ] , k = 0, N 1 of the signal xp (t) of the greater number of samples xp [kT ] , k = 0, m 1,(m N ) with the small period of sample T = (Tc /m). Thus the formula (2) can be written as: p p p Xm = SmN XN = Sm,0 , Sm,1 , . . . , Sm,N 1 XN N N N (3)

= sinc (t/T k).

p p where Xm = [xp [0], xp [T ], xp [2T ], ..., xp [(m 1)T ]]T is the m-dimensional vector; XN = [xp [0], xp [T ], xp [2T ], ..., xp [(N 1)T ]]T is the N -dimensional vector; SmN is the matrix of bordering functions. The columns of matrix SmN are the vectors:
k k k mT mT , SN , ..., SN Sm,k = SN , SN N
N 2N N(m1) k mT

, Sm,k Rm . N

According to the expression (3) the formula of the restoration ( the interpolation) of a nal signal xp (t) in time, with properties mentioned above, can be submitted as the product of the matrix SmN p of bordering functions on the vector XN of the samples measured on conditions that T < /h . The expression (3) can be considered, as a new interpolation formula as a kind of the decomposition of the signal (the function) xp (t) by columns Sm,k of the matrix SmN . Its feature is that the samples N xp [kT ], k = 0, N 1 of the signal xp (t) are the factors of decomposition, so their calculation is not required. References 1. A. V. Sedov, The specication of the sampling theorem and the formula of the restoration of a signal on discrete samples, Electromechanics, 2001. Vol. 2, pp. 5259.

73

The property of the quasi-orthogonality of vectors and the quasi-unitary matrixes at the modelling of signals A. V. Sedov The Southern Scientic Centre of Russian Academy of Sciences, 132, Prosveshchenia Str., Novocherkassk, Rostov Region, 346428, Russia sedov a.v@mail.ru We shall stop on some properties of the matrix of bordering functions SmN (see the previous report). The particular properties of this matrix allow to receive a number of the general properties and denitions, namely, the property of the quasi-orthogonality of vectors and the denition of the quasi-unitary matrix. 1.The property of the quasi-orthogonality of vectors. We shall construct Grams matrix GN N for the vectors Sm,k of the matrix SmN of bordering N functions: GN N = SmN
T

SmN =

s1 IN N

s1 IN N , for N odd; s2 TN N = EN N , for N even,

where IN N is an unitary matrix of dimension N N ; TN N = (1)i j i,j=1,N is Toeplitzs matrix, it elements are equal along each diagonal; s1 = m/N ; s2 = (1 (2 N /2 + 1)/N ) m/N . In case of the odd value N the matrix GN N is the diagonal one, and the vectors Sm,k are N N N is not the diagonal one, and the orthogonal ones. In case of the even value N the matrix G the vectors Sm,k are not the orthogonal ones. However, if in the latter case the samples xp [kT ], N k = 0, N 1 of the signal xp (t) are chosen from the condition T < /h , and it is easy to show, that p p the formula Xm = SmN XN also is true, as in case of the orthogonal vectors Sm,k . We shall dene N m,k mN or the harmonious the given property as the quasi-orthogonality vectors SN of the matrix S quasi-orthogonality. 2. The denition of the quasi-unitary matrixes. N The quasi-unitary matrix for the square matrix Xp N of dimension N N , where N is an even value, which we shall name a matrix EN N = N /m EN N such as
N N Xp N = EN N Xp N .

(1)

The distinctive feature of the quasi-unitary matrixes EN N is that it is not a diagonal matrix with units on a diagonal, but at the same time it possesses its basic property (1) for the N N matrix Xp N . Thus any columns of the matrix Xp N are vectors XN , l = 0, N 1 consisting of pl the samples xpl [kT ], k = 0, N 1 of dierent signals xpl (t) with the spectra limited by frequency h , and the condition T < /h is satised for the sampling interval. The expression (1) is true and 1 for the columns XN = EN N XN . The inverse matrix EN N also is the quasi-unitary matrix IN N for the matrix The equality (1) is broken, if in the spectrum of the signals xpl (t) making the columns XN of the pl N matrix Xp N , there is a harmonic with number M = N /2 , and this harmonic is not a sine with a zero phase. N It is possible to sum up that we can nd for the matrixes of a kind Xp N besides the unitary matrix IN N not a diagonal matrix for which the condition (1) is also satised. References 74
pl N N . Xp pl

1. A. V. Sedov, The specication of the sampling theorem and the formula of the restoration of a signal on discrete samples, Electromechanics, 2001. Vol. 2, pp. 5259.

Automorphisms of the semigroup of invertible matrices with nonnegative elements over commutative partially ordered rings Bunina E. I., Semenov P. P. Department of Mechanics and Mathematics, Moscow State University, 119992, Moscow, Russia helenbunina@yandex.ru In the paper [1] A. V. Mikhalev and M. A. Shatalova described all automorphisms of the semigroup Gn (R) of invertible matrices with nonnegative elements in the case when R is a skeweld and n 2. In [2] E. I. Bunina and A. V. Mikhalev described all automorphisms of the semigroup Gn (R), if R is an arbitrary linearly ordered associative ring with 1/2, n 3. In this paper we describe automorphisms of the semigroup Gn (R) in the case when R is a commutative partially ordered ring with 1/2, 1/3 0, n 3. Let n be the symmetric group of the order n, S be the matrix of transposition n , Sn = {S | n }. By Dn (R) we denote the group of all invertible diagonal matrices from Gn (R). By Bij (x) we denote the matrix E + xeij . Let P be the subsemigroup in Gn (R), generated by all matrices S ( n ), Bij (x) (x R+ , i = j) and Dn (R). Two matrices A, B Gn (R) are called P-equivalent (see [1]), if there exist matrices Aj Gn (R), j = 0, . . . , k, A = A0 , B = Ak , and matrices Pi , Pi , Qi , Qi P, i = 0, . . . , k 1 such that Pi Ai Pi = + Qi Ai+1 Qi . By GEn (R) we denote the subsemigroup in Gn (R), generated by all matrices, that are P-equivalent to the matrices from P. For every matrix M n (R) let M denote an automorphism of Gn (R) such that X Gn (R) M (X) = M XM 1 . For every y() Aut(R+ ) by y we denote an automorphism of Gn (R) such that X = (xij ) Gn (R) y (X) = y ((xij )) = (y(xij )). Theorem. Let be an automorphism of the semigroup Gn (R), n 3, 1/2, 1/3 R+ . Then + on the semigroup GEn (R) = M c , where M n (R), c() Aut(R+ ), () is a central + homotethy of GEn (R). References 1. A. V. Mikhalev, M. A. Shatalova, Automorphisms and antiautomorphisms of the semigroup of invertible matrices with nonnegative elements, Math. Sbornik, 1970, 81(4), 600609. 2. E. I. Bunina, A. V. Mikhalev, Automorphisms of the semigroup of invertible matrices with nonnegative elements, Fundamentalnaya i prikladnaya matematika, 2005, 11(2), 323.

On extensions of the closure operation for matrices over the max-plus semiring Serge Sergeev 75

Department of Physics, Sub-Department of Quantum Statistics and Field Theory, Moscow State University, Moscow, 119992 Leninskie Gory, Russia The research was supported by the RFBR grant 05-01-00824 and the joint RFBR/CNRS grant 05-01-02807. sergiej@gmail.com Algebraic closure of a square matrix A over the max-plus semiring Rmax = R {} is the series I A A2 . . ., where I is the identity matrix. This series converges i (A) 1, where (A) is the maximal cycle mean of A. I show that the closure operation can be extended, by some kind of renormalization, over the max-plus matrices which have at least one nite diagonal (whose weight is not equal to ). This has been published in [2]. I also show that further extensions can be obtained by making use of the cellular decomposition and other concepts considered by Develin and Sturmfels [1]. References 1. M. Develin and B. Sturmfels, Tropical convexity, Documenta Math., 9:127, 2004. Also arXiv:math.MG/0308254. 2. S. Sergeev, Max-plus denite matrix closures and their eigenspaces, Linear Algebra Appl., 421:182201, 2007. E-print arXiv:math.MG/0506177.

Tridiagonal canonical matrices of bilinear and sesquilinear forms Vladimir V. Sergeichuk Institute of Mathematics, Tereshchenkivska 3, Kiev, Ukraine sergeich@imath.kiev.ua This is joint work with Vyacheslav Futorny and Roger A. Horn. We give tridiagonal canonical forms of matrices of (i) bilinear forms and sesquilinear forms, (ii) pairs of forms, in which each form is either symmetric or skew-symmetric, and (iii) pairs of Hermitian forms over an algebraically closed eld of characteristic dierent from 2. Our canonical forms are direct sums of matrices or pairs of matrices of the form a 0 c 0 b d 0 a ; c 0 b .. . d 0 .. .. . . 0 they employ relatively few dierent types of canonical direct summands. 76

References 1. V. Futorny, R. A. Horn, V. V. Sergeichuk, Tridiagonal canonical matrices of bilinear or sesquilinear forms and of pairs of symmetric, skew-symmetric, or Hermitian forms, Preprint RT-MAT 2006-16, Universidade de So Paulo, 2006. Submitted for publication. a

The computing of the stationary magnetic eld in the presence of surface with innite permeability K. S. Shaposhnikov Southern Scientic Centre of Russian Academy of Science cyr @mail.ru We consider the problem of stationary magnetic eld computation in presence of piecewisesmooth Lipschitz surface with innite permeability. The media is uniform and isotropic, i.e. = const. The problem of magnetic eld computing in classic formulation can be reduced to the following boundary problem: = 0 outside ; = C 0 on ; (M ) 0,
M

(1)

where is scalar potential of intensity of reaction eld H , 0 is a potential of non-perturbed eld intensity H0 , H = H + H0 , C = const. We use the orthogonal projections method for the problem (1) solving [1]. We denote is the closure of to the closed surface = , is the adjacent boundary between and . is space portion, bounded by , + its outside, = + . We orient an outward normal with respect to . We consider a Hilbert space H () of the real functions that is introduced and studied in [2]. The elements of H () have a nite Dirichlet integral, they are harmonic out of ball R : R , and vanish at innity [2]. Inner product and norm in H () are following: (1 , 2 )H =

1 2 d,

(, )H .

According [2], there is the orthogonal decomposition for this space: H () = H () H0 (). The subspace H () consists of the functions with constant value on . H0 () consists of the functions that representable by potentials of simple layer with the zero mean of densities on . We choose the potentials of H0 () that have zero densities on . The set of such functions 0(1) form a subspace in H0 () [1] that is denoted as H (). We have shown that the orthogonal 0 () is the subspace that consists of potentials with the constant complement for its subspace in H 0(2) values on with the zero mean of densities on . We denote this subspace as H (). We shall interpret 0 as the potential of equivalent non-perturbed eld that is potentially in and has the nite energy in . The equivalence is understood in sence of [3]. This mean that of potentials of the initial and the equivalent elds have coincidetnd values on . We have shown that 0(1) if 0 is in the H () then belongs to H () and the potential of the equivalent resulting eld 77

= 0 + is in H (). Thus, the solution of the problem is reduced to 0 projection to subspace 0(1) H () that can be found by the calculation of 0 coordinates {ck }n in some basis {k }n for k=1 k=1 n 0(1) is represented by the following: ck k . Coordinates are determined by H (). Then =
k=1

0(2)

We use coordinate functions that are represented by potentials of simple layer with nite piecewise constant densities on . Densities is zero on . We have proved that sush system of functions is 0(1) complete in H () in sence of Ritzss approximate sequence convergence. In this case the elements of SLAE (2) are dened by the following formulas: (i , k )H = 1 4
i

the solution of the system of linear algebraic equations with the Grams matrix in H (): 0 , 1 H c1 (1 , 1 )H . . . (1 , n )H ... = . ... ... ... ... 0, (n , 1 )H . . . (n , n )H cn n H

(2)

i
k

k dd, r

0 , i

=
i

0 i d.

The computing algorithm and eective software package for its realization are developed. References 1. Astakhov V. I., Kochubey T. V., Shaposhnikov K. S., The orthogonal projections method in tasks of stationary magnetic elds calculation, Studies of the Southern scientic centre of the Russian academy of sciences, Rostov-on-Don, SSC RAS Publishing house, 2007, pp. 5172. (in Russian) 2. Astakhov V. I. Surface potentials and operators of potentials theory in Dirichlets spaces, Electromechanics, 2000, 2, pp. 318. (in Russian). 3. Astakhov V. I., About an admissibility of idealization of borders of polarizable bodies and some energetic identities for stationary magnetic and electrostatic elds, Electromechanics, 2000, 1, p. 314. (in Russian).

On discrete projection methods for rotating incompressible ow with Coriolis force Andriy Sokolov , Maxim A. Olshanskii and Stefan Turek for Applied Mathematics, University of Dortmund, Germany Department of Mechanics and Mathematics, Lomonosov Moscow State University, Moscow, Russia {asokolow,stefan.turek}@math.uni-dortmund.de, maxim.olshanskii@mtu-net.ru
Institute

The construction of an ecient solver for the incompressible system of Navier-Stokes equations is a long-term purpose of CFD researchers. Since decades an evident progress is observed: large variety of methods and algorithms were proposed and implemented into commercial and open-source code. In many physical and industrial processes there is a necessity of numerical simulations of models with moving boundaries in a 3-dimensional space. As proposed in the literature, approaches like ctitious boundary [4] or Arbitrary Lagrangian Eulerian methods require large amount of CPU time 78

to simulate even a 2D benchmark models. Moreover, they could provide a source of additional errors in velocity and pressure elds. At the same time, there is a large class of rotating models, where the use of the above methods can be avoided by some modications of Navier-Stokes equations and/or transformations of the coordinate system. While performing numerical simulation for the Stirred Tank Reactor problem (g. 1, left), we can avoid moving boundary parts (blades of the rotating propeller) by transforming the coordinate system in such a way that the propeller is considered to be motionless, but the wall of the tank rotates. By doing so, we are required to perform a numerical simulation for the system of NavierStokes equations with rotational forces:
u t

+ (u )u u + 2 u + ( r) + p = f u=0

where 2 u and ( r) are the so-called Coriolis and centrifugal forces, respectively.

Figure 1: STR benchmarks. (LEFT) mesh (RIGHT) simulation Thus, in our talk we present a numerical analysis and algorithmic details for treating the system of Stokes and Navier-Stokes equations with Coriolis force term. Using a Discrete Projection Method with modied Pressure Schur Complement operator, we examine the inuence of the Coriolis force on every step of the algorithm: modication of the momentum equation and building of a new block diagonal preconditioner, construction of a new Pressure Schur Complement operator by inserting the o-diagonal parts, which are due to the rotating terms and an explicit formation of its inverse, and nally, correction of pressure and velocity to satisfy the incompressibility constraints. Detailed numerical studies of the improvements in the convergence rate, its dependence on the magnitude of the time step and angular velocity are provided. We also consider the addition of the convection term into the Schur Complement, discuss its inuence on the iterative behaviour of the solver and propose a coriolis-convection oriented Pressure Schur Complement. Finally, we provide numerical simulations for polygonal geometries and check the eciency of our solver by comparing simulation results (g. 1, right) with those obtained by the ctitious boundary method for moving boundary parts. The used solver is based on the 3-dimensional PP3D CFD-code from the Featow package (www.featflow.de), [2,3]. References 1. M. A. Olshanskii, An iterative Solver for the Oseen Problem and Numerical Solution of Incompressible Navier-Stokes Equations, Numer. Linear Algebra Appl., Vol. 6, page 353378, 1999. 2. S. Turek, On discrete projection methods for the incompressible Navier-Stokes equations: an algorithmical approach, Comput. Methods Appl. Mech. Engrg., Vol. 143, page 271288, 1997. 3. S. Turek, Ecient Solvers for Incompressible Flow Problems: An Algorithmic and Computational Approach, LNCSE 6, Springer, 1999. 79

4. S. Turek, D. C. Wan and L. P. Rivkind, The Fictitious Boundary Method for the implicit treatment of Dirichlet boundary conditions with applications to incompressible ow simulations, In E. Bnsch editor, Challenges in Scientic Computing CISC 2002, LNCSE, page 3768, a Berlin, 2002. Springer, Berlin.

A New Method for a Quadratic Inverse Eigenvalue Problem in Structural Dynamics Vadim Sokolov (with Biswa Datta, Sien Deng, Moody Chu and Daniil Sarkissian) Northern Illinois University, DeKalb, USA The research of Biswa Datta has been supported by NSF grant #DMS-0505784. sokolov83@gmail.com In structural dynamics, it is very often required that an analytical model be updated using the eigenvalues and eigenvectors measured from a real-life or a prototype structure. Mathematically, the problem is an inverse eigenvalue problem for the associated quadratic matrix pencil. Despite much research done by scientists and practicing engineers, the problem has not been satisfactorily solved. Most of the industrial solutions are ad hoc in nature and lack solid mathematical foundation. Furthermore in practice almost always the problem is solved for the undamped model, thus the underlying inverse problem is the one for a linear matrix pencil, which makes problem much easier to solve. Most of these methods also assume that the measured data is accurate and proceed to updating the model under that assumption. In reality, the measured data has always some noise and damping should not be neglected. In this talk, we present a new optimization method for a damped nite element model updating. The method consists of two stages. In Stage I, the measured data is updated so that the updated eigenvectors satisfy a recently established orthogonality condition involving mass and damping matrix. In stage II the stiness matrix is updated so that the quadratic pencil has the prescribed eigendata.

An approach to feature-based basis construction for image analysis via Karhunen-Loeve transform A. Yu. Solodovschikov Keldysh Institute of Applied Mathematics RAS, Moscow anton mipt@7ka.mipt.ru The paper is devoted to an investigation of discrete and continuous variants of the KarhunenLoeve transform. An algorithm of analytical construction of a basis with non-correlated expansion coecients for analyzing digital images based on this expansion is presented. The results obtained are compared with the existing results in the problem of edge detection. As an example, the problem of formal construction of operators similar to the Hueckel operator by this method is solved. The characteristics of the formed basis functions are compared. 80

Usually, natural scenes consist of a set of three-dimensional objects. TV digital or photo images are two-dimensional projections of these scenes and consist of regions with dierent intensity characteristics of particular pixels. To obtain a vector description of the source digital image of three-dimensional scenes, the operator proposed by M. Hueckel that transforms a digital image to a set of linear segments of a xed length can be successfully applied. In his paper, Hueckel solved a more particular problem of mapping a digital intensity image in a circular window of a xed radius into the equation of the straight line that approximates the position and direction of the line of maximums of the intensity gradients at points of the window. For this purpose, he proposed a totality of smooth basis functions and their discrete representation for a two-dimensional expansion within the window. The detailed investigation of these functions has shown that they satisfy the criterion of the minimum error between the elements of the source image and the most probable position of the line of intensity gradients. The important idea of the statement of the problem of analyzing an image that states that it is necessary to nd a convenient basis of a certain integral transform that is adequate to the problem solved is contained in. This approach forestalls the later method of wavelet transforms, which, currently, have noticeably pressed the two-dimensional Fourier transforms. Indeed, to obtain a concise description of the scene content, it is reasonable to select from its image only necessary (depending on the desired content) elements of the spatial interrelations of pixel intensities eliminating unnecessary (excessive) information and compensating dropped data (as a result of noise in sensors). The problem of constructing a method of image description such that description elements are weakly dependent arises. One of the most promising methods for solving this non-simple problem is to nd a two-dimensional orthogonal basis for an integral transform of a digital image that is adequate to the problem of scene recognition. It is important that this basis should provide that the desired information is to be concentrated in the minimum number of the expansion coecients. As a rule, these higher harmonics of the expansion spectrum contain the desired information about the integral intensity discontinuities, while the high-frequency components of the spectrum reect only excessive information, e.g., in the problem of nding contours of image regions. The performed testing has conrmed the adequacy of the Hueckel basis to the problem posed by him. However, it is extremely desirable to obtain a formal method for constructing a basis that is adequate to the solved problem of image processing. We have managed to obtain this method applying another Kahrunen-Loeve transform (KL-transform), which is well-known in the recognition theory. In contrast to a number of papers that develop a similar approach, the investigation of the method is not limited to the discrete variant of the transform. The use of the continuous KLtransform allows one to describe the structure of the basis analytically, which is also convenient in the scale analysis. In literature, the KL-transform is described within the probabilistic signal model. However, unconditionally, the Kahrunen-Loeve transform has more general nature. It is applicable for a totality of image elements that are not mandatory random, for which there exists knowledge of an analogue of the correlation function of their intensity. The KL-transform has a number of properties that distinguish it among the other linear transforms. The main drawback of the KL-transform is that there are no fast algorithms for its computation. The method obtained in the paper is described by examples of constructing operators for transforming images similar to the Hueckel operator. A formal derivation of the basis for the Hueckel operator is presented and the signicance of its harmonics is analyzed.

81

On some fundamental properties and applications of Chebyshv continued fractions e Sergey P. Suetin Steklov Mathematical Institute of RAS The research was carried out with the nancial support of the Russian Foundation for Basic Research (grant no 05-01-01027) and the Programme of Support of Leading Scientic Schools (grant no NSh-4466.2006.1). suetin@mi.ras.ru 1. Let be a positive Borel measure with a compact support on R, function (z) := d(x) , zx z supp , / (1)

be a Markov function. In 1855 P. Chebyshv [1] assosiated to the function of type (1) the continued e fraction a1 a2 an b (z) C (z) = , () z b2 z bn z b1
b where aj > 0, bj R, a1 = (R). The n-th convergent Cn (z) = Pn (z)/Qn (z) to () has the following fundamental property b Cn (z) =

s0 s1 s2n1 + 2 + + 2n + , z z z

sk =

xk d(x),

(2)

polynomial Qn is of n degree and is orthogonal with respect to . b 2. By A. Markov theorem [2] Cn , n , locally uniformly in C \ supp . Let supp = [, ], > 0 a.e. on [, ] and rational function r C(z) H[, ]. Denote by G = {f : f = + r for a segment [, ]}. Let f be given by its power series expansion at innity

f=
k=0

ck . k+1 z

(3)

f A. Gonchar [3] proved that if f G then Cn f in spherical metric locally uniformly in C \ [, ]. f In particularly each pole of f of multiplycity m attractes exactly m poles of Cn . When measure f is complex but smooth enough the m poles of Cn , f = + r, asymptoticaly form the vertices of a regular m-polygon (A. Gonchar, S. Suetin [4]). The last result is valid for the functions of type f (z) = r1 (z)/ (z )(z ) + r2 (z), where r1 , r2 C(z). From the Gonchars theorem it follows that for f = + r G

f (z) =

f Cn0 (z)

+
k=n0

Ak , (Qk Qk+1 )(z)

z C \ supp ,

(4)

i.e. the Chebyshv continued fraction C f (z) provides a nonlinear method of summation of power e series (3). f cap Let f be an algebraic function given by (3). Then by H. Stahl [5] theorem Cn f on the compact subsets of D := C \ S, where f has the single-valued meromorphic continuation in D and the compact S has a special symmetric property (as a segment on R). From Stahls theorem it follows that Ak cap f , z C \ S. f (z) = Cn0 (z) + (5) (Qk Qk+1 )(z)
k=n0

82

f 3. The n-th convergent Cn to Chebyshv continued fraction C f depends only on the rst 2n e f coecients c0 , c1 , . . . , c2n1 of (3) just as the partial sum S2n . But from (4) and (5) it follows that f f in numerical analysis the {Cn }nN has a great advantage before the partial sums {S2n }nN . To illustrate that fact we calculated numerically [6] the limit cycle of the very well-known free van der Pol equation U + U = U (1 U 2 ), U = U (t, ),

for some values of , where is a small parameter, [0, +). References 1. P. Tchebycheff, Sur les fractions continues, Journ. de math. pures et appl. Sr. 2, 3 e (1858), 289323. 2. A. Markoff, Deux dmonstrations de la convergence de certaines fractions continues, Acta e Math., 19 (1895), 93104. Mathscinet 1554864. 3. A. A. Gonchar, On convergence of Pade approximants for some classes of meromorphic functions, Mat. Sb., 97(139): 4(8) (1975), 607629. Mathscinet 387552. 4. A. A. Gonchar, S. P. Suetin, On the Pade approximants to Markov type meromorphic functions, Sovremennye problemy matematiki (In Russian), 5, Steklov Mathematical Institute, Moscow, 2004. Mathscinet 2141293. 5. H. Stahl, Orthogonal polynomials with complex valued weight function, I, II, Constr. Approx., 2 (1986), 225240; 241251. Mathscinet 891973. 6. A. A. Gonchar, S. P. Suetin, Numerical analysis of some parameters of limit cycle of the free van der Pol equation, in Russian, in preparation.

Asymptotic solution of large Toeplitz systems with an innitely dierentiable symbol Sergey Sukhoverkhov, Igor B. Simonenko Southern Federal University, Rostov-na-Donu, Russia ssukhoverkhov@gmail.com We consider the asymptotics of the solutions of large linear systems with Toeplitz matrices generated by a complex valued symbol which is innitely dierentiable, has no zeroes and has a n zero argument increment. We construct an asymptotic inverse of n n Toeplitz matrix T , which approaches exact inverse faster then any positive degree of 1/n as n tends to plus innity. If the right hand side vector b is quasi-polynomial, i.e., b(k) = eik (k k0 )m we obtain an asymptotic formula n for the solution of T x = b. Using this formula dierent unknowns xk could be found independently of each other. That is, in order to nd particular components of the solution there is no necessity to solve the whole system. Special attention will be paid to the asymptotic formula of the central part of the solution, because it has more simple form.

Matrix approach to model of polarized radiation transfer in heterogeneous systems 83

Tamara A. Sushkevich, Sergey A. Strelkov, Sveta V. Maksakova Keldysh Institute for Applied Mathematics, Russian Academy of Sciences 4, Miusskaya Ploschadj, Moscow, 125047 Russia The work have been supported by the Russian Foundation for Basic Research (projects 06-01-00666, 05-01-00202). tamaras@keldysh.ru The mathematical models of the matrix-vector transfer operators are suggested to compute Stokes parameters vector as the solution of the general boundary-value problem in polarized radiation transfer theory for the 1d, 2d, 3d plane layers and spherical shells. Space and angular distributions of polarized radiation inside the relevant layer of these media as well as reected and passed through the layer radiation are formed as a result of multiple scattering and absorption. A new approach is proposed to radiation transfer modeling in optically thick layers that are represented as a heterogeneous multi-layer system each layer of which is described by dierent optical characteristics. The kernels of the matrix-vector transfer operators are the tensors of the inuence functions or the tensors of the spatial-frequency characteristics. The tensors of the inuence functions of each layer are determined by solutions of the radiation transfer rst boundary vector problem with an vector external source function while the contribution of multiple scattering and absorption in the layer is taking into account (Sushkevich, 2005). The basic models of the tensors of the inuence functions and the tensors of the spatial-frequency characteristics are stated. The matrix-vector operation, which describe one act of polarized radiation interaction with internal and external boundaries and take into account multiply scattering in the layers through tensor of the inuence functions, is determined. The boundary vector problem for each layer can be solved depending on optical thickness, scattering and absorption characteristics by one of the following techniques: i) as a solution of transfer equation with an azimuth dependence; ii) as a solution of the problem with azimuth symmetry; iii) as a solution in two-ux approach; iv) as an approximate solution in asymptotic approach. The matrix operators of radiation transmittance and reectance on the boundaries between the layers are formulated based on the collision integrals and the separate layers are united in a system by these operators. To calculate the total radiation inside or on the boundaries of the system with radiation exchange between the layers a matrix-vector operation is constructed for the operators the kernels of which are given by the tensor of the inuence functions of the layers. The representation of the solution to boundary-value vector problem as a functional is the transfer operator of the radiation transfer system which establishes the explicit relationship between the recording radiation and the scenarios (the optical image) at the dividing boundaries of media. In turn, by the use of the tensors of the inuence functions, the scenarios is described clearly through the characteristics of the reection and transmission of the dividing boundaries at the given its illumination. The tensors of the inuence functions are invariant about the conditions of the illumination and the properties of the dividing boundaries. This approach enables to simulate radiation elds in a wide range of variations of the optical characteristics of these layers and to analyze mechanisms of radiation characteristics formation inside and outside the layers as well as to estimate any contribution of each region. Reected radiation and distributions of the polarized radiation characteristics inside the layer near to the border illuminated by an external ux are thus calculated with higher accuracy than in common-used techniques. Original mathematical tools are proposed for the rst time to model radiation transfer in multilayer non-homogeneous heterogeneous media with dierent radiation regimes within separate regions of the system. The proposed approach is based on construction of generalized solutions in the form of matrix functionals, tensors of inuence functions for each layer of the system serve as kernels of 84

the functionals. The tensors of the inuence functions of the layers with dierent characteristics of scattering and absorption and radiation regimes can thus be calculated by various methods in dierent approaches of radiation transfer theory. Development of new mathematical tools to be realized on high-output multi-processor computer systems is needed to solve direct and inverse problems of radiation transfer theory for natural heterogeneous media as related to international cooperation on air-space global monitoring of the Earth as well as to an international global project on research. New opportunities of the proposed model are linked with verication of engineering applied techniques and plain layers approach which are often used for an express-analysis of space data in remote sensing and in radiation blocks for models of climate, circulation, forecast, photo-chemical kinetics, dynamics of ozone-sphere, trans-boundary displacement of pollutions for an air basin, etc. References 1. Sushkevich T.A., Mathematical models of radiation transfer, Moscow, BINOM. Knowledge Laboratory, 2005. 661 p.

Spectral ltering methods with Reective/Anti-Reective boundary conditions Cristina Tablino-Possio Dipartimento di Matematica e Applicazioni, Universit` di Milano Bicocca, via Cozzi 53, 20125 Milano, Italy a The work of the author is supported in part by MIUR grant no. 2006017542. cristina.tablinopossio@unimib.it In this talk, we analyze and compare some spectral ltering methods as truncated singular/eigenvalue Decompositions and Tikhonov/Re-blurring regularizations in the case of the recently proposed Reective [10] and Anti-Reective [11] boundary conditions. Numerical evidence is given to the fact that spectral decompositions (SDs) provide a good image restoration quality and in particular the Anti-Reective spectral decomposition, despite its loss of orthogonality. The related computational cost is comparable with previously known spectral decompositions. The model extension to the cross-channel blurring phenomenon of color images is also considered and the quoted spectral ltering methods are suitable adapted. ` 1. A. Arico, M. Donatelli, S. Serra Capizzano, Spectral analysis of the anti-reective algebra, Linear Algebra Appl., to appear. ` 2. A. Arico, M. Donatelli, J. Nagy, S. Serra-Capizzano, The anti-reective transform and regularization by ltering, 2006, submitted. 3. M. Bertero, P. Boccacci, Introduction to inverse problems in imaging, Inst. of Physics Publ. London, UK, 1998. 4. M. Donatelli, C. Estatico, A. Martinelli, S. Serra Capizzano, Improved image deblurring with anti-reective boundary conditions and re-blurring, Inverse Problems, 22, pp. 2035 2053, 2006. 5. M. Donatelli, S. Serra Capizzano, Anti-reective boundary conditions and re-blurring, Inverse Problems, 21, pp. 169182, 2005.

85

6. H. Engl, M. Hanke, A. Neubauer, Regularization of Inverse Problems, Kluwer Academic Publishers, Dordrecht, The Netherlands, 2000. 7. C.W. Groetsch, The Theory of Tikhonov Regularization for Fredholm Integral Equations of the First Kind, Pitman, Boston, 1984. 8. P.C. Hansen, Rank-decient and discrete ill-posed problems, SIAM, Philadelphia, PA, 1997. 9. P.C. Hansen, J. Nagy, D.P. OLeary, Deblurring Images Matrices, Spectra and Filtering, SIAM Publications, Philadelphia, 2006. 10. M.K. Ng, R.H. Chan, W.C. Tang, A fast algorithm for deblurring models with Neumann boundary conditions, SIAM J. Sci. Comput., 21, no. 3, pp. 851866, 1999. 11. S. Serra Capizzano, A note on anti-reective boundary conditions and fast deblurring models, SIAM J. Sci. Comput., 253, pp. 13071325, 2003.

Monotonicity of operator with respect to preordering generated by coupling of subadditive measures L. Truet Ecole des Mines de Nantes, Dpt Automatique et Productique, 4, rue Alfred Kastler BP. 20722, 44307 Nantes, Cedex 3, France Laurent.Truffet@emn.fr We are studying coupling techniques of particular subadditive measures. These particular measures we are studying are set functions with values in an idempotent semiring (a semiring which addition is idempotent). Such measures are used for modelling nonprobabilistic approaches of uncertainity not as a contrast but as a generalization of the probabilistic approach. They are called pseudomeasures. Inspired by the probabilistic approach, we dene the (increasing) coupling pseudo-measure of two pseudo-measures on a dicrete space. Then we derive the necessary and sucient condition of the existence of a coupling measure between two given pseudo-measures. We also chracterize the existence of an increasing pseudo-measure of two given pseudo-measures as a preorder between the two given pseudo-measures. Next we dene and characterize the monotonicity of a linear operator over an idempotent semiring with respect to the preordering generated by an increasing coupling of pseudo-measures.

Stable rational functions and Hurwitzlike formulas for Hurwitz matrices Yu. S. Barkovskii , M. Yu. Tyaglov , M. V. Shubnyi Southern Federal University, Rostov-on-Don Technische Universitt Berlin, Institut f r Mathematik a u tyaglov@gmail.com

86

Lets consider rational function R(z) = h(z) = t0 z rm + t1 z rm1 + t2 z rm2 + . . . . g(z) (1)

Here h(z) and g(z) are real polynomials: h(z) = b0 z r + b1 z r1 + . . . + br1 z + br , g(z) = c0 z m + c1 z m1 + . . . + cm1 z + cm , and r, m N {0} (r + m > 0). b0 , b1 , . . . , br R, b0 > 0; (2)

c0 , c1 , . . . , cm R, c0 > 0.

Denition. Real rational function is called stable if all its roots lye in an open left half-plane of complex plane and all its poles lye in an open right half-plane. It is proved that Theorem 1. Real rational function R(z) (1)(2) is stable i all successive principal minors k (R) of innite Hurwitz matrix t1 t3 t5 t7 t9 . . . t0 t2 t4 t6 t8 . . . 0 t1 t3 t5 t7 . . . H(R) = 0 t0 t2 t4 t6 . . . 0 0 t1 t3 t5 . . . 0 0 t0 t2 t4 . . . ...................... are positive up to order n = r + m: 1 (R) > 0, 2 (R) > 0, . . . , n (R) > 0.

There is the following Theorem 2. Real rational function R(z) (1)(2) is stable i 2 > 0, 4 > 0, . . . , 2n > 0, where 2j are determinants of order 2j formed of coecients of polynomials h(z) and g(z): c0 b0 0 0 0 0 ... 0 0 c1 b1 0 b0 0 0 ... 0 0 c2 b2 c0 b1 0 b0 ... 0 0 c3 b3 c1 b2 0 b1 ... 0 0 c4 b4 c2 b3 c0 b2 ... 0 0 c5 b5 c3 b4 c1 b3 ... 0 0 ... ... ... ... ... ... ... ... ... cj1 bj1 cj3 bj2 cj5 bj3 ... 0 b0 cj bj cj2 bj1 cj4 bj2 ... 0 b1 ... ... ... ... ... ... ... ... ... c2j2 b2j2 c2j4 b2j3 c2j6 b2j4 ... c0 bj1 c2j1 b2j1 c2j3 b2j2 c2j5 b2j3 ... c1 bj

2j

We assume that bk = 0 whenever k > r and cj = 0 whenever j > m.

87

On the Inner and Outer Geometry of the Field of Values using Geometric Computing Methods Frank Uhlig Department of Mathematics, Auburn University Auburn, AL 368495310, USA uhligfd@auburn.edu The eld of values F (A) = {x Ax C | x Cn , x = 1} of a matrix A Cn,n is a compact and convex subset of the complex plane. We develop an algorithm to decide whether a given point p C lies inside or outside the eld of values of a given matrix and that computes the distance, or generalized Crawford number, of an arbitrary point p C from the boundary of F (A). Moreover for a point p F (A) we describe a further algorithm for the inverse problem of nding a unit vector x Cn with x Ax = p. These algorithms use simple and ecient methods of geometric computing and lead to questions about the covering number of points p inside F (A). Some of the questions are answered and others are formulated as conjectures on the inner geometry of the quadratic map : x Cn , x = 1 x Ax C.

SSA-based approach to decomposition of two-dimensional scalar elds Konstantin Usevich , Nina Golyandina St.Petersburg State University, Mathematical Department usevich.k.d@gmail.com, nina@gistatgroup.com Singular Spectrum Analysis (SSA) has been approved in many areas as an eective modelfree technique of analysis of one-dimensional time series [1,3] (the word time is not essential here as the method deals with arbitrary equidistant measurements). The SSA approach (another name is Caterpillar [2]) can be applied to such problems as decomposition of time series into sum of trend, oscillations, and noise, detection of periodicities, smoothing, signal denoising, forecasting extracted time series components, imputation of missing data, change-point detection (see http://www.gistatgroup.com/cat/ (in English) and http://www.gistatgroup.com/gus/ (in Russian)). In addition, the SSA technique has much in common with a family of signal processing methods that are devoted to signal denoising for damped/undamped sinusoids and estimation of signal parameters. Let us shortly describe the basic algorithm of SSA applied to the time series FN = (f0 , . . . , fN 1 ). For a given window length L, 1 < L < N , there exists a one-to-one correspondence between sequences of length N and Hankel L K-matrices (called trajectory matrices), where K = N L + 1: FN X = (xij )L,K , xij = fi+j2 . First, we perform the singular value decomposition of the trajectory i,j=1 matrix constructed from the time series FN . Then we group the SVD matrix components, sum them and obtain a decomposition of the matrix X = Xi . Finally, by means of projection on the set
i

of Hankel (trajectory) matrices in Frobenius norm, we come to a decomposition of the initial time series. 88

and xed window sizes (Lr , Lc ), 1 Lr Nr , 1 Lc Nc , 1 < Lr Lc < Nr Nc , we introduce two equivalent representations of the trajectory matrix analogue. The rst representation is the block Hankel matrix with rectangular blocks; let us call it the block trajectory matrix: A1 A2 . . . AK r A2 A3 . . . AKr +1 W= . (1) , . . .. . . . . . . . ALr ALr +1 . . . A Nr where Kr = Nr Lr + 1 and Ai is the Lc -trajectory fi,1 fi,2 fi,2 fi,3 Ai = . . . . . . fi,Lc matrix for the i-th row of the matrix F: . . . fi,Kc . . . fi,Kc +1 , . .. . . . fi,Nc

If we solve the problem of signal denoising for nite-rank signals (i.e., signals with rank-decient trajectory matrices) by SSA, we should group the d leading SVD components, where d is equal to the signal rank. The corresponding problem of signal processing is solved via approximation of the trajectory full-rank Hankel matrix by a low-rank Hankel matrix. SSA has extensions to simultaneous analysis of several time series (MSSA [2,1]) and to analysis of 2D scalar elds (2D-SSA, [2]). The 2D-SSA algorithm coincides with the basic (one-dimensional) SSA algorithm up to construction of an analogue of the trajectory matrix. For a given eld f1,1 f1,2 . . . f1,Nc f2,1 f2,2 . . . f2,Nc F= . . . .. . . . . . . . fNr ,1 fNr ,2 . . . fNr ,Nc

fi,Lc +1 . . .

In the above notation, the decomposition step of 2D-SSA (an analogue of the SVD step in the one-dimensional case) consists in the SVD of the block trajectory matrix (1) or, that is equivalent, the tensor Kronecker-Product SVD of the 2D-trajectory matrix (3). 89

Therefore, the other natural form of W is the so called 2D-trajectory matrix F1,1 F1,2 . . . F1,Kc F2,1 F2,2 . . . F2,Kc X= . . . . .. . . . . . . . FKr ,1 FKr ,2 . . . FKr ,Kc

where Kc = Nc Lc + 1. Note that the column vectors of the block trajectory matrix (1) can be expressed as the vectorized moving 2D-windows fi,j fi,j+1 . . . fi,j+Lc1 fi+1,j fi+1,j+1 . . . fi+1,j+Lc1 Fi,j = . (2) . . . .. . . . . . . . fi+Lr 1,j fi+Lr 1,j+1 . . . fi+Lr 1,j+Lc 1

(3)

There are some methods in a sense similar to 2D-SSA. For example, there are results on extraction and parameter estimation for sums of products of damped sinusoids from noisy 2D elds by means of approximation of a matrix like (1), with no references to SSA [4]. Also, there is a method of image processing that uses the KP-SVD of matrices having the form (3); however, the submatrices corresponding to (2) are non-overlapping [5], unlike 2D-SSA. The aim of this paper is to describe the 2D-SSA approach to 2D elds (images) decomposition, including the transfer of the main SSA concepts (separability, time series of nite rank, continuation) [3] to the two-dimensional case and an overview of related two-dimensional problems that can be solved by 2D-SSA (e.g. smoothing, image/eld denoising, image reconstruction, periodic noise removal, lling in missing data). References 1. Elsner, J., Tsonis, A., Singular Spectrum Analysis. A New Tool in Time Series Analysis, New York, Plenum Press, 1996. 2. Danilov, D. and Zhigljavsky, A. (Eds.), Principal Components of Time Series: the Caterpillar method, St.-Petersburg, SPbSU Press, 1997. (in Russian). 3. Golyandina, N., Nekrutkin, V., and Zhigljavsky, A., Analysis of Time Series Structure: SSA and Related Techniques, London, Chapman & Hall/CRC, 2001. 4. Rouquette, S. and Najim, M., Estimation of frequencies and damping factors by twodimensional ESPRIT-type methods, IEEE Transactions on Signal Processing, 49(1) (2001), 237245. 5. Zhang, H. et al., Representing Images by Multiple Kronecker Product Sum. CAD/Graphics, 2003.

Factor rank of non-negative matrices with xed classical rank Olga A. Vaysman Lomonosov Moscow State University science@land.ru The theory of non-negative matrices is a classical branch of algebra which is actively developing nowadays, see for example [1] and its references. As regards the matrices over a eld, rank is one of the most eective invariants. Semiring matrix theory has the same situation with one distinction: there are several dierent rank functions which can be investigated together or separately. Also if the semiring under consideration is embedded in a eld we can investigate the dierence between classical rank over the eld and ranks over the semiring. Let us consider non-negative real numbers R+ . One of the most important in dierent applications rank functions for matrices over R+ is dened in the following way: Denition 1. Matrix A Mm,n (R+ ), A = 0, has factor rank equal to k (f (A) = k) if k is the least non-negative integer such that there exists a decomposition A = BC where B Mm,k (R+ ), C Mk,n (R+ ). By denition, it is said that f (0) = 0. It is easy to see that for any matrix A Mm,n (R+ ) it holds that rk(A) f (A). The following theorem shows that non-negative matrices with xed classical rank can have an arbitrary large factor rank. That is, in a certain way there is no dependence between classical and factor rank of matrices with non-negative real entries. 90

Theorem 1. For all natural numbers k r 3 there exists a natural number n and a matrix A Mn,n (R+ ) such that rk(A) = r and f (A) > k where rk(A) denotes the classical rank of A. I wish to thank my scientic advisor Dr. Alexander E. Guterman for problem statement and helpful hints. References 1. R. Brualdi and H. Ryser, Combinatorial Matrix Theory, Cambridge University Press, Cambridge, 1991.

Product/quotient SVD using rank structured matrices Yvette Vanberghen, Marc Van Barel, Paul Van Dooren Katholieke Universiteit Leuven, Department of Computer Science, Belgium marc.vanbarel@cs.kuleuven.be In this talk we reduce the computation of the singular values of a general product/quotient of matrices to the computation of the singular values of an upper triangular semiseparable matrix. Compared to the reduction into a bidiagonal matrix the reduction into semiseparable form exhibits a nested subspace iteration. Hence, when there are large gaps between the singular values, these gaps manifest themselves already during the reduction algorithm in contrast to the bidiagonal case.

Band plus Algebra Preconditioners for BTTB Toeplitz Systems D. Noutsos and P. Vassalos Department of Mathematics, University of Ioannina, GR-451 10 Ioannina, Greece Department of Informatics, Athens University of Business and Economics, GR-104 34, Athens, Greece dnoutsos@cc.uoi.gr, pvassal@aueb.gr We are concerned with the fast and ecient solution of nm nm symmetric ill-conditioned block Toeplitz with Toeplitz blocks (BTTB) linear systems of the form Tnm (f )x = b where the generating function f is a priori known and in particular is real valued, nonnegative, having isolated roots of even order. The preconditioner that we propose is a product of a band Toeplitz matrix and matrices that belong to a certain trigonometric algebra. The underline idea of the proposed scheme, is to embody the well known advantages that each component of the product presents when it is used alone. As a result we obtain a exible preconditioner which can be applied to the system Tnm (f )x = b infusing fast convergence to the PCG method. For that, we compare the proposed method with techniques already employed in the literature. Our results fully conrm the eectiveness of the proposed strategy and the adherence to the theoretical analysis.

91

Iterative technologies for sti problems with jumping anisotropic diusion tensors Yu. Vassilevski , S. Goreinov and V. Chugunov Institute of Numerical Mathematics, Russian Academy of Sciences, 8, Gubkina str., 119991, Moscow, Russia vasilevs@dodo.inm.ras.ru Finite element discretizations of the diusion equation produce large and sti systems with sparse matrices. Spectral properties of the matrices deteriorate with the mesh renement. The degradation is even stronger in the case of jumping and anisotropic diusion tensors. Several blackbox techniques for the iterative solution of the above systems are reviewed. Numerical comparisons for a 3D model problem are presented.

An Enestrm-Kakeya theorem for hermitian polynomial matrices o Harald K. Wimmer Mathematisches Institut, Universitt Wrzburg, D-97074 Wrzburg, Germany a u u wimmer@mathematik.uni-wuerzburg.de The Enestrm-Kakeya theorem states that a real polynomial h(z) = am z m + + a1 z + a0 with o am a1 a0 0, am > 0, has all its zeros in the closed unit disc and that the zeros on the unit circle are simple. We extend the Enestrm-Kakeya theorem and its renement by Hurwitz to o polynomial matrices H(z) with positive semidenite coecients. We determine an annular region containing the zeros of det H(z). A stability result for systems of linear dierence equations is given as an application. (Joint work with G. Dirr.)

Positivity criteria generalizing the leading principal minors criterion Nadya Zharko Mech.-Math. Faculty, Kiev National University, Vladimirskaya 64, Kiev, Ukraine n.zharko@mail.ru This is joint work with Vyacheslav Futorny and Vladimir V. Sergeichuk. An nn Hermitian matrix is positive denite if and only if all leading principal minors 1 , . . . , n are positive. We show that certain sums l of l l principal minors can be used instead of l in this criterion. We describe all suitable sums l for 3 3 Hermitian matrices. For an n n Hermitian matrix A partitioned into blocks Aij with square diagonal blocks, we prove that A is positive denite if and only if the following numbers l are positive: l is the sum of all l l principal minors that contain the leading block submatrix [Aij ]k1 (if k > 1) and that are contained in [Aij ]k , where i,j=1 i,j=1 k is the index of the block Akk containing the (l, l) diagonal entry of A. We also prove that l can be used instead of l in other inertia problems. 92

References 1. V. Futorny, V.V. Sergeichuk, N. Zharko, Positivity criteria generalizing the leading principal minors criterion, Positivity 11 (1) (2007) 191199.

The Method of Regularization of Tikhonov Based on Augmented Systems A. I. Zhdanov , T. G. Parchaikina Samara State Aerospace University, Moskovskoe sh. 34, Samara, 443086 Russia zhdanov@ssau.ru Formulation of the problem. Solving an approximate systems of linear algebraic equations (SLAEs) is fundamental problem of numerical analysis. Let the exact SLAE Au = f (1) be specied by the (a priori unknown) initial data d = {A, f }, where A = (aij ) Rmn , f = (f1 , . . . , fm ) Rm and u = (u1 , . . . , un ) Rn . In the general case, the solution to exact SLAE (1) is understood as its normal pseudosolution u = A+ f, where A+ is the pseudoinverse or the MoorePenrose generalized of A. Then, the inconsistency measure of the exact SLAE (1) is dened by = infn Au f = Au f 0.
uR

(2)

Throughout the report, the vector norms in Rm and Rn are the Euclidean (quadratic) norms; i.e., r = r 2 , where r = f Au is the residual. The information on system (1) is given by the approximate data d = {A, f } (i.e., by an individual = f ) such that approximate SLAE Au A A h, f f ,

where the scalars h 0 and 0 specify the errors in the assignment of the approximate data d and A is the spectral norm of A. If rank A < min(m, n), then solving system (1) (as specied in (2)) on the basis of approximate data d with h > 0 is an ill-posed problem in the sense of Hadamard, because the approximate normal pseudosolution u = A+ f is unstable for innitesimal perturbations in the initial data. To determine stable solutions to system (1) from approximate data d, various regularization methods are used. The regularization method of A.N. Tikhonov is among the most universal ones. It is well known that, in this method, the regularized solution u is determined as a (unique) solution to the Euler equation (A A + En )u = A f , > 0, (3) 93

where En is the identity matrix of order n and is the regularization parameter. Method of augmented systems. In this report, a approach is proposed for systems with > 0. This approach is based on the reduction of the original system to an equivalent consistent augmented system. Moreover, the proposed approach makes it possible to obtain ecient numerical algorithms for solving the problem under discussion. The normal system of equations is equivalent to the augmented system r + Au = f , r = 0 A Em A A 0 r u = f 0 Gz = b. (4)

where z = ( , u ) Rm+n . r Hence, the regularized solution z to system (4) is determined as a (unique) solution to the Euler equation b. (G2 + Em+n )z = G (5) Thus, using equivalent augmented systems, we can directly extend the basic results obtained for consistent systems ( = 0) to the class of inconsistent systems ( > 0). Using the symmetry of G, we can reduce the condition number of the regularized (5). To this end, we apply the method of an imaginary shift of the spectrum (Faddeevas method). According to this method, we replace Eq. (5) by the equation b, (6) (G + i E)z = where i = 1 is the imaginary unit. An immediate implication of (6) is that x = Re z is a solution to the equation (5). Condition number of system (6) is estimated by inequality cond2 (G + i E) where 1 is the maximal singular value of A. References 1. Zhdanov A. I., Regularization of Unstable Finite-Dimensional Linear Problems Based on Augmented Systems, Comp. Math. and Math. Physics, Vol. 45, No. 11, 2005, pp. 18451853. 1+ 1 4
2

1+

1 + 41 2

The radioactive pollution transport modeling in air nearby Volgodonsk nuclear station V. N. Zubov, L. A. Krukier , G. V. Muratova, T. N. Subbotina Computer Center of Southern Federal University, Rostov-on-Don, Russia zubov@rsu.ru, krukier@rsu.ru, tsub@rsu.ru The importance of the atmosphere radio nuclear pollution modeling directly connected with the development of nuclear industry. Like in many other regions of Russia, the lack of hydrocarbon energy exists in the North Caucasus. Hence Volgodonsk nuclear station (VNS) has great economical 94

value in the South of Russia. The rst module of APP was included into the common power grid in 2001, and in June of 2006 the decision of building the second module was taken. At the same time VNS brings extra risks in the ecosystem of the region. And such risks are as important as ordinary energy supply problems. Thus radioactive safety around the nuclear station and its environmental territory is of great importance. Modeling of dierent out of order and extreme situation in atmosphere nearby nuclear station can be considered as investigation of large-scale transport processes with point source of studied substance. This problem can be divided into several parts: prediction of pollution movement in air, deposition on water and soil, analysis of possible eects on activity of population. The prediction supposes the possibility of taking decisions immediately with certain accuracy. Each model of pollution transport consists of two main components: Meteorological model of the wind eld Transport model which determines concentration of considered substance according to diusion, advection, deposition, interaction between substances and other important processes. The main purpose in modeling of radioactive pollution transport is estimation of radioactive inuence risks. Thus it is rather convenient to use activity of radioactive substances instead of its quantity in calculations. Model of radionuclide transport in atmosphere is based on 3-dimensional turbulent diusion equation in Cartesian coordinates. The medium can be considered as incompressible with regard to permanency of air density in the domain of calculations. = 1 v1 + (v1 ) + v2 + (v2 ) + v3 + (v3 ) + t 2 x x y y z z 2 2 +Kx x2 + Ky y2 + z Kz z + +E (x, y, z, t) vd kr , | = (s, t) , s D. + t=0 = 0 (x, y, z) , n
D

The results received in numerical experiments allow to estimate ecological situation nearby VNS working in regular and irregular regimes.

Experiments can be applied in taking precautions against unfavorable inuence of radio nuclear substances and making a shield plan in case of unpredictable situations.

95

96

You might also like