VIEWS: 45 PAGES: 26 POSTED ON: 8/6/2011 Public Domain
Curl-Curl-Eigenvalue Equation jS ≡ 0 ( no current excitation ) σ = 0, ε , µ real ( loss-free ) M ε 1 CM µ−1 C e = ω 2 e − eigenvalue problem Ax = λ x − A CC = Mε 1 CM µ−1 C ε −1 curl µ-1 curl E = ω 2 E Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics For the special case where no exciting currents and no conducting materials (σ = 0) are present and all materials are loss-free (ε, µ real) the discrete wave equation of FIT resultss in the equation displayed above. This is an algebraic eigenvalue problem of the form Ax = λx. The solutions of such kind of equations are non-trivial eigenvalues xj ≠0 and the according real or complex eigenvalues λj. A matrix of dimension n has maximally n linear independent eigensolutions. The eigenvalue problem is called discrete Curl-Curl-eigenvalue equation. It´s system matrix is denoted as ACC. The Curl-Curl-eigenvalues equation corresponds to the continuous eigenvalue equation for resonators. Curl-Curl-Eigenvalue Equation Use transformation e ' = M 1/ 2 e ε to derive an equation with symmetric system matrix M 1/ 2 CM µ −1 CM ε 1/ 2 e ' = ω 2 e ' ε − − system matrix A ' = M 1/ 2 A CC M ε 1/ 2 ε − − = M ε 1/ 2 CM µ −1CM ε 1/ 2 ( )( ) T − = M ε 1/ 2 CM −1/ 2 µ − M ε 1/ 2 CM −1/ 2 µ is symmetric Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Eigenvalues of both equations are the quadratic resonance frequencies ω2 and the eigenvectors correspond each to the fields of the according resonator modes. Now we can study the properties of the FIT discretization by analyzing the algebraic properties of the system matrix ACC. In the actual form of the curl-curl-eigenvalue equation the matrix ACC is non- symmetric. Yet, using the transformation given above for the electric grid voltage we receive the above formulation with real symmetric system matrix A’. The eigenvalues of ACC and A’ are identical. Solution Space of Eigenvalue Problem λi = ω i2 ≥ 0 ∀ λi = eigenvalue of A ' Consistency between mathematical model and physics is based upon • Symmetry of the curl operators: C = CT • Possibility to split the material operators: M ε = M1/ 2 M1/ 2 and M µ = M1/ 2 M1/ 2 ε ε µ µ (real diagonal matrices with non-negative entries) Consider SC M µ−1 C e = ω 2S M ε e = ω 2S d = 0 =0 ⇒ either ω = 0 or S d = 0 Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics The solutions of the discrete eigenvalue equation are not only of interest if we directly want to determine the oscillating modes in resonators but they also hold information about the behaviour of time-variable fields obtained by FIT-discretization. Using purely algebraic reasoning we can get information about the location of the eigenvalues: Because of it´s symmetry the matrix A’ (and thus also ACC) has purely real eigenvalues. From it´s representation as product of two matrices transposed to each other it also follows that all eigenvalues are non-negative (A’ is positive semi-definite). Thus, also the eigenfrequencies ωj are real and non-negative. This corresponds to the physical property that loss-free resonators have real eigenfrequencies, only. This consistency between the mathematical model and physics is not self-evident. In this case it is based on the properties of • the symmetry of the discrete curl-operators of FIT, • the possibility to decompose the material operators because they are real diagonal matrices with non-negative entries. A further hint about the solution space of the eigenvalue problem can be obtained by the following physical consideration: If we multiply the curl-curl-eigenvalue equation from left by Mε and build the divergence next (multiplying from left the discrete divergence operator) we get the equation displayed above. This means that for all eigensolutions either the eigenfrequency ω or the divergence of the electric flux density have to vanish. Solution Space of Eigenvalue Problem Solution space L can be decomposed: L = LS ∪ Ld ω2 = 0 ω2 ≠ 0 1. Subspace L S of electrostatic solutions with ω = 0 : i potential approach: ES = −grad ϕ and eS = − −ST Φ ( ) ( ) T i because of curl grad = 0 and CST = SC =0 ⇒ curl ES = 0 and C eS = 0 ⇒ ω =0 i in general: div DS ≠ 0 and S d S ≠ 0 Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics The solution space L of the Curl-Curl-eigenvalue equation thus can be split into two disjoint subspaces: LS is the space of static solutions since it holds all eigensolutions λ=ω²=0. Ld is the space of dynamic solutions which all have resonance frequencies above zero and thus correspond to harmonic oscillating fields. 1. For the eigenvectors, i.e. the discrete field vectors, of both solution classes we find, using the properties of FIT – discretization: Electrostatic fields with ω=0: Such fields may be represented as gradient of a scalar potential. Because of curl grad=0 the curl of the electric field and electric grid voltage is vanishing for these modes. Inserting this into the curl-curl-eigenvalue equations we obtain ω=0 as it should be. In a grid of NP points there exist about Np linear independent potential vectors ϕ and thus about NP linear independent static eigenmodes (ignoring points with fixed potential, as e.g. at the boundary). This means that the eigenvalue λ=0 is (about) NP-fold degenerate. In general, the potential distribution does not fulfil the Laplace equation such that we need to assume charges in the grid points for the existence of such solutions. This means that the divergence of the flux density is not equal to zero. Note: In problems with multiple conductors (structures with several disjoint ideal conductors), there exist also some modes with ω=0 which are divergence- and curl-free. We will not treat these cases here. Solution Space of Eigenvalue Problem L = LS ∪ Ld ω2 = 0 ω2 ≠ 0 2. Subspace L d of dynamic solutions with ω ≠ 0 : i from Faraday's law we have: D5 ≠ 0 curl Ed ≠ 0 and C ed ≠ 0 i from Ampere's law we get: 1 1 Dd = curl H d and d d = C hd jω jω i because of div curl = 0 and SC = 0 we have div Dd = 0 and S d d = 0 Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Dynamic eigen-oscillations with ω≠0: According to Faraday´s induction law these modes possess a non- vanishing curl-contribution. The corresponding electric grid flux can be represented with help of Ampère´s law. Because of div curl=0 the dynamic modes are divergence – free. The exact separation of both subspaces with the provable curl-and divergence-freeness of the corresponding modes thus holds in the discrete case just because the matrix operators fulfil the necessary conditions. This is not at all self-evident and for some other discretization methods it does not hold! According to our previous considerations, the eigenvalue problem of dimension 3NP has the about NP – fold degenerated eigenvalue ω² = 0 and only about 2Np non-zero eigenvalues. The static modes, however, are meaningful solutions of Maxwell´s equations if and only if the charge distribution which results from div Ds ≠ 0 is allowed for the solution domain. Since this is generally not the case as usually div D = 0 is assumed for harmonic problems our system matrix turns out to be „too large“ by about 1/3. Solution Space of Eigenvalue Problem Im {λ } N-fold eigenvalue 0 (static solutions) Re {λ } ω12 ω 22 … ω 22N smallest (non-zero) eigenvalue = fundamental oscillation Symmetry and positive semi-definiteness of A' ⇒ orthogonality of the modes: e 'i ⋅ e ' j = 0 ∀i, j with λi ≠ λ j Orthogonality condition after back-transformation: ei ⋅ M ε e j = 0 and ei ⋅ d j = 0 (i ≠ j ) 4 x stored electric energy (non-zero only for i = j ) Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Consequently, we should search for some transformation which reduces the dimension of the system matrix from 3NP to 2NP. Building in the discrete Gauss law for electricity with vanishing right hand side (no charges) would yield such a transformation but destroy the band structure of the matrix affording higher computational effort for it´s storage. Instead of this it is more elegant to control the static modes by a modified formulation of the wave equation and/or by specific algebraic solution methods which will be treated later. Another property of the solution of the Curl-Curl matrix A´. For two eigenvectors, each, of this matrix orthogonality holds if the corresponding eigenvalues are not equal to each other. Hence, orthogonality holds for all pairs of non-degenerated modes – also for each pair build by one dynamic and one static mode. For degenerated dynamic modes (with λi=λj≠0) orthogonal linear combinations can always be found. After back-transformation from A` to Acc we obtain the orthogonality relation for dynamic eigenmodes given above. The scalar product of the vectors of electric grid voltage and electric grid flux formally corresponds to four times of the stored electric energy which is non-vanishing only if both vectors belong to the same field. Grid Dispersion Equation ( j ω t − k ⋅r ) Plane waves: E = E0 e ω2 Dispersion equation: k 2 = k x2 + k y + k z2 = 2 c2 − j ky y With e − j k ⋅ r = e − j k x x e e − j kz z we find the phase factor − j ky ∆ y T x = e− j kx ∆ x , Ty =e , T z = e− j kz ∆ z by which the complex grid voltage is multiplied when the wave proceeds by one step size in x-, y - or z -direction Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Based on the discrete wave equation we will introduce next the so–called grid dispersion equation which allows for an extensive analysis of the properties of the Maxwell-Grid-Equation (MGE) for time- dependant fields. The basis for the derivation is Fourier´s theorem according to which each wave may be represented as sum of elementary waves. In electrical engineering, this theorem is mainly applied to purely time- dependant signals. It may also be applied to space-dependant electromagnetic fields: Then, elementary waves are fields which can be represented in space and time by harmonic functions, i.e. sine, cosine or the complex exponential function. Physically, these are just plane waves. In homogeneous space (vacuum) the plane waves are solution of (continuous) Maxwell´s equations for which the frequency ω or and the wave number k = |k| fulfil the dispersion equation. The idea of the following analysis is to check under which conditions plane waves exist on the grid, i.e. we can derive an according representation for the discrete field vectors which solve the MGE. First, we will study the time-harmonic case. In order to eliminate the influence of the boundary conditions we assume an indefinitely large Cartesian computation grid and equidistant step size in all three coordinate direction, i.e ∆u=∆v=∆w for all i=1,..., I, j=1,...,J, K=1, ..., K. Further, we assume constant material coefficients ε0, µ0 and σ=0 (vacuum), Taking the continuous definition of a plane wave we impress the given electric field of the plane wave into the vector components of the electric grid voltage by evaluation of the corresponding path integral for each edge in the grid. Then, we find that the complex grid voltage has to be multiplied with the phase factors given above as the wave proceeds one grid step in x-, y- and z-direction, respectively. Thus, the complete discrete field vector can be explained by one voltage component per direction in space which will be denoted as shown here (a hat as sign for the wave amplitudes). As arbitrary origin for these discrete wave amplitudes we choose the three voltage components associated to some grid point n0 (usual indexing). Grid Dispersion Equation Regard Faraday's induction law: e3 = E y ⋅ T z z b1 = B x e4 = E z e2 = E z ⋅ T y x y e1 = E y We have − jω b1 = e1 − e3 + e2 − e4 = e1 (1 − T z ) + e4 ( −1 + T y ) and − jω B x = E y (1 − T z ) + E z ( −1 + T y ) for the wave amplitudes Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics In the next step we will complete the time derivative of the magnetic flux component according to Faraday´s induction law for the corresponding grid areas. As we study the frequency domain, the time derivative is given by multiplication by јω. With the notations of the sketch we get the relations shown here. Grid Dispersion Equation Discrete induction law: 0 − (T z − 1) T y −1 E x Bx T z −1 0 − (T x − 1) E y = − jω B y − (T y − 1) T x −1 0 Ez B z C ∆x H x µ ∆ y ∆z Bx ∆y ⇒ Hy = By ∆x =∆x , etc. µ ∆z ∆x homogeneous Hz Bz material ∆z µ ∆x ∆ y −1 Mµ Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Thus, for all three directions in space, we obtain a mapping of the discrete induction law to the 3 x 3 system of equations as displayed above. So we have determined the magnetic flux components. Next, we can use the material equations to determine the wave amplitudes of the magnetic grid voltages. The discrete relation for the simple case of a homogeneous equidistant grid is given here. Grid Dispersion Equation Regard Ampere's law: 1 Hy Tx 1 Hx jω D z = ( ) ( ) Ty Dz −1 −1 Hx − 1− T y H x + 1− T x H y Hy Discrete law: 0 ( − 1− T z −1 ) 1− T y −1 Hx Dx 1 − T −1 z 0 ( − 1− T x −1 ) H y = jω D y Hz ( − 1 − T −1 y ) 1− T x −1 0 Dz C Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Next, we will apply the discrete Ampère´s law to the corresponding dual area as shown in this sketch. The inverse phase factors in the resulting equation just have the effect of a displacement in negative coordinate direction. Again, we can summarize the equations for all three directions in space obtaining the system of equations given here. As we can easily verify the system matrices („curl operators“) again fulfil the relation that the one on the primary grid equals the Hermitian one the dual grid (Hermitian matrix = transposed plus conjugate complex). Grid Dispersion Equation Electric material matrix: ∆x Ex ε ∆ y ∆z Dx ∆y Ey = Dy ε ∆z ∆x E D z ∆z z ε ∆x ∆ y −1 Mε Ex Amplitude vector: e = E y Ez Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Finally , the electric material matrix is applied in order to transform the electric grid fluxes into grid voltages. Thus we have transformed all equations now to the three-dimensional space of wave amplitudes. Grid Dispersion Equation The local Curl-Curl-wave equation −1 H −1 Mε C M µ C e = ω 2 e has to hold for e ! Using the relation e j x = cos ( x ) + j sin ( x ) the lengthy solution finally yields λ 1 = 0 (the static solution), 2 2 kx ∆ x k y ∆ y kz ∆ z 2 sin sin sin 2 2 + 2 + 2 λ 2 = λ 3 = c ∆x ∆y ∆z 2 2 2 Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Sequentially applying the equations previously derived, a condition may be formulated under which the discrete plane wave fulfils the Maxwell-Grid-Equations as demanded before (thus justifying the approach used here). The electric amplitude vector has to fulfil the local Curl-Curl-Wave equation. This is an algebraic 3x3 eigenvalue equation for the electric amplitude vector with eigenvalue λ=ω². After some lengthy calculation and paying attention to the relation ejx = cos(x) + j sin (x) one eigenvalue λ1=0 and a double eigenvalue λ2= λ3 as given above are found. The solution λ=ω²=0 is denoted as static solution. It is not regarded here. The two solutions λ2,3≠0 describe the discrete waves, searched for. The corresponding two-dimensional space of their eigenvectors indicates the two polarisation directions of the plane wave. Grid Dispersion Equation The grid dispersion equation 2 kx ∆ x k y ∆ y kz ∆ z 2 2 sin 2 sin 2 sin 2 2 + + = ω ∆x ∆y ∆z c 2 2 2 is the condition for the existence of the discrete waves belonging to the eigenvalues λ 2,3 ≠ 0. Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Finally, the condition for the existence of these waves is given as so-called grid dispersion equation. It connects the properties of the spatial discretization on the left hand side with the second time derivative of the wave equation which arises as factor –ω² in frequency domain (the negative sign is compensated by that in Faraday´s induction law). The grid dispersion equation fixes how plane waves propagate in the computational FIT-grid. For example, we can see that the propagation depends on the step size as well as on the propagation direction. Very important in that is that the grid dispersion equation tends to the analytic dispersion equation as we build the limits ∆x, ∆y, ∆z, ∆t = 0. This is another proof of the convergence of FIT. Please note that we used some relevant assumptions on the grid for the derivation above: an equidistant discretization in all spatial directions, homogeneous material and no boundary conditions. Nevertheless, the grid dispersion equation is of great importance for the analysis of FIT´s properties as will get more obvious in the following chapters. Solution of the Eigenvalue Equation Loss-free (real ε , σ =0): M ε 1 CM µ −1 C e = ω 2 e − eliminate the N P solutions with λi = 0: dynamic modes with: div(ε E ) = 0 grad div D = 0 − ST SM ε e = 0 Transformation e ' = M ε 2 e: 1/ -M1/2ST SM1/2 e ' = 0 ε ε (M −1/ 2 ε CM µ−1 CM ε 1/ 2 + M1/ 2ST SM1/ 2 e ' = ω 2 e ' − ε ε ) (ε -1 curl µ−1 curl- grad div ε ) E = ω 2 E "Nabla-Squared Eigenvalue Eq." ( curl curl - grad div = -∇ 2 ) Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics The eigenvalue equation of FIT for time-harmonic fields in closed structures (resonators) has been formulated in the previous chapter. For the loss-free case (real material coefficients and vanishing conductivity) it is repeated here. Usually, the dynamic modes with the smallest non-zero eigenvalues λ=ω²≠0 are of technical interest. As previously found, in this regime the spectrum of the system matrix is determined by the multiple degenerated eigenvalue λ=0. Since this situation poses great numerical difficulties for most eigenvalue solvers we will derive an alternative formulation in the sequel. Solution methods which search for the smallest eigenvalues and the corresponding eigenvectors of a matrix would first compute all eigenvalues with ω²=0 and only then they could compute the fundamental mode as second-smallest eigenvalue. Therefore, our goal is to eliminate the NP-fold eigenvalue zero in the matrix ACC. For all dynamic modes searched for we have div (ε E)=0 and the corresponding discrete equation in grid space. Now we multiply the discrete equation from left with the discrete gradient operator (continuously, we build the gradient of the left hand side.) Replacing now the electric grid voltage by its transformed form we find a symmetric form of the last equation. Now we subtract this equation from the Curl-Curl-eigenvalue equation. The resulting equation is given in the original and in its symmetric form. Also the continuous form is given. Solution of the Eigenvalue Equation „Ghostmodes“ in inhomogeneous waveguide Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Because of the vector identity curl curl – grad div = -∇² this equation is also denoted as Nabla- Squared-Eigenvalue equation. Only the dynamic modes are searched for, but not the static fields (with div D≠0) fulfil this equation. Since the solution manifold has to be preserved we will study next which solutions replace the static solutions. One can show that, again, we obtain two types of solution: 1. eigensolutions with ω² >0, div D=0 and curl E≠0: These are the modes searched for. (Here, we also count the modes in multiple conductor systems for which still holds: ω²=0, div D=0, curl E=0) 2. eigensolutions with ω² >0, div D≠0, curl E=0: These modes obey to the equation - grad div E=ω² ω µ E and therefore they are no physically meaningful solutions. The elimination of the static solutions thus leads to the appearance of unphysical solutions, the so- called „Ghost Modes“. They always appear when we solve the Curl-Curl-eigenvalue equation in the latter form. After the solution of the eigenvalue problem we thus need to distinguish the ghost modes from the physical dynamic modes. Solution of the Eigenvalue Equation (C‘ted) „Ghostmodes“ in inhomogeneous waveguide Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Numerical approaches and discretization methods which cannot fulfil the relation div curl=0 on the grid space have no straightforward possibility to distinguish the two solution types. They often apply the so-called penalty method where the grad div- term in the eigenvalue equation is weighted by some factor p (penalty factor). After multiple solution with different values of p the correct (physical) eigensolutions can be distinguished because they do not show any dependence on p. With FIT this effort of multiple solution of the eigenvalue problem is not necessary since the product of the div and curl operator vanishes and thus it is sufficient to apply the dual divergence operator to the electric grid flux to distinguish the two types of solution. Memory Usage Example: mesh points: N P = 10 6 (double precision: 8 bytes/number) FIT: 39 bands of N P entries (21 have to be stored, symmetry) ⇒ 170 MB (+ some add. vectors) no band structure: fill-in-entries by solver (still symmetric!) ⇒ 0 .5 (3 N P ) 2 = 3 .6 ∗ 1013 Byte Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Our eigenvalue problem has the dimension N=3 NP and thus also N eigensolutions. The number NP of grid points may be very large in some applications (up to a few millions) which involves that a complete solution is generally hardly possible. In the following, we will describe an iteration method for the solution of such kind of eigenvalue problems. In the practically interesting case of only a few 10-100 eigenvalues searched for, this method provides good results with acceptable effort. All methods described here exploit the fact that the system matrix is very large but sparse: In the iterative solution of the eigenvalue problem only so-called „matrix-vector-operations“ are used for which the matrix pattern can directly be exploited for an efficient implementation. In general, methods which afford a decomposition of the matrix should not be applied since they destroy the band structure. As example let us assume a computational grid of NP = 106 grid points. According to the previously found structure the Curl-Curl-System matrix has 39 bands of length NP where only 21 of those have to be stored because of symmetry reasons. For a double-precision calculation we need 8 Byte per floating point number which yields a storage requirement of about 170 MB. Additionally, we need to store some auxiliary vectors (depends on the solution method). This is easily affordable on modern PC´s or workstations. Yet, if we would loose our band structure and would need to store the complete matrix (because of fill-ins produced by the solution method) we would end up with 0.5* (3NP)² = 4.5 * 1012 numbers or 3.6 * 1013 Byte (still for a symmetric matrix!) von Mises-Iteration A is real symmetric: Ax = λ x dim ( A ) = N λ1 > λ 2 ≥ λ 3 ≥ …≥ λ N x( ) filled with random numbers, expressed as linear combination of ei : 0 N x = ∑ α i ei ( 0) i =1 Iteration scheme x( v ) = Ax( v −1) leads to: N N λi v x = ∑ α i λi ei = λ1 α1 e1 + ∑ α i ei (v) v v i =1 i = 2 λ1 → 0 for v →∞ Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics As introduction we will study the von Mises method, also called „power iteration“, to determine a specific solution for an eigenvalue problem of the form Ax = λx, dim (A) = N. We assume that the matrix A is real symmetric with eigenvalues | λ1 | > | λ2 | ≥ | λ3 | ≥ ... ≥ | λN |. We search for the (uniquely defined) eigenvalue λ1 with large absolute value and the corresponding eigenvector x1 . We start with an initial vector x(0) filled with random numbers. Since the eigenvectors of A build a complete orthogonal system we can represent x(0) as linear combination of eigenvectors e1, ..., eN as displayed above. Now we use the iteration scheme x(v) = A x (v-1) And get the expression given above. The iteration converges against the eigenvector e1 belonging to the eigenvalue λ1 with larges absolute value. Eigenvalue Solver Rayleigh-quotient: ε v +∞ x( ) ⋅ Ax( ) v v λε ℜ x ( ) (v) = ( v ) ( v ) = λ∞ ∞ + Ο x ⋅x λ∞ λ1 = ℜ x ( v ) ( ) Normalization during iteration: x( v) (v) ( v −1) (v) x = Ax , x = x( v) Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics A very good estimation of this eigenvalue can be obtained by the Rayleigh- quotient given above. In order to avoid a fast increase of the vector components of eigenvalues with | λi | > 1 during the iteration a normalization is carried out after each step of the von Mises- iteration. The vector norm used in the normalization is arbitrary. Iteration Scheme ( 0) 1. Choose randomly filled vector x 2. For ν=1,2,3,… build: x( ) = Ax( v v −1) , x( v) (v) x = x( v) 3. Build Rayleigh quotient λ1 = ℜ x (v ) ( ) 4. Repeat step 2-3 until the desired accuracy is reached Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics The complete iteration is given above. Spectral Shift The von Mises-iteration yields the eigenvector with largest eigenvalue. Often the eigenvector corresponding to the smallest eigenvalue is of special interest. Define A ' with: A ' x = Ax + sx = ( λ + s ) x ( x: Eigenvector of A ) System matrix A' with eigenvectors x , but eigenvalues are shifted: λ'=λ+s Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics With the von Mises-iteration in the previous form only the maximal eigenvalue (absolute value) of a matrix and it´s corresponding eigenvector can be found. Yet, for practical applications just the smallest eigenvalues are of interest (the fundamental mode of a cavity resonator, e.g.). By help of a spectral shift the same method can also be used to compute the smallest eigenvalues of a matrix. To do so the iteration is carried out with the slightly modified matrix A‘ = A + sI which has the same eigenvectors as A but shifted eigenvalues λ‘= λ + s as the above given consideration shows. If the spectrum of A is known in a good approximation the spectral shift s can now be chosen such that the smallest eigenvalue λk searched for gets the eigenvalue λk’ of A’ with largest absolute value and can thus be computed with the von Mises-iteration. However, with this procedure we have to face the problem that numerical cancellation might occur for finite computational precision in the final back-transformation λk = λk’- s. Another problem arises in case that an eigenvalue λi is chosen for s since the transformed matrix A’ will get singular then. Inverse Iteration Iteration of A −1 : 1 Ax = λ x ⇒ A −1 x = x λ In connection with spectral shift, each arbitrary eigenvalue may be computed: A ' = ( A − sI ) −1 Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics If we take the inverse A-1 instead of A for our iteration the algorithm will converge against that eigenvalue of A with smallest absolute value because of Ax = λx ⇒ A-1x = 1/ λ · x . But this advantage has to be paid by the much higher numerical effort of the inverse iteration: Either we have to build A-1 once a priori or we have to solve a linear system in each step. Nevertheless this method is often used since in connection with a spectral shift A’ = (A – S I) -1 each arbitrary eigenvalue of the spectrum may be computed. Sub-Space-Iteration Very large, sparse matrix A, dim( A) = N , λ ≥ 0 real Look for basis vectors (not necessarily eigenvectors): p v ( ) = ∑ cij ei ( j = 1… p ) j i =1 Eliminate all components of unwanted eigenvectors by iteration with A − rk I ( k = 1… n ) Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics The sub-space iteration described in the following simultaneously computes several eigenvectors in each application. The basic idea is to compute several basis vectors during the iteration which together span a subspace of the complete solution space where the searched eigensolution lie in this subspace. Given a very large sparse real matrix A of dimension N x N with eigenvalues λ. Further we assume that all eigenvalues are real and non-negative ( symmetric positive semi-definite matrix.) We are searching for the „first“ p eigensolution, i.e. those eigenvectors belonging to the p smallest eigenvalues ( with p=10; ....100) Sub-Space Iteration The computation of the eigensolution is done in two steps: 1. The p so-called basis vectors spanning the sub-space containing the eigenvectors, searched for, are computed. 2. Using these basis vectors, the given N x N- eigenvalue problem is transformed into a smaller eigenvalue problem of dimension p x p. The spectrum of this smaller problem holds the searched eigenvectors and can be solved by a direct solution method. Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics The computation of the eigensolution is done in two steps: 1. The p so-called basis vectors spanning the sub-space containing the eigenvectors, searched for, are computed. 2. Using these basis vectors, the given N x N- eigenvalue problem is transformed into a smaller eigenvalue problem of dimension p x p. the spectrum of this smaller problem holds the searched eigenvectors and can be solved by a direct solution method. Sub-Space-Iteration N v0 = v ( 0) = ∑αi ei i =1 N v1 = ( A − r1I ) v (0) = ∑ αi ( λi − r1 ) e i i =1 n N n v n = ∏ ( A − rk I ) v ( ) = ∑α i ∏ ( λi − rk ) e i 0 k =1 i =1 k =1 N = ∑ αi Pn ( λi ) e i i =1 n with polynomial of the order n: Pn ( λ ) = ∏ ( λ − rk ) k =1 Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics Thus we need to search for a set of basis vector, i.e. linear independent vectors v(j) which each can be represented as linear combination of all searched eigenvectors ej (i =1…p). (The vectors v(j) themselves need not to be eigenvectors of A.) Then, an arbitrary vector of this sub-space of p basis vectors does not hold any component of the remaining (not interesting) eigenvectors ei with i = p+1, …, N. Therefore, our goal in the computation of the basis vectors is to eliminate all components of these undesired eigenvectors starting with some arbitrary initial vector v(0) . This can be reached by repeated multiplication of the initial vector v(0) with the matrix A - rk I (k = 1, …, n). The spectral shift by rk can vary now in each iteration step. Analogous to the procedure in the von Mises-iteration we receive the representation given above with a polynomial Pn of order n, i.e. the number of iterations. Solution of the Eigenvalue Equation 20 15 10 5 0 Suppressed domain −5 0 1 2 3 4 5 Tschebyscheff polynomial of degree n=20, shifted and scaled Ursula van Rienen, Universität Rostock, Institut für Allgemeine Elektrotechnik, AG Computational Electrodynamics We can interpret the real values ri as zeros of the polynomial Pn. We see that the components of the single eigenvectors ei contained in the initial vector v(0) are suppressed or “accelerated” according to the polynomial value Pn (λi). Now, the so-called accelerating polynomial and it´s zeros rk, respectively, may be chosen such that after sufficiently long iteration the basis vectors only hold components of the searched eigenvectors. Appropriate are e.g. the Tschebyscheff polynomials which are well-known from filtering techniques. Their degree n is fitted to the prescribed accuracy conditions.