You are on page 1of 8

MATLAB PROGRAMS FOR MATRIX EXPONENTIAL

FUNCTION DERIVATIVE EVALUATION
Lubomír Brančík
Institute of Radio Electronics, Faculty of Electrical Engineering and Communication
Brno University of Technology

Abstract
The paper deals with six approaches how to determine a derivative of the matrix
exponential function in the Matlab language environment. Namely, a Taylor series
expansion, an augmented matrix utilization, an eigenvalues decomposition, a Laplace
transform approach, a convolution integral evaluation and a Padé approximation
method are discussed in the paper. Some of the above methods are connected with a
scaling and squaring process to improve the stability. Besides, possible Matlab listings
are shown at each method as well.

1 Introduction & Motivation Comment
Six possible methods how to compute a first derivative of the matrix exponential function are
discussed while using the Matlab language environment. A usefulness for such the computation
appears in many branches of the electrical engineering simulation. For example, in the field of the
circuit theory, it arises in the process of a sensitivity determination at the electrical circuits containing
multiconductor transmission lines (MTL) [1, 2]. The properties of an (n+1)-conductor TL in the
s–domain can be expressed using a chain matrix
Φ ( x, s ) = e M ( s ) x , (1)
where x denotes a geometric coordinate along the MTL and M(s) is the 2n-order square matrix
⎡ 0 −Z0 ( s) ⎤
M(s) = ⎢ , (2)

⎣ 0 Y ( s ) 0 ⎥⎦
with
Z0 ( s ) = R 0 + sL 0 , Y0 ( s ) = G 0 + sC0 (3)
as the n-order series impedance and shunting admittance matrices, respectively, R0, L0, G0 and C0 as
MTL per-unit-length matrices, and 0 as a zero matrix. The sensitivities can be got resulting from the
first derivative of (1), with respect to a parameter γ∈M(s) [1]. Thus, γ is the element of one of the per-
unit-length matrices stated above, while x is considered as a constant in this case. As will be shown
later, for all the techniques the first derivative of the matrix at the exponent is required. In case of our
application example, depending on γ and considering (2) and (3), it is given by Table 1 [1].

TABLE 1. DERIVATIVES ∂M(S)/∂γ NEEDED FOR DETERMINING ∂Φ(X,S)/∂γ
γ ∈ R0 γ ∈ L0 γ ∈G 0 γ ∈C0
⎡ ∂R 0 ⎤ ⎡ ∂L 0 ⎤ ⎡ 0 0⎤ ⎡ 0 0⎤
⎢ 0 − ∂γ ⎥ ⎢ 0 − s ∂γ ⎥ ⎢ ∂G ⎥ ⎢ ∂C ⎥
⎢ ⎥ ⎢ ⎥ ⎢− 0
0⎥ ⎢−s 0 0⎥
⎢⎣ 0 0 ⎥⎦ ⎢⎣ 0 0 ⎥⎦ ⎢⎣ ∂γ ⎥⎦ ⎢⎣ ∂γ ⎥⎦

As is obvious finding the above derivatives is already straightforward. Some examples of sensitivities
computation at the MTL systems can be found e.g. in [1-3].
To generalize our approach for any type of matrix, however, let us denote the matrix in view as
M ≡ M(γ), when its dependence on γ is evident. The matrix exponential function is then formulated as
Φ ( γ, x) = eM ( γ ) x , (4)
with x standing for any real parameter, and ∂Φ(γ,x)/∂γ is what we want to determine.

. (5) k =0 k ! Then. end %********************************************************************************************* Calling the function as [F. the infinite sum (6) will only be evaluated for some finite number of terms. n=1. (6) ∂γ k =1 k ! ∂γ where a derivative of the k–th power of the matrix is stated as follows. l=1. F=FM+M. r=max(0. (9) successively for m = M. leads to the evaluation of both matrix exponential function (stored in F) and its derivative (stored in dF). In practice. a recursive formula to restore ∂Φ(γ. a restoring Φ(γ. l=l*n. A procedure is based on so-called scaling and squaring [4].2 Taylor Series Expansion with Scaling & Squaring As is well known. x) ∞ x k ∂M k ( γ ) =∑ . First.e]=log2(norm(M. x) = Φ r ( γ.x)/∂γ from ∂Φ(γ. x 2m ) ∂Φ ( γ . we can write ∂Φ ( γ.dF]=dexpmt(M. with M and its derivative dM as arguments. incF=sum(sum(abs(FM)+abs(dFM))).…. x 2 m ) = Φ ( γ .x/r) can fast be performed by a squaring process Φ ( γ. end % Undo scaling by repeated squaring for k=1:r dF=dF*F+F*dF. 2008 % Scale M by power of 2 so that its norm is < 1/2 [f. F=F+FM*M/l. The convergence sufficiently rapid is ensured if the product M(γ)x is close enough to a null matrix. considering a formula based on (4) as Φ ( γ. an identity Mk(γ) = M(γ)Mk-1(γ) the first derivative is written as ∂M k ( γ ) ∂M ( γ ) k -1 ∂M k -1 (γ ) = M (γ )+M ( γ ) . successively for m = M. dM=dM/2^r. This recursive formula contains (9) as its inner part.x) from Φ(γ. x r ) . FM=FM*M. the matrix exponential function (4) can be expanded into the Taylor series as ∞ xk k Φ( γ.dF]=dexpmt(M. M-1.20) (the values are recommended in [4] for 3 × 3 matrices). Then. This temporal result is transformed back to get the original derivative.g.…. while incF>1e-16 n=n+1.1. x) = ∑ M (γ) . %********************************************************************************************* function [F. x 2m ) . x 2 m −1 ) ∂Φ ( γ . x 2m −1 ) = Φ 2 ( γ. M=M/2^r. the derivative is found for a transformed matrix exponential function Φ(γ. incF=1. dFM=dM*FM+M*dFM. for such the r in order to meet the Euclidian norm ||M(γ)x/r)|| < (0.dM). (7) ∂γ ∂γ ∂γ that can be processed recursively. Therefore.1. dF=dF+dFM/l. A possible Matlab function dexpmt listing is shown below. (8) then if r = 2M.dM) % by Lubomír Brančík.x/r).x/r)/∂γ follows (9) as ∂Φ ( γ .'inf')). (10) ∂γ ∂γ ∂γ processing it M-times. dFM=dM.15 to 0. % Taylor series expansion of exp(M) and diff[exp(M)] FM=eye(size(M)). F=F*F. x 2m ) + Φ ( γ . dF=dM. M integer. M-1. starting with k = 2.e+1). x 2m ) . If we consider e. During evaluating (6) the successive increments are used to obtain temporal errors enabling to stop the summation process.

x) ⎥ = exp ⎜ ⎢ ∂M ( γ ) ⎥ x ⎟ ⋅ ⎢ ∂W( γ . ∂Φ(γ. When putting together (12) with its first derivative with respect to γ of the form ∂W ( γ . A procedure is based on a concept in [5]. with associated . Besides.M].x). calling the function corresponds to the preceding method. x) + M ( γ ) . as a transient matrix. can be written as ⎡ W ( γ. x) = M ( γ ) W( γ . x) . x) ⎜ ⎟= W ( γ. x) ∂Φ ( γ .0) ⎤ ⎡ W( γ . the expm built-in function based on Padé approximation and scaling and squaring can be utilized with advantages to determine the transient matrix Ψ(γ.j = 1. x) ⎤ ⋅ ⎢ ∂W ( γ . x ) . 2008 N=length(M).0) ⎥ .x). A possible Matlab function dexpma listing is shown below. (17) ⎢ ⎥ ⎜⎢ M ( γ ) ⎥ ⎟⎟ ⎢ ⎥ ⎢ Ψ ( γ . x) W ( γ .x) = [Ψij(γ. combining (11) with its first derivative with respect to γ of the form d ⎛ ∂W ( γ . % Extracting F and dF from PSI F=PSI(1:N. (11) dx namely W ( γ . dF=PSI(N+1:2*N.x)].0) ⎤ ⎢ ∂W ( γ . x) Ψ12 ( γ. (13) ∂γ ∂γ ∂γ then only one. x) 0 ⎤ ⎡ W ( γ .0) ⎥ .0) . Let us assume M(γ) having distinct eigenvalues qj(γ). x) = Φ ( γ . (16) dx ⎢ M(γ)⎥ ⎢ ⎥ ⎣⎢ ∂γ ⎦⎥ ⎣⎢ ∂γ ⎦⎥ ⎣⎢ ∂γ ⎦⎥ The solution of the last differential equation can be written as ⎡ W( γ.0) + Φ ( γ .zeros(N). x) ⎤ ⎛ ⎡ M(γ) 0 ⎤ ⎞ ⎡ W ( γ . Now denoting Ψ(γ. by using a Matlab language.0) = W ( γ .0) as an initial condition.0) ⎤ ⎢ ∂W ( γ . % Creating augmented matrix AM=[M. x) ⎥⎥ = ⎢⎢ ∂M ( γ ) ⎥ ⋅ ⎢ ∂W ( γ . in the form ⎡ W( γ.dM. x) ⎥⎦ ⎢ ⎥ ⎢⎣ ∂γ ⎥⎦ ⎜ ⎢ ∂γ ⎝⎣ ⎥⎦ ⎠ ⎢⎣ ∂γ ⎥⎦ ⎣ 21 ⎢⎣ ∂γ ⎥⎦ with the same initial conditions as are in (14). (15) dx ⎝ ∂γ ⎠ ∂γ ∂γ we get only one. x) Ψ22 ( γ . and comparing it with the system matrix in (14). an augmented matrix equation.2. (14) ⎢ ⎥ ⎢ Φ( γ. (12) with W(γ.dM) % by Lubomír Brančík. see the right side of (17).1:N). x) ⎥ . %********************************************************************************************* 4 Eigenvalues Decomposition An another Matlab function can be utilized to get a matrix exponential function derivative. x) ⎥ = ⎢ ∂Φ ( γ .x)/∂γ is equal to the homothetic submatrix Ψ21(γ.0) ⎥ = ⎡ Ψ11 ( γ. % Matrix exponential function of augmented matrix – usage of Matlab expm PSI=expm(AM).1:N). %********************************************************************************************* function [F. namely the eig built-in function for the matrix eigenvalues and eigenvectors calculation. x) ⎤ ⎡ M ( γ ) 0 ⎤ ⎡ W ( γ. The derivative ∂M(γ)/∂γ can be stated easily. A further technique is based on the concept in [4]. x) ⎞ ∂M ( γ ) ∂W ( γ. x) ⎤ ⎡ Φ ( γ. x) ⎥ ⋅ ⎢ ∂W ( γ .dF]=dexpma(M. x) ∂W ( γ . x) ⎤ d ⎢ ∂W ( γ. i. x) ⎥ ⎢ ⎥ ⎣⎢ ∂γ ⎦⎥ ⎣⎢ ∂γ ⎦⎥ ⎣⎢ ∂γ ⎦⎥ Similarly.3 Augmented Matrix Utilization On principle the matrix exponential function (4) figures in the solution of the first-order matrix ordinary differential equation dW( γ. an augmented matrix differential equation.

Based on (23) and (24). (20) ∂γ In the last equation. j = 1. x) ⎧ −1 ∂M ( γ ) −1 ⎫ = L−x1 ⎨( qI -M ( γ ) ) ( qI -M ( γ ) ) ⎬ . i ≠ j . %********************************************************************************************* function [F.q2(γ). P(γ. it happens rarely in practice [4.j)).usage of Matlab eig [U.i)*RQ. H=U\dM*U.. The calling Matlab eig function with the argument M(γ) will return both Q(γ) and U(γ) matrices. A possible Matlab function dexpme listing is shown below. e q2 ( γ ) x .dF]=dexpme(M.. a similarity transform leads to M ( γ ) = U ( γ )Q( γ ) U -1 ( γ ) . 2008 % Eigenvalues/eigenvectors decomposition .*eQr). P=diag(diag(H). eQr=exp(diag(Q)).j)-th elements of the matrix ∂M (γ ) H ( γ ) = U −1 ( γ ) U( γ ) . then differentiated and finally inverted into the x–domain again. %********************************************************************************************* 5 Laplace Transform Approach This technique is based on the fact that the derivative ∂Φ(γ. In case of repeated eigenvalues a more complex method based on a Jordan canonical decomposition should be applied instead.. A commutativity property of the integration and derivative operations leads to the equality ∂Φ ( γ. x)}⎬ . its calling corresponds to the first method.. (18) Then. and pii ( γ .qN(γ)) and a modal matrix U(γ).j)=H(i. (21) qi ( γ ) − q j ( γ ) where hij(γ) are the (i.Q]=eig(M). (19) and its first derivative as [5] ∂Φ ( γ .x) is a matrix built-up from the elements q (γ) x e qi ( γ ) x − e j pij ( γ . the matrix exponential function (4) can be expressed as ( ) Φ ( γ . x)U -1 ( γ ) . x) = U( γ )eQ ( γ ) x U -1 ( γ ) = U ( γ )diag e q1 ( γ ) x .. however. P(i..N.2.x)/∂γ is easier to find after a Laplace transform of the matrix exponential function with respect to x is performed. P(j. It means its q–domain image is first found. dF=U*P/U.j)*RQ. for i=1:length(H)-1 for j=i+1:length(H) RQ=(eQr(i)-eQr(j))/(Q(i. e qN ( γ ) x U -1 ( γ ) . and doing some arrangements. (22) ∂γ The derivative ∂M(γ)/∂γ can be stated easily. x) = U ( γ )P ( γ .dM) % by Lubomír Brančík. end end F=U*diag(eQr)/U. i = j . x)} = ∫ eM ( γ ) x e− qx dx = ∫ e−( qI −M ( γ ) ) x dx = ( qI -M ( γ ) ) −1 .K. x) = hii ( γ ) xe qi ( γ ) x .i)-Q(j.i)=H(j. (25) ∂γ ⎩ ∂γ ⎭ . x) = hij ( γ ) . 5]. the result is ∂Φ ( γ .. x) ⎧∂ ⎫ = L−x1 ⎨ L x {Φ ( γ . Now creating a diagonal matrix Q(γ) = diag(q1(γ). (23) ∂γ ⎩ ∂γ ⎭ where the Laplace transform can analytically be stated as ∞ ∞ L x {Φ( γ.. (24) 0 0 with I as an identity matrix. composed of uj(γ) as its columns.eigenvectors uj(γ).

5)*pi. [6. an argument of the inverse Laplace transform in (25) has the form of a product of matrices. however. an efficient technique is developed whose accuracy depends on evaluation of respective matrix exponential function Φ(γ. Taking beside (24) into account the result is (see also [5]) x ∂Φ ( γ. N. It is possible to use various types of NILT techniques.j = 1. x) N N ∂mij ( γ ) x = ∑∑ ∂γ ∫0 Φ ( γ. i. 7]. The derivative ∂M(γ)/∂γ can be stated easily. % number of terms in a basic sum ndif=10. M=M/2^r. u ) du .g.5. for n=1:nsum qr=a+j*n*pi. %********************************************************************************************* function [F. % error boundary proportional to exp(-4*a) nd=1:ndif. end %********************************************************************************************* 6 Convolution Integral Evaluation This method is a consequence of the Laplace transform approach in the last section. In such a case the inverse Laplace transform is solved for x = 1. however. end F=exp(a)/2*(F+Fe/2^ndif). F=F*F. Here./factorial(nd-1))). dM=dM/2^r. In [5]. dF=2*F*dM*F. % number of terms for Euler transform a=8. therefore it can formally be transformed into the original domain through a convolution integral. dFe=zeros(N). for n=nsum+1:nsum+ndif qr=a+j*n*pi. (27) ∂γ i =1 j =1 ./factorial(ndif+1-nd). qi=a+j*(n-0. To guarantee the sufficiently fast convergence the scaling and squaring are applied in according to (8) – (10). x − u )Bij Φ ( γ . 2008 N=length(M). its calling follows the first method. u )du . F=F+kn(n)*(real(LFr)+imag(LFi)).e]=log2(norm(M. x) ∂M (γ) = ∫ Φ( γ. The matrix exponent part is scaled to obtain its Euklidian norm ||M(γ)x/r)|| < 0.dM) % by Lubomír Brančík. a NILT based on approximation of exp(qx) in the ILT Bromwich integral by a specially created infinite series. A possible Matlab function dexpml listing is shown below. The derivative ∂M(γ)/∂γ can be stated easily. LFr=inv(qr*eye(N)-M).'inf')).…. see e. (26) ∂γ 0 ∂γ where u stands for an integration variable. Namely. this function is applicable for real matrices M only. the last equation can be expressed as ∂Φ ( γ.e+1). end Fe=zeros(N). kn=(-1). Fe=Fe+kn(n)*(real(LFr)+imag(LFi))*eul(n-nsum). dFe=dFe+kn(n)*(real(LFr*dM*LFr)+imag(LFi*dM*LFi))*eul(n-nsum). qi=a+j*(n-0. dF=exp(a)/2*(dF+dFe/2^ndif). % Scale M by power of 2 so that its norm is < 1/2 [f.5)*pi. r=max(0. The inverse Laplace transform has to be done numerically in general. % Setting up parameters of the NILT method nsum=200. is applied [7]. LFi=inv(qi*eye(N)-M).dF]=dexpml(M. Namely. Due to the NILT used. % Undo scaling by repeated squaring for k=1:r dF=dF*F+F*dF. LFr=inv(qr*eye(N)-M). F=inv(a*eye(N)-M)/2. The above integral could then be evaluated numerically. denoting M(γ) = [mij(γ)]. eul=fliplr(cumsum(factorial(ndif). x − u ) Φ ( γ . in conjunction with Euler transform.x) and on inversion of a specially built-up matrix. dF=dF+kn(n)*(real(LFr*dM*LFr)+imag(LFi*dM*LFi)). LFi=inv(qi*eye(N)-M).^(1:nsum+ndif). with an accuracy dependent on the method applied.2.

:))).:)])). B(i.j)=1.DC(r. end end dF=IS\SS. end SS=SS+dM(i. or as the sums of certain subdeterminants of M(γ). ⎡⎛ N ∂R N ( γ ) ⎞ N −1 ⎛ N ⎞ ⎤ ⎢⎜⎜ ∑ (−1) N − k N − k +1 M k −1 ( γ ) ⎟⎟ xΦ( γ.K-1).x) .Φ(γ.j)*(Sk*F-Su). C=B*F-F*B.DC(r.x)Bij is a commutator. else DC=nchoosek([1:min(i.:). end SS=zeros(N).dF]=dexpmc(M.:)].j)+1:N]. either by expm directly or by eig through formula (19).i+1:N]. see [5] for details. for u=1:N-1 Sk=zeros(N). while 0 elsewhere. end end end Sk=Sk+(-1)^(N-k)*dR*M^(k-1). Denoting the integral in (27) as Sij(γ. IS=zeros(N).j)-th element 1. end end else if K==1 dR=0. and RkN(γ) coefficients and their derivatives w. for k=1:N IS=IS+k*Rs(N-k+1)*M^(k-1). The Φ(γ. x)M u −1 ( γ ) ⎥ ⎣⎢⎝ k =1 ∂mij ( γ ) ⎠ u =1 ⎝ k = u +1 ⎠ ⎦⎥ where Cij(γ.1) dR=dR+det(M(DC(r.i). This condition is fulfilled if the eigenvalues of M(γ) are distinct [5].dM) % by Lubomír Brančík. if i==j if K==1 dR=1. elseif K==2 dR=-M(j. it is valid −1 ⎡N ⎤ Sij ( γ. dR=0.x). for k=u+1:N Sk=Sk+(k-u)*Rs(N-k+1)*M^(k-u-1).x) can be evaluated by Matlab built-in functions.DC(r.max(i. Su=zeros(N). end Su=Su+Sk*C*M^(u-1).1) dR=dR-det(M([j. dR=0. A possible Matlab function dexpmc listing is shown below.x) = BijΦ(γ.j)+1:max(i. The inverse matrix in (28) must exist to be able to apply the method. %********************************************************************************************* function [F. N=length(M).j)-1. %********************************************************************************************* .with Bij as a square matrix with the (i. else DC=nchoosek([1:i-1.r. for i=1:N for j=1:N B=zeros(N). its calling corresponds to the first method. for k=1:N K=N-k+1. for r=1:size(DC. end Sk=zeros(N). Rs=poly(M).[i. 2008 F=expm(M).K-2). x) = ⎢∑ (−1) N − k kRNN− k ( γ )M k −1 ( γ ) ⎥ × (28) ⎣ k =1 ⎦ .j)-1. x) − ∑ ⎜ ∑ (−1) N − k (k − u ) RNN− k ( γ )M k −u −1 ( γ ) ⎟ Cij ( γ.min(i. to mij(γ) can be found via processing the characteristic polynomial of M(γ) (Matlab function poly can be utilized). for r=1:size(DC.

2008 % Scale M by power of 2 so that its norm is < 1/2 [f. D=eye(size(M))-c*M. %********************************************************************************************* function [F. x) + D−pq1 ( γ . q=6. cX=c*X. r=max(0.5. x) ⎟ . it is convenient to choose p = q in the sums (30) and (33) which leads to a good numerical stability and a simplification. Here the matrix exponent part is scaled to get its Euklidian norm ||M(γ)x/r)|| < 0. % Pade approximation of exp(M) and diff[exp(M)] X=M. dF=dF+cY. x) =& N pq ( γ . x) . Besides. dD=dD-cY. p = q = 6 is usually sufficient choice. F=F*F.7 Padé Approximation Method with Scaling & Squaring As was pointed out earlier the Matlab built-in function expm uses a Padé approximation method in conjunction with scaling and squaring techniques [8]. (32) ∂γ ⎝ ∂ γ ∂ γ ⎠ where ∂N pq ( γ. % Undo scaling by repeated squaring for k=1:r dF=dF*F+F*dF. D pq ( γ. A similar method can be applied to dedermine the matrix exponential function derivative. x) ⎛ ∂N pq ( γ . if p D=D+cX. Y=dM. x) ∂D pq ( γ . (31) ∂γ ∂γ ∂γ The substitution for ∂Dpq-1(γ. x) = ∑ ( -M( γ ) x ) k k (30) k = 0 ( p + q )!k !( p − k )! k = 0 ( p + q )!k !( q − k )! are polynomials of the matrix argument. The generalized Matlab function dexpm has been developed based on the above theory. p=1. for k=2:q c=c*(q-k+1)/(k*(2*q-k+1)).e]=log2(norm(M. end p=~p. x) p ( p + q − k )! p ! ∂M k ( γ ) k ∂D pq ( γ. (29) where p ( p + q − k )! p ! q ( p + q − k )!q ! N pq ( γ . end %********************************************************************************************* . =∑ (− x)k (33) ∂γ k = 0 ( p + q )!k !( p − k )! ∂γ ∂γ k = 0 ( p + q )!k !( q − k )! ∂γ result directly from (30) and ∂Mk(γ)/∂γ can be stated by the recursive formula (7). x) =& D−pq1 ( γ. Now we can follow (29) to get the derivative w. Calling the function corresponds to the first method. X=M*X. dF=c*dM.dM) % by Lubomír Brančík.dF]=dexpm(M. In practice. M=M/2^r. F=F+cX. to γ as −1 ∂Φ ( γ . the Matlab expm built-in function uses. x ) ⎜ − Φ ( γ. F=eye(size(M))+c*M.r. x) = ∑ ( M(γ) x ) . dD=-c*dM. c=1/2. as e. dF=D\(dF-dD*F). end F=D\F. dD=dD+cY.g. x) ∂N pq ( γ . dM=dM/2^r.'inf')). x) . x) ⎞ =& D−pq1 ( γ . x) q ( p + q − k )!q ! ∂M k ( γ ) =∑ x . Y=dM*X+M*Y. cY=c*Y. To guarantee the sufficiently fast convergence the scaling and squaring are applied in according to (8) – (10). else D=D-cX. Let us consider the Padé approximation of the matrix exponential function (4) as Φ ( γ.x)/∂γ and further arrangements lead to a formula ∂Φ ( γ . x)N pq ( γ.e+1). x) ∂D pq ( γ .

Voltage/current waves sensitivity in hybrid circuits with nonuniform MTLs. Przeglad Elektrotechniczny . CSc. 2. The matrices M can be complex in general. while taking respective matrix (M) and its first derivative (dM) as the arguments. Brančík. S. the techniques are stated more precisely. All the methods give both matrix exponential function (F) and its first derivative (dF) as the results.g. [10] L. Brančík.cz. Moreover. 2001. vol. 1998. p. The only exception was the Matlab listing in the Laplace transform approach. while describing them in more general notation. C. 177-180. . Techniques of matrix exponential function derivative for electrical engineering simulations. C. (45) 2003. Procedures for matrix exponential function derivative in Matlab. Phone: +420541149128. Fung. no. Fax: +420541149244. K. Moler . Univ. 7 . Technical report #311. FEEC BUT in Brno. [9] L. An error analysis based on the second-order chain matrices (2). Mumbai (India): Indian Institute of Technology Bombay. 2006. vol. Brančík. Tsai. WSEAS Transactions on Communications. 11. with applications to continuous-time modeling.8 Concluding Comments The paper has followed rather theoretical work [9] where the above stated methods have been collectively summarized for purposes of MTL systems simulation. Here. Nineteen dubious ways to compute the exponential of a matrix. p. Ing. [8] C. in Proceedings of IMAPS CS International Conference EDS'06. A note on parameter differentiation of matrix exponentials. International Journal for Numerical Methods in Engineering. [3] L. can be found in [9]. Brančík. Chain matrix derivative via Laplace transform method with applications to distributed circuits simulation. Computation of the matrix exponential and its derivatives by scaling and squaring. Lubomír Brančík. Acknowledgments This work was supported by the Czech Ministry of Education under the MSM 0021630503 Research Project MIKROSYN. 612 00 Brno. 1. 153 . in [10].2613. 11. Approximate formulae for numerical inversion of Laplace transforms. no. twenty- five years later. Brančík. which is just the case M ≡ M(s) needed in the MTL systems simulation. 53 .Konferencje. Brančík. 2007. [6] L. [4] T. Berlin (Germany). 59 (2004). no.166. p. Loan. [2] L. SIAM Review. 58 – 63. Table 2 shows the mean values of relative errors obtained. 83. 1–22. __________________________________________________________________________________ Doc. 2007. MEAN VALUES OF RELATIVE ERRORS Taylor series expansion Augmented matrix utilization Eigenvalues decomposition ≈ 10−13 ≈ 10−13 ≈ 10−15 Laplace transform approach Convolution integral evaluation Padé approximation method ≈ 10−12 ≈ 10−15 ≈ 10−13 References [1] L. but it can also be modified to enable it. International Journal of Numerical Modelling: Electronic Networks. 4 (2005). in Proceedings of the 10th IEEE Workshop on SPI’06. L. Modified technique of FFT-based numerical inversion of Laplace transforms with applications. of Iowa.10. 1–46. in Proceedings of IEEE International Conference on Industrial Technology. No. V. E-mail: brancik@feec. Brančík. Novel techniques for sensitivity evaluation in multiconductor transmission line systems. with known analytical solution expressed through the hyperbolic functions. 1273–1286. 3. possible Matlab listings of all these techniques programmed in the form of M-file functions are presented. 5. Brno (Czech Republic).56. Valsa. 2006. Przeglad Elektrotechniczny. 216–223. Czech Republic. One can see another generally usable LT approach e. Department of Radio Electronics. 2006. vol. Devices and Fields.vutbr. [7] J. TABLE 2. p. 2608 . No. Chan. Purkyňova 118. 5. [5] H.