General Linear Methods for Integrated Circuit Design
Steffen Voigtmann
Oberwolfach, April 2006
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Integrated circuit design
design constraints
design specifications
schematic generation
simulation and characterisation
specification met? release to
manufacturing
General linear methods for integrated circuit design – St. Voigtmann, p. 2
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Classical methods
I BDF
. artificial damping
-1 0 1
20 time 10 0
I Trapezoidal rule
. undesired oscillations
−2
−1 0 1 2
2π time 3π π 2 π 0 2
I Runge-Kutta methods . high computational costs
0 30 60 90
problem size 30
20 10 0
computingtime
Runge Kutta BDF
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Classical methods
I BDF
. artificial damping
-1 0 1
20 time 10 0
I Trapezoidal rule
. undesired oscillations
−2
−1 0 1 2
2π time 3π π 2 π 0 2
I Runge-Kutta methods . high computational costs
0 30 60 90
problem size 30
20 10 0
computingtime
Runge Kutta BDF
General linear methods for integrated circuit design – St. Voigtmann, p. 2
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Classical methods
I BDF
. artificial damping
-1 0 1
20 time 10 0
I Trapezoidal rule
. undesired oscillations
−2
−1 0 1 2
2π time 3π π 2 π 0 2
I Runge-Kutta methods . high computational costs
0 30 60 90
problem size 30
20 10 0
computingtime
Runge Kutta BDF
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Classification of methods
I Linear multistep methods . low costs
. very successful (BDF) . not A-stable forp>2 I Runge-Kutta methods
. very good stability properties . stepsize change is easy . high costs
I General linear methods (GLM) . combine advantages
of both classes
. make new methods possible . provide unifying framework
for known methods
General linear methods for integrated circuit design – St. Voigtmann, p. 4
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Contents
Differential Algebraic Equations
General Linear Methods
Practical General Linear Methods
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Index2 Index1 Index0
ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs circuit simulation
prop. stated index-2DAEs
General linear methods for integrated circuit design – St. Voigtmann, p. 5
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Mx0=f(x,·)
Index2 Index1
Index0 ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs circuit simulation
prop. stated index-2DAEs
y0 =f(y,t) ordinary differential equations
I well understood (theoretically, numerically)
I Butcher, Dahlquist, Gear, Hairer, Petzold, . . .
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Mx0=f(x,·)
Index2 Index1
Index0 ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs circuit simulation
prop. stated index-2DAEs
E x0+Fx=q A[Dx]0+Bx=q
linear DAEs
I standard form (Hairer/Wanner, Kunkel/Mehrmann)
I prop. stated (M ¨arz, Balla, Kurina)
General linear methods for integrated circuit design – St. Voigtmann, p. 5
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Mx0=f(x,·)
Index2 Index1
Index0 ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs circuit simulation
prop. stated index-2DAEs
y0=f(y,z) 0=g(z)
DAEs in Hessenberg form
I Runge-Kutta (Hairer/Wanner, Kværnø)
I lin. multistep (Campbell, Gear, Petzold)
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Mx0=f(x,·)
Index2 Index1
Index0 ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs
circuit simulation
prop. stated index-2DAEs
A d0(x,·)+b(x,·) =0 nonlinear index-1 DAEs (prop. stated)
I extension of decoupling procedure
I M ¨arz, Higueras
General linear methods for integrated circuit design – St. Voigtmann, p. 5
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Mx0=f(x,·)
Index2 Index1
Index0 ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs circuit simulation
prop. stated index-2DAEs
A d0(x,·)+b(Ux,·) +BTx=0
DAEs appearing in electrical circuit simulation
I index-2, no Hessenberg form
I Tischendorf, Est ´evez Schwarz (initialisation)
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs of increasing complexity
A[Dx]0+Bx=q A[Dx]0+b(x,·) =0 A d0(x,·)+b(x,·) =0
Mx0=f(x,·)
Index2 Index1
Index0 ODEs
lin. DAEs
Hessenberg
nonlin. index-1DAEs circuit simulation
prop. stated index-2DAEs A d0(x,·)+b(x,·) =0 nonlinear DAEs with properly stated leading terms
I existence and uniqueness of solutions
I convergence results for numerical methods
General linear methods for integrated circuit design – St. Voigtmann, p. 6
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs in electrical circuit simulation
I Modified Nodal Analysis Aq(x,˙ ·)+b(x,·) =0
I analysis: tractability index . low smoothness requirements . use projectors (Pi,Qi,U,T, . . . )
and matrix sequences
I index can be determined topologically . look for CV loops and LI cutsets I index-2componentsTxare given by
. currents ofV-sources inCVloops . voltages of inductors andI-sources
inLIcutsets
I index-2componentsTxenter linearly Aq(x,˙ ·)+b(Ux,·) +BTx=0
V1
V2
VBB
VDD u4
u12 u1
u3
u2
u7
u6 u5
u11
u10
u9 u8
C
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs in electrical circuit simulation
I Modified Nodal Analysis Aq(x,˙ ·)+b(x,·) =0
I analysis: tractability index . low smoothness requirements . use projectors (Pi,Qi,U,T, . . . )
and matrix sequences
I index can be determined topologically . look for CV loops and LI cutsets I index-2componentsTxare given by
. currents ofV-sources inCVloops . voltages of inductors andI-sources
inLIcutsets
I index-2componentsTxenter linearly Aq(x,˙ ·)+b(Ux,·) +BTx=0
V1
V2
VBB
VDD u4
u12 u1
u3
u2
u7
u6 u5
u11
u10
u9 u8
C
General linear methods for integrated circuit design – St. Voigtmann, p. 6
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs in electrical circuit simulation
Est ´evez Schwarz (2000) Cons. initialization for index-2 DAEs and it’s application to circuit simulation, PhD thesis
I Modified Nodal Analysis Aq(x,˙ ·)+b(x,·) =0
I analysis: tractability index . low smoothness requirements . use projectors (Pi,Qi,U,T, . . . )
and matrix sequences
I index can be determined topologically . look for CV loops and LI cutsets I index-2componentsTxare given by
. currents ofV-sources inCVloops . voltages of inductors andI-sources
inLIcutsets
I index-2componentsTxenter linearly Aq(x,˙ ·)+b(Ux,·) +BTx=0
V1
V2
VBB
VDD u4
u12 u1
u3
u2
u7
u6 u5
u11
u10
u9 u8
C
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs in electrical circuit simulation (cont.)
A[Dx]0+b(Ux,·) +BTx=0 ➪ index-2 components enter linearly
I Idea: Introduce new variables
u=DP1x, w=P1D−(Dx)0+ (Q0+Q1)x.
For a solutionxthis implies
x=D−u+ (P0Q1+Q0P1)w+Q0Q1D−(Dx)0.
I Consequences:
Ux=D−u+ (P0Q1+UQ0)w, A[Dx]0+BTx= (AD+BT)w
➪ F(u,w,·) =A[Dx]0+b(Ux,·) +BTx=0.
General linear methods for integrated circuit design – St. Voigtmann, p. 7
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs in electrical circuit simulation (cont.)
A[Dx]0+b(Ux,·) +BTx=0 ➪ index-2 components enter linearly
I Idea: Introduce new variables
u=DP1x, w=P1D−(Dx)0+ (Q0+Q1)x.
For a solutionxthis implies
x=D−u+ (P0Q1+Q0P1)w+Q0Q1D−(Dx)0.
I Consequences:
Ux=D−u+ (P0Q1+UQ0)w, A[Dx]0+BTx= (AD+BT)w
➪ F(u,w,·) =A[Dx]0+b(Ux,·) +BTx=0.
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
DAEs in electrical circuit simulation (cont.)
A[Dx]0+b(Ux,·) +BTx=0 ➪ index-2 components enter linearly
I Idea: Introduce new variables
u=DP1x, w=P1D−(Dx)0+ (Q0+Q1)x.
For a solutionxthis implies
x=D−u+ (P0Q1+Q0P1)w+Q0Q1D−(Dx)0.
I Consequences:
Ux=D−u+ (P0Q1+UQ0)w, A[Dx]0+BTx= (AD+BT)w
➪ F(u,w,·) =A[Dx]0+b(Ux,·) +BTx=0.
General linear methods for integrated circuit design – St. Voigtmann, p. 8
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Local existence and uniqueness of solutions
V. (2004) General linear methods for nonlinear DAEs in circuit simulation, SCEE
Theorem.
I The properly stated index-2DAE
F(u,w,·) =A[Dx]0+b(Ux,·) +BTx=0, F(u0,w0,t0) =0,
is locally equivalent to w(u0,t0) =w0, F u,w(u,t),t
=0.
I For everyx0∈IRm, the initial value problem
A[Dx]0+b(Ux,·) +BTx=0, DP1x(t0) =DP1x0.
is uniquely solvable. The solutionx=D−u+z0+z1satisfies u0 =f u,w(u,t),t
, z1=P0Q1w(u,t), z0 =g u,(Dz1)0,t .
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Local existence and uniqueness of solutions
Theorem.
I The properly stated index-2DAE
F(u,w,·) =A[Dx]0+b(Ux,·) +BTx=0, F(u0,w0,t0) =0,
is locally equivalent to w(u0,t0) =w0, F u,w(u,t),t
=0.
I For everyx0∈IRm, the initial value problem
A[Dx]0+b(Ux,·) +BTx=0, DP1x(t0) =DP1x0.
is uniquely solvable. The solutionx=D−u+z0+z1satisfies u0 =f u,w(u,t),t
, z1=P0Q1w(u,t), z0 =g u,(Dz1)0,t .
General linear methods for integrated circuit design – St. Voigtmann, p. 8
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Local existence and uniqueness of solutions
V. (2004) General linear methods for nonlinear DAEs in circuit simulation, SCEE
Theorem.
I The properly stated index-2DAE
F(u,w,·) =A[Dx]0+b(Ux,·) +BTx=0, F(u0,w0,t0) =0,
is locally equivalent to w(u0,t0) =w0, F u,w(u,t),t
=0.
I For everyx0∈IRm, the initial value problem
A[Dx]0+b(Ux,·) +BTx=0, DP1x(t0) =DP1x0.
is uniquely solvable. The solutionx=D−u+z0+z1satisfies u0 =f u,w(u,t),t
, z1=P0Q1w(u,t), z0 =g u,(Dz1)0,t
➩ .
inherent ordinary differential equation
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Properly stated index-2 DAEs
A q(x, ˙ ·)+ b(x, ·) = 0
I split solution into characteristic parts x=D−u+z+w
x∗
P0x∗ z0∗=Q0x∗
D−DP¯1x∗ z1∗=P0Q¯1x∗ U Q0x∗ T Q0x∗ D−u z=z1∗+U z0∗ w=T z0∗
I split equations similarly u0=f(u,v0,t)
v=g(u,t)
. x =D−u+z(u,·) +w(u,v0,·) z=z(u,t) w=w(u,v0,t)
F(u, w, z, η, ζ, t) = 0
Fˆ1(u, z, t) = 0 u0= (u, w, t) Fˆ2(u, w, ζ, t) = 0
z=(u, t)
v=(u, t) w=(u, ζ, t)
u0=f(u, v0, t) v=g(u, t)
¯ ZG¯−1
2 D¯P1G¯−1
2 T¯G−1
2
I I−fv0gu remains non-singular (locally) Implicit Index-1 System
General linear methods for integrated circuit design – St. Voigtmann, p. 9
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Properly stated index-2 DAEs
V. (2004) Accessible criteria for the local existence and uniqueness of DAE solutions, MATHEON
A q(x, ˙ ·)+ b(x, ·) = 0
I split solution into characteristic parts x=D−u+z+w
x∗
P0x∗ z0∗=Q0x∗
D−DP¯1x∗ z1∗=P0Q¯1x∗ U Q0x∗ T Q0x∗
D−u z=z1∗+U z0∗ w=T z0∗
I split equations similarly u0=f(u,v0,t)
v=g(u,t)
. x =D−u+z(u,·) +w(u,v0,·) z=z(u,t) w=w(u,v0,t)
F(u, w, z, η, ζ, t) = 0
Fˆ1(u, z, t) = 0 u0= (u, w, t) Fˆ2(u, w, ζ, t) = 0
z=(u, t)
v=(u, t) w=(u, ζ, t)
u0=f(u, v0, t) v=g(u, t)
¯ ZG¯−1
2 D¯P1G¯−1
2 T¯G−1
2
I I−fv0gu remains non-singular (locally) Implicit Index-1 System
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Properly stated index-2 DAEs
A q(x, ˙ ·)+ b(x, ·) = 0
I split solution into characteristic parts x=D−u+z+w
x∗
P0x∗ z0∗=Q0x∗
D−DP¯1x∗ z1∗=P0Q¯1x∗ U Q0x∗ T Q0x∗
D−u z=z1∗+U z0∗ w=T z0∗
I split equations similarly u0=f(u,v0,t)
v=g(u,t)
. x =D−u+z(u,·) +w(u,v0,·) z=z(u,t) w=w(u,v0,t)
F(u, w, z, η, ζ, t) = 0
Fˆ1(u, z, t) = 0 u0= (u, w, t) Fˆ2(u, w, ζ, t) = 0
z=(u, t)
v=(u, t) w=(u, ζ, t)
u0=f(u, v0, t) v=g(u, t)
¯ ZG¯−1
2 D¯P1G¯−1
2 T¯G−1
2
I I−fv0gu remains non-singular (locally) Implicit Index-1 System
General linear methods for integrated circuit design – St. Voigtmann, p. 10
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Properly stated index-2 DAEs (cont.)
V. (2006) General linear methods for integrated circuit design, PhD thesis
A q(x, ˙ ·)+ b(x, ·) = 0
⇔ u0=f(u,v0,t) v=g(u,t)x=D−u+z(u,·) +w(u,v0,·) z=z(u,t) w=w(u,v0,t)
I new decoupling procedure
I existence and uniqueness results
I only mild smoothness assumptions
I covers/extends results on
. linear DAEs (Balla, M ¨arz, Kurina) . nonlinear index-1 DAEs (Higueras, M ¨arz)
. DAEsA[Dx]0+b(Ux,·) +BTx=0 (Tischendorf, Est ´evez Schwarz) . Hessenberg DAEs (Hairer, Lubich, Roche, Wanner)
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Properly stated index-2 DAEs (cont.)
A q(x, ˙ ·)+ b(x, ·) = 0
⇔ u0=f(u,v0,t) v=g(u,t)x=D−u+z(u,·) +w(u,v0,·) z=z(u,t) w=w(u,v0,t)
I new decoupling procedure
I existence and uniqueness results
I only mild smoothness assumptions
I covers/extends results on
. linear DAEs (Balla, M ¨arz, Kurina) . nonlinear index-1 DAEs (Higueras, M ¨arz)
. DAEsA[Dx]0+b(Ux,·) +BTx=0 (Tischendorf, Est ´evez Schwarz) . Hessenberg DAEs (Hairer, Lubich, Roche, Wanner)
General linear methods for integrated circuit design – St. Voigtmann, p. 11
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Decoupling and discretisation
Aq(x,˙ t) +b(x,t) =0 index-2 DAE
decoupling u0=f(u,v0,t) v=g(u,t) +constraints
discretisation discretisation
discretised index-2 DAE
decoupling discretised index-1 DAE +discretised constraints
I If two subspaces associated with the DAE,DN1andDS1are constant, then this diagram commutes.
I It is always assumed thatN0∩S0does not depend onx.
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Decoupling and discretisation
Aq(x,˙ t) +b(x,t) =0 index-2 DAE
decoupling u0=f(u,v0,t) v=g(u,t) +constraints
discretisation discretisation
discretised index-2 DAE
decoupling discretised index-1 DAE +discretised constraints
I If two subspaces associated with the DAE,DN1andDS1are constant, then this diagram commutes.
I It is always assumed thatN0∩S0does not depend onx.
General linear methods for integrated circuit design – St. Voigtmann, p. 11
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Decoupling and discretisation
Aq(x,˙ t) +b(x,t) =0 index-2 DAE
decoupling u0=f(u,v0,t) v=g(u,t) +constraints
discretisation discretisation
discretised index-2 DAE
decoupling discretised index-1 DAE +discretised constraints
I If two subspaces associated with the DAE,DN1andDS1are constant, then this diagram commutes.
I It is always assumed thatN0∩S0does not depend onx.
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Contents
Differential Algebraic Equations
General Linear Methods
Practical General Linear Methods
General linear methods for integrated circuit design – St. Voigtmann, p. 13
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for ODEs y
0= f (y)
I Linear multistep: yn = hβ0f(yn) + Pk
i=1αiyn−i
I Runge-Kutta: Yi = hPs
j=1aijf(Yj) + y[n−1], y[n] = hPs
i=1 bif(Yi) + y[n−1]
I General linear: Yi = hPs
j=1aijf(Yj) + Pr
j=1uijy[n−1]j , y[n]i = hPs
j=1 bijf(Yj) + Pr
j=1vijy[n−1]j
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for ODEs y
0= f (y)
I Linear multistep: yn = hβ0f(yn) + Pk
i=1αiyn−i
I Runge-Kutta: Yi = hPs
j=1aijf(Yj) + y[n−1], y[n] = hPs
i=1 bif(Yi) + y[n−1]
I General linear: Yi = hPs
j=1aijf(Yj) + Pr
j=1uijy[n−1]j , y[n]i = hPs
j=1 bijf(Yj) + Pr
j=1vijy[n−1]j
General linear methods for integrated circuit design – St. Voigtmann, p. 13
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for ODEs y
0= f (y)
I Linear multistep: yn = hβ0f(yn) + Pk
i=1αiyn−i I Runge-Kutta: Yi = hPs
j=1aijf(Yj) + y[n−1], y[n] = hPs
i=1 bif(Yi) + y[n−1]
I General linear: Yi = hPs
j=1aijf(Yj) + Pr
j=1uijy[n−1]j , y[n]i = hPs
j=1 bijf(Yj) + Pr
j=1vijy[n−1]j
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for ODEs y
0= f (y)
I Linear multistep: yn = hβ0f(yn) + Pk
i=1αiyn−i I Runge-Kutta: Yi = hPs
j=1aijf(Yj) + y[n−1], y[n] = hPs
i=1 bif(Yi) + y[n−1]
I General linear: Yi = hPs
j=1aijf(Yj) + Pr
j=1uijy[n−1]j , y[n]i = hPs
j=1 bijf(Yj) + Pr
j=1vijy[n−1]j
General linear methods for integrated circuit design – St. Voigtmann, p. 13
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for ODEs y
0= f (y)
I Linear multistep: yn = hβ0f(yn) + Pk
i=1αiyn−i I Runge-Kutta: Yi = hPs
j=1aijf(Yj) + y[n−1], y[n] = hPs
i=1 bif(Yi) + y[n−1]
I General linear: Yi = hPs
j=1aijf(Yj) + Pr
j=1uijy[n−1]j , y[n]i = hPs
j=1 bijf(Yj) + Pr
j=1vijy[n−1]j
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
An example method
"
A U B V
#
=
1
4 0 0 0 1 0 −321 −1921
49 25
1
4 0 0 1−171100 −10049 −60043
123
1225 −225392 14 0 1 13631400 1394139200 784005379
−9584 −5984 367 14 1 4318 3142 33637
−9584 −5984 367 14 1 4318 3142 33637
0 0 0 1 0 0 0 0
−26821 8621 −289 4 0 709 1021 −215
−3221 8821 −2249 16 0 569 2021 −1021
(Butcher, 2004) I diagonally implicit
I orderp=3and stage orderq=3
I stiffly accurate, A-stable, L-stable
I Nordsieck form, y[n]i+1 ≈hiy(i)(tn)
General linear methods for integrated circuit design – St. Voigtmann, p. 14
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
An example method
"
A U B V
#
=
1
4 0 0 0 1 0 −321 −1921
49 25
1
4 0 0 1−171100 −10049 −60043
123
1225 −225392 14 0 1 13631400 1394139200 784005379
−9584 −5984 367 14 1 4318 3142 33637
−9584 −5984 367 14 1 4318 3142 33637
0 0 0 1 0 0 0 0
−26821 8621 −289 4 0 709 1021 −215
−3221 8821 −2249 16 0 569 2021 −1021
(Butcher, 2004) I diagonally implicit
I orderp=3and stage orderq=3
I stiffly accurate, A-stable, L-stable
I Nordsieck form, y[n]i+1 ≈hiy(i)(tn)
Diagonally implicit methods with high stage order
are possible!
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Why general linear methods?
"
Y y[n]
#
=
"
A U B V
#
·
"
hf(Y) y[n−1]
#
I improve stability . damping properties
similar to RK methods -1
0 1
20 time 10
0 −2
−1 0 1 2
2π time 3π π 2 π 0 2
I improve efficiency
. diagonally implicit schemes . solve stages sequentially
A=
λ 0
... . .. aij · · · λ
I benefit from high stage order . no order reduction
. cheap and reliable error estimates
1.0·10−2 1.5·10−2
2.0·10−2
0 5 10 15 time
p= 3 p= 2 p= 1
General linear methods for integrated circuit design – St. Voigtmann, p. 15
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Why general linear methods?
"
Y y[n]
#
=
"
A U B V
#
·
"
hf(Y) y[n−1]
#
I improve stability . damping properties
similar to RK methods -1
0 1
20 time 10
0 −2
−1 0 1 2
2π time 3π π 2 π 0 2
I improve efficiency
. diagonally implicit schemes . solve stages sequentially
A=
λ 0
... . ..
aij · · · λ
I benefit from high stage order . no order reduction
. cheap and reliable error estimates
1.0·10−2 1.5·10−2
2.0·10−2
0 5 10 15 time
p= 3 p= 2 p= 1
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
Why general linear methods?
"
Y y[n]
#
=
"
A U B V
#
·
"
hf(Y) y[n−1]
#
I improve stability . damping properties
similar to RK methods -1
0 1
20 time 10
0 −2
−1 0 1 2
2π time 3π π 2 π 0 2
I improve efficiency
. diagonally implicit schemes . solve stages sequentially
A=
λ 0
... . ..
aij · · · λ
I benefit from high stage order . no order reduction
. cheap and reliable error estimates
1.0·10−2 1.5·10−2
2.0·10−2
0 5 10 15 time
p= 3 p= 2 p= 1
General linear methods for integrated circuit design – St. Voigtmann, p. 16
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for index-2 DAEs
A q x(t),t 0+b x(t),t
=0
| {z }
↑ ↑ ↑
singular charges/ voltages/
fluxes currents
I input quantities
q[n−1]k+1 ≈hk dd tkkq x(t),t
I
"
q(Xn,tc) q[n]
#
= A U
B V
·
" hQ0n q[n−1]
#
such that
AQ0n+b(Xn,tc) =0
I solve for the stagesXn Remark
I use implicit methods (Anon-singular)
I charge conservation is guaranteed
I only charges / fluxes are passed on from step to step
I analysis usesimplicit index-1 systemy0=f(y,z0), z=g(y)
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for index-2 DAEs
A q x(t),t 0+b x(t),t
=0
| {z }
↑ ↑ ↑
singular charges/ voltages/
fluxes currents
I input quantities
q[n−1]k+1 ≈hk dd tkkq x(t),t
I
"
q(Xn,tc) q[n]
#
= A U
B V
·
"
hQ0n q[n−1]
#
such that
AQ0n+b(Xn,tc) =0
I solve for the stagesXn
Remark
I use implicit methods (Anon-singular)
I charge conservation is guaranteed
I only charges / fluxes are passed on from step to step
I analysis usesimplicit index-1 systemy0=f(y,z0), z=g(y)
General linear methods for integrated circuit design – St. Voigtmann, p. 16
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for index-2 DAEs
A q x(t),t 0+b x(t),t
=0
| {z }
↑ ↑ ↑
singular charges/ voltages/
fluxes currents
I input quantities
q[n−1]k+1 ≈hk dd tkkq x(t),t
I
"
q(Xn,tc) q[n]
#
= A U
B V
·
"
hQ0n q[n−1]
#
such that
AQ0n+b(Xn,tc) =0
I solve for the stagesXn Remark
I use implicit methods (Anon-singular)
I charge conservation is guaranteed
I only charges / fluxes are passed on from step to step
I analysis usesimplicit index-1 systemy0=f(y,z0), z=g(y)
Motivation DAEs General Linear Methods Practical General Linear Methods Summary
GLMs for index-1 DAEs
Apply M=
A U
B V
to implicit index-1 DAEs y0=f(y,z0), z=g(y) Y=hAf(Y,Z0) + Uy[n] g(Y) =hAZ0+Uz[n]
y[n+1]=hBf(Y,Z0) +Vy[n] z[n+1]=hBZ0+ Vz[n]