The document discusses applying alternating direction implicit (ADI) methods to solve tensor structured equations. ADI methods were originally developed to solve linear systems related to Poisson problems on uniform grids. The document outlines how ADI can be generalized to solve systems with tensor structure, such as those arising from tensor train decompositions of multidimensional problems. By exploiting the tensor structure, the ADI method can solve large problems with significantly less computational cost and storage than solving the equivalent vectorized problem directly.
1. Workshop on Matrix Equations and
Tensor Techniques 2011 Aachen,
21 November 2011
ADI for Tensor Structured Equations
Thomas Mach and Jens Saak
Max Planck Institute for Dynamics of Complex Technical Systems
Computational Methods in Systems and Control Theory
MAX PLANCK INSTITUTE
FOR DYNAMICS OF COMPLEX
TECHNICAL SYSTEMS
MAGDEBURG
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 1/37
2. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Classic ADI [Peaceman/Rachford ’55]
Developed to solve linear systems related to Poisson problems
−∆u = f in Ω ⊂ Rd , d = 2
u=0 on ∂Ω.
uniform grid size h, centered differences, d = 1,
⇒ ∆1,h u = h2 f
2 −1
−1 2 −1
∆1,h =
.. .. .. .
. . .
−1 2 −1
−1 2
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 2/37
3. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Classic ADI [Peaceman/Rachford ’55]
Developed to solve linear systems related to Poisson problems
−∆u = f in Ω ⊂ Rd , d = 2
u=0 on ∂Ω.
uniform grid size h, 5-point difference star, d = 2,
⇒ ∆2,h u = h2 f
K −I 4 −1
−I K −I −1 4 −1
∆2,h =
.. .. .. and K =
.. .. .. .
. . . . . .
−I K −I −1 4 −1
−I K −1 4
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 2/37
4. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Classic ADI [Peaceman/Rachford ’55]
Observation
∆2,h = (∆1,h ⊗ I ) + (I ⊗ ∆1,h ).
=:H =:V
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 3/37
5. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Classic ADI [Peaceman/Rachford ’55]
Observation
∆2,h = (∆1,h ⊗ I ) + (I ⊗ ∆1,h ).
=:H =:V
˜
Solve ∆2,h u = h2 f =: f exploiting structure in H and V .
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 3/37
6. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Classic ADI [Peaceman/Rachford ’55]
Observation
∆2,h = (∆1,h ⊗ I ) + (I ⊗ ∆1,h ).
=:H =:V
˜
Solve ∆2,h u = h2 f =: f exploiting structure in H and V .
For certain shift parameters perform
˜
(H + pi I ) ui+ 1 = (pi I − V ) ui + f ,
2
˜
(V + pi I ) ui+1 = (pi I − H) ui+ 1 + f ,
2
until ui is good enough.
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 3/37
7. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
ADI and Lyapunov Equations [Wachspress ’88]
Lyapunov Equation
FX + XF T = −GG T
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 4/37
8. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
ADI and Lyapunov Equations [Wachspress ’88]
Lyapunov Equation
FX + XF T = −GG T
Vectorized Lyapunov Equation
(I ⊗ F ) + (F ⊗ I ) vec(X ) = −vec(GG T )
=:HF =:VF
Same structure ⇒ apply ADI
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 4/37
9. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
ADI and Lyapunov Equations [Wachspress ’88]
Lyapunov Equation
FX + XF T = −GG T
Vectorized Lyapunov Equation
(I ⊗ F ) + (F ⊗ I ) vec(X ) = −vec(GG T )
=:HF =:VF
Same structure ⇒ apply ADI
(F + pi I ) Xi+ 1 = −GG T − Xi F T − pi I
2
(F + pi I ) Xi+1 = −GG T − Xi+ 1 F T − pi I
T
2
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 4/37
10. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
LR-ADI for Lyapunov Equations
Lyapunov Equation
FX + XF T = −GG T
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 5/37
11. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
LR-ADI for Lyapunov Equations
Lyapunov Equation
FX + XF T = −GG T
Often singular values of X decay rapidly when G is “thin”.
X = ZZ T with Z “thin”.
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 5/37
12. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
LR-ADI for Lyapunov Equations
Lyapunov Equation
FX + XF T = −GG T
Often singular values of X decay rapidly when G is “thin”.
X = ZZ T with Z “thin”.
LR-ADI [Penzl ’99, Li/White ’02]
Z0 = [] V1 = −2 Re (p1 )(F + p1 I )−1 G
Re (pi )
Vi = I − (pi + pi−1 )(F + pi I )−1 Vi−1
Re (pi−1 )
Zi = [Zi−1 , Vi ]
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 5/37
13. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Generalizing Matrix Equations
∆2,h vec(X ) = vec(B)
I ⊗ ∆1,h + ∆1,h ⊗ I vec(X ) = vec(B)
=H =V =u =f
∆µa a
Xa c
+ = Ba c
c ∆µc
Xa c
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 6/37
14. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Generalizing Matrix Equations
∆4,h vec(X ) = vec(B)
I ⊗ I ⊗ I ⊗ ∆1,h + I ⊗ I ⊗ ∆1,h ⊗ I + I ⊗ ∆1,h ⊗ I ⊗ I + ∆1,h ⊗ I ⊗ I ⊗ I vec(X ) = vec(B)
=H =V =R =Q =u =f
∆µa a
Xabcd + Xabcd
∆µb b
+ = Babcd
c ∆µc
Xabcd + Xabcd
d ∆µd
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 6/37
15. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Generalizing ADI
I ⊗ ∆1,h + ∆1,h ⊗ I vec(X ) = vec(B)
=H =V =u =f
(H + I ⊗ pi,1 I )Xi+ 1 = (pi,1 I − V )Xi + B
2
(V + pi,2 I ⊗ I )Xi+ 1 = (pi,2 I − H)Xi+ 1 + B
2 2
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 7/37
16. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Generalizing ADI
I ⊗ ∆1,h + ∆1,h ⊗ I vec(X ) = vec(B)
=H =V =u =f
(H + I ⊗ pi,1 I )Xi+ 1 = (pi,1 I − V )Xi + B
2
(V + pi,2 I ⊗ I )Xi+ 1 = (pi,2 I − H)Xi+ 1 + B
2 2
I ⊗ I ⊗ I ⊗ ∆1,h + I ⊗ I ⊗ ∆1,h ⊗ I + I ⊗ ∆1,h ⊗ I ⊗ I + ∆1,h ⊗ I ⊗ I ⊗ I vec(X ) = vec(B)
=H =V =R =Q =u =f
(H + I ⊗ I ⊗ I ⊗ pi,1 I )Xi+ 1 = (pi,1 I − V − R − Q)Xi +B
4
(V + I ⊗ I ⊗ pi,2 I ⊗ I )Xi+ 1 = (pi,2 I − H − R − Q)Xi+ 1 +B
2 4
(R + I ⊗ pi,3 I ⊗ I ⊗ I )Xi+ 3 = (pi,3 I − H − V − Q)Xi+ 1 +B
4 2
(Q + pi,4 I ⊗ I ⊗ I ⊗ I )Xi+1 = (pi,4 I − H − V − R)Xi+ 3 +B
4
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 7/37
17. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Goal
Solve AX = B
A = I ⊗ I ⊗ · · · ⊗ I ⊗ I ⊗ A1 +
I ⊗ I ⊗ · · · ⊗ I ⊗ A2 ⊗ I +
... +
Ad ⊗ I ⊗ · · · ⊗ I ⊗ I ⊗ I
B is given in tensor train decomposition
⇒ X is sought in tensor train decomposition.
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 8/37
18. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
r1 ,...,rd−1
T (i1 , i2 , . . . , id ) = G1 (i1 , α1 )G2 (α1 , i2 , α2 )
α1 ,...,αd−1 =1
· · · Gj (αj−1 , ij , αj ) · · ·
Gd−1 (αd−2 , id−1 , αd−1 )Gd (αd−1 , id ).
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 9/37
19. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
r1 ,...,rd−1
T (i1 , i2 , . . . , id ) = G1 (i1 , α1 )G2 (α1 , i2 , α2 )
α1 ,...,αd−1 =1
· · · Gj (αj−1 , ij , αj ) · · ·
Gd−1 (αd−2 , id−1 , αd−1 )Gd (αd−1 , id ).
G1 (i1 , α1 ) α1 G2 (α1 , i2 , α2 ) α2 ··· Gd (αd−1 , id )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 9/37
20. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
Tensor trains are
computable, and
d
require only O(dnr 2 ) storage, with TT-rank r and T ∈ Rn .
Canonical representation
T (i1 , i2 , . . . , id ) = G1 (i1 , α) · · · Gd (id , α)
α
Tucker decomposition
T (i1 , i2 , . . . , id ) = C (α1 , . . . , αd )G1 (i1 , α1 ) · · · Gd (id , αd )
α1 ,...,αd
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 10/37
21. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
(I ⊗ · · · ⊗ I ⊗ A1 ) T
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 11/37
22. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
(I ⊗ · · · ⊗ I ⊗ A1 ) T
A1 (β, i1 )
i1
G1 (i1 , α1 ) α1 G2 (α1 , i2 , α2 ) α2 ··· Gd (αd−1 , id )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 11/37
23. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
(I ⊗ · · · ⊗ I ⊗ A1 ) T
A1 (β, i1 )
i1
G1 (i1 , α1 ) α1 G2 (α1 , i2 , α2 ) α2 ··· Gd (αd−1 , id )
T (i1 , i2 , . . . , id ) ×1 A1 = A1 β,i1 G1 (i1 , α1 )G2 (α1 , i2 , α2 )
α1 ,...,αd−1
· · · Gd−1 (αd−2 , id−1 , αd−1 )Gd (αd−1 , id )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 11/37
24. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
(I ⊗ · · · ⊗ I ⊗ A1 ) T
˜
= G (β, α1 ) = A1 G1
A1 (β, i1 )
i1
G1 (i1 , α1 ) α1 G2 (α1 , i2 , α2 ) α2 ··· Gd (αd−1 , id )
T (i1 , i2 , . . . , id ) ×1 A1 = A1 β,i1 G1 (i1 , α1 )G2 (α1 , i2 , α2 )
α1 ,...,αd−1
· · · Gd−1 (αd−2 , id−1 , αd−1 )Gd (αd−1 , id )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 11/37
25. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Tensor Trains [Oseledets, Tyrtyshnikov ’09]
(I ⊗ · · · ⊗ I ⊗ A1 ) T
˜
= G (β, α1 ) = A1 G1
A1 (β, i1 )
i1
G1 (i1 , α1 ) α1 G2 (α1 , i2 , α2 ) α2 ··· Gd (αd−1 , id )
T (i1 , i2 , . . . , id ) ×1 A1 −1 = A1 −1 β,i1 G1 (i1 , α1 )G2 (α1 , i2 , α2 )
α1 ,...,αd−1
· · · Gd−1 (αd−2 , id−1 , αd−1 )Gd (αd−1 , id )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 11/37
26. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Eigenvalues
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
St´phanos’ theorem:
e
⇒ λi (A) = λi1 (A1 ) + λi2 (A2 ) + · · · + λid (Ad ),
d−1
with i = i1 + i2 n1 + · · · + id nj .
j=1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 12/37
27. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Eigenvalues
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
St´phanos’ theorem:
e
⇒ λi (A) = λi1 (A1 ) + λi2 (A2 ) + · · · + λid (Ad ),
d−1
with i = i1 + i2 n1 + · · · + id nj .
j=1
d
AX = B ⇔ X ×j Aj = B
j=1
A is regular ⇔ λi (A) = 0 ∀i ⇐ Ai Hurwitz ∀i
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 12/37
28. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Algorithm
Input: {A1 , . . . , Ad }, tensor train B, accuracy
Output: tensor train X , with AX = B
forall j ∈ {1, . . . , d} do
(0)
Xj := zeros(n, 1, 1)
end
while r (i) > do
Choose shift pi
forall k ∈ {1, . . . , d} do
d
×j Aj ×k (Ak + pi I )−1
k k−1 k−1
X (i+ d ) := B +pi X (i+ d ) − X (i+ d )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 13/37
29. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Algorithm
r (i) := B
Input: {A1 , . . . , Ad }, tensor train B, accuracy
forall j ∈ {1, . . . , d} do
Output: tensor train X , with AX(i) B =
r (i) := r − Xi ×j Aj
forall j ∈ {1, . . . , d} do
(0) end
Xj := zeros(n, 1, 1)
end
while r (i) > do
Choose shift pi
forall k ∈ {1, . . . , d} do
d
×j Aj ×k (Ak + pi I )−1
k k−1 k−1
X (i+ d ) := B +pi X (i+ d ) − X (i+ d )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 13/37
30. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Algorithm
Input: {A1 , . . . , Ad }, tensor train B, accuracy
Output: tensor train X , with AX = B
forall j ∈ {1, . . . , d} do
(0)
Xj := zeros(n, 1, 1)
end
while r (i) > do
Choose shift pi
forall k ∈ {1, . . . , d} do
d
×j Aj ×k (Ak + pi I )−1
k k−1 k−1
X (i+ d ) := B +pi X (i+ d ) − X (i+ d )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 13/37
31. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Algorithm
Input: {A1 , . . . , Ad }, tensor train B, accuracy
Output: tensor train X , with AX = B
forall j ∈ {1, . . . , d} do
(0)
Xj := zeros(n, 1, 1)
end (I ⊗ I ⊗ · · · ⊗ I ⊗ Aj ⊗ I ⊗ · · · ⊗ I ) Xi+ k−1
d
while r (i) > do
Choose shift pi
forall k ∈ {1, . . . , d} do
d
×j Aj ×k (Ak + pi I )−1
k k−1 k−1
X (i+ d ) := B +pi X (i+ d ) − X (i+ d )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 13/37
32. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Improvements
better shifts, e.g. Wachspress/Penzl
test residual in innermost loop
replace inner loop by random k
use tensor train truncation and start with low accuarcy
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 14/37
33. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Improvements
better shifts, e.g. Wachspress/Penzl
test residual in innermost loop
replace inner loop by random k
use tensor train truncation and start with low accuarcy
[...]
while r (i) > do
Choose shifts pi,j
forall k ∈ {1, . . . , d} do
k −1
d
k k −1
X (i+ d ) := B + pi,k X (i+ d ) − X (i+ d ) ×j Aj ×k (Ak − pi,k I )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 14/37
34. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Improvements
better shifts, e.g. Wachspress/Penzl
test residual in innermost loop
replace inner loop by random k
use tensor train truncation and start with low accuarcy
[...]
while r (i) > do
Choose shifts pi,j
forall k ∈ {1, . . . , d} do
k −1
d
k k −1
X (i+ d ) := B + pi,k X (i+ d ) − X (i+ d ) ×j Aj ×k (Ak − pi,k I )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 14/37
35. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Improvements
better shifts, e.g. Wachspress/Penzl
test residual in innermost loop
replace inner loop by random k
use tensor train truncation and start with low accuarcy
[...]
while r (i) > do
Choose shifts pi,j
forall = 1, . . . , 5 do
k := random(1, . . . , d)
−1
d
−1
X (i+ 5 ) := B + pi, X (i+ 5 ) − X (i+ 5 ) ×j Aj ×k (Ak − pi, I )
j=1
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 14/37
36. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Improvements
better shifts, e.g. Wachspress/Penzl
test residual in innermost loop
replace inner loop by random k
use tensor train truncation and start with low accuarcy
Additional remark:
[...] The improvement with ran-
while r (i) > do dom chosen directions works
for the Lyapunov case with
Choose shifts pi,j
the special right hand side as
forall = 1, . . . , 5 do
k := random(1, . . . , d) in the examples. The investi-
gation of the convergence be-
−1
d
−1
X (i+ 5 ) := B + pi, X (i+ 5havior for(i+ 5 ) ×j Arhs×k (Ak − pi, I )
)
− X random j is work
j=1
in progress.
j=k
end
end
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 14/37
37. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Improvements
better shifts, e.g. Wachspress/Penzl
test residual in innermost loop
replace inner loop by random k
use tensor train truncation and start with low accuarcy
1 · 105
constant truncation error 10−2
i
Storage in Double
80,000 tightened truncation error
Truncation Error
60,000 10−8
40,000
10−14
20,000
0 10−20
0 5 10 15 20 25 30
Iteration
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 14/37
38. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Lemma
Lemma [Grasedyck ’04]
The tensor equation
d
j=1 X ×j Aj = B
with λi (A) = 0 ∀i and Ak Hurwitz has the solution
∞
X =− 0 B ×1 exp(A1 t) ×2 · · · ×d exp(Ad t)dt
Z (t) = B ×1 exp(A1 t) ×2 · · · ×d exp(Ad t)
d ∞
˙
Z (t) = Z (t) ×j Aj Z (∞) − Z (0) = ˙
Z (t)dt,
j=1 0
d ∞
0−B = Z (t)dt ×j Aj
j=1 0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 15/37
39. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Theorem
Theorem
{A1 , . . . , Ad } ⇒ A, Λ(A) ⊂ [−λmax , −λmin ] ⊕ ı [−µ, µ] ⊂ C− .
Let k ∈ N and use the quadrature points and weights:
√
hst := √k , tj := log e jhst + 1 + e 2jhst , wj := √ hst−2jh .
π
1+e st
Then the solution X can be approximated by
r1 ,...,rd−1
˜
X (i1 , i2 , . . . , id ) = − H1 (i1 , α1 ) · · · Hd (αd−1 , id ),
α1 ,...,αd−1 =1
2tj
2wj Ap
with Hp (αp−1 , ip , αp ) := k
j=−k λmin βp e
λmin
ip ,βp
Gp (αp−1 , βp , αp )
with the approximation error
2µλ−1 +1 √
(λI − 2A/λmin )−1
min
˜
X −X ≤ Cst −π k
πλmin e dΓ λ B 2.
2 π
Γ 2
extending [Grasedyck ’04] (X and B of low Kronecker rank) to low TT-rank
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 16/37
40. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Theorem
Theorem
{A1 , . . . , Ad } ⇒ A, Λ(A) ⊂ [−λmax , −λmin ] ⊕ ı [−µ, µ] ⊂ C− .
Let k ∈ N and use the quadrature points and weights:
√
hst := √k , tj := log e jhst + 1 + e 2jhst , wj := √ hst−2jh .
π
1+e st
Then the solution X can be approximated by
r1 ,...,rd−1
˜
X (i1 , i2 , . . . , id ) = − H1 (i1 , α1 ) · · · Hd (αd−1 , id ),
α1 ,...,αd−1 =1
2tj
2wj Ap
with Hp (αp−1 , ip , αp ) := k
j=−k λmin βp e
λmin
ip ,βp
Gp (αp−1 , βp , αp )
with the approximation error
2µλ−1 +1 √
(λI − 2A/λmin )−1
min
˜
X −X ≤ Cst −π k
πλmin e dΓ λ B 2.
2 π
Γ 2
extending [Grasedyck ’04] (X and B of low Kronecker rank) to low TT-rank
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 16/37
41. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Theorem
Theorem
{A1 , . . . , Ad } ⇒ A, Λ(A) ⊂ [−λmax , −λmin ] ⊕ ı [−µ, µ] ⊂ C− .
Let k ∈ N and use the quadrature points and weights:
√
hst := √k , tj := log e jhst + 1 + e 2jhst , wj := √ hst−2jh .
π
1+e st
Then the solution X can be approximated by
r1 ,...,rd−1
˜
X (i1 , i2 , . . . , id ) = − H1 (i1 , α1 ) · · · Hd (αd−1 , id ),
α1 ,...,αd−1 =1
2tj
2wj Ap
with Hp (αp−1 , ip , αp ) := k
j=−k λmin βp e
λmin
ip ,βp
Gp (αp−1 , βp , αp )
with the approximation error
2µλ−1 +1 √
(λI − 2A/λmin )−1
min
˜
X −X ≤ Cst −π k
πλmin e dΓ λ B 2.
2 π
Γ 2
extending [Grasedyck ’04] (X and B of low Kronecker rank) to low TT-rank
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 16/37
42. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Theorem
Theorem
{A1 , . . . , Ad } ⇒ A, Λ(A) ⊂ [−λmax , −λmin ] ⊕ ı [−µ, µ] ⊂ C− .
Let k ∈ N and use the quadrature points and weights:
√
hst := √k , tj := log e jhst + 1 + e 2jhst , wj := √ hst−2jh .
π
1+e st
B (i1 , i2 , . . . , id ) = G1 (i1 , α1 )G2 (α1 , i2 , α2 )
α1 ,...,αd−1
Then the solution X can be approximated by
· · · Gj (αj−1 , ij , αj ) · · ·
r1 ,...,rd−1
˜
X (i1 , i2 , . . . , id ) = − Gd−1 (αd−2 ,(i , α d−1·)GH (α , id ). i
H id−1 , α ) · · d (αd−1 ,
1 1 1 d d−1 d ),
α1 ,...,αd−1 =1
2tj
2wj Ap
with Hp (αp−1 , ip , αp ) := k
j=−k λmin βp e
λmin
ip ,βp
Gp (αp−1 , βp , αp )
with the approximation error
2µλ−1 +1 √
(λI − 2A/λmin )−1
min
˜
X −X ≤ Cst −π k
πλmin e dΓ λ B 2.
2 π
Γ 2
extending [Grasedyck ’04] (X and B of low Kronecker rank) to low TT-rank
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 16/37
43. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Theorem
Theorem
{A1 , . . . , Ad } ⇒ A, Λ(A) ⊂ [−λmax , −λmin ] ⊕ ı [−µ, µ] ⊂ C− .
Let k ∈ N and use the quadrature points and weights:
√
hst := √k , tj := log e jhst + 1 + e 2jhst , wj := √ hst−2jh .
π
1+e st
Then the solution X can be approximated by
r1 ,...,rd−1
˜
X (i1 , i2 , . . . , id ) = − H1 (i1 , α1 ) · · · Hd (αd−1 , id ),
α1 ,...,αd−1 =1
2tj
2wj Ap
with Hp (αp−1 , ip , αp ) := k
j=−k λmin βp e
λmin
ip ,βp
Gp (αp−1 , βp , αp )
with the approximation error
2µλ−1 +1 √
(λI − 2A/λmin )−1
min
˜
X −X ≤ Cst −π k
πλmin e dΓ λ B 2.
2 π
Γ 2
extending [Grasedyck ’04] (X and B of low Kronecker rank) to low TT-rank
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 16/37
44. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 1
The quadrature formula (tj , wj ) can be found in [Stenger ’93,
Example 4.2.11], with d = π/2, α = β = = 1, n = N = M = k.
[Hackbusch ’09, D.4.3] shows, that d = π/2 is optimal. The
quadrature formula is used to approximate 1 resp. the inverse of
r
matrices by
∞ k
1 tr
= e dt ≈ wj e tj r .
−r 0 j=−k
The quadrature error is bounded by [Stenger ’93, (4.2.60)]
∞ k √
tr
e dt − wj e tj r ≤ C3 e −π k
,
0 j=−k
with [Grasedyck, Hackbusch, Khoromskij ’03],
C3 ≤ Cst e | (z)|/π
.
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 17/37
45. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 2
The lemma
∞
X =− B ×1 exp(A1 t) ×2 · · · ×d exp(Ad t)dt
0
2
together with scaling λmin we get the formula
r1 ,...,rd−1
˜
X (i1 , i2 , . . . , id ) = − H1 (i1 , α1 ) · · · Hd (αd−1 , id ),
α1 ,...,αd−1 =1
k 2tj
2wj Ap
where Hp (αp−1 , ip , αp ) := λmin e λmin ip ,βp
Gp (αp−1 , βp , αp ).
j=−k βp
∞ k
1
−r = e tr dt ≈ wj e tj r
0 j=−k
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 18/37
46. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 3
For the error holds
r1 ,...,rd−1 d
˜
X −X = Gp (αp−1 , βp , αp )
2
α1 ,...,αd−1 =1 p=1 βp
∞ k 2tj
2t
A
λmin p
A
λmin p
− e + wj e .
0 j=−k
ip ,βp 2
The Dunford-Cauchy formula and the quadrature error give
∞ k 2tj
2t Ap Ap
− e λmin + wj e λmin
0 j=−k
∞ −1 k −1
1 2Ap 2Ap
≤ − e tλ λI − dΓ λ + wj e tj λ λI − dΓ λ
2π Γ 0 λmin Γ λmin
j=−k
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 19/37
47. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 3
Dunford-Cauchy representation of the matrix exponential:
e tAp = 1
2πı Γe
tλ (λI − Ap )−1 dΓ λ
The Dunford-Cauchy formula and the quadrature error give
∞ k 2tj
2t Ap Ap
− e λmin + wj e λmin
0 j=−k
∞ −1 k −1
1 2Ap 2Ap
≤ − e tλ λI − dΓ λ + wj e tj λ λI − dΓ λ
2π Γ 0 λmin Γ λmin
j=−k
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 19/37
48. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 3
The Dunford-Cauchy formula and the quadrature error give
∞ k 2tj
2t Ap Ap
− e λmin + wj e λmin
0 j=−k
∞ −1 k −1
1 2Ap 2Ap
≤ − e tλ λI − dΓ λ + wj e tj λ λI − dΓ λ
2π Γ 0 λmin Γ λmin
j=−k
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 19/37
49. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 3
The Dunford-Cauchy formula and the quadrature error give
∞ k 2tj
2t Ap Ap
− e λmin + wj e λmin
0 j=−k
∞ −1 k −1
1 2Ap 2Ap
≤ − e tλ λI − dΓ λ + wj e tj λ λI − dΓ λ
2π Γ 0 λmin Γ λmin
j=−k
∞ k −1
1 2Ap
≤ − e tλ + wj e tj λ λI − dΓ λ
2π 0 Γ λmin
j=−k
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 19/37
50. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 3
The Dunford-Cauchy formula and the quadrature error give
∞ k 2tj
2t Ap Ap
− e λmin + wj e λmin
0 j=−k
∞ −1 k −1
1 2Ap 2Ap
≤ − e tλ λI − dΓ λ + wj e tj λ λI − dΓ λ
2π Γ 0 λmin Γ λmin
The quadrature error is bounded by [Stenger ’93, (4.2.60)]
j=−k
∞ k −1
1 2Ap
≤ − e tλ + w∞tj λ
je λI −
k dΓ λ √
2π 0 tr Γ λmin t r −π k
j=−k
e dt − wj e j
≤ C3 e ,
0 j=−k
with [Grasedyck, Hackbusch, Khoromskij ’03],
C3 ≤ Cst e | (z)|/π
.
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 19/37
51. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Proof – Part 3
The Dunford-Cauchy formula and the quadrature error give
∞ k 2tj
2t Ap Ap
− e λmin + wj e λmin
0 j=−k
∞ −1 k −1
1 2Ap 2Ap
≤ − e tλ λI − dΓ λ + wj e tj λ λI − dΓ λ
2π Γ 0 λmin Γ λmin
j=−k
∞ k −1
1 2Ap
≤ − e tλ + wj e tj λ λI − dΓ λ
2π 0 Γ λmin
j=−k
−1
1 | (z)| −π
√ 2Ap
≤ Cst e π e k λI − dΓ λ.
2π Γ λmin
2
Summing over p completes the proof.
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 19/37
52. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Numerical Results
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 20/37
53. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Example: Laplace – Ai = ∆1, 11
1
Ai = ∆1, 1
11
B = 0 0 ... 0 1
Shifts:
pi := e1 (∗1 ) + . . . + ed (∗d ) — random chosen eigenvalue
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 21/37
54. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Numerical Results – Ai = ∆1, 11
1
d t in s residual mean(#it)
2 0.3887 e+00 7.015 e−10 112.8
5 5.3975 e+00 7.467 e−10 45.8
8 6.0073 e+00 6.936 e−10 12.8
10 3.6624 e+00 7.685 e−10 6.8
25 3.1421 e+01 2.437 e−10 5.0
50 2.2682 e+02 2.049 e−10 5.0
75 7.1918 e+02 4.036 e−10 5.0
100 1.6997 e+03 1.864 e−10 5.0
150 5.5375 e+03 1.801 e−10 5.0
200 1.2795 e+04 1.472 e−10 5.0
250 2.4991 e+04 1.816 e−10 5.0
300 4.2979 e+04 2.535 e−10 5.0
500 1.9515 e+05 2.039 e−10 5.0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 22/37
55. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Numerical Results – Ai = ∆1, 11
1
sparse dense
d TADI MESS Penzl’s sh. lyap
2 0.310 0.0006 0.024 0.003 0.0003 0.0005
4 3.130 0.1695 0.011 0.049 6.331 0.012
6 8.147 — 0.076 0.094 — 7.165
8 5.458 — 5.863 1.097 — 13 698.212
10 5.306 — 3 445.523 249.464 — —
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 23/37
56. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Numerical Results – Ai = ∆1, 11
1
105
104
Computation Time in s
103
102
Tensor ADI
101 sparse
MESS
100
Penzl’s shifts
10−1 dense
lyap
10−2
10 100 300
Dimension d
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 24/37
57. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Numerical Results – Ai = ∆1, 11
1
105
104
Computation Time in s
103
102
Tensor ADI
101 sparse
MESS
100
Penzl’s shifts
10−1 dense
lyap
10−2
10 100 300
Dimension d
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 24/37
58. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Single Shift and Convergence
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
We assume Λ(Ak ) ⊂ R− .
Error Propagation, Single Shift
p− λk + λl λk
d d
k k
G1 2 ≤ max = 1 − .
λk ∈Λ(Ak ), p + λl p + λl
k=1,...,d l=0 l=0
If G1 2 < 1, then the ADI iteration converges.
p < 0 and p > −∞
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 25/37
59. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Single Shift and Convergence
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
We assume Λ(Ak ) ⊂ R− .
Error Propagation, Single Shift
p− λk + λl λk
d d
k k
G1 2 ≤ max = 1 − .
λk ∈Λ(Ak ), p + λl p + λl
k=1,...,d l=0 l=0
If G1 2 < 1, then the ADI iteration converges.
p < 0 and p > −∞
d
p < λi (A) = k=1 λk (Ak ) ∀i
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 25/37
60. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Single Shift and Convergence
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
We assume Λ(Ak ) ⊂ R− .
Error Propagation, Single Shift
p− λk + λl λk
d d
k k
G1 2 ≤ max = 1 − .
λk ∈Λ(Ak ), p + λl p + λl
k=1,...,d l=0 l=0
If G1 2 < 1, then the ADI iteration converges.
p < 0 and p > −∞
d
p < λi (A) = k=1 λk (Ak ) ∀i
d−2
Lyapunov case (Ak = A0 ∀k): p < 2 λmin (A0 )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 25/37
61. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Single Shift and Convergence
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
We assume Λ(Ak ) ⊂ R− .
Error Propagation, Single Shift
p− λk + λl λk
d d
k k
G1 2 ≤ max = 1 − .
λk ∈Λ(Ak ), p + λl p + λl
k=1,...,d l=0 l=0
If G1 2 < 1, then the ADI iteration converges.
p < 0 and p > −∞
d
p < λi (A) = k=1 λk (Ak ) ∀i
2−2
Lyapunov case (Ak = A0 ∀k): p < 2 λmin (A0 ) =0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 25/37
62. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Shifts
Min-Max-Problem
d
pi,k − j=k λj
min max
{p1,1 ,...,p ,d }⊂C λk ∈Λ(Ak ) ∀k pi,k + λk
i=0 k=0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 26/37
63. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Shifts
Min-Max-Problem
d
pi,k − j=k λj
min max
{p1,1 ,...,p ,d }⊂C λk ∈Λ(Ak ) ∀k pi,k + λk
i=0 k=0
Min-Max-Problem, Lyapunov case (Ak = A0 ∀k, A0 Hurwitz)
d
pi,k − j=k λj
min max
{p1,1 ,...,p ,d }⊂C λk ∈Λ(A0 ) ∀k pi,k + λk
i=0 k=0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 26/37
64. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Shifts
Min-Max-Problem
d
pi,k − j=k λj
min max
{p1,1 ,...,p ,d }⊂C λk ∈Λ(Ak ) ∀k pi,k + λk
i=0 k=0
Min-Max-Problem, Lyapunov case (Ak = A0 ∀k, A0 Hurwitz)
d
pi − j=k λj
min max
{p1 ,...,p }⊂C λk ∈Λ(A0 ) ∀k pi + λk
i=0 k=0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 26/37
65. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Shifts
Min-Max-Problem
d
pi,k − j=k λj
min max
{p1,1 ,...,p ,d }⊂C λk ∈Λ(Ak ) ∀k pi,k + λk
i=0 k=0
Min-Max-Problem, Lyapunov case (Ak = A0 ∀k, A0 Hurwitz)
d
pi − j=k λj
min max
{p1 ,...,p }⊂C λk ∈Λ(A0 ) ∀k pi + λk
i=0 k=0
λk = λ0 ∀k
Penzl’s idea: {p1 , . . . , p } ⊂ (d − 1)Λ(A0 )
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 26/37
66. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Random Example
seed := 1;
R := rand(10);
R := R + R ;
R := R − λmin + 0.1;
A0 = −R;
Λ(A0 ) = {−0.1000, −0.2250, −1.1024, −1.7496, −2.0355,
−2.4402, −3.1330, −3.3961, −3.9347, −11.9713}
⇒ The random shifts do not lead to convergence.
p0 = λ10 (A0 )(d − 1)
p1 = λ9 (A0 )(d − 1)
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 27/37
67. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Numerical Results – Ai = −R
random k, test residual every 5 inner iterations, max. 250 iterations
d t in s residual #it
2 0.3627 2.6327 e−06 250
5 17.6850 1.4517 e−07 250
8 62.4336 9.3164 e−09 200
10 44.1547 8.5963 e−09 125
15 12.2231 5.0356 e−09 60
20 15.7506 3.3142 e−09 50
25 25.2221 3.6501 e−09 45
50 49.2004 5.4141 e−09 35
75 118.1297 6.8682 e−09 30
100 614.4017 2.4598 e−09 30
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 28/37
68. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Conditioning of the Problem
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 29/37
69. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Eigenvalues
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
St´phanos’ theorem:
e
⇒ λi (A) = λi1 (A1 ) + λi2 (A2 ) + · · · + λid (Ad ),
d−1
with i = i1 + i2 n1 + · · · + id nj .
j=1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 30/37
70. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Eigenvalues
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
St´phanos’ theorem:
e
⇒ λi (A) = λi1 (A1 ) + λi2 (A2 ) + · · · + λid (Ad ),
d−1
with i = i1 + i2 n1 + · · · + id nj .
j=1
d d
max |λl (A)| ≥ λmk (Ak ) ≥ Re (λmk (Ak )) , if ∀k : Ak Hurwitz
l
k=1 k=1
d
min |λl (A)| ≤ min |λlk (Ak )| .
l lk
k=1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 30/37
71. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Eigenvalues
A = I ⊗ · · · ⊗ I ⊗ A1 + I ⊗ · · · ⊗ I ⊗ A2 ⊗ I + . . . + Ad ⊗ I ⊗ · · · ⊗ I
St´phanos’ theorem:
e
⇒ λimk = λi1 (A1 argmax 2 ) + ·|λ·j (Ak )| (Ad ),
(A) = ) + λi2 (A · + λid
j∈{1,...,nk : Im (λi (Ak ))≥0}
d−1
with i = i1 + i2 n1 + · · · + id nj .
j=1
d d
max |λl (A)| ≥ λmk (Ak ) ≥ Re (λmk (Ak )) , if ∀k : Ak Hurwitz
l
k=1 k=1
d
min |λl (A)| ≤ min |λlk (Ak )| .
l lk
k=1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 30/37
72. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Normal Matrices Aj
Lemma
If all Ai , i = 1, . . . , d are normal, then also A is normal.
Proof.
Here d = 3, the extension to larger d is obvious.
AAT = A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 31/37
73. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Normal Matrices Aj
Lemma
If all Ai , i = 1, . . . , d are normal, then also A is normal.
Proof.
Here d = 3, the extension to larger d is obvious.
AAT = A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 AT ⊗ I ⊗ I + I ⊗ AT ⊗ I + I ⊗ I ⊗ AT
3 2 1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 31/37
74. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Normal Matrices Aj
Lemma
If all Ai , i = 1, . . . , d are normal, then also A is normal.
Proof.
Here d = 3, the extension to larger d is obvious.
AAT = A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 AT ⊗ I ⊗ I + I ⊗ AT ⊗ I + I ⊗ I ⊗ AT
3 2 1
= A3 AT ⊗ II ⊗ II
3 + A3 I ⊗ IAT ⊗ II
2 + A3 I ⊗ II ⊗ IAT
1
+ IAT ⊗ A2 I ⊗ II
3 + II ⊗ A2 AT ⊗ II
2 + II ⊗ A2 I ⊗ IAT
1
+ IAT ⊗ II ⊗ A1 I
3 + II ⊗ IAT ⊗ A1 I
2 + II ⊗ II ⊗ A1 AT
1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 31/37
75. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Normal Matrices Aj
Lemma
If all Ai , i = 1, . . . , d are normal, then also A is normal.
Proof.
Here d = 3, the extension to larger d is obvious.
AAT = A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 AT ⊗ I ⊗ I + I ⊗ AT ⊗ I + I ⊗ I ⊗ AT
3 2 1
= AT A3 ⊗ II ⊗ II
3 + IA3 ⊗ AT I ⊗ II
2 + IA3 ⊗ II ⊗ AT I
1
+ AT I ⊗ IA2 ⊗ II
3 + II ⊗ AT A2 ⊗ II
2 + II ⊗ IA2 ⊗ AT I
1
+ AT I ⊗ II ⊗ IA1 + II ⊗ AT I ⊗ IA1 + II ⊗ II ⊗ AT A1
3 2 1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 31/37
76. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Normal Matrices Aj
Lemma
If all Ai , i = 1, . . . , d are normal, then also A is normal.
Proof.
Here d = 3, the extension to larger d is obvious.
AAT = A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 AT ⊗ I ⊗ I + I ⊗ AT ⊗ I + I ⊗ I ⊗ AT
3 2 1
= AT A3 ⊗ II ⊗ II
3 + IA3 ⊗ AT I ⊗ II
2 + IA3 ⊗ II ⊗ AT I
1
+ AT I ⊗ IA2 ⊗ II
3 + II ⊗ AT A2 ⊗ II
2 + II ⊗ IA2 ⊗ AT I
1
+ AT I ⊗ II ⊗ IA1 + II ⊗ AT I ⊗ IA1 + II ⊗ II ⊗ AT A1
3 2 1
T
= A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1
T
=A A
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 31/37
77. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Normal Matrices Aj
Lemma
If all Ai , i = 1, . . . , d are normal, then also A is normal.
Proof. max |λl (A)|
σmax (A)
κ3, the= A 2 A−1 larger d is obvious.l
Here d = 2 (A) extension to 2 = =
σmin (A) min |λl (A)|
l
AAT = A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 AT ⊗ I ⊗ I + I ⊗ AT ⊗ I + I ⊗ I ⊗ AT
3
max |λlu (A0 )|
2 1
∀i lu
= AT A3 ⊗ II ⊗ II + IA3 ⊗ AT I ⊗ II + IA3 ⊗ II ⊗ AT I
3 2 = 1
Ai =A0 min |λll (A0 )|
+ A3 I ⊗ IA2 ⊗ II + II ⊗ A2 A2 ⊗ II + II ⊗ IAl2 ⊗ AT I
T T
l 1
+ AT I ⊗ II ⊗ IA1 + II ⊗ AT I ⊗ IA1 + II ⊗ II ⊗ AT A1
3 2 1
T
= A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1 A3 ⊗ I ⊗ I + I ⊗ A2 ⊗ I + I ⊗ I ⊗ A1
T
=A A
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 31/37
78. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Lower Bounds for Non-Normal Matrices Aj
FX + XF T = −G (i.e., d = 2, A1 = A2 = F Hurwitz) [Zhou ’02]
max |λl (F )|
l
κ2 (A) ≥
min |λl (F )|
l
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 32/37
79. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Lower Bounds for Non-Normal Matrices Aj
FX + XF T = −G (i.e., d = 2, A1 = A2 = F Hurwitz) [Zhou ’02]
max |λl (F )|
l
κ2 (A) ≥
min |λl (F )|
l
by observing
σmax (A) = A 2 = sup Ay 2
y 2 =1
≥ sup Ay 2
y 2 =1,y EV
= max |λl (A)|
l
= 2 max |λl (F )|
l
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 32/37
80. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Lower Bounds for Non-Normal Matrices Aj
FX + XF T = −G (i.e., d = 2, A1 = A2 = F Hurwitz) [Zhou ’02]
max |λl (F )|
l
κ2 (A) ≥
min |λl (F )|
l
and
−1 A−1 y 2 −1
σmin (A) = A−1 2
= sup
y 2
y 2 Ay 2
= inf −1 y
= inf
A 2 y 2
Ay 2
≤ inf
y EV y 2
= min |λl (A)| = 2 min |λl (F )|
l l
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 32/37
81. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Lower Bounds for Non-Normal Matrices Aj
Tensor Lyapunov Equations
(i.e., ∀i = 1, . . . , d, Ai = A0 Hurwitz) [M./S. ’11]
max |λl (A0 )|
l
κ2 (A) ≥
min |λl (A0 )|
l
σmax (A) ≥ max |λl (A)| = d max |λl (A0 )|
l l
σmin (A) ≤ min |λl (A)| = d min |λl0 (A0 )|
l l0
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 33/37
82. ADI ADI for Tensors Numerical Results and Shifts Conditioning of the Problem Conclusions
Lower Bounds for Non-Normal Matrices Aj
Tensor Silvester Equations
(i.e., ∀i = 1, . . . , d, Ai Hurwitz) [M./S. ’11]
d
λmk (Ak )
k=1
κ2 (A) ≥ d
min |λl (Ak )|
k=1 l
mk as before
d
σmax (A) ≥ max |λl (A)| ≥ λmk (Ak )
l
k=1
d
σmin (A) ≤ min |λl (A)| ≤ min |λlk (Ak )|
l lk
k=1
Max Planck Institute Magdeburg Thomas Mach, Jens Saak, Tensor-ADI 33/37