model reduction method for solving parameter-dependent...
TRANSCRIPT
Model reduction method for solvingparameter-dependent dynamical systems
Marie Billaud-Friess?,†
Joint work with : A. Nouy?
Work supported by GDR MOMAS
† [email protected]?GeM, Ecole Centrale de Nantes
October 6th, 2015
Outline
1 Motivations
2 Low-rank approximation methods
3 Proposed method
4 Numerical experiments
M. Billaud-Friess — Workshop MOMAS 2015 2/28
General context
Parameter and time dependent problem
Find u(t, ξ) : ξ × [0, T ]→ Rd such that
F(u(t, ξ), t, ξ) = 0
where ξ ∈ Ξ ⊂ Rp are parameters or random variables in measure space (Ξ, µ).
Example : Groundwater flows : benchmark couplex with uncertainties 1
Ω
D
ΩD
ΩA
ΩC
ΩMω
(∂C
∂t+ λ(ξ)C
)−∇·(de(ξ)∇C)+u(ξ)·∇C = 0
with u(ξ) = −Ki(ξ)∇H s.t.
−∇ · (K(ξ)∇H) = 0, K(ξ) =∑i
1ΩiKi(ξ)
Finite discretization Algebraic system with d very large
1. https://www.ljll.math.upmc.fr/cances/gdrmomas/ex_qualifications.html
M. Billaud-Friess — Workshop MOMAS 2015 Motivations 3/28
Need of model reduction
Issue : Multiple parameter simulations can be expensive and intractable
• Variable of interest : s(ξ) = `(u(ξ))
• Quantity of interest : statistical moments, failure probability, sensitivity indices
Derive a surrogate problem cheaper to solve
Strategy : Model order reduction methods
• Proper generalized decomposition,
• Reduced basis with POD-Greedy strategy,
• Dynamical low-rank approach,
• Dynamical Orthogonal approximation ...
Low-rank approximation methods
u ≈r∑i=1
vi ⊗ αi, with r << d small.
M. Billaud-Friess — Workshop MOMAS 2015 Motivations 4/28
Problem setting
Parameter-dependent dynamical system
Find u : ξ × [0, T ]→ X such thatu′(t, ξ) = f(u(t, ξ), t, ξ) t ∈ (0, T ),u(0, ξ) = u0(ξ).
(1)
• ξ ∈ Ξ ⊂ Rp a (random) parameter vector ,
• X = Rd equiped with the L2 scalar product 〈·, ·〉X ,
• the state u is differentiable,
• u0 : ξ → X is the initial datum,
• f : X × [0, T ]×Ξ→ X the flux that is Lipschitz continuous with respect to u.
M. Billaud-Friess — Workshop MOMAS 2015 Low-rank approximation methods 5/28
Low-rank approximation methods : a sub-space point of view
À Ideal rank-r approximation for parametric dynamical system (1)
minv(t)∈Mr
‖v(t)− u(t, ·)‖Ξ,q, t > 0
with Mr =v =
∑ri=1 vi ⊗ αi,vi ∈ X,αi ∈ Lpµ(Ξ)
and r ∈ N such that r d.
Á Subspace point of view
minXr ⊂ X
dimXr = r
minv(t,·)∈L
pµ⊗Xr
‖v(t, ·)− u(t, ·)‖Ξ,p, t > 0
u unknown a priori Practical algorithms for computing sub-optimal approx. ur
L2-norm : Dynamical low-rank approximation [Koch07]
L∞-norm : Reduced basis approach
1 Greedy construction of Xr ⊂ X using “snapshots” of the solution
2 Low-rank approximation ur ∈ Xr by (Petrov-)Galerkin projection of (1) in Xr
M. Billaud-Friess — Workshop MOMAS 2015 Low-rank approximation methods 6/28
Low-rank approximation with RB methods I
Low-rank approximation with time dependent reduced space Xr(t)
u(t, ξ) ≈ ur(t, ξ) =
r∑i=1
vi(t)αi(t, ξ)
Method 1 : Weak-greedy algorithm
For each t > 0, while maxξ∈Ξtrain
∆r(t, ξ) > ε or r < rmax do.
1 Find ξr+1(t) = maxξ∈Ξtrain
∆r(t, ξ).
2 Compute u(t, ξr+1(t)).
3 Set Xr+1(t) = Xr(t) + spanu(t, ξr+1(t)).
3 ∆r surrogate of the exact projection error s.t.
cr∆r(t, ·) ≤ ‖u(t, ·)− P r(t)u(t, ·)‖X ≤ Cr∆r(t, ·), cr, CR > 0
3 At each instant t, the weak-greedy algorithm converges [Bin11, Buff12]
7 Costly strategy since ξr(t) is time dependent Many calls to (1).
M. Billaud-Friess — Workshop MOMAS 2015 Low-rank approximation methods 7/28
Low-rank approximation with RB methods II
Low-rank approximation with time independent reduced space Xr
u(t, ξ) ≈ ur(t, ξ) =
r∑i=1
viαi(t, ξ)
Strategies : Snapshot method, POD based method, POD-greedy method
Method 2 : POD-greedy algorithm [Haas11]
While maxξ∈Ξtrain
‖∆r(·, ξ)‖(0,T ),2 > ε or r < rmax do.
1 Find ξr+1 = maxξ∈Ξtrain
‖∆r(·, ξ)‖(0,T ),2.
2 Compute t 7→ u(t, ξr+1) and S(·, ξr) : t 7→ u(t, ξr+1)− P ru(t, ξr+1).
3 Set X(r+1)` = Xr` + spanPOD(S(·, ξr+1), `
).
3 Global error estimate ‖∆r(·, ξ)‖(0,T ),2
Time independent ξr with at most r calls to (1) Parameter ξr selected for given objective (e.g. GO norms)
3 Quasi-optimal convergence rates for Algorithm 2 [Haas13]7 The approximation ur is not an interpolation of u at points ξ1, . . . , ξr7 Large r when u has a rich spectral content (e.g. u solution of the advection equation)
M. Billaud-Friess — Workshop MOMAS 2015 Low-rank approximation methods 8/28
Proposed approach
Goal : Design a time-dependent low-rank approximation method for solvingparametric dynamical systems
u(t, ξ) ≈ ur(t, ξ) =r∑i=1
vi(t)αi(t, ξ)
Specifications : A method combining the best features of method 1 & 2
1 Time dependent reduced space Xr(t) ⊂ X by snapshots of the solutionselected with a greedy algorithm
2 Low-rank approximation ur(t) by solving the reduced dynamical systemobtained by (Petrov-)Galerkin projection of (1) on Xr(t)
3 Rigorous a-posteriori error estimation ‖ur(t, ·)− u(t, ·)‖X ≤ ∆r(t, ·)
M. Billaud-Friess — Workshop MOMAS 2015 Proposed method 9/28
Constructing the reduced space
Method 3 : T-greedy algorithm
While maxξ∈Ξtrain
‖∆r(·, ξ)‖(0,T ),2 > ε or r < rmax do.
1 Find ξr+1 = arg maxξ∈Ξtrain
‖∆r(t, ξ)‖(0,T ),2.
2 Compute t 7→ u(t, ξr+1) and S(·, ξr) : t 7→ u(t, ξr+1)− P ru(t, ξr+1).
3 For each time t > 0 proceed as follows.
• Set vr+1(t) =Sr+1(t)
‖Sr+1(t)‖X.
• Construct Xr+1(t) = Xr(t) + spanvr+1(t).
Proposition
Assuming that dimXr(t) = r for t > 0, then
1 the applications vi : t 7→ vi(t), i = 1, . . . , r exists and are differentiable,
2 the resulting family of functions vi(t)ri=1 is orthonormal for each t > 0.
M. Billaud-Friess — Workshop MOMAS 2015 Proposed method 10/28
Projection strategy
Notations : Here ξ ∈ Ξ is fixed
• V r(t) ∈ Rd×r = [v1(t), . . . ,vr(t)] s.t. V Tr (t)V r(t) = Ir, P r(t) = V r(t)V
Tr (t),
• αr ∈ Rr,ur ∈ Rd s.t. ur(t, ξ) = V r(t)αr(t, ξ)⇔ αr(t, ξ) = V r(t)Tur(t, ξ).
¬ The Galerkin projection
V r(t)Tu′r(t, ξ) = V r(t)
Tf(ur(t), ξ, t)
The low-rank approximation ur
u′r(t, ξ) = P r(t)f(ur(t, ξ), ξ, t) + (Id − P r(t))V′r(t)αr(t, ξ) ur(0) = P r(0)u0
Remark : Dependence in time of Xr(t) taken into account by means of V ′r(t)
® The reduced dynamical system
α′r(t, ξ) = fr(αr(t, ξ), ξ, t) α0r = V T
r (0)u0 a
with fr(αr(t, ξ), ξ, t) = V r(t)Tf(ur(t, ξ), ξ, t)− V r(t)
TV ′r(t)αr(t, ξ)
a. For parameter independent initial condition α0r = (1, 0, . . . , 0).
M. Billaud-Friess — Workshop MOMAS 2015 Proposed method 11/28
Error analysisLinear case
Summary : For f(u(t), t) = Au(t), the approximation error e(t) = ur(t)− u(t) fort > 0 is
e′(t) = Ae(t)︸ ︷︷ ︸flux error
+Qr(t)Aur(t)︸ ︷︷ ︸flux projection
−Qr(t)V′r(t)αr(t)︸ ︷︷ ︸
basis dynamics
e(0) = Qr(0)u0.
where Qr(t) = P r(t)− Id ∈ Rd×d
Proposition
The approximation error e satisfies ‖e(t)‖X ≤ ∆r(t),∀t ∈ [0, T ], where
∆r(t) :=
[sups∈[0,t]
‖r(s)‖X1− e−‖A‖X t
‖A‖X+ ‖e(0)‖X
]et‖A‖X
with ‖A‖X = λmax(AAT ) and r(s) = Qr(t)(Aur(t)− V ′r(t)αr(t))
Remark :
• Error estimate valid for time independent reduced space [Haas11]
• Coarse error estimate valid only for linear cases
M. Billaud-Friess — Workshop MOMAS 2015 Proposed method 12/28
Error analysisGeneral case
Summary : The approximation error e(t) = ur(t)− u(t) for t > 0 ise′(t) = (f(ur(t), t)− f(u(t), t))︸ ︷︷ ︸
flux error
+Qr(t)f(ur(t), t)︸ ︷︷ ︸flux projection
−Qr(t)V′r(t)αr(t)︸ ︷︷ ︸
basis dynamics
e(0) = Qr(0)u0.
Proposition (general case)
The approximation error e satisfies ‖e(t)‖X ≤ ∆r(t), ∀t ∈ [0, T ], where
∆r(t) :=∫ t
0γ(s)e
∫ ts β(τ)dτds + ‖e(0)‖Xe
∫ t0 β(τ)dτds,
γ(t) = ‖Qr(t)(f(ur(t), t)− V ′r(t)αr(t))‖X ,
β(t) = LX [f ](ur(t)).
In practice : A first order linearization of f is used [Wir14]
LX [f ](yr(t)) ≤ LX [J(ur(t))] + o(‖e‖X)
with J(ur) ∈ Rd×d the Jacobian matrix of f at ur.
M. Billaud-Friess — Workshop MOMAS 2015 Proposed method 13/28
Test case 1
Deterministic dvection equation with periodic boundary conditions
∂
∂tu(x, t) + a
∂
∂xu(x, t) = 0, u0(x) =
1√2πe− (x+0.06)2
0.052
• Spatial domain Ω = (−1, 1) and time interval I = (0, 1).
• The advection velocity is a = 1.325.
Discretization :
1 Finite difference upwind scheme on a regular mesh Ω = xid+1i=0 where
xi = −1 + iδx and δx = 2d+1
with d = 200 nodes.
d
dtu(t) = Au(t), u(0) = (u0)di=1.
2 Time integration with explicit Euler scheme on uniform discretizationI = tkKk=1, with tk = kδt and δt = δx/2
uk+1 = (Id + δtA)uk.
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 14/28
Deterministic problem
Comparison : Confrontation of two projection methods with
1 fixed reduced space Xr = spanPOD`(u0, . . . ,uK)2 time dependent reduced space Xk
r = spanuk
Method n1
dim(Xr) = ` E2 E∞
1 0.99862 15 0.99592 110 0.95843 0.4123220 0.070717 0.01070550 1.6532e-10 2.9288e-11100 4.9305e-13 4.2383e-14
Tab.1 : Relative error
Ep = ‖ur − u‖(0,T ),p/‖u‖(0,T ),p
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−1
0
1
2
3
4
5
x
uur for ` = 1ur for ` = 5ur for ` = 10ur for ` = 20ur for ` = 50ur for ` = 100
Fig.1 : Approximation compared tothe exact solution at T = 1
1 For large r, fine precision limits due to POD
2 With r = 1, E2 = 1.9235e-15, E∞ =3.9074 e-15 interpolant method
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 15/28
Test case 2
Advection-diffusion eq. with homogeneous Dirichlet boundary conditions
∂
∂tu(x, t, ξ) + a(ξ)
∂
∂xu(x, t, ξ)− µ(ξ)
∂
∂x2u(x, t, ξ) = 0, u0(x) = x(2− x)2e2x.
• Spatial domain Ω = (0, 2) and time interval I = (0, 1),
• µ(ξ) = 0.5(2 + cos(πξ)2) and a(ξ) = 0.1 sin(πξ) with ξ ∼ N (0, 1).
Discretization :
1 Finite difference scheme on a regular mesh with d = 200 nodes
d
dtu(t) = A(ξ)u(ξ, t), u(0) = u0
2 Time integration with implicit Euler scheme with K = 200
(IX − δtA(ξ))uk+1(ξ) = uk(ξ)
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 16/28
First results
Results for T-greedy algorithm r = 10
ξ = 0 ξ = 0.2 ξ = −0.5
0 2 · 10−24 · 10−26 · 10−28 · 10−2 0.10
0.5
1
1.5
2
t
x
0 2 · 10−24 · 10−26 · 10−28 · 10−2 0.10
0.5
1
1.5
2
tx
0 2 · 10−24 · 10−26 · 10−28 · 10−2 0.10
0.5
1
1.5
2
t
x
ξ = 0 ξ = 0.2 ξ = −0.5
0 0.5 1 1.5 20
2
4
6
8
x
uur
0 0.5 1 1.5 20
2
4
6
8
x
uur
0 0.5 1 1.5 20
2
4
6
8
x
uur
Fig.3 : Surface contours of ur (top) and comparison to the exact solution at T = 1 (bottom)
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 17/28
T-greedy vs. POD-greedyGreedy algorithm stopped for r = rmax
Data : Use of ∆NL, with #Ξtrain = 30 and ` = 1 for POD-greedy
0 5 10 15 2010−14
10−11
10−8
10−5
10−2
rank r
sup E2
sup E∞mean E2
mean E∞
0 5 10 15 20
10−7
10−5
10−3
10−1
rank r
sup E2
sup E∞mean E2
mean E∞
Fig. 4 : Relative error with respect to r : for T-greedy (left) and POD-greedy (right) methods
• Methods provide approximation with good keeping r small.
• POD-greedy converges more slowly than T-greedy.
• Coarser approximation for POD-greedy that requires richer reduced space.
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 18/28
Results for r = 10 and ξ = 0.2
0 50 100 150 20010−11
10−10
10−9
10−8
10−7
10−6
10−5
10−4
time index k
‖e‖X‖∆NL‖X‖∆L‖X
0 50 100 150 200
2
4
6
8
10
time index k
κNLκL
0 50 100 150 20010−4
10−3
10−2
10−1
100
time index k
‖e‖X‖∆NL‖X‖∆L‖X
0 50 100 150 200100
101
102
103
time index k
κNLκL
Fig. 5 : T-greedy (top) and POD-greedy (above)algorithm
Denoting the effectivity index as
κL,NL = ‖∆L,NL‖X/‖e‖X
¬ For T-greedy algorithm : efficienterror estimates for with ∆NL thesharpest.
maxt‖κL(k)‖X max
k‖κNL(k)‖X
9.9393 3.72878
For the POD-greedy algorithm :coarse error estimates with ∆NL
the sharpest.
maxk‖κL(k)‖X max
k‖κNL(k)‖X
572.101 47.0514
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 19/28
T-greedy vs. POD-greedyGreedy alg. with stopping criterion in precision ε
Data : Use of ∆NL, with #Ξtrain = 30 and ` = 1 for POD-greedy
Strong greedy algorithmε T-G POD-G
1e-1 5 71e-2 6 111e-3 8 151e-4 9 191e-5 11 231e-6 12 27
weak-greedy algorithmε T-G POD-G
1e-1 6 121e-2 7 171e-3 8 211e-4 11 251e-5 12 -1e-6 14 -
Tab. 3 : Reduced space dimension with respect to ε with exact error criterion (left), witherror estimate criterion (right)
• Small reduced spaces with T-greedy.
• Influence of the error estimator especially for the POD-greedy algorithm
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 20/28
Test case 3
Burgers’s equation with homogeneous Dirichlet boundary conditions
∂
∂tu(x, t, ξ)− µ(ξ)
∂2
∂x2u(x, t, ξ) + u(t, x, ξ)
∂
∂xu(x, t, ξ) = b(x)Ts(t, ξ), u(x, 0) = 0,
• Domain Ω = (0, 1) and time interval I = (0, 1)
• µ(ξ) = ξ with ξ ∼ U(0.01, 0.06).
• Source terme s(t) = (s1(t), s2(t)), b(x) = (b1(x), b2(x)) with
s1(t) = sin(4πt) s2(t) = 1[0.2,0.4](t),
b1(x) = 4e−( x−0.20.03
)21[0.1,0.3](x), b2(x) = 4 · 1[0.6,0.7](x)
Discretization : Finite difference scheme on a regular mesh with d = 200 nodes andtime integration with semi-implicit Euler scheme on uniform discretization of I withK = 300
(IX − dtA(ξ))uk+1(ξ) = uk(ξ) + dt[f(uk) +Bs(tk)]
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 21/28
First results
Results for T-greedy algorithm r = 10
ξ = 0.01 ξ = 0.035 ξ = 0.06
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
t
x
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
tx
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
t
x
ξ = 0.01 ξ = 0.035 ξ = 0.06
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
uur
0 0.2 0.4 0.6 0.8 1
0
5 · 10−2
0.1
0.15
uur
0 0.2 0.4 0.6 0.8 1
0
5 · 10−2
0.1
uur
Fig.7 : Surface contours of ur (top) and comparison to the exact solution at T = 1 (bottom)
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 22/28
Study of the T-greedy algorithmGreedy algorithm stopped for r = rmax
Data : #Ξtrain = 30 and use of ∆NL
Results for T-greedy algorithm r = 10
0 2 4 6 8 10 12 1410−12
10−9
10−6
10−3
100
rank r
sup E2sup E∞mean E2mean E∞
0 100 200 30010−14
10−11
10−8
10−5
10−2
time index k
ξ = 0.01ξ = 0.035ξ = 0.06
0 100 200 3000
50
100
time index k
κN
L
ξ = 0.01ξ = 0.035ξ = 0.06
Fig. 8 Relative error with respect to r Fig.9 : Error est. (left) and effectivity index (right)
• Good convergence with rank
• Error estimate more pessimistic forthis problem
• Influence of ξ on the error estimate
ξ maxt‖κNL(t)‖X
0.01 110.0640.035 26.85780.06 25.7706
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 23/28
Study of the T-greedy algorithmGreedy alg. with stopping criterion in precision ε
Strong greedy alg. weak-greedy alg.ε ξ = 0.01 ξ = 0.035 ξ = 0.06 ξ = 0.01 ξ = 0.035 ξ = 0.06
1e-1 4 4 4 7 7 61e-2 5 5 5 8 7 61e-3 7 8 7 9 9 81e-4 8 9 8 11 10 91e-5 10 11 9 12 12 101e-6 10 12 10 13 11 11
Tab. 6 : Reduced space dimension with respect to ε
• Richer reduced spaces needed (e.g. for ξ = 0.01)
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 24/28
Test case 4
Multi-parameter convection-reaction-diffusion with dirichlet boundary conditions
∂
∂tu(x, t, ξ)−µ(ξ)4u(x, t, ξ)+c(x, ξ)·∇u(x, t, ξ)+κ(ξ)u(x, t, ξ) = b(x), u(x, 0) = 0
• Domain Ω = (0, 1)2 and time interval I = (0, 0.03)
• µ(ξ) = 1 + 0.2ξ1, c(x, ξ) = 250(1 + 0.2ξ2)(x2− 12, 1
2−x1)T , κ(ξ) = 10(1 + 0.2ξ3)
ξ = (ξ1, . . . , ξ3) where ξi are i.i.d. ξi ∼ U(−1, 1).
• Source term b(x) = 1(0.7,0.8)2(x).
Discretization :
• P1 finite element approximation with d = 2212 nodes
M d
dtu(t, ξ) = A(ξ)u(t, ξ) + b⇒ d
dtu(t, ξ) = A(ξ)u(t, ξ) + b,
with M = LT L and Lu = u.
• Semi-implicit Euler scheme with K = 100
(Id − dtA(ξ))uk+1(ξ) = uk(ξ) + dt b
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 25/28
First results, r = 10 and ξ = (−0.2244, 0.6744, 0.5326)
Fig.10 : ur for t = T2
(left) and t = T (right)
Fig.11 : Absolute error u− ur
M. Billaud-Friess — Workshop MOMAS 2015 Numerical experiments 26/28
Conclusions & perspectives
T-greedy method : A low-rank approximation method, with sub-space point of view,for solving parameter dependent dynamical systems that can be interpreted as a RBmethod involving time-dependent reduced space.
3 Practical method with efficient error estimates, validated for both linear andnon-linear cases
3 An alternative to POD-greedy method
Improvement and further explorations
• Efficient treatment of non-linearity : combining EIM and T-greedy strategies
• Analysis of the T-greedy algorithm
Thank you for your attention
Any questions ?
M. Billaud-Friess — Workshop MOMAS 2015 Conclusions 27/28
Conclusions & perspectives
T-greedy method : A low-rank approximation method, with sub-space point of view,for solving parameter dependent dynamical systems that can be interpreted as a RBmethod involving time-dependent reduced space.
3 Practical method with efficient error estimates, validated for both linear andnon-linear cases
3 An alternative to POD-greedy method
Improvement and further explorations
• Efficient treatment of non-linearity : combining EIM and T-greedy strategies
• Analysis of the T-greedy algorithm
Thank you for your attention
Any questions ?
M. Billaud-Friess — Workshop MOMAS 2015 Conclusions 27/28
Bibliography
[Bin11] P. Binev, A. Cohen, W. Dahmen, R. De Vore, G. Petrova, P. WojtaszczykConvergence Rates for Greedy Algorithms in Reduced Basis Algorithms. SIAMJournal on Mathematical Analysis, 2011.
[Buff12] A. Buffa, Y. Maday, A. Patera, C. Prud’homme, G. Turicini A priori convergence ofthe greedy algorithm for the parametrized reduced basis. ESAIM-Math. Model.Numer. Anal., 2012
[Haas11] B. Haasdonk, M. Ohlberger, M. Efficient reduced models and a posteriori errorestimation for parametrized dynamical systems by offline/online decomposition.Mathematical and Computer Modelling of Dynamical Systems, 2011.
[Haas13] B. Haasdonk Convergence Rates of the POD-Greedy method. M2AN Math.Model. Numer. Anal., 2013.
[Hom07] C. Homescu, L.R. Petzold, R. Serban Error Estimation for Reduced-Order Modelsof Dynamical Systems. j-SIAM-REVIEW 2007.
[Koch07] O. Koch, C. Lubich Dynamical Low-Rank Approximation. SIAM Journal onMatrix Analysis and Applications, 2007.
[Sap09] T.P. Sapsis, P.F.J. Lermusiaux Dynamically orthogonal field equations forcontinuous stochastic dynamical systems. Physica D : Nonlinear Phenomena, 2009.
[Wir14] D. Wirtz, D.C. Sorensen, B. Haasdonk A-posteriori error estimation for DEIMreduced nonlinear dynamical systems. SIAM J. Sci. Comput., 2014.
M. Billaud-Friess — Workshop MOMAS 2015 Conclusions 28/28