Hugo de la Cruz
Escola de Matemática Aplicada (FGV-EMAp) Rio de Janeiro
joint work with: J. C. Jimenez, P.Zubelli
x
00(
t
) +
f
(
t
,
x
(
t
)
,
x
0
(
t
)) =
∑
mj=1
g
j(t
)
η
jtdx
(
t
) =
y
(
t
)
dt
dy
(
t
) =
f
(
t
,
x
(
t
)
,
y
(
t
))
dt
+
m∑
j=1g
j(t)dw
tjx2Rd,y=x0
dx
2=
y
2dy
1= (
α
1y
1(
1
x
12)
x
1+
β
1x
2)
dt
+
γ
1(t
)dw
1dy
2=
(
α
2y
2(
1
x
22)
x
2+
β
2x1
)
dt
+
γ
2(t
)dw
2IC:(x0
1,x20,y10,y20)
d
x
(
t
)
y
(
t
)
=
0
1
ω
2β
x
(
t
)
y
(
t
)
dt
+
0
ε
u
(
x
(
t
))
dt
+
d
x
(
t
)
y
(
t
)
=
0
1
ω
2β
x
(
t
)
y
(
t
)
dt
+
0
ε
u
(
x
(
t
))
dt
+
0
σ
dw
t,
K. Burrage, G. Lythe, Accuratestationary densitieswith partitioned numerical methods for stochastic di¤erential Equations, SIAM J. Num. Anal. 47 (3) (2009) 1601–1618.
D. Cohen, M. Sigg, Convergence analysis of trigonometric methodsforsti¤ second-order stochastic di¤erential equations, Num. Math., 121 (2012) 1-29.
A. H. Strømmen, D. J. Higham, Numerical simulation of alinear oscillatorwith additive noise, Appl. Numer. Math. 51 (2004) 89-99.
A. Tocino, On preserving long-time features of alinear stochastic oscillator, BIT Num. Math. 47 (2007) 189-196.
J. Hong, R. Scherer, L. Wang,Predictor-Corrector Methodsfor aLinear Stochastic Oscillatorwith Additive Noise, Mathematical and Computer Modelling. 46 (2007) 738-764.
D. Cohen, On thenumerical discretisationofstochastic oscillators, Math. and Comput in Simulation, 82 (2012) 1478-1495.
d
y(t) = Ω2 0 x(t)dt+ Σ dwt,
x(t) = x(t)
d
y(t) = Ω2 0 x(t)dt+ Σ dwt,
x(t) = x(t)
y(t) 2R2d, Ω2Rd d, Σ2Rd m
-This is a Hamiltonian system,
dx(t) = ∂H
∂ydt
dy(t) = ∂H ∂xdt+
m
∑
j=1∂Hj
∂x dw
j t
with
H(x,y) = 1 2 kΩxk
2+
kyk2
P1)growth rate of the energy: E x2(t) +ω2y2(t) =E(x2(t
0) +ω2y2(t0)) +σ2(t t0),
P2)oscillatory behavior around 0 : x(t) has in…nitely many zeros on [t0,∞),
P3)symplectic structure:dx(t)^dy(t) =dx(t0)^dy(t0), for allt t0,
-L. Markus. Stochastic oscillators, J. Di¤. Eq. 71 (1988)
P1)growth rate of the energy: E x2(t) +ω2y2(t) =E(x2(t
0) +ω2y2(t0)) +σ2(t t0),
P2)oscillatory behavior around 0 : x(t) has in…nitely many zeros on [t0,∞),
P3)symplectic structure:dx(t)^dy(t) =dx(t0)^dy(t0), for allt t0,
-L. Markus. Stochastic oscillators, J. Di¤. Eq. 71 (1988)
To study certain properties of the nonlinear oscillator "is enough" to study the linear one (with respect to a new brownian motion)
Previous works:
- The analysis focused on the simple (d=1) harmonic oscillator - No method fordamping coupled oscillator mimics these 3 properties
du(t) = 0d Id
fx(tn,xn,yn) fy(tn,xn,yn) (u(t) xn) +
0
ft (t tn) +
yn
du(t) = 0d Id
fx(tn,xn,yn) fy(tn,xn,yn) (u(t) xn) +
0
ft (t tn) +
yn
f(tn,xn,yn) dt
du(t) = 0d Id
fx(tn,xn,yn) fy(tn,xn,yn) (u(t) xn) +
0
ft (t tn) +
yn
f(tn,xn,yn) dt
u(t) =xn+LeCn(t tn)r
with
L = [I2d 02d 2],
r| = h0
(2d+1) 11
i
and
Cn=
2
6 6 4
0d Id 0d yn
fx(tn,xn,yn) fy(tn,xn,yn) ft(tn,xn,yn) f(tn,xn,yn)
01 d 01 d 0 1
0 0 0 0
3
7 7
x(t) =xn+LeCn(t tn)r+z(t)
dz(t) =q(t,z(t))dt+
m
∑
j=10
gj(t) dw j t,
x(t) =xn+LeCn(t tn)r+z(t)
dz(t) =q(t,z(t))dt+
m
∑
j=10
gj(t) dw j t,
z(tn) =0,
q(t,z) = a(t,xn+LeCn(t tn)r+z) ax(tn,xn)LeCn(t tn)r at(tn,xn) (t tn) a(tn,xn)
a(t,z) = [0d Id] z|
xn+1=xn+LeCnhr+zn+1
LL-Euler:
zn=
2
4
0
m
∑
j=1
gj(tn)∆wnj
3
5
LL-BE:
zn+1=q(tn+1,zn+1)h+
2 4 0 m ∑ j=1
gj(tn)∆wnj
3
5
LL-MR:
zn+1=q tn+1
+tn
2 ,
zn+1
2 h+
2 4 0 m ∑ j=1
gj(tn)∆wnj
3
xn+1=xn+Le r+zn+1
LL-PC(E,BE):
zn+1=q tn+1, "
0,
m
∑
j=1gj(tn)∆wnj
#! h+ 2 4 0 m ∑ j=1
gj(tn)∆wnj
3
5
LL-PC(E,MR):
zn+1=q tn+tn+1
2 ,
"
0,1 2
m
∑
j=1gj(tn)∆wnj
#! h+ 2 4 0 m ∑
j=1gj
(tn)∆wnj
3 5 LL-PE: 2 4 z1 n+1 3
5=q tn, z1n+1,0 h+ 2
4
0
m
∑g (t )∆wj 3
Theorem
For coupled harmonic oscillators
d x(t) y(t) =
0 I
Ω2 0 x(t)dt+ Σ0 dwt, with initial conditionx(t0) = (x0,y0), the Locally Linearized integrators satisfy
E kyn+1k2+kΩxn+1k2 =E ky0k2+kΩx0k2 + (kQ2k2+kΩQ1k2) (tn+1 t0),
Proof: It follows that
vn+1=Mvn+Q∆wn, where
vn+1= Ωyxn+1
n+1 , M=
cos(Ωh) sin(Ωh)
sin(Ωh) cos(Ωh) , and Q=
ΩQ1
Proof: It follows that
vn+1=Mvn+Q∆wn, where
vn+1= Ωyxn+1
n+1 , M=
cos(Ωh) sin(Ωh)
sin(Ωh) cos(Ωh) , and Q=
ΩQ1
Q2 .
FromE(Q∆wn) =0follows that
E kyn+1k2+kΩxn+1k2 =E Mvn 2+2 Mvn,B∆wn + Q∆wn 2
=E v|
nM
|
Mvn +E ∆w|nQ
|
Q∆wn
Theorem
Proof: We have that
xn+1=Mn+1x0+ n
∑
k=0MkQ∆wn k, where
Proof: We have that
xn+1=Mn+1x0+ n
∑
k=0MkQ∆wn k, where
Mk = cosΩ (kΩh) khsinc(kΩh) sin(kΩh) cos(kΩh) . LetQ=[q1, ...,qm], then
xn+1=cos((n+1)Ωh)x0+sin((n+1)Ωh)Ω 1y0+ n
∑
k=0m
∑
i=1Vki !
with
Vki= q1icos(kΩh) + (q2iΩ 1)sin(kΩh) ∆win k, and
Vki sN 0,σ2ik , σ2ik= qi1cos(kΩh) + (qi2Ω 1)sin(kΩh)
2
Proof: We have that
xn+1=Mn+1x0+ n
∑
k=0MkQ∆wn k, where
Mk = cosΩ (kΩh) khsinc(kΩh) sin(kΩh) cos(kΩh) . LetQ=[q1, ...,qm], then
xn+1=cos((n+1)Ωh)x0+sin((n+1)Ωh)Ω 1y0+ n
∑
k=0m
∑
i=1Vki !
with
Vki= q1icos(kΩh) + (q2iΩ 1)sin(kΩh) ∆win k, and
Vki sN 0,σ2ik , σ2ik= qi1cos(kΩh) + (qi2Ω 1)sin(kΩh)
2
h
What about
n
∑
k=0m
∑
i=1Vki
Proof (Cont): Let us compute
sn2= n
∑
k=0m
∑
i=1σ2ik
!
Proof (Cont): Let us compute
sn2= n
∑
k=0m
∑
i=1σ2ik
!
.
σ2ik can be rewritten as σ2ik = h qi
2
cos2 kΩh αi ,
with αi = arctan q2iΩ 1
qi 1
forqi
16=0, andαi=
π 2 forq
i
1=0, qi= q i 1
Proof (Cont): Let us compute
sn2= n
∑
k=0m
∑
i=1σ2ik
!
.
σ2ik can be rewritten as σ2ik = h qi
2
cos2 kΩh αi ,
with αi = arctan q2iΩ 1
qi 1
forqi
16=0, andαi=
π 2 forq
i
1=0, qi= q i 1
qi 2Ω 1
Then
s2 n = h
m
∑
i=1qi 2
∑
n k=0cos2 kΩh αi
!
= h
m
∑
qi2
n+sin(nΩh+Ωh)cos nΩh 2α
i !
Proof (Cont): ...Since limn!∞sn2=∞andσ2in h qi 2, then lim n!∞ σ2 in s2 n =0.
The Law of the Iterated Logarithm may be applied to the sequence
n
∑
k=0 m∑
i=1Vki. Then,
i) P 0
B B B @lim supn!∞
0 B B B @ n
∑
k=0m
∑
i=1Vki
p
2sn(log logsn) 1 2 1 C C C A=1
1
C C C A=1,
ii) P 0
B B B @lim infn!∞
0 B B B @ n
∑
k=0 m∑
i=1 Vi k p2sn(log logsn) 1 2 1 C C C
A= 1
1
Proof (Cont): ...Since limn!∞sn2=∞andσ2in h qi 2, then lim n!∞ σ2 in s2 n =0.
The Law of the Iterated Logarithm may be applied to the sequence
n
∑
k=0 m∑
i=1Vki. Then,
i) P 0
B B B @lim supn!∞
0 B B B @ n
∑
k=0m
∑
i=1Vki
p
2sn(log logsn) 1 2 1 C C C A=1
1
C C C A=1,
ii) P 0
B B B @lim infn!∞
0 B B B @ n
∑
k=0 m∑
i=1 Vi k p2sn(log logsn) 1 2 1 C C C
A= 1
1
C C C A=1.
Fromi), for 0<ε<1
Proof (Cont): Thus,
xn+1=cos((n+1)Ωh)x0+sin((n+1)Ωh)Ω 1y0+ n
∑
k=0m
∑
i=1Vki
!
Proof (Cont): Thus,
xn+1=cos((n+1)Ωh)x0+sin((n+1)Ωh)Ω 1y0+ n
∑
k=0m
∑
i=1Vki
!
>0 a.s, in…nitely often.
From
ii) P 0
B B B @lim infn!∞
0 B B B @ n
∑
k=0 m∑
i=1 Vi k p2sn(log logsn) 1 2 1 C C C
A= 1
1
C C C A=1,
for 0<ε<1
n
∑
k=0m
∑
i=1Vki <( 1+ε)
p
2sn(log logsn) 1
2 a.s, for in…nite values ofn,
which implies that
1
∑
n∑
m iTheorem
For the coupled harmonic oscillators with initial conditionx(t0) = (x0,y0)2R2d, the Locally
Proof: Let us consider the di¤erential 2-form
w2:=
d
∑
i=1dxni ^dyni,
where
dxni^dyni(ξ,η) =det ξ i 1 ηi1
ξi2 ηi2
, ξ= ξ1 ξ2 2R
2d
,η= η1 η2 2R
2d
In matrix notation,
w2(ξ,η) =ξ|Jη
, withJ= 0 I
Proof: Let us consider the di¤erential 2-form
w2:=
d
∑
i=1dxni ^dyni,
where
dxni^dyni(ξ,η) =det ξ i 1 ηi1
ξi2 ηi2
, ξ= ξ1 ξ2 2R
2d
,η= η1 η2 2R
2d
In matrix notation,
w2(ξ,η) =ξ|Jη
, withJ= 0 I
I 0 .
Letψ(t,t0)the ‡ow determined by the LL integrators, i.e.,[xnyn]|=ψ(tn,t0) [x0y0]|. All we need to prove is thatψ0=∂(xn,yn)
∂(x0,y0) satis…es
ψ0 | Jψ0=J.
Since the noise is additive, we have thatψ0=M.
d
x
(
t
)
y
(
t
)
=
0
1
ω
20
x
(
t
)
y
(
t
)
dt
+
2650 2660 2670 2680 2690 2700 2710 - 15 - 10 - 5 0 5 10 15
390 400 410 420 430 440 450 460 470 - 4 - 3 - 2 - 1 0 1 2 3 4 h2
750 760 770 780 790 800
- 6 - 4 - 2 0 2 4 6
180 190 200 210 220 230
- 5 0 5
-10 -5 0 5 10 15 -8
-6 -4 -2 0 2 4 6
Exact Sol. LL-P(EC)¹ P(EC)¹ T
1= 30
T
2= 70
T
Pendulum without damping perturbed by additive noise:Ex. (Gitterman05)
dX1(t) =X2(t)dt,
dX2(t) = sin X1(t) dt+σdW
t.
- 2 - 1 0 1 2 - 2 - 1 0 1 2 L L -PC (E, B E)
- 1 0 1 2
- 2 - 2 - 1 0 1 2
- 2 - 1 0 1 2
- 2 - 1 0 1 2
- 2 - 1 0 1 2
- 2 - 1 0 1 2 P C( E ,BE ) h3
- 2 - 1 0 1 2
- 2 - 1 0 1 2 2 h2
- 2 - 1 0 1 2
Ex. (Schurz09, Cohen12)
dx(t) =y(t)dt
dy(t) = (w2x(t) +x4(t))dt+σdw
Ex. (Schurz09, Cohen12)
dx(t) =y(t)dt
dy(t) = (w2x(t) +x4(t))dt+σdw
t
Expected value of the energy along the solution
E(1 2(y
2(t) +w2x2(t)) +ε 5x
5(t)) =1 2(y
2
0 50 100 150 0
10 20 30 40 50 60 70 80
E
ner
gy
Coupled Van der Pol oscillators driven by external random forces:(Artemiev et.al)
dx1=y1
dx2=y2 dy1= (1
2y1(1 x 2
1) x1+0.1x2)dt+1.5dw1 dy2=(5y2(1 x22) x2+0.1x1)dt+2.5dw2
σ1=1.5andσ2=2.5, initial condition(x0 1,x0
2,y0 1,y0
2) = (1,1,0,0).
- 2 - 1 0 1 2 3 - 10 - 5 0 5 10 T=17.5 L L P C( E ,M R)
- 3 - 2 - 1 0 1 2 3
- 15 - 10 - 5 0 5 10 T=19.375
- 3 - 2 - 1 0 1 2 3
- 15 - 10 - 5 0 5 10 15 T=100
- 2 - 1 0 1 2 3
- 8 - 6 - 4 - 2 0 2 4 6 8 P C( E ,M R)
- 20 - 15 - 10 - 5 0 5 - 50 0 50 100 150 200 250
- 15 - 10 - 5 0 5
x 10137 - 1 0 1 2 3 4 5 6 7x 10
- 2 - 1 0 1 2 3 - 8 - 6 - 4 - 2 0 2 4 6 8 10 T 1=17.5
- 3 - 2 - 1 0 1 2 3 - 15 - 10 - 5 0 5 10 T 2=19.37
- 3 - 2 - 1 0 1 2 3 - 15 - 10 - 5 0 5 10 15 T
A Scienti…c Canvas of Nonlinearity and Complex Dynamics Understanding Complex Systems, 2013, 539-557.
D. Cohen, On the numerical discretization of stochastic oscillators, Math. Comput. Simul., 82 (2012) 1478–1495.
H. de la Cruz, R.J. Biscay, J.C. Jimenez, F. Carbonell, HOLL methods: An approach for constructing A-stable explicit schemes for SDEs with additive noise, BIT, 50 (2010) 509-539.
M. Gitterman, The noisy oscillator, World Scienti…c, 2005.
J.C. Jimenez, H. de la Cruz, Convergence rate of strong local linearization schemes for SDEs with additive noise, BIT, 52 (2012) 357–382.
H. Kunita, Stochastic ‡ows and stochastic di¤erential equations. Cambridge University Press. 1990
L. Markus, A. Weerasinghe, Stochastic oscillators, J. Di¤erential Equations. 71 (2) (1988) 288-314.
G.N. Milstein , M.V. Tretyakov, Stochastic Numerics for Mathematical Physics, Springer, 2004.