Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
(hypo)coercive schemes for the Fokker-Planck equation F. Hérau (LMJL Nantes) with G. Dujardin (Inria Lille Europe) and P. Lafitte (Centrale Supélec) Workshop on kinetic and fluid PDEs Paris Descartes and Paris Diderot
Paris - 6 mars 2018
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Table of contents
1
Introduction
2
Continuous inhomogeneous case
3
Semi-discrete homogeneous case
4
Poincaré inequalities
5
All cases
6
Numerics
Poincaré inequalities
All cases
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Introduction We are interested in the study of the following inhomogeneous Fokker-Planck equation ∂t F + v ∂x F − ∂v (∂v + v )F = 0,
F |t=0 = F 0 ,
where 0 ≤ F = F (t, x , v ),
(t, x , v ) ∈ R+ × T × R
ZZ
Fdxdv = 1.
We focus in this talk on the case d = 1. The now rather standard hypocoercive methods give that F (t, x , v ) −→ M(x , v ), t→+∞
exponentially fast (for a large family of similar kinetic equations), where here the Maxwellian is given by 2 1 M(x , v ) = µ(v ) = √ e−v /2 . 2π
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
A much simpler equation (homogeneous kinetic equation) is ∂t F − ∂v (∂v + v )F = 0,
F |t=0 = F 0 ,
where 0 ≤ F = F (t, v ),
(t, v ) ∈ R+ × R,
Z
F dv = 1,
for which this is very easy to get (by "coercive" methods) that F (t, x , v ) −→ µ(v ). t→+∞
(This is just the heat equation for the harmonic oscillator.)
(1)
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Functional framework and proof for the homogeneous problem : set F = µ + µf , the equation is ∂t f + (−∂v + v )∂v f = 0 with f |t=0 = f 0 , consider f ∈ L2 (dµ) ⊂ L1 (dµ) (strictly smaller), R def R note that hf i = f dµ = f0 dµ = 0,
note that f ∈ L1 (dµ) ⇔ F ∈ L1 (dv ), compute 2 2 d dt kf kL2 (dµ) = −2 h(−∂v + v )∂v f , f iL2 (dµ) = −2 k∂v f kL2 (dµ) , 2
2
use Poincaré inequality kf kL2 (dµ) ≤ k∂v f kL2 (dµ) , so that d 2 2 kf kL2 (dµ) ≤ −2 kf kL2 (dµ) , dt
use Gronwall inequality kf kL2 (dµ) ≤ e−t f 0 L2 (dµ) ,
synthesis kF − Mk 1 ≤ kf k 2 ≤ e−t f 0 2 L (dv )
L (dµ)
L (dµ)
.
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
Many ingredients were involved in the short previous proof : Hilbertian framework, coercivity, Poincaré inequality, Gronwall lemma, existence of a Maxwellian ...
Aim of this talk : Explain how to adapt to the inhomogeneous case ✄ well understood and robust theory Explain how to discretize and numerically implement the problems ✄ new even in the homogeneous case
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
The continuous inhomogeneous case Perform the same change of variables F = µ + µf , and work in H 1 (dµdx ) ֒→ L2 (dµdx). The inhomogeneous equation reads ∂t f + v ∂x f + (−∂v + v )∂v f = 0, f |t=0 = f 0 , ZZ
def hf i = f dµdx = f 0
very partial biblio (Guo, Villani-Desvillettes ... H. 06-07, Mouhot-Neumann 06, H.-Nier 04, Villani 07, Dolbeault-Mouhot-Schmeiser 15, etc...) robust proof (Boltzmann, Landau, fractional etc...) and methods (hypocoercivity, hypoellipticity) wide applications (hydro limits, perturbative NL solutions, VPFP, Landau damping, enlargement theory, statistical mechanics, low temperature, UQ etc
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
The continuous inhomogeneous case Perform the same change of variables F = µ + µf , and work in H 1 (dµdx ) ֒→ L2 (dµdx). The inhomogeneous equation reads ∂t f + v ∂x f + (−∂v + v )∂v f = 0, f |t=0 = f 0 , ZZ
def hf i = f dµdx = f 0 commutator identity [∂v , v ∂x ] = ∂x (hypoellipticity results by Hörmander, Kohn, developped by Helffer, Nourrigat, mention subunit balls by Fefferman and subelliptic geometry etc... ).
how to discretize such an equality and equation ? fundamental point : have the simplest proofs and techniques in order to adapt them to the discretized cases.
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
The modified entropy
We define the entropy functional for C > D > E > 1, to be defined later on H : f 7→ C kf k2 + D k∂v f k2 + E h∂v f , ∂x f i + k∂x f k2 .
(2)
Then for C, D, E well chosen, we will prove that t 7→ H(f (t)) is nonincreasing when f solves the rescaled equation with initial datum f 0 ∈ H 1 (dµ). First note that if E 2 < D, H is equivalent to the H 1 (dµdx)-norm : 1 kf k2H 1 ≤ H(f ) ≤ 2C kf k2H 1 2 We have modified the norm in H 1 .
(3)
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
✄ First term d kf k2 = 2 h∂t f , f i = −2 hv ∂x f , f i − 2 h(−∂v + v )∂v f , f i = −2 k∂v f k2 dt ✄ Second term d 2 k∂v f k = 2 h∂v (∂t f ), ∂v f i dt = −2 h∂v (v ∂x f + (−∂v + v )∂v f ), ∂v f i
= −2 hv ∂x ∂v f , ∂v f i − 2 h[∂v , v ∂x ] f , ∂v f i − 2 h∂v (−∂v + v )∂v f , ∂v f i . = −2 h∂x f , ∂v f i − 2 k(−∂v + v )∂v f k
✄ Last term
2
d 2 2 k∂x f k = −2 k∂v ∂x f k dt
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
✄ Third important term d h∂x f , ∂v f i dt = − h∂x (v ∂x f + (−∂v + v )∂v f ), ∂v f i − h∂x f , ∂v (v ∂x f + (−∂v + v )∂v f )i = − hv ∂x (∂x f ), ∂v f i − h(−∂v + v )∂v f , ∂x ∂v f i − h∂x f , [∂v , v ∂x ] f i − h∂x f , v ∂x ∂v f i − h∂x f , [∂v , (−∂v + v )] ∂v f i − h(−∂v + v )∂v f , ∂x ∂v f i .
we have hv ∂x ∂x f , ∂v f i + h∂x f , v ∂x ∂v f i = 0. and [∂v , (−∂v + v )] = 1 so that d 2 h∂x f , ∂v f i = − k∂x f k + 2 h(−∂v + v )∂v f , ∂x ∂v f i − h∂x f , ∂v f i . dt
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
✄ Entropy dissipation inequality d 2 2 2 2 H(f ) = −2C k∂v f k − 2D k(−∂v + v )∂v f k − E k∂x f k − 2 k∂x ∂v f k dt − 2(D + E) h∂x f , ∂v f i − 2E h(−∂v + v )∂v f , ∂x ∂v f i . Therefore, using Cauchy-Schwartz : for 1 < E < D < C well chosen, E d H(f ) ≤ −C k∂v f k2 − (E − 1/2) k∂x f k2 ≤ − (k∂v f k2 + k∂x f k2 ). dt 2 Using the Poincaré inequality in space and velocity E E E cp d 2 2 2 H(f ) ≤ − (k∂v f k + k∂x f k ) − cp kf k ≤ − H(f ). dt 4 4 4 2C ✷
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
✄ Synthesis We pose 2κ =
E cp and we get by Gronwall lemma 4 2C
Theorem For all f 0 ∈ H 1 , the solution f to the rescaled inhomogeneous Fokker-Planck equation satisifies for all t ≥ 0,
2 1 1 2 2 kf kL2 ≤ kf kH 1 ≤ H(f (t)) ≤ e−2κt H(0) ≤ 2Ce−2κt f 0 H 1 . 2 2
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
We want to discretize the equation in space, velocity and time, with preservation of the long time behavior, (hypo)coercivity, the notion of Maxwellian. Keywords and the discrete case Equation ? derivative ? Hilbert space ? Maxwellian ? Gronwall ? Poincaré ? commutators ? local ? Huge literature : (Herda, Bessemoulin, AP, Jin, etc etc...). In the spirit of hypocoercivity for short time : Poretta-Zuazua
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
The semi-discrete homogeneous case
We want to discretize (and implement) the equation ∂t F − ∂v (∂v + v )F = 0. We look for a discretization only in velocity. ✄ The velocity derivative : for F ∈ ℓ1 (Z), define Dv F ∈ ℓ1 (Z∗ ) by (Dv F )i =
Fi − Fi−1 for i > 0, h
(Dv F )i =
Fi+1 − Fi for i < 0. h
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
✄ The Maxwellian : solving equation (Dv + v )µh = 0 yields µhi = Q|i|
ch
l=0 (1
+ hvi )
,
i ∈ Z.
Then µh is even, positive, in ℓ1 . Proof by direct computation : (Dv + v )µh = 0 writes h µ − µhi−1 i + vi µhi = 0 h h h µi+1 − µi + v µh = 0 i i h
which gives the expression of µh ∈ ℓ1 .
for i > 0 for i < 0,
All cases
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
✄ The "adjoint" : for G ∈ ℓ1 (Z∗ ), define Dv♯ F ∈ ℓ1 (Z) (Dv♯ G)i =
Gi+1 − Gi for i > 0, h (Dv♯ G)0 =
(Dv♯ G)i =
Gi − Gi−1 for i < 0, h
G1 − G−1 . h
✄ The Hilbert spaces : we pose F = µh + µh f and consider X f ∈ ℓ2 (µh ) ⇔ fi2 µhi < +∞ i
then denoting µ♯i = µhi−1 for i > 0 and µ♯i = µhi+1 for i < 0 −Dv♯ ((Dv + v )µh f ) = Dv♯ (µ♯ (Dv f )i ) = µhi (−Dv♯ + v )Dv f .
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
The discrete equation is then ∂t F − Dv♯ (Dv + v )F = 0,
F |t=0 = F 0 .
With F = µh + µh f , we have Proposition The equation satisfied by f is the following ∂t f + (−Dv♯ + v )Dv f = 0,
f |t=0 = f 0 .
The operator (−Dv♯ + v )Dv is selfadjoint non-negative in ℓ2 (µh ) : D E D E (−Dv♯ + v )Dv f , g = hDv f , Dv gi♯ = f , (−Dv♯ + v )Dv g , where
2
ϕ ∈ ℓ2 (µ♯ ) ⇔ kϕk♯ =
X i6=0
ϕ2i µ♯i < ∞.
Constant sequences are the equilibrium states of the equation.
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
We can do the same proof as in the continuous case P def P note that hf i = fi µhi = fi0 µhi = 0, D E 2 2 compute dtd kf k = −2 (−Dv♯ + v )Dv f , f = −2 kDv f k♯ , use Poincaré inequality kf k2 ≤ kDv f k2♯ ,
use Gronwall inequality kf k ≤ e−t f 0 ,
synthesis F − µh ℓ1 ≤ kf k ≤ e−t f 0 .
? ? Poincare inequality ? ?
All cases
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
We can do the same proof as in the continuous case P def P note that hf i = fi µhi = fi0 µhi = 0, D E 2 2 compute dtd kf k = −2 (−Dv♯ + v )Dv f , f = −2 kDv f k♯ , use Poincaré inequality kf k2 ≤ kDv f k2♯ ,
use Gronwall inequality kf k ≤ e−t f 0 ,
synthesis F − µh ℓ1 ≤ kf k ≤ e−t f 0 .
? ? Poincare inequality ? ?
All cases
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Poincaré inequality Lemma (adapted proof of that by H. Poincare (1912)) 2
2
For all f ∈ H 1 (µ) with hf i = 0, we have kf kL2 (dµ) ≤ k∂v f kL2 (dµ) Proof. Denote f (v ) = f , f (v ′ ) = f ′ , dµ = µ(v )dv and dµ′ = µ(v ′ )dv ′ .
Z
1 f dµ = 2 R 2
ZZ
1 (f − f ) dµdµ = 2 R2 ′
′
2
ZZ
R2
Z
v′
v
∂v f (w)dw
!2
From Cauchy Schwartz 1 f dµ ≤ 2 R
Z
2
ZZ
R2
Z
v
v′
2
|∂v f (w)| dw
!
(v ′ − v )dµdµ′
dµdµ′
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
Rv 2 Denote F (v ) = a |∂v f (w)| dw. Then ZZ Z 1 (F ′ − F ) (v ′ − v )dµdµ′ f 2 dµ ≤ 2 2 R R ZZ ZZ ZZ ZZ 1 ′ ′ F ′ v dµdµ′ Fv ′ dµdµ′ − Fv dµdµ′ − F v dµdµ′ + = 2 R2 R2 R2 R2 Z Fv dµ, = R
Note that ∂v µ = −v µ and perform an integration by parts Z Z Z Z Z 2 f dµ ≤ Fv µdv = − F ∂v µdv = ∂v F µdv = |∂v f |2 dµ. R
R
R
R
R
✷
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Discrete Poincaré inequality Proposition (Discrete Poincaré inequality) Let f be a sequence in H 1 . Then, 2
2
kf − hf ikℓ2 (µh ) ≤ kDv f kℓ2 (µ♯ ) . Proof. Assume hf i = 0 and write X X 1X fi2 µi = (fj − fi )2 µi µj = (fj − fi )2 µi µj 2 i
i,j
i0
i D > E > 1 well chosen. There exists k , t0 , h0 > 0 such that for all f 0 ∈ H 1 with f 0 = 0, all δt ∈ (0, t0 ), and all h ∈ (0, h0 ) the sequence defined by the implicit Euler scheme satisfies for all n ∈ N,
2 1 n 2 kf kH 1 ≤ H(f n ) ≤ H(f 0 )e−knδt ≤ 2C f 0 H 1 e−knδt . 2
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Numerics
Elements of proof Consider 2
2
2
H(f ) = C kf k + D kDv f k♯ + E hDv f , SDx f i♯ + kDx f k .
(4)
with S = [Dv , v ] and therefore SDx = [Dv , vDx ] : (Sg)i = gi−1 for i ≥ 1
(Sg)i = gi+1 for i ≤ −1.
We have for example Dv (−Dv♯ + v )S − S(−Dv♯ + v )Dv = S + δ, where δ is the singular operator from ℓ2 to ℓ2♯ defined for f ∈ ℓ2 by (δf )j = 0 if |j| ≥ 2,
(δf )−1 =
f1 − f0 , h2
(δf )1 = −
f0 − f−1 . h2
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Poincaré inequalities
All cases
Especially for the singular term involving δ, we have for all ε > 0,
2 1
2 ♯ hδDx f , Dv f i♯ ≤ (−Dv + v )Dv f + ε kDv Dx f k♯ . ε
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Numerics
Poincaré inequalities
All cases
Numerics
Introduction
Continuous inhomogeneous case
Semi-discrete homogeneous case
Thank you !
Poincaré inequalities
All cases
Numerics