Section: Research Program
Research Program
The MINGuS project is devoted to the mathematical and numerical analysis of models arising in plasma physics and nanotechnology. The main goal is to construct and analyze numerical methods for the approximation of PDEs containing multiscale phenomena. Specific multiscale numerical schemes will be proposed and analyzed in different regimes (namely highlyoscillatory and dissipative). The ultimate goal is to dissociate the physical parameters (generically denoted by $\epsilon $) from the numerical parameters (generically denoted by $h$) with a uniform accuracy. Such a task requires mathematical prerequisite of the PDEs.
Then, for a given stiff (highlyoscillatory or dissipative) PDE, the methodology of the MINGuS team will be the following

Mathematical study of the asymptotic behavior of multiscale models.
This part involves averaging and asymptotic analysis theory to derive asymptotic models, but also longtime behavior of the considered models.

Construction and analysis of multiscale numerical schemes.
This part is the core of the project and will be deeply inspired for the mathematical prerequisite. In particular, our ultimate goal is the design of Uniformly Accurate (UA) schemes, whose accuracy is independent of $\epsilon $.

Validation on physically relevant problems.
The last goal of the MINGuS project is to validate the new numerical methods, not only on toy problems, but also on realistic models arising in physics of plasmas and nanotechnologies. We will benefit from the Selalib software library which will help us to scaleup our new numerical methods to complex physics.
Dissipative problems
In the dissipative context, the asymptotic analysis is quite well understood in the deterministic case and multiscale numerical methods have been developed in the last decades. Indeed, the socalled AsymptoticPreserving schemes has retained a lot of attention all over the world, in particular in the context of collisional kinetic equations. But, there is still a lot of work to do if one is interesting in the derivation high order asymptotic models, which enable to capture the original solution for all time. Moreover, this analysis is still misunderstood when more complex systems are considered, involving non homogeneous relaxation rates or stochastic terms for instance. Following the methodology we aim at using, we first address the mathematical analysis before deriving multiscale efficient numerical methods.
A simple model of dissipative systems is governed by the following differential equations
$\left\{\begin{array}{cccc}{\displaystyle \frac{d{x}^{\epsilon}\left(t\right)}{dt}}\hfill & =\hfill & \mathcal{G}({x}^{\epsilon}\left(t\right),{y}^{\epsilon}\left(t\right)),\hfill & {\displaystyle \phantom{\rule{1.em}{0ex}}{x}^{\epsilon}\left(0\right)={x}_{0},}\hfill \\ \hfill \\ {\displaystyle \frac{d{y}^{\epsilon}\left(t\right)}{dt}}\hfill & =\hfill & {\displaystyle \frac{{y}^{\epsilon}\left(t\right)}{\epsilon}+\mathscr{H}({x}^{\epsilon}\left(t\right),{y}^{\epsilon}\left(t\right)),}\hfill & \phantom{\rule{1.em}{0ex}}{y}^{\epsilon}\left(0\right)={y}_{0},\hfill \end{array}\right.$  (3) 
for given initial condition $({x}_{0},{y}_{0})\in {\mathbb{R}}^{2}$ and given smooth functions $\mathcal{G},\mathscr{H}$ which possibly involve stochastic terms.
Asymptotic analysis of dissipative PDEs (F. Castella, P. Chartier, A. Debussche, E. Faou, M. Lemou)
Derivation of asymptotic problems
Our main goal is to analyze the asymptotic behavior of dissipative systems of the form (3) when $\epsilon $ goes to zero. The center manifold theorem [35] is of great interest but is largely unsatisfactory from the following points of view

a constructive approach of $h$ and ${x}_{0}^{\epsilon}$ is clearly important to identify the highorder asymptotic models: this would require expansions of the solution by means of Bseries or wordseries [37] allowing the derivation of error estimates between the original solution and the asymptotic one.

a better approximation of the transient phase is strongly required to capture the solution for small time: extending the tools developed in averaging theory, the main goal is to construct a suitable change of variable which enables to approximate the original solution for all time.
Obviously, even at the ODE level, a deep mathematical analysis has to be performed to understand the asymptotic behavior of the solution of (3). But, the same questions arise at the PDE level. Indeed, one certainly expects that dissipative terms occurring in collisional kinetic equations (2) may be treated theoretically along this perspective. The key new point indeed is to see the center manifold theorem as a change of variable in the space on unknowns, while the standard point of view leads to considering the center manifold as an asymptotic object.
Stochastic PDEs
We aim at analyzing the asymptotic behavior of stochastic collisional kinetic problems, that is equation of the type (2). The noise can describe creation or absorption (as in (2)), but it may also be a forcing term or a random magnetic field. In the parabolic scaling, one expect to obtain parabolic SPDEs at the limit. More precisely, we want to understand the fluid limits of kinetic equations in the presence of noise. The noise is smooth and non delta correlated. It contains also a small parameter and after rescaling converges formally to white noise. Thus, this adds another scale in the multiscale analysis. Following the pioneering work [38], some substantial progresses have been done in this topic.
More realistic problems may be addressed such as high field limit describing sprays, or even hydrodynamic limit. The full Boltzmann equation is a very long term project and we wish to address simpler problems such as convergences of BGK models to a stochastic Stokes equation.
The main difficulty is that when the noise acts as a forcing term, which is a physically relevant situation, the equilibria are affected by the noise and we face difficulties similar to that of high field limit problems. Also, a good theory of averaging lemma in the presence of noise is lacking. The methods we use are generalization of the perturbed test function method to the infinite dimensional setting. We work at the level of the generator of the infinite dimensional process and prove convergence in the sense of the martingale problems. A further step is to analyse the speed of convergence. This is a prerequisite if one wants to design efficient schemes. This requires more refined tools and a good understanding of the Kolmogorov equation.
Numerical schemes for dissipative problems (All members)
The design of numerical schemes able to reproduce the transition from the microscopic to macroscopic scales largely matured with the emergence of the Asymptotic Preserving schemes which have been developed initially for collisional kinetic equations (actually, for solving (2) when $\beta \to 0$). Several techniques have flourished in the last decades. As said before, AP schemes entail limitations which we aim at overcoming by deriving

AP numerical schemes whose numerical cost diminishes as $\beta \to 0$,

Uniformly accurate numerical schemes, whose accuracy is independent of $\beta $.
Time diminishing methods
The main goal consists in merging MonteCarlo techniques [33] with AP methods for handling automatically multiscale phenomena. As a result, we expect that the cost of the soobtained method decreases when the asymptotic regime is approached; indeed, in the collisional (i.e. dissipative) regime, the deviational part becomes negligible so that a very few number of particles will be generated to sample it. A work in this direction has been done by members of the team.
We propose to build up a method which permits to realize the transition from the microscopic to the macroscopic description without domain decomposition strategies which normally oblige to fix and tune an interface in the physical space and some threshold parameters. Since it will permit to go over domain decomposition and AP techniques, this approach is a very promising research direction in the numerical approximation of multiscale kinetic problems arising in physics and engineering.
Uniformly accurate methods
To overcome the accuracy reduction observed in AP schemes for intermediate regimes, we intend to construct and analyse multiscale numerical schemes for (3) whose error is uniform with respect to $\epsilon $. The construction of such a scheme requires a deep mathematical analysis as described above. Ideally one would like to develop schemes that preserve the center manifold (without computing the latter!) as well as schemes that resolve numerically the stiffness induced by the fast convergence to equilibrium (the socalled transient phase). First, our goal is to extend the strategy inspired by the central manifold theorem in the ODE case to the PDE context, in particular for collisional kinetic equations (2) when $\beta \to 0$. The design of Uniformly Accurate numerical schemes in this context would require to generalize twoscale techniques introduced in the framework of highlyoscillatory problems [36].
Multiscale numerical methods for stochastic PDEs
AP schemes have been developed recently for kinetic equations with noise in the context of Uncertainty Quantification UQ [41]. These two aspects (multiscale and UQ) are two domains which usually come within the competency of separate communities. UQ has drawn a lot of attention recently to control the propagation of data pollution; undoubtedly UQ has a lot of applications and one of our goals will be to study how sources of uncertainty are amplified or not by the multiscale character of the model. We also wish to go much further and by developing AP schemes when the noise is also rescaled and the limit is a white noise driven SPDE, as described in section (3.1.1.1). For simple nonlinear problem, this should not present much difficulties but new ideas will definitely be necessary for more complicated problems when noise deeply changes the asymptotic equation.
Highlyoscillatory problems
As a generic model for highlyoscillatory systems, we will consider the equation
$\frac{d{u}^{\epsilon}\left(t\right)}{dt}=\mathcal{F}(t/\epsilon ,{u}^{\epsilon}\left(t\right)),\phantom{\rule{1.em}{0ex}}{u}^{\epsilon}\left(0\right)={u}_{0},$  (4) 
for a given ${u}_{0}$ and a given periodic function $\mathcal{F}$ (of period $P$ w.r.t. its first variable) which possibly involves stochastic terms. Solution ${u}^{\epsilon}$ exhibits highoscillations in time superimposed to a slow dynamics. Asymptotic techniques resorting in the present context to averaging theory [45] allow to decompose
${u}^{\epsilon}\left(t\right)={\Phi}_{t/\epsilon}\circ {\Psi}_{t}\circ {\Phi}_{0}^{1}\left({u}_{0}\right),$  (5) 
into a fast solution component, the $\epsilon P$periodic change of variable ${\Phi}_{t/\epsilon}$, and a slow component, the flow ${\Psi}_{t}$ of a nonstiff averaged differential equation. Although equation (5) can be satisfied only up to a small remainder, various methods have been recently introduced in situations where (4) is posed in ${\mathbb{R}}^{n}$ or for the Schrödinger equation (1).
In the asymptotic behavior $\epsilon \to 0$, it can be advantageous to replace the original singularly perturbed model (for instance (1) or (2)) by an approximate model which does not contain stiffness any longer. Such reduced models can be derived using asymptotic analysis, namely averaging methods in the case of highlyoscillatory problems. In this project, we also plan to go beyond the mere derivation of limit models, by searching for better approximations of the original problem. This step is of mathematical interest per se but it also paves the way of the construction of multiscale numerical methods.
Asymptotic analysis of highlyoscillatory PDEs (All members)
Derivation of asymptotic problems
We intend to study the asymptotic behavior of highlyoscillatory evolution equations of the form (4) posed in an infinite dimensional Banach space.
Recently, the stroboscopic averaging has been extended to the PDE context, considering nonlinear Schrödinger equation (1) in the highlyoscillatory regime. A very exciting way would be to use this averaging strategy for highlyoscillatory kinetic problem (2) as those encountered in strongly magnetized plasmas. This turns out to be a very promising way to rederive gyrokinetic models which are the basis of tokamak simulations in the physicists community. In contract with models derived in the literature (see [34]) which only capture the average with respect to the oscillations, this strategy allows for the complete recovery of the exact solution from the asymptotic (non stiff) model. This can be done by solving companion transport equation that stems naturally from the decomposition (5).
Longtime behavior of Hamiltonian systems
The study of longtime behavior of nonlinear Hamiltonian systems have received a lot of interest during the last decades. It enables to put in light some characteristic phenomena in complex situations, which are beyond the reach of numerical simulations. This kind of analysis is of great interest since it can provide very precise properties of the solution. In particular, we will focus on the dynamics of nonlinear PDEs when the initial condition is close to a stationary solution. Then, the longtime behavior of the solution is studied through mainly three axis

linear stability: considering the linearized PDE, do we have stability of a stationary solution ? Do we have linear Landau damping around stable non homogeneous stationary states?

nonlinear stability: under a criteria, do we have stability of a stationary solution in energy norm like in [42], and does this stability persist under numerical discretization? For example one of our goals is to address the question of the existence and stability of discrete travelling wave in space and time.

do we have existence of damped solutions for the full nonlinear problem ? Around homogeneous stationary states, solutions scatter towards a modified stationary state (see [43], [39]). The question of existence of Landau damping effects around non homogeneous states is still open and is one of our main goal in the next future.
Asymptotic behavior of stochastic PDEs
The study of SPDEs has known a growing interest recently, in particular with the fields medal of M. Hairer in 2014. In many applications such as radiative transfer, molecular dynamics or simulation of optical fibers, part of the physical interactions are naturally modeled by adding supplementary random terms (the noise) to the initial deterministic equations. From the mathematical point of view, such terms change drastically the behavior of the system.

In the presence of noise, highlyoscillatory dispersive equations presents new problems. In particular, to study stochastic averaging of the solution, the analysis of the long time behavior of stochastic dispersive equations is required, which is known to be a difficult problem in the general case. In some cases (for instance highlyoscillatory Schrödinger equation (1) with a time white noise in the regime $\epsilon <\phantom{\rule{0.166667em}{0ex}}\phantom{\rule{0.166667em}{0ex}}<1$), it is however possible to perform the analysis and to obtain averaged stochastic equations. We plan to go further by considering more difficult problems, such as the convergence of a stochastic KleinGordonZakharov system to as stochastic nonlinear Schrödinger equation.

The longtime behavior of stochastic Schrödinger equations is of great interest to analyze mathematically the validity of the Zakharov theory for wave turbulence (see [44]). The problem of wave turbulence can be viewed either as a deterministic Hamiltonian PDE with random initial data or a randomly forced PDEs where the stochastic forcing is concentrated in some part of the spectrum (in this sense it is expected to be a hypoelliptic problem). One of our goals is to test the validity the Zakharov equation, or at least to make rigorous the spectrum distribution spreading observed in the numerical experiments.
Numerical schemes for highlyoscillatory problems (All members)
This section proposes to explore numerical issues raised by highlyoscillatory nonlinear PDEs for which (4) is a prototype. Simulating a highlyoscillatory phenomenon usually requires to adapt the numerical parameters in order to solve the period of size $\epsilon $ so as to accurately simulate the solution over each period, resulting in a unacceptable execution cost. Then, it is highly desirable to derive numerical schemes able to advance the solution by a time step independent of $\epsilon $. To do so, our goal is to construct Uniformly Accurate (UA) numerical schemes, for which the numerical error can be estimated by $C{h}^{p}$ ($h$ being any numerical parameters) with $C$ independent of $\epsilon $ and $p$ the order of the numerical scheme.
Recently, such numerical methods have been proposed by members of the team in the highlyoscillatory context [36]. They are mainly based on a separation of the fast and slow variables, as suggested by the decomposition (5). An additional ingredient to prove the uniformly accuracy of the method for (4) relies on the search for an appropriate initial data which enables to make the problem smooth with respect to $\epsilon $.
Such an approach is assuredly powerful since it provides a numerical method which enables to capture the high oscillations in time of the solution (and not only its average) even with a large time step. Moreover, in the asymptotic regime, the potential gain is of order $1/\epsilon $ in comparison with standard methods, and finally averaged models are not able to capture the intermediate regime since they miss important information of the original problem. We are strongly convinced that this strategy should be further studied and extended to cope with some other problems. The ultimate goal being to construct a scheme for the original equation which degenerates automatically into a consistent approximation of the averaged model, without resolving it, the latter can be very difficult to solve.

When rapidly oscillating coefficients in space (i.e. terms of the form $a(x,x/\epsilon )$) occur in elliptic or parabolic equations, homogenization theory and numerical homogenization are usually employed to handle the stiffness. However, these strategies are in general not accurate for all $\epsilon \in \phantom{\rule{0.277778em}{0ex}}]0,1]$. Then, the construction of numerical schemes which are able to handle both regimes in an uniform way is of great interest. Separating fast and slow spatial scales merits to be explored in this context. The delicate issue is then to extend the choice suitable initial condition to an appropriate choice of boundary conditions of the augmented problem.

For more complex problems however, the recent proposed approaches fail since the main oscillations cannot be identified explicitly. This is the case for instance when the magnetic field $B$ depends on $t$ or $x$ in (2) but also for many other physical problems. We then have to deal with the delicate issue of spacetime oscillations, which is known to be a very difficult problem from a mathematical and a numerical point of view. To take into account the spacetime mixing, a periodic motion has to be detected together with a phase $S$ which possibly depends on the time and space variables. These techniques originate from geometric optics which is a very popular technique to handle highlyfrequency waves.

The questions related to the geometric aspects of multiscale numerical schemes are of crucial importance, in particular when longtime simulations are addressed (see [40]). Indeed, one of the main questions of geometric integration is whether intrinsic properties of the solution may be passed onto its numerical approximation. For instance, if the model under study is Hamiltonian, then the exact flow is symplectic, which motivates the design of symplectic numerical approximation. For practical simulations of Hamiltonian systems, symplectic methods are known to possess very nice properties (see [40]). It is important to combine multiscale techniques to geometric numerical integration. All the problems and equations we intend to deal with will be addressed with a view to preserve intrinsic geometric properties of the exact solutions and/or to approach the asymptotic limit of the system in presence of a small parameter. An example of a numerical method developed by members of the team is the multirevolution method.

So far, numerical methods have been proposed for the periodic case with single frequency. However, the quasiperiodic case (replacing $t/\epsilon $ by $t\omega /\epsilon $ in (4), with $\omega \in {\mathbb{R}}^{d}$ a vector of nonresonant frequencies) is still misunderstood although many complex problems involve multifrequencies. Even if the quasiperiodic averaging is doable from a theoretical point of view in the ODE case (see [45]), it is unclear how it can be extended to PDEs. One of the main obstacle being the requirement, usual for ODEs like (4), for $\mathcal{F}$ to be analytic in the periodic variables, an assumption which is clearly impossible to meet in the PDE setting. An even more challenging problem is then the design of numerical methods for this problem.

All these questions will be revisited within the stochastic context. The mathematical study opens the way to the derivation of efficient multiscale numerical schemes for this kind of problems. We believe that the theory is now sufficiently well understood to address the derivation and numerical analysis of multiscale numerical schemes. Multirevolution composition methods have been recently extended to highlyoscillatory stochastic differential equations The generalization of such multiscale numerical methods to SPDEs is of great interest. The analysis and simulation of numerical schemes for highlyoscillatory nonlinear stochastic Schrödinger equation under diffusionapproximation for instance will be one important objective for us. Finally, an important aspect concerns the quantification of uncertainties in highlyoscillatory kinetic or quantum models (due to an incomplete knowledge of coefficients or imprecise measurement of datas). The construction of efficient multiscale numerical methods which can handle multiple scales as well as random inputs have important engineering applications.