betaFeedback

WorksMathematics

Full-horizon magnetic control of the Vlasov–Poisson system for plasma confinement

Abstract

We study the full-horizon control of a two-dimensional Vlasov–Poisson system through an externally imposed magnetic field. The control is represented by time-dependent coefficients associated with smooth, spatially localized basis functions, and is designed to steer selected position and velocity moments of the distribution function while limiting particle dispersion and the control effort. Starting from the kinetic constrained optimization problem, we derive the continuous first-order optimality conditions, including a backward Vlasov equation, an elliptic adjoint equation, and the variational inequality associated with bounded controls. We then project this system onto a particle–grid representation and identify the resulting particle adjoints and control gradient. The full-horizon field is then computed on a reduced particle ensemble, initialized by an instantaneous control, and subsequently applied to a larger set of particles. Numerical experiments indicate that the instantaneous control field already provides an effective baseline, while the full-horizon correction can further reduce the relative thermal energy near the boundaries and improve confinement over the prescribed time interval.

Introduction

Plasma dynamics is governed by the collective motion of charged particles and by their interaction with both externally imposed and self-consistent electromagnetic fields [6, 35, 50]. A particularly important application is magnetic confinement fusion, whose objective is to confine a high-temperature plasma for sufficiently long times to sustain thermonuclear reactions while limiting energy and particle losses [35, 50, 36]. Depending on the physical regime, plasmas may be thermal or non-thermal and may exhibit phenomena spanning a wide range of spatial and temporal scales. This multiscale character has led to different levels of description, including fluid, kinetic, and hybrid models [19, 32, 25, 8, 24]. Fluid descriptions, such as magnetohydrodynamics, provide an effective representation of the macroscopic evolution of the plasma and play a central role in the analysis of equilibrium and large-scale stability in magnetic-fusion devices [46, 7]. However, several processes relevant to tokamaks and stellarators, including non-equilibrium transport, energetic-particle dynamics, velocity-space anisotropies, and losses near the plasma boundary, require a kinetic description. Kinetic models describe the evolution of particle distribution functions in phase space and can therefore represent effects that cannot be adequately captured through a finite set of macroscopic quantities [15, 45, 16, 18]. Among kinetic descriptions, the Vlasov equation constitutes a fundamental model for collisionless plasmas. It describes the transport of the particle distribution under the action of electromagnetic forces, generally coupled to the distribution itself through Maxwell’s or Poisson’s equations. The high dimensionality of phase space, together with the nonlinear coupling between particles and fields, makes the numerical approximation of Vlasov-type systems particularly demanding. These challenges have motivated the development of a broad range of numerical techniques, including semi-Lagrangian, conservative, multiscale, and asymptotic-preserving methods [12, 17, 20, 23, 48, 33, 24, 25, 30, 31, 34, 52, 47, 51]. In this context, particle-in-cell (PIC) methods are especially attractive for high-dimensional problems. They approximate the distribution function through a finite ensemble of numerical particles, whose trajectories are advanced in a Lagrangian fashion, while the electromagnetic fields are computed on a spatial grid [10, 39, 38, 22]. This combination considerably reduces the computational cost associated with a direct discretization of the full phase space while retaining the essential kinetic features of the model.

The control of plasma dynamics is a central issue in magnetic confinement fusion. In both tokamaks and stellarators, externally generated magnetic fields can be used to modify particle trajectories, improve confinement, reduce cross-field transport, and limit the flux of energetic particles toward boundaries [21, 28, 36, 37]. The design of effective controls is nevertheless challenging because the magnetic field acts on an evolving particle distribution and must account for the nonlinear coupling between the plasma and its self-consistent fields. Furthermore, a successful strategy should not only reduce instantaneous losses but also anticipate the subsequent evolution of the plasma and preserve confinement over relevant time scales. Both the analytical foundations and the numerical realization of kinetic plasma control have therefore attracted increasing attention in recent years [9, 5, 42, 49, 27, 26, 11, 44, 43, 29]. In particular, instantaneous magnetic control strategies for deterministic and uncertain Vlasov–Poisson models were developed in [2, 3]. These approaches replace optimization over the entire time horizon with a sequence of local-in-time control problems. As a result, they are computationally inexpensive and can be readily coupled with kinetic solvers. Their local character, however, also represents an important limitation: the control is determined using only information available at the current time and does not explicitly account for the future evolution of the distribution function, the self-consistent field, or the resulting plasma position. This lack of predictive capability may become particularly relevant in magnetic confinement applications, where the plasma response to external actuation may be delayed and the primary objective is to maintain confinement and prevent particle losses over extended time intervals.

Motivated by the limitations of local in time control strategies, in this work we consider a deterministic, collisionless kinetic model and seek an optimal magnetic control over a prescribed time horizon. We describe the evolution of a population of negatively charged particles, identified with electrons, while the ions are treated as a fixed neutralizing background.

In dimensionless variables, the governing Vlasov–Poisson system reads

∂tf(t,x,v)+v⋅∇xf(t,x,v)+F(t,x,v)⋅∇vf(t,x,v)=0,−Δxϕ(t,x)=ρ(t,x)−ρˉ,E(t,x)=−∇xϕ(t,x),(1)\partial_t f(t,\mathbf{x},\mathbf{v})+\mathbf{v}\cdot\nabla_{\mathbf{x}}f(t,\mathbf{x},\mathbf{v})+F(t,\mathbf{x},\mathbf{v})\cdot\nabla_{\mathbf{v}}f(t,\mathbf{x},\mathbf{v})=0, \\ -\Delta_{\mathbf{x}}\phi(t,\mathbf{x})=\rho(t,\mathbf{x})-\bar{\rho},\qquad E(t,\mathbf{x})=-\nabla_{\mathbf{x}}\phi(t,\mathbf{x}), \tag*{(1)}

supplemented with the initial condition f(0,x,v)=f0(x,v)f(0,\mathbf{x},\mathbf{v})=f^0(\mathbf{x},\mathbf{v}). The electron density is obtained as the velocity marginal of the distribution function

ρ(t,x)=∫Rdvf(t,x,v) dv,\rho(t,\mathbf{x})=\int_{\mathbb{R}^{d_v}}f(t,\mathbf{x},\mathbf{v})\,\mathrm{d}\mathbf{v},

and the ion density is assumed to be homogeneous and stationary. Under this assumption, the net charge density on the right-hand side of the Poisson equation reduces to ρ−ρˉ\rho-\bar{\rho}, where ρˉ\bar{\rho} is the spatial mean of the electron density, which is conserved in time by the Vlasov equation and plays the role of the neutralizing ion background. With the sign convention adopted here, the effective Lorentz force acting on the electrons is given by

F(t,x,v)=E(t,x)+v×Bext(t,x),(2)F(t,\mathbf{x},\mathbf{v})=E(t,\mathbf{x})+\mathbf{v}\times\mathbf{B}_{\mathrm{ext}}(t,\mathbf{x}), \tag*{(2)}

where Bext(t,x)\mathbf{B}_{\mathrm{ext}}(t,\mathbf{x}) denotes the externally applied magnetic field, which plays the role of the control variable, while E(t,x)E(t,\mathbf{x}) is the self-consistent electric field determined through the Poisson equation in (1).

Our starting point is the instantaneous magnetic control framework introduced in [2, 3]. In the deterministic setting of [2], the control was obtained through a sequence of local in time optimization problems and represented by a spatially piecewise constant, time-dependent magnetic field. This spatial localization reflects the physical constraint that arbitrarily localized magnetic configurations cannot be generated in practice. The approach was subsequently extended in [3] to systems affected by uncertainty and collisions. Although these strategies are computationally efficient, their instantaneous nature prevents them from explicitly anticipating the future evolution of the plasma.

The aim of the present work is consequently to move beyond this instantaneous formulation and determine an open loop optimal magnetic control over the entire time interval [0,T][0,T]. Full-horizon magnetic control of the Vlasov–Poisson system has been studied analytically in [42], for fields generated by fixed coils, and numerically in [5], where state and adjoint equations are solved by a particle-in-cell method. With respect to these works, we consider moment-based tracking and dispersion objectives in a 2D–2V geometry with reflecting walls, derive a particle–grid optimality system consistent with the discretize-then-optimize approach, initialize the optimization with the instantaneous control of [2], thus allowing a direct comparison between the two strategies, and compute the control on a reduced particle ensemble before validating it on a larger one.

To this end, we study the associated first-order optimality conditions through a coupled forward–backward system [44, 5, 42]. Accounting for the full time horizon allows the control to anticipate the subsequent evolution of the particle distribution and to balance immediate confinement with longer-term objectives. This predictive capability comes however at a substantially higher computational cost, since the state equation must be coupled with a backward adjoint equation and an optimality condition for the control. More precisely, we consider the constrained optimization problem

min⁡Bext∈BadJ(f,Bext),subject to xref‘(1.1)‘.(3)\min_{\mathbf{B}_{\mathrm{ext}}\in\mathcal{B}_{\mathrm{ad}}}\mathcal{J}(f,\mathbf{B}_{\mathrm{ext}}),\qquad\text{subject to }{xref}`(1.1)`. \tag*{(3)}

Here, Bad\mathcal{B}_{\mathrm{ad}} denotes the set of admissible magnetic controls. At this stage, we introduce the cost functional in the general form

J(f,Bext)=∫0T(D(f)(t)+γ2R(f,Bext)(t)) dt+Ψ(f(T)),(4)\mathcal{J}(f,\mathbf{B}_{\mathrm{ext}})=\int_{0}^{T}\left(\mathcal{D}(f)(t)+\frac{\gamma}{2}\mathcal{R}(f,\mathbf{B}_{\mathrm{ext}})(t)\right)\,\mathrm{d}t+\Psi(f(T)), \tag*{(4)}

where γ>0\gamma>0. The running cost D(f)\mathcal{D}(f) measures the deviation of selected position- and velocity-space observables from their prescribed targets and controls their dispersion. The regularization term R(f,Bext)\mathcal{R}(f,\mathbf{B}_{\mathrm{ext}}) penalizes the control effort, whereas Ψ(f(T))\Psi(f(T)) promotes the desired plasma configuration at the final time TT. The precise definitions of these terms are introduced in § 2, after specifying the spatial parameterization of the external magnetic field. In the following, the first-order optimality conditions are derived through a Lagrangian (adjoint) approach. Solving the complete optimality system directly at the kinetic scale would, however, be prohibitively expensive. We therefore synthesize the control using a reduced particle representation of the plasma dynamics. The reduced system is employed to compute the open-loop optimal field, which is subsequently reconstructed and applied to a larger particle simulation. In this way, the expensive optimization is performed on a representative particle ensemble, while the effectiveness and transferability of the resulting control are assessed on the higher-dimensional dynamics.

The remainder of the paper is organized as follows. § 2 introduces the model and the optimal control objective. § 3 derives the continuous optimality system and its particle–grid representation. § 4 describes the numerical approximation and the optimization method. § 5 presents the computational experiments. A final section summarizes the main conclusions and discusses possible future developments. The appendices provide the discretize-then-optimize derivation and the instantaneous control strategy employed to initialize the full-horizon optimization.

Problem setting

We consider a reduced dx×dv=2×2d_x \times d_v=2\times2 setting, with particle positions x=(x,y)∈Ωx⊂R2\mathbf{x}=(x,y)\in\Omega_x\subset\mathbb{R}^2 and velocities v=(vx,vy)∈Ωv⊂R2\mathbf{v}=(v_x,v_y)\in\Omega_v\subset\mathbb{R}^2. The dynamics is studied in the plane orthogonal to the dominant component of the externally applied magnetic field. To obtain a physically and computationally realizable control, we partition the spatial domain into Ncx×NcyN_c^x\times N_c^y cells, with

Ng=NcxNcy.N_g=N_c^xN_c^y.

Denoting the cells by CkC_k, k=1,…,Ngk=1,\ldots,N_g, the partition satisfies

Ωx=⋃k=1NgCk,Ck∩Cj=∅for k≠j.\Omega_x=\bigcup_{k=1}^{N_g}C_k,\qquad C_k\cap C_j=\varnothing\quad\text{for }k\ne j.

We then associate with each cell a time-dependent control coefficient bk(t)b_k(t) and introduce the control vector

b(t):=(b1(t),…,bNg(t)).\mathbf{b}(t):=(b_1(t),\ldots,b_{N_g}(t)).

For notational convenience, we set

Kk(x):=K(x,xck),k=1,…,Ng,\mathcal{K}_k(\mathbf{x}):=\mathcal{K}(\mathbf{x},\mathbf{x}_c^k),\qquad k=1,\ldots,N_g,

where xckx_{\mathrm{c}}^{k} denotes the center of CkC_k. The scalar component of the external magnetic field is reconstructed from the control coefficients as

Bb(t,x)=∑k=1Ngbk(t)Kk(x).(5)B_b(t,\mathbf{x}) = \sum_{k=1}^{N_g} b_k(t)\mathcal{K}_k(\mathbf{x}). \tag*{(5)}

Each kernel Kk\mathcal{K}_k is assumed to be nonnegative and provides a smooth approximation of the indicator function of the corresponding cell. More precisely, it is approximately constant inside CkC_k and smoothly decays to zero outside it. Consequently, the resulting magnetic field is nearly piecewise constant over the partition while remaining globally smooth across cell interfaces. The corresponding three-dimensional external magnetic field is defined as

Bext[b](t,x):=(0,0,Bb(t,x)).(6)\mathbf{B}_{\mathrm{ext}}[\mathbf{b}](t,\mathbf{x}) := (0,0,B_b(t,\mathbf{x})). \tag*{(6)}

The scalar component BbB_b is subject to the pointwise amplitude constraint

∣Bb(t,x)∣≤Bmax⁡,|B_b(t,\mathbf{x})| \le B_{\max},

where Bmax⁡>0B_{\max} > 0 denotes the maximum admissible magnetic-field intensity. With this convention, the planar component of the magnetic force is given by

v×Bext[b]=(vyBb,−vxBb).\mathbf{v} \times\mathbf{B}_{\mathrm{ext}}[\mathbf{b}] = (v_yB_b,-v_xB_b).

The precise form of the kernel, the choice of the number NgN_g of control cells, and a sensitivity analysis with respect to this number will be discussed in the numerical experiments section. The corresponding two-dimensional Vlasov–Poisson system reads

∂tf(t,x,v)+v⋅∇xf(t,x,v)+(E(t,x)+v×Bext[b](t,x))⋅∇vf(t,x,v)=0,−Δxϕ(t,x)=ρ(t,x)−ρˉ,E(t,x)=−∇xϕ(t,x),f(0,x,v)=f0(x,v).(7)\begin{aligned} \partial_t f(t,\mathbf{x},\mathbf{v}) + \mathbf{v}\cdot\nabla_{\mathbf{x}}f(t,\mathbf{x},\mathbf{v}) + \left(\mathbf{E}(t,\mathbf{x}) + \mathbf{v}\times\mathbf{B}_{\mathrm{ext}}[\mathbf{b}](t,\mathbf{x})\right)\cdot\nabla_{\mathbf{v}}f(t,\mathbf{x},\mathbf{v}) &= 0,\\ -\Delta_{\mathbf{x}}\phi(t,\mathbf{x}) &= \rho(t,\mathbf{x})-\bar{\rho},\qquad\mathbf{E}(t,\mathbf{x})=-\nabla_{\mathbf{x}}\phi(t,\mathbf{x}),\qquad f(0,\mathbf{x},\mathbf{v})=f_0(\mathbf{x},\mathbf{v}). \tag*{(7)} \end{aligned}

Here, f0(x,v)f_0(\mathbf{x},\mathbf{v}) is the prescribed initial distribution, whereas f(t,x,v)f(t,\mathbf{x},\mathbf{v}) denotes the electron distribution function in phase space. The quantity

ρˉ=1∣Ωx∣∫Ωxρ(t,x) dx\bar{\rho}=\frac{1}{|\Omega_x|}\int_{\Omega_x}\rho(t,\mathbf{x})\,\mathrm{d}\mathbf{x}

is conserved in time by the Vlasov equation and plays the role of the neutralizing ion background. In both equations, periodic boundary conditions are imposed in the xx-direction, while reflective boundary conditions are prescribed in the yy-direction to model the presence of material walls. The control objective is to steer the plasma by means of the external magnetic field and, in particular, to reduce particle transport toward and interaction with these walls. We now reformulate the continuous optimal control problem (3) in terms of the finite set of control coefficients. Let

I:={x,v},Ω:=Ωx×Ωv.\mathcal{I}:=\{\mathbf{x},\mathbf{v}\},\qquad\Omega:=\Omega_x\times\Omega_v.

The set of admissible controls is defined as

Uad:={b∈L2(0,T;RNg):∣Bb(t,x)∣≤Bmax⁡ for a.e. (t,x)∈(0,T)×Ωx}.(8)U_{\mathrm{ad}}:=\left\{\mathbf{b}\in L^2(0,T;\mathbb{R}^{N_g}): |B_b(t,\mathbf{x})|\le B_{\max}\ \text{for a.e. }(t,\mathbf{x})\in(0,T)\times\Omega_x\right\}. \tag*{(8)}

The reduced optimal control problem is therefore

min⁡b∈UadJ(f,b),subject to xref‘(2.3)‘.(9)\min_{\mathbf{b}\in U_{\mathrm{ad}}} J(f,\mathbf{b}),\qquad\text{subject to }{xref}`(2.3)`. \tag*{(9)}

We first specify the regularization of the control coefficients. The regularization of the reconstructed magnetic field would naturally lead to

Rfield(f,b)(t):=∫Ω∣Bb(t,x)∣2f(t,x,v) dx dv=∑k=1Ng∣bk(t)∣2∫ΩKk(x)2f(t,x,v) dx dv+2∑1≤k<j≤Ngbk(t)bj(t)∫ΩKk(x)Kj(x)f(t,x,v) dx dv.(10)\mathcal{R}_{\mathrm{field}}(f,\mathbf{b})(t) := \int_{\Omega} \lvert\mathbf{B}_{b}(t,\mathbf{x}) \rvert^{2} f(t,\mathbf{x},\mathbf{v}) \,d\mathbf{x}\,d\mathbf{v} = \sum_{k=1}^{N_g} \lvert b_k(t) \rvert^{2} \int_{\Omega} \mathcal{K}_k(\mathbf{x})^{2} f(t,\mathbf{x},\mathbf{v}) \,d\mathbf{x}\,d\mathbf{v} + 2 \sum_{1 \le k < j \le N_g} b_k(t)b_j(t) \int_{\Omega} \mathcal{K}_k(\mathbf{x})\mathcal{K}_j(\mathbf{x})f(t,\mathbf{x},\mathbf{v}) \,d\mathbf{x}\,d\mathbf{v}. \tag*{(10)}

Since the kernels are smooth approximations of disjoint indicator functions, they satisfy

Kk2≃Kk,KkKj≃0for k≠j,(11)\mathcal{K}_k^{2} \simeq\mathcal{K}_k,\qquad\mathcal{K}_k\mathcal{K}_j \simeq0 \quad\text{for } k \ne j, \tag*{(11)}

except within the narrow transition regions near cell interfaces. Motivated by this localization property, we neglect the overlap terms and adopt the diagonal approximation

Rdiag(f,b)(t):=∑k=1Ng∣bk(t)∣2∫ΩKk(x)f(t,x,v) dx dv.(12)\mathcal{R}_{\mathrm{diag}}(f,\mathbf{b})(t) := \sum_{k=1}^{N_g} \lvert b_k(t) \rvert^{2} \int_{\Omega} \mathcal{K}_k(\mathbf{x})f(t,\mathbf{x},\mathbf{v}) \,d\mathbf{x}\,d\mathbf{v}. \tag*{(12)}

This approximation penalizes the control coefficients independently and avoids coupling between controls associated with different cells. It will be used throughout the remainder of the paper. The cost functional is defined by

J(f,b):=∫0TL(f,b)(t) dt+Ψ(f(T)),(13)\mathcal{J}(f,\mathbf{b}) := \int_{0}^{T} \mathcal{L}(f,\mathbf{b})(t)\,dt + \Psi(f(T)), \tag*{(13)}

where the running cost is

L(f,b)(t):=∑q∈IDq(f)(t)+γ2Rdiag(f,b)(t),(14)\mathcal{L}(f,\mathbf{b})(t) := \sum_{q \in\mathcal{I}} \mathcal{D}_q(f)(t) + \frac{\gamma}{2}\mathcal{R}_{\mathrm{diag}}(f,\mathbf{b})(t), \tag*{(14)}

with γ>0\gamma> 0. The first contribution measures the deviation from the prescribed target configuration, whereas the second penalizes the magnitude of the localized control coefficients. For each q∈Iq \in\mathcal{I}, the tracking term is defined as

Dq(f)(t):=αq2∫Ω∣ψq(x,v)−ψ^q b∣2f(t,x,v) dx dv+βq2∫Ω∣ψq(x,v)−mq[f](t)∣2f(t,x,v) dx dv.(15)\mathcal{D}_q(f)(t) := \frac{\alpha_q}{2} \int_{\Omega} \left\lvert\psi_q(\mathbf{x},\mathbf{v}) - \widehat{\psi}_q^{\,b} \right\rvert^{2} f(t,\mathbf{x},\mathbf{v})\,d\mathbf{x}\,d\mathbf{v} + \frac{\beta_q}{2} \int_{\Omega} \left\lvert\psi_q(\mathbf{x},\mathbf{v}) - m_q[f](t) \right\rvert^{2} f(t,\mathbf{x},\mathbf{v})\,d\mathbf{x}\,d\mathbf{v}. \tag*{(15)}

where αq,βq≥0\alpha_q,\beta_q \ge0. The functions ψq\psi_q represent the selected physical- or velocity-space observables, while ψ^q b\widehat{\psi}_q^{\,b} denotes the corresponding prescribed target. The first term in (15) promotes the concentration of the distribution around the target value ψ^q b\widehat{\psi}_q^{\,b}. The second term penalizes the dispersion of the observable ψq\psi_q around its instantaneous mean. More precisely, the mean of the observable ψq\psi_q is defined by

mq[f](t):=1M∫Ωψq(x,v)f(t,x,v) dx dv,M:=∫Ωf0(x,v) dx dv.(16)m_q[f](t) := \frac{1}{\mathcal{M}} \int_{\Omega} \psi_q(\mathbf{x},\mathbf{v})f(t,\mathbf{x},\mathbf{v})\,d\mathbf{x}\,d\mathbf{v},\qquad\mathcal{M} := \int_{\Omega} f_0(\mathbf{x},\mathbf{v})\,d\mathbf{x}\,d\mathbf{v}. \tag*{(16)}

Finally, the terminal cost is defined as

Ψ(f(T)):=∑q∈IDq(f)(T).(17)\Psi(f(T)) := \sum_{q \in\mathcal{I}} \mathcal{D}_q(f)(T). \tag*{(17)}

It has the same structure as the running cost, but is evaluated at the final time TT. Its purpose is to promote the desired plasma configuration at the end of the control horizon.

Derivation of the optimality system

The aim of this section is to introduce and compare two formally equivalent strategies for solving the minimization problem (9). We first derive the continuous optimality system through a Lagrangian formulation and then project the resulting forward-backward equations onto a particle approximation of the distribution function. This procedure yields a particle-level optimality system involving the characteristic variables, their corresponding adjoint variables, and the gradient with respect to the control coefficients. We subsequently relate this construction to the alternative procedure in which the particle-grid approximation is introduced before deriving the optimality conditions. When the same kernel interpolation, quadrature rules, and particle discretization are employed, the two approaches formally lead to the same particle-level optimality system, up to the errors introduced by the discretization.

Continuous first-order optimality conditions

We derive here the continuous first-order optimality conditions associated with the constrained minimization problem (9) through a Lagrangian approach. Let p(t,x,v)p(t,\mathbf{x},\mathbf{v}) and λ(t,x)\lambda(t,\mathbf{x}) denote the adjoint variables associated with the Vlasov equation and the Poisson equation in (7), respectively. For a given control vector b=(b1,…,bNg)\mathbf{b}=(b_1,\ldots,b_{N_g}), we recall the notation

Fb(t,x,v):=−∇xϕ(t,x)+v×Bext[b](t,x).\mathbf{F}_{\mathbf{b}}(t,\mathbf{x},\mathbf{v}) := -\nabla_{\mathbf{x}}\phi(t,\mathbf{x})+\mathbf{v}\times\mathbf{B}_{\mathrm{ext}}[\mathbf{b}](t,\mathbf{x}).

To distinguish the Lagrangian functional from the running cost L\mathcal{L} introduced in (14), we denote the former by L\mathscr{L}. It is defined as

L(f,ϕ,p,λ,b):=∫0TL(f,b)(t) dt+Ψ(f(T))+∫0T∫Ωp(t,x,v)(∂tf(t,x,v)+v⋅∇xf(t,x,v)+Fb(t,x,v)⋅∇vf(t,x,v)) dx dv dt+∫0T∫Ωxλ(t,x)(−Δxϕ(t,x)−∫Ωvf(t,x,v) dv+ρˉ) dx dt,(18)\begin{aligned} \mathscr{L}(f,\phi,p,\lambda,\mathbf{b}) := {}& \int_0^T \mathcal{L}(f,\mathbf{b})(t)\,\mathrm{d}t+\Psi(f(T)) \\ &+ \int_0^T\int_{\Omega} p(t,\mathbf{x},\mathbf{v})\left(\partial_t f(t,\mathbf{x},\mathbf{v})+\mathbf{v}\cdot\nabla_{\mathbf{x}}f(t,\mathbf{x},\mathbf{v})\right.\\ &\left.\qquad\qquad\qquad\qquad+\mathbf{F}_{\mathbf{b}}(t,\mathbf{x},\mathbf{v})\cdot\nabla_{\mathbf{v}}f(t,\mathbf{x},\mathbf{v})\right)\,\mathrm{d}x\,\mathrm{d}v\,\mathrm{d}t \\ &+ \int_0^T\int_{\Omega_x}\lambda(t,\mathbf{x})\left(-\Delta_{\mathbf{x}}\phi(t,\mathbf{x})-\int_{\Omega_v}f(t,\mathbf{x},\mathbf{v})\,\mathrm{d}v+\bar{\rho}\right)\,\mathrm{d}x\,\mathrm{d}t, \tag*{(18)} \end{aligned}

where Ω=Ωx×Ωv\Omega=\Omega_x\times\Omega_v. The running cost L\mathcal{L}, the tracking terms Dq\mathcal{D}_q, the diagonal control regularization, and the terminal cost are defined in (14), (15), (12), and (17), respectively. In deriving the optimality conditions, we assume that all functions are sufficiently regular and that the boundary terms arising from integration by parts vanish under the prescribed state boundary conditions and the corresponding compatible adjoint boundary conditions. Let (f⋆,ϕ⋆,p⋆,λ⋆,b⋆)(f^\star,\phi^\star,p^\star,\lambda^\star,\mathbf{b}^\star) denote the optimal state-adjoint-control. Stationarity, together with the variational optimality condition for the constrained control, gives

{DpL=0,DλL=0,DfL=0,DϕL=0,⟨DbL,c−b⋆⟩≥0for every c∈Uad,(19)\begin{cases} D_p\mathscr{L}=0,\\ D_\lambda\mathscr{L}=0,\\ D_f\mathscr{L}=0,\\ D_\phi\mathscr{L}=0,\\ \langle D_{\mathbf{b}}\mathscr{L},\mathbf{c}-\mathbf{b}^\star\rangle\ge0\quad\text{for every }\mathbf{c}\in\mathcal{U}_{\mathrm{ad}}, \tag*{(19)} \end{cases}

where all derivatives are evaluated at the optimal value. If b⋆\mathbf{b}^\star lies in the interior of Uad\mathcal{U}_{\mathrm{ad}}, the last condition reduces to DbL=0D_{\mathbf{b}}\mathscr{L}=0. The first two conditions in (19) recover the forward

Vlasov–Poisson system

∂tf+v⋅∇xf+Fb⋅∇vf=0,−Δxϕ=ρ−ρˉ,(20)\begin{aligned} \partial_t f+\mathbf{v}\cdot\nabla_x f+F_b\cdot\nabla_v f&=0,\\ -\Delta_x\phi&=\rho-\bar{\rho}, \tag*{(20)} \end{aligned}

where, for brevity, the dependence on t,x,vt,\mathbf{x},\mathbf{v} has been omitted and the forward system is supplemented with the initial condition f(0,x,v)=f0(x,v)f(0,\mathbf{x},\mathbf{v})=f_0(\mathbf{x},\mathbf{v}). The variations with respect to ff and ϕ\phi yield the backward system

−(∂tp+v⋅∇xp+Fb⋅∇vp)−λ+G=0,−Δxλ−∇x⋅(∫Ωvf∇vp dv)=0,(21)\begin{aligned} -\left(\partial_t p+\mathbf{v}\cdot\nabla_xp+F_b\cdot\nabla_vp\right)-\lambda+\mathcal{G}&=0,\\ -\Delta_x\lambda-\nabla_x\cdot\left(\int_{\Omega_v}f\nabla_vp\,\mathrm{d}v\right)&=0, \tag*{(21)} \end{aligned}

where the source term G\mathcal{G} is the derivative of the running cost with respect to ff. It is given by

G(t,x,v)=∑q∈T[αq2∣ψq(x,v)−ψ^q∣2+βq2∣ψq(x,v)−mq[f](t)∣2]+γ2∑k=1Ng∣bk(t)∣2Kk(x).(22)\mathcal{G}(t,\mathbf{x},\mathbf{v})=\sum_{q\in\mathcal{T}}\left[\frac{\alpha_q}{2}\left|\psi_q(\mathbf{x},\mathbf{v})-\widehat{\psi}_q\right|^2+\frac{\beta_q}{2}\left|\psi_q(\mathbf{x},\mathbf{v})-m_q[f](t)\right|^2\right]+\frac{\gamma}{2}\sum_{k=1}^{N_g}|b_k(t)|^2\mathcal{K}_k(\mathbf{x}). \tag*{(22)}

The backward Vlasov equation is supplemented with a terminal condition obtained from

p(T,x,v)+δΨ(f(T))δf(T,x,v)=0.p(T,\mathbf{x},\mathbf{v})+\frac{\delta\Psi(f(T))}{\delta f(T,\mathbf{x},\mathbf{v})}=0.

Therefore,

p(T,x,v)=−GT(x,v),(23)p(T,\mathbf{x},\mathbf{v})=-\mathcal{G}_T(\mathbf{x},\mathbf{v}), \tag*{(23)}

where

GT(x,v):=∑q∈T[αq2∣ψq(x,v)−ψ^q∣2+βq2∣ψq(x,v)−mq[f](T)∣2].(24)\mathcal{G}_T(\mathbf{x},\mathbf{v}):=\sum_{q\in\mathcal{T}}\left[\frac{\alpha_q}{2}\left|\psi_q(\mathbf{x},\mathbf{v})-\widehat{\psi}_q\right|^2+\frac{\beta_q}{2}\left|\psi_q(\mathbf{x},\mathbf{v})-m_q[f](T)\right|^2\right]. \tag*{(24)}

The multiplier λ(t,x)\lambda(t,\mathbf{x}), on the other hand, does not satisfy a terminal condition, since it is determined at each time by the elliptic adjoint Poisson equation. In addition, periodic boundary conditions are imposed on pp, while reflective boundary conditions are applied to qq. Finally, we compute the derivative with respect to each control coefficient. For any k=1,…,Ngk=1,\ldots,N_g, the kk-th component of the control gradient is

gk(t):=DbkL(f,ϕ,p,λ,b)=∫Ωx[γbk(t)ρ(t,x)+∫Ωv(v×∇vp(t,x,v))f(t,x,v) dv]Kk(x) dx,(25)\begin{aligned} g_k(t)&:=D_{b_k}\mathcal{L}(f,\phi,p,\lambda,\mathbf{b})\\ &=\int_{\Omega_x}\left[\gamma b_k(t)\rho(t,\mathbf{x})+\int_{\Omega_v}\left(\mathbf{v}\times\nabla_vp(t,\mathbf{x},\mathbf{v})\right)f(t,\mathbf{x},\mathbf{v})\,\mathrm{d}v\right]\mathcal{K}_k(\mathbf{x})\,\mathrm{d}x, \tag*{(25)} \end{aligned}

where, for two planar vectors a=(ax,ay)\mathbf{a}=(a_x,a_y) and c=(cx,cy)\mathbf{c}=(c_x,c_y), we denote by a×c:=axcy−aycx\mathbf{a}\times\mathbf{c}:=a_xc_y-a_yc_x the scalar (out-of-plane) component of their cross product.

The control optimality condition can thus be written as

∑k=1Ng∫0Tgk∗(t)(ck(t)−bk∗(t)) dt≥0for every c∈Uad.(26)\sum_{k=1}^{N_g}\int_0^Tg_k^*(t)\left(c_k(t)-b_k^*(t)\right)\,\mathrm{d}t\geq0\qquad\text{for every }\mathbf{c}\in U_{\mathrm{ad}}. \tag*{(26)}

For an interior optimum, this condition reduces to gk∗(t)=0g_k^*(t)=0 for every k=1,…,Ngk=1,\ldots,N_g and almost every t∈(0,T)t\in(0,T). Altogether, equations (20), (21), (23), and (26) define the continuous forward–backward optimality system. The state equations evolve forward from the prescribed initial distribution, whereas the adjoint Vlasov equation evolves backward from the terminal condition. The elliptic state and adjoint Poisson equations are solved at each time, and the control coefficients are characterized by the variational inequality (26).

From the continuous to the particle–grid optimality system

We now derive a particle–grid counterpart of the continuous optimality system (3.3)–(3.9). Throughout this subsection, NgN_g continues to denote the number of control cells, whereas NhN_h denotes the number of nodes of the spatial grid used by the Particle-in-Cell (PIC) discretization [14, 30, 31]. Let NN be the number of particles and consider the regularized empirical measure

fN(t,x,v):=∑i=1NωiSh(x−xi(t))δ(v−vi(t)),(27)f^N(t,\mathbf{x},\mathbf{v}) := \sum_{i=1}^{N} \omega_i S_h(\mathbf{x}-\mathbf{x}_i(t))\delta(\mathbf{v}-\mathbf{v}_i(t)), \tag*{(27)}

where the particle weights satisfy ωi>0\omega_i > 0, ∑i=1Nωi=MN\sum_{i=1}^{N}\omega_i = \mathcal{M}^N, and ShS_h is a normalized, compactly supported shape function of width hh [40, 14, 30]. The particle positions and velocities are denoted by (xi(t),vi(t))(\mathbf{x}_i(t),\mathbf{v}_i(t)), with initial values chosen consistently with f0f_0.

We introduce spatial grid nodes {xℓ}ℓ=1Nh\{\mathbf{x}_\ell\}_{\ell=1}^{N_h}, quadrature weights {ηℓ}ℓ=1Nh\{\eta_\ell\}_{\ell=1}^{N_h}, a discrete Laplacian AhA_h, and a discrete gradient GhG_h. We then have

ρℓN(t):=∑i=1NωiSh(xℓ−xi(t)),ℓ=1,…,Nh,(28)\rho_\ell^N(t) := \sum_{i=1}^{N}\omega_i S_h(\mathbf{x}_\ell-\mathbf{x}_i(t)), \qquad\ell=1,\ldots,N_h, \tag*{(28)}

while the discrete Poisson problem becomes

−(Ahϕ)ℓ=ρℓN−ρˉ,ℓ=1,…,Nh.(29)-(A_h\phi)_\ell= \rho_\ell^N-\bar{\rho}, \qquad\ell=1,\ldots,N_h. \tag*{(29)}

The grid electric field and its interpolation to the particle positions are defined by

Eh:=−Ghϕ,Eih:=∑ℓ=1Nhηℓ(Eh)ℓSh(xℓ−xi).(30)\mathbf{E}_h := -G_h\phi, \qquad\mathbf{E}_i^h := \sum_{\ell=1}^{N_h}\eta_\ell(\mathbf{E}_h)_\ell S_h(\mathbf{x}_\ell-\mathbf{x}_i). \tag*{(30)}

The particle characteristics associated with the Vlasov equation are

x˙i=vi,v˙i=Fih:=Eih+vi×Bi,Bi:=Bb(t,xi)=∑k=1Ngbk(t)Kk(xi),(31)\dot{\mathbf{x}}_i = \mathbf{v}_i,\qquad \dot{\mathbf{v}}_i = \mathbf{F}_i^h := \mathbf{E}_i^h+\mathbf{v}_i\times\mathbf{B}_i,\qquad \mathbf{B}_i := \mathbf{B}_b(t,\mathbf{x}_i) = \sum_{k=1}^{N_g}b_k(t)\mathbf{K}_k(\mathbf{x}_i), \tag*{(31)}

for i=1,…,Ni=1,\ldots,N while vi×Bi=(vy,iBi,−vx,iBi)\mathbf{v}_i\times\mathbf{B}_i=(v_{y,i}B_i,-v_{x,i}B_i). To project the continuous adjoint equation onto the particle trajectories, set

P:=∇xp,Q:=∇vp.\mathcal{P}:=\nabla_{\mathbf{x}}p,\qquad\mathcal{Q}:=\nabla_{\mathbf{v}}p.

Taking the spatial and velocity gradients of the first equation in (21) yields

∂tP+v⋅∇xP+Fb⋅∇vP+(∇xFb)TQ=∇xG−∇xλ,∂tQ+v⋅∇xQ+Fb⋅∇vQ+P+(∇vFb)TQ=∇vG.(32)\begin{aligned} \partial_t\mathcal{P}+\mathbf{v}\cdot\nabla_{\mathbf{x}}\mathcal{P}+\mathbf{F}_b\cdot\nabla_{\mathbf{v}}\mathcal{P}+(\nabla_{\mathbf{x}}\mathbf{F}_b)^{T}\mathcal{Q} &= \nabla_{\mathbf{x}}\mathcal{G}-\nabla_{\mathbf{x}}\lambda,\\ \partial_t\mathcal{Q}+\mathbf{v}\cdot\nabla_{\mathbf{x}}\mathcal{Q}+\mathbf{F}_b\cdot\nabla_{\mathbf{v}}\mathcal{Q}+\mathcal{P}+(\nabla_{\mathbf{v}}\mathbf{F}_b)^{T}\mathcal{Q} &= \nabla_{\mathbf{v}}\mathcal{G}. \tag*{(32)} \end{aligned}

Defining the particle adjoints by

pi(t):=P(t,xi(t),vi(t)),qi(t):=Q(t,xi(t),vi(t)),(33)\mathbf{p}_i(t):=\mathcal{P}(t,\mathbf{x}_i(t),\mathbf{v}_i(t)),\qquad \mathbf{q}_i(t):=\mathcal{Q}(t,\mathbf{x}_i(t),\mathbf{v}_i(t)), \tag*{(33)}

and applying the chain rule along (31), we obtain

p˙i=−(∇xiFih)Tqi+∇xiGiN−∇xiλih,q˙i=−pi+∇viGiN+qi×Bi.(34)\begin{aligned} \dot{\mathbf{p}}_i &= -(\nabla_{\mathbf{x}_i}\mathbf{F}_i^h)^{T}\mathbf{q}_i+\nabla_{\mathbf{x}_i}\mathcal{G}_i^N-\nabla_{\mathbf{x}_i}\lambda_i^h,\\ \dot{\mathbf{q}}_i &= -\mathbf{p}_i+\nabla_{\mathbf{v}_i}\mathcal{G}_i^N+\mathbf{q}_i\times\mathbf{B}_i. \tag*{(34)} \end{aligned}

Now we set

ψx(x,v)=y,ψv(x,v)=vy,ψ^x=y^,ψ^v=v^y.\psi_x(\mathbf{x},\mathbf{v})=y,\qquad\psi_v(\mathbf{x},\mathbf{v})=v_y,\qquad\widehat{\psi}_x=\widehat{y},\qquad\widehat{\psi}_v=\widehat{v}_y.

and

y‾N:=1MN∑j=1Nωjyj,v‾yN:=1MN∑j=1Nωjvy,j.\overline{y}^{N}:=\frac{1}{\mathcal{M}^{N}}\sum_{j=1}^{N}\omega_j y_j,\qquad\overline{v}_y^{N}:=\frac{1}{\mathcal{M}^{N}}\sum_{j=1}^{N}\omega_j v_{y,j}.

The gradients of the particle approximation of the source G\mathcal{G} in (22) become then

∇xiGiN=[αx(yi−y^)+βx(yi−y‾N)]ey+γ2∑k=1Ng∣bk(t)∣2∇Kk(xi),(35)\nabla_{\mathbf{x}_i}\mathcal{G}_i^{N}=\left[\alpha_x(y_i-\widehat{y})+\beta_x(y_i-\overline{y}^{N})\right]\mathbf{e}_y+\frac{\gamma}{2}\sum_{k=1}^{N_g}|b_k(t)|^2\nabla\mathcal{K}_k(\mathbf{x}_i), \tag*{(35)}
∇viGiN=[αv(vy,i−v^y)+βv(vy,i−v‾yN)]ey,\nabla_{\mathbf{v}_i}\mathcal{G}_i^{N}=\left[\alpha_v(v_{y,i}-\widehat{v}_y)+\beta_v(v_{y,i}-\overline{v}_y^{N})\right]\mathbf{e}_y,

with ey=(0,1)T\mathbf{e}_y=(0,1)^T. The discretized elliptic adjoint equation becomes

−Ah∗λ~+Gh∗Q=0.(36)-\mathcal{A}_h^*\widetilde{\lambda}+\mathcal{G}_h^*Q=0. \tag*{(36)}

where

Qℓ:=∑i=1NωiSh(xℓ−xi)qi,ℓ=1,…,Nh,(37)Q_\ell:=\sum_{i=1}^{N}\omega_iS_h(\mathbf{x}_\ell-\mathbf{x}_i)\mathbf{q}_i,\qquad\ell=1,\ldots,N_h, \tag*{(37)}

and λ~=(λ~1,…,λ~Nh)T\widetilde{\lambda}=(\widetilde{\lambda}_1,\ldots,\widetilde{\lambda}_{N_h})^T. The adjoints Ah∗\mathcal{A}_h^* and Gh∗\mathcal{G}_h^* use the same quadrature-weighted grid inner product used in (29)–(30), meaning that on a uniform grid one has Ah∗=AhT=Ah\mathcal{A}_h^*=\mathcal{A}_h^T=\mathcal{A}_h and Gh∗=GhT\mathcal{G}_h^*=\mathcal{G}_h^T. Thus, the particle adjoint equations in backward form are

\begin{aligned} -\dot{\mathbf{p}}_i&=(\nabla_{\mathbf{x}_i}F_i^h)^T\mathbf{q}_i-\nabla_{\mathbf{x}_i}\mathcal{G}_i^N+\nabla_{\mathbf{x}_i}\lambda_i^h,\\ -\dot{\mathbf{q}}_i&=\mathbf{p}_i-\nabla_{\mathbf{v}_i}\mathcal{G}_i^N-\mathbf{q}_i\times\mathbf{B}_i, \end{aligned} \qquad i=1,\ldots,N. \tag*{(38)}\end{aligned}

The terminal condition follows from (23):

pi(T)=−∇xiGT,iN,qi(T)=−∇viGT,iN,(39)\mathbf{p}_i(T)=-\nabla_{\mathbf{x}_i}\mathcal{G}_{T,i}^N,\qquad\mathbf{q}_i(T)=-\nabla_{\mathbf{v}_i}\mathcal{G}_{T,i}^N, \tag*{(39)}

where GT,iN\mathcal{G}_{T,i}^N is obtained from the tracking part of GiN\mathcal{G}_i^N, evaluated at TT; in particular, it contains no control-regularization term. Finally, inserting (27) into the continuous gradient (25) gives, for k=1,…,Ngk=1,\ldots,N_g,

gkN,h(t)=∑i=1Nωi∫ΩxSh(x−xi)[γbk(t)+vi×qi]Kk(x) dx.(40)g_k^{N,h}(t)=\sum_{i=1}^{N}\omega_i\int_{\Omega_x}S_h(\mathbf{x}-\mathbf{x}_i)\left[\gamma b_k(t)+\mathbf{v}_i\times\mathbf{q}_i\right]\mathcal{K}_k(\mathbf{x})\,\mathrm{d}x. \tag*{(40)}

If ShS_h has approximation order rr, then, for every sufficiently smooth function φ\varphi,

∫ΩxSh(x−xi)φ(x) dx=φ(xi)+O(hr).\int_{\Omega_x}S_h(\mathbf{x}-\mathbf{x}_i)\varphi(\mathbf{x})\,\mathrm{d}x=\varphi(\mathbf{x}_i)+O(h^r).

Hence the particle control gradient is

gkN(t)=∑i=1Nωi[γbk(t)+vi(t)×qi(t)]Kk(xi(t))+O(hr).(41)g_k^N(t)=\sum_{i=1}^{N}\omega_i\left[\gamma b_k(t)+\mathbf{v}_i(t)\times\mathbf{q}_i(t)\right]\mathcal{K}_k(\mathbf{x}_i(t))+O(h^r). \tag*{(41)}

The discrete control optimality condition is therefore

∑k=1Ng∫0T(gkN)∗(t)(ck(t)−bk∗(t)) dt≥0for every c∈Uadh,(42)\sum_{k=1}^{N_g}\int_0^T\left(g_k^N\right)^*(t)\left(c_k(t)-b_k^*(t)\right)\,\mathrm{d}t\geq0\qquad\text{for every }c\in\mathcal{U}_{ad}^h, \tag*{(42)}

where UadhU_{\mathrm{ad}}^h is defined in (A.13). In the unconstrained case, or when the discrete amplitude constraint is inactive, this reduces formally to gkN(t)=0g_k^N(t)=0. Equations (31), (29), (38), (36), (39), (42) form the particle–grid optimality system. If the same shape function and quadrature are used, and if the discrete differential operators are paired with their adjoints, this projected optimize-then-discretize system agrees formally with the system obtained by discretizing the particle–grid dynamics first and then differentiating it, up to particle, grid, and time-discretization errors. A detailed comparison is provided in Appendix § 7.

Numerical approximation and optimization method

In this section, we describe the fully discrete strategy used to solve the particle–grid optimality system (31), (29), (38), and (36), with terminal conditions (39) and control optimality condition (42). Optimisation is performed using a projected-gradient method with Armijo backtracking, following standard approaches for particle and mean-field control problems [1, 4].

Let

jN,h,Δt(b):=JN,h,Δt[fN,h,Δt[b],b]j_{N,h,\Delta t}(\mathbf{b}) := \mathcal{J}_{N,h,\Delta t}\left[f_{N,h,\Delta t}[\mathbf{b}],\mathbf{b}\right]

denote the reduced, fully discrete cost functional, where fN,h,Δt[b]f_{N,h,\Delta t}[\mathbf{b}] is the particle approximation obtained by solving the discrete state equations for the control b\mathbf{b}. The time integrals in J\mathcal{J} are approximated using the same time quadrature used to define the discrete gradient below. We initialize the optimization with

bk(0)(t)=bk0,k=1,…,Ng,b_k^{(0)}(t)=b_k^0,\qquad k=1,\ldots,N_g,

where the coefficients bk0b_k^0 are obtained from the instantaneous control described in Appendix § 8. In optimisation iteration mm, the current coefficients b(m)\mathbf{b}^{(m)} are kept fixed while the particle–grid state system is integrated forward in time. This produces (xi(m),vi(m))(\mathbf{x}_i^{(m)},\mathbf{v}_i^{(m)}). The adjoint equations are then solved backward along the same trajectory, producing (pi(m),qi(m))(\mathbf{p}_i^{(m)},q_i^{(m)}), and the gradient g(m)\mathbf{g}^{(m)} is evaluated from (41). In particular, the magnetic field acting on particle ii is

Bi(m)(t)=Bb(m)(t,xi(m)(t))=∑k=1Ngbk(m)(t)Kk(xi(m)(t)).(43)B_i^{(m)}(t)=B_{\mathbf{b}^{(m)}}(t,\mathbf{x}_i^{(m)}(t))=\sum_{k=1}^{N_g}b_k^{(m)}(t)K_k(\mathbf{x}_i^{(m)}(t)). \tag*{(43)}

After time discretization, let Uadh,ΔtU_{\mathrm{ad}}^{h,\Delta t} denote the discrete-time counterpart of UadhU_{\mathrm{ad}}^h, namely

Uadh,Δt:={(bn)n=0Nt−1:∣∑k=1NgbknKk(xℓ)∣≤Bmax⁡, ℓ=1,…,Nh, n=0,…,Nt−1}.U_{\mathrm{ad}}^{h,\Delta t}:=\left\{(\mathbf{b}^n)_{n=0}^{N_t-1}:\left|\sum_{k=1}^{N_g}b_k^nK_k(\mathbf{x}_\ell)\right|\le B_{\max},\ \ell=1,\ldots,N_h,\ n=0,\ldots,N_t-1\right\}.

We denote by ΠUadh,Δt\Pi_{U_{\mathrm{ad}}^{h,\Delta t}} the orthogonal projection onto this closed convex set with respect to the discrete L2L^2 inner product. For a trial step size α>0\alpha>0, define

bα(m):=ΠUadh,Δt(b(m)−αg(m)).(44)\mathbf{b}_{\alpha}^{(m)}:=\Pi_{U_{\mathrm{ad}}^{h,\Delta t}}\left(\mathbf{b}^{(m)}-\alpha\mathbf{g}^{(m)}\right). \tag*{(44)}

The accepted step size is

α(m)=α0ηrm,η∈(0,1),(45)\alpha^{(m)}=\alpha_0\eta^{r_m},\qquad\eta\in(0,1), \tag*{(45)}

where rmr_m is the smallest nonnegative integer for which

jN,h,Δt(bα(m)(m))≤jN,h,Δt(b(m))+c⟨g(m),bα(m)(m)−b(m)⟩Δt,c∈(0,1),(46)\begin{aligned} j_{N,h,\Delta t}\left(\mathbf{b}_{\alpha^{(m)}}^{(m)}\right)&\le j_{N,h,\Delta t}\left(\mathbf{b}^{(m)}\right)\\ &\quad+c\left\langle\mathbf{g}^{(m)},\mathbf{b}_{\alpha^{(m)}}^{(m)}-\mathbf{b}^{(m)}\right\rangle_{\Delta t},\qquad c\in(0,1), \tag*{(46)} \end{aligned}

is satisfied, where

⟨g,b⟩Δt=Δt∑n=0Nt−1gn⋅bn\langle\mathbf{g},\mathbf{b}\rangle_{\Delta t}=\Delta t\sum_{n=0}^{N_t-1}\mathbf{g}^{n}\cdot\mathbf{b}^{n}

denotes the discrete time inner product approximating the continuous L2(0,T)L^2(0,T) inner product via a Riemann sum, with gn,bn∈RNxc×Nyc\mathbf{g}^{n},\mathbf{b}^{n}\in\mathbb{R}^{N_x^c\times N_y^c} denoting the gradient and control arrays at time tnt^n respectively. The control update is then

b(m+1)=bα(m)(m).(47)\mathbf{b}^{(m+1)}=\mathbf{b}_{\alpha^{(m)}}^{(m)}. \tag*{(47)}

Each trial evaluation in the line search requires a new forward solve, but the adjoint system and the gradient are computed only once at the beginning of iteration mm [13, 41]. The iteration stops either when the maximum number of iterations is reached, or when the cost functional stagnates, that is if

∣JN,h,Δt(b(m))−JN,h,Δt(b(m−1))∣≤εstall,\left|J_{N,h,\Delta t}\left(\mathbf{b}^{(m)}\right)-J_{N,h,\Delta t}\left(\mathbf{b}^{(m-1)}\right)\right|\leq\varepsilon_{\mathrm{stall}},

holds for nstalln_{\mathrm{stall}} consecutive iterations.

To reduce the computational cost, the state, adjoint, and gradient are evaluated on a representative subset Is⊂{1,…,N}\mathcal{I}_s\subset\{1,\ldots,N\}, with ∣Is∣=Ns≪N|\mathcal{I}_s|=N_s\ll N. Assuming equal particle weights and relabeling the particles in the reduced ensemble by j=1,…,Nsj=1,\ldots,N_s, we assign the rescaled weights

ω^j:=MN/Ns,j=1,…,Ns,with∑j=1Nsω^j=MN.\widehat{\omega}_j:=\mathcal{M}^N/N_s,\qquad j=1,\ldots,N_s,\qquad\text{with}\qquad\sum_{j=1}^{N_s}\widehat{\omega}_j=\mathcal{M}^N.

Equivalently, if the full ensemble has weights ωi=MN/N\omega_i=\mathcal{M}^N/N, then ω^j=N/Nsωj\widehat{\omega}_j=N/N_s\omega_j. Omitting for simplicity in the following the dependence on optimization iteration (m)(m), we partition [0,T][0,T] into NtN_t intervals of length Δt=T/Nt\Delta t=T/N_t. The controls are then represented as piecewise constant functions in time, with bkn≃bk(tn)b_k^n\simeq b_k(t^n) while the Poisson problem is solved according to

ρℓn=∑j=1Nsω^jSh(xℓ−xjn),−(Ahϕn)ℓ=ρℓn−1,ℓ=1,…,Nh,Ehn=−Ghϕn,Eih,n=∑ℓ=1Nhη(Eℓn)ℓSh(xℓ−xin).(48)\begin{aligned} \rho_\ell^n&=\sum_{j=1}^{N_s}\widehat{\omega}_j S_h(x_\ell-x_j^n),\\ -(A_h\phi^n)_\ell&=\rho_\ell^n-1,\qquad\ell=1,\ldots,N_h,\\ \mathbf{E}_h^n&=-G_h\phi^n,\qquad\mathbf{E}_i^{h,n}=\sum_{\ell=1}^{N_h}\eta(\mathbf{E}_\ell^n)_\ell S_h(x_\ell-x_i^n). \tag*{(48)} \end{aligned}

The controlled magnetic field at the particle position becomes

Bin=∑k=1NgbknKk(xin).(49)\mathbf{B}_i^n=\sum_{k=1}^{N_g}b_k^n\mathcal{K}_k(x_i^n). \tag*{(49)}

The state characteristics in the reduced system (Ns≪NN_s\ll N) are advanced using a first-order semi-implicit scheme [30]:

\begin{aligned} \mathbf{v}_i^{n+1}&=\mathbf{v}_i^n+\Delta t\left(\mathbf{E}_i^{h,n}+\mathbf{v}_i^{n+1}\times\mathbf{B}_i^n\right),\\ \mathbf{x}_i^{n+1}&=\mathbf{x}_i^n+\Delta t\,\mathbf{v}_i^{n+1}, \end{aligned} \qquad i=1,\ldots,N_s. \tag*{(50)}\end{aligned}

In the same way, the particle adjoints are initialized by

piNt=−∇xiGT,iN,qiNt=−∇viGT,iN,i=1,…,Ns.(51)\mathbf{p}_i^{N_t}=-\nabla_{x_i}G_{T,i}^N,\qquad\mathbf{q}_i^{N_t}=-\nabla_{v_i}G_{T,i}^N,\qquad i=1,\ldots,N_s. \tag*{(51)}

while for n=Nt−1,…,0n=N_t-1,\ldots,0, we employ a semi-implicit backward step for the particle adjoint equations:

{pin=pin+1+Δt[JF,in+1qin−∇xiGiN,n+1+Lin+1],qin=qin+1+Δt[pin−∇viGiN,n+1−qin×Bin+1],−Ah∗λ~n+1+Gh∗Qn+1=0.(52)\begin{cases} \mathbf{p}_i^n=\mathbf{p}_i^{n+1}+\Delta t\left[J_{F,i}^{n+1}\mathbf{q}_i^n-\nabla_{\mathbf{x}_i}G_i^{N,n+1}+\mathbf{L}_i^{n+1}\right],\\ \mathbf{q}_i^n=\mathbf{q}_i^{n+1}+\Delta t\left[\mathbf{p}_i^n-\nabla_{\mathbf{v}_i}G_i^{N,n+1}-\mathbf{q}_i^n\times\mathbf{B}_i^{n+1}\right],\\ -A_h^*\widetilde{\boldsymbol{\lambda}}^{n+1}+G_h^*\mathbf{Q}^{n+1}=0. \tag*{(52)} \end{cases}

with

JF,in+1:=(∇xiFih,n+1)T,Fih,n+1=Eih,n+1+vin×Bin+1.J_{F,i}^{n+1}:=\left(\nabla_{\mathbf{x}_i}F_i^{h,n+1}\right)^{T},\qquad F_i^{h,n+1}=\mathbf{E}_i^{h,n+1}+\mathbf{v}_i^n\times\mathbf{B}_i^{n+1}.

and

Qℓn+1=∑i=1Nsω^iSh(xℓ−xin+1)qin+1,ℓ=1,…,Nh,(53)\mathbf{Q}_{\ell}^{n+1}=\sum_{i=1}^{N_s}\widehat{\omega}_iS_h(\mathbf{x}_{\ell}-\mathbf{x}_i^{n+1})\mathbf{q}_i^{n+1},\qquad\ell=1,\ldots,N_h, \tag*{(53)}
Lin+1:=∇xiλih,n+1=∑ℓ=1Nhηℓλ~ℓn+1∇xiSh(xℓ−xin+1).(54)\mathbf{L}_i^{n+1}:=\nabla_{\mathbf{x}_i}\lambda_i^{h,n+1}=\sum_{\ell=1}^{N_h}\eta_{\ell}\widetilde{\lambda}_{\ell}^{n+1}\nabla_{\mathbf{x}_i}S_h(\mathbf{x}_{\ell}-\mathbf{x}_i^{n+1}). \tag*{(54)}

The discrete gradient with respect to the control coefficients is finally computed as

gk,n:=∑i=1Nsω^i[γkn+vin×qin]Kk(xin),(55)g_{k,n}:=\sum_{i=1}^{N_s}\widehat{\omega}_i\left[\gamma_k^n+\mathbf{v}_i^n\times\mathbf{q}_i^n\right]K_k(\mathbf{x}_i^n), \tag*{(55)}

We emphasize that (55) is obtained by time-discretizing the semidiscrete particle adjoint system. Its relation with the gradient obtained by differentiating the fully discrete state map is discussed in Appendix A.

While the first-order scheme (50) yields a simple and efficient formulation for the iterative optimization phase, it exhibits numerical dissipation over long times. To avoid artificial damping when computing physical quantities, all validation tests on the full ensemble are performed using an energy-preserving second-order semi-implicit scheme [30]. This reads

\begin{cases} \mathbf{v}_i^{(1)}=\mathbf{v}_i^n+\dfrac{\Delta t}{2}\left(\mathbf{E}_i^{h,n}+\mathbf{v}_i^{(1)}\times\mathbf{B}_i^n\right),\\ \mathbf{x}_i^{(1)}=\mathbf{x}_i^n+\dfrac{\Delta t}{2}\mathbf{v}_i^{(1)}, \end{cases} \qquad i=1,\ldots,N, \tag*{(56)}\end{cases}

where the apex (1) denotes the first stage, while the second stage reads

\begin{cases} \mathbf{v}_i^{(2)}=\mathbf{v}_i^n+\dfrac{\Delta t}{2}\left(\mathbf{E}_i^{h,*}+\mathbf{v}_i^{(2)}\times\mathbf{B}_i^*\right),\\ \mathbf{x}_i^{(2)}=\mathbf{x}_i^n+\dfrac{\Delta t}{2}\mathbf{v}_i^{(2)}, \end{cases} \qquad i=1,\ldots,N, \tag*{(57)}\end{cases}

where the electric field Eih,∗\mathbf{E}_i^{h,*} is computed as in (48) by solving the Poisson problem, whereas the magnetic field Bi∗\mathbf{B}_i^* is computed as in (49) by replacing xin\mathbf{x}_i^n with xi∗=2xi(1)−xin\mathbf{x}_i^*=2\mathbf{x}_i^{(1)}-\mathbf{x}_i^n. Finally, the numerical solution at time tn+1t^{n+1} reads,

\begin{cases} \mathbf{v}_i^{n+1}=\mathbf{v}_i^{(1)}+\mathbf{v}_i^{(2)}-\mathbf{v}_i^n,\\ \mathbf{x}_i^{n+1}=\mathbf{x}_i^{(1)}+\mathbf{x}_i^{(2)}-\mathbf{x}_i^n, \end{cases} \qquad i=1,\ldots,N. \tag*{(58)}\end{cases}

Numerical experiments

In this section, we assess the performance of the proposed control strategy on a set of standard benchmark problems. For the Armijo line search in (45)–(46), we set the parameters to α0=10\alpha_0 = 10, η=0.5\eta= 0.5, and c=10−4c = 10^{-4}. The stopping tolerance for the optimization algorithm is set to εstall=10−5\varepsilon_{\mathrm{stall}} = 10^{-5} over a stall limit of nstall=50n_{\mathrm{stall}} = 50 iterations. A full ensemble of N=107N = 10^7 particles is sampled from the initial distribution, whereas the forward–backward optimality system and the corresponding gradient are evaluated using a fixed representative random subset of Ns=103N_s = 10^3 particles at each optimization iteration. The rest of the parameters are specified separately for each test case. As described in § 2, the control is represented on a coarse rectangular partition consisting of Ng=NcxNcyN_g = N_c^x N_c^y cells, with NcxN_c^x and NcyN_c^y cells along the xx- and yy-directions, respectively. We denote the corresponding cell centers by xcj,k=(xcj,yck)\mathbf{x}_{c}^{j,k} = (x_c^j, y_c^k). The control-cell widths are denoted by hxch_x^c and hych_y^c. They are chosen substantially larger than the spacings Δx\Delta x and Δy\Delta y of the finer grid used in the Particle-in-Cell reconstruction, namely hxc≫Δxh_x^c \gg\Delta x, hyc≫Δyh_y^c \gg\Delta y. Using a double-index representation bj,k(t)b_{j,k}(t) of the control coefficients introduced in § 2, the magnetic field evaluated at the ii-th particle position is given by

Bi(t):=B(t,xi(t))=∑j=1Ncx∑k=1Ncybj,k(t)Kj,k(xi(t)).(59)B_i(t) := \mathbf{B}(t,\mathbf{x}_i(t)) = \sum_{j=1}^{N_c^x}\sum_{k=1}^{N_c^y} b_{j,k}(t)\mathcal{K}_{j,k}(\mathbf{x}_i(t)). \tag*{(59)}

The two-dimensional control kernels are defined as tensor products,

Kj,k(x,y)=κjx(x)κky(y),(60)\mathcal{K}_{j,k}(x,y) = \kappa_j^x(x)\kappa_k^y(y), \tag*{(60)}

where

κjx(x)=σ(x−xj−12cεc)−σ(x−xj+12cεc),(61)\kappa_j^x(x) = \sigma\left(\frac{x-x_{j-\frac{1}{2}}^c}{\varepsilon_c}\right) - \sigma\left(\frac{x-x_{j+\frac{1}{2}}^c}{\varepsilon_c}\right), \tag*{(61)}
κky(y)=σ(y−yk−12cεc)−σ(y−yk+12cεc),(62)\kappa_k^y(y) = \sigma\left(\frac{y-y_{k-\frac{1}{2}}^c}{\varepsilon_c}\right) - \sigma\left(\frac{y-y_{k+\frac{1}{2}}^c}{\varepsilon_c}\right), \tag*{(62)}

with σ(z)=11+exp⁡(−z)\sigma(z) = \frac{1}{1+\exp(-z)} and εc=10−10\varepsilon_c = 10^{-10}. Once the optimal coefficients b∗(t)\mathbf{b}^*(t) have been computed over the time interval [0,T][0,T] using the reduced ensemble of NsN_s particles, the resulting control is validated by evolving the full ensemble of N=107N = 10^7 particles using the second order scheme in (56)–(57)–(58).

To assess the effectiveness of the control strategy near the material walls, we introduce a boundary strip

Ωb=[ax,bx]×([ay,ay+δb]∪[by−δb,by])⊂Ωx,\Omega_b = [a_x,b_x] \times([a_y,a_y+\delta_b] \cup[b_y-\delta_b,b_y]) \subset\Omega_x,

where δb\delta_b denotes the thickness of the strip. Unless otherwise specified, we set δb=Δy\delta_b = \Delta y. Let {xℓ}ℓ=1Nh\{\mathbf{x}_{\ell}\}_{\ell=1}^{N_h} denote the nodes of the fine spatial grid used for the PIC reconstruction, and define

Ib:={ℓ∈{1,…,Nh}:xℓ∈Ωb},Nb:=∣Ib∣.\mathcal{I}_b := \{\ell\in\{1,\ldots,N_h\} : \mathbf{x}_{\ell} \in\Omega_b\}, \qquad N_b := |\mathcal{I}_b|.

The particle mass contained in the boundary strip is approximated by

Mb(t)=ΔxΔy∑ℓ∈Ib∑i=1NωiSh(xℓ−xi(t)),(63)\mathcal{M}_b(t) = \Delta x\Delta y \sum_{\ell\in\mathcal{I}_b}\sum_{i=1}^{N}\omega_i S_h(\mathbf{x}_{\ell}-\mathbf{x}_i(t)), \tag*{(63)}

where ShS_h is the two-dimensional tensor-product quadratic B-spline shape function scaled by the spatial resolution parameter h=ΔxΔyh = \Delta x \Delta y. The corresponding mean velocity is

Ub(t)=ΔxΔyMb(t)∑ℓ∈Ib∑i=1Nωivi(t)Sh(xℓ−xi(t)).(64)U_b(t) = \frac{\Delta x \Delta y}{\mathcal{M}_b(t)} \sum_{\ell\in\mathcal{I}_b} \sum_{i=1}^{N} \omega_i \mathbf{v}_i(t) S_h(\mathbf{x}_\ell- \mathbf{x}_i(t)). \tag*{(64)}

We then define the thermal energy contained in the boundary strip as

Eb(t)=ΔxΔy2∑ℓ∈Ib∑i=1Nωi∣vi(t)−Ub(t)∣2S(xℓ−xi(t)).(65)\mathcal{E}_b(t) = \frac{\Delta x \Delta y}{2} \sum_{\ell\in\mathcal{I}_b} \sum_{i=1}^{N} \omega_i \lvert\mathbf{v}_i(t) - U_b(t) \rvert^2 S(\mathbf{x}_\ell- \mathbf{x}_i(t)). \tag*{(65)}

For comparison, the total reconstructed mass and the global mean velocity are

M(t)=ΔxΔy∑ℓ=1Nh∑i=1NωiS(xℓ−xi(t)),(66)\mathcal{M}(t) = \Delta x \Delta y \sum_{\ell=1}^{N_h} \sum_{i=1}^{N} \omega_i S(\mathbf{x}_\ell- \mathbf{x}_i(t)), \tag*{(66)}
U(t)=ΔxΔyM(t)∑ℓ=1Nh∑i=1Nωivi(t)S(xℓ−xi(t)).(67)U(t) = \frac{\Delta x \Delta y}{\mathcal{M}(t)} \sum_{\ell=1}^{N_h} \sum_{i=1}^{N} \omega_i \mathbf{v}_i(t) S(\mathbf{x}_\ell- \mathbf{x}_i(t)). \tag*{(67)}

The total thermal energy is consequently defined as

E(t)=ΔxΔy2∑ℓ=1Nh∑i=1Nωi∣vi(t)−U(t)∣2S(xℓ−xi(t)).(68)\mathcal{E}(t) = \frac{\Delta x \Delta y}{2} \sum_{\ell=1}^{N_h} \sum_{i=1}^{N} \omega_i \lvert\mathbf{v}_i(t) - U(t) \rvert^2 S(\mathbf{x}_\ell- \mathbf{x}_i(t)). \tag*{(68)}

Finally, the percentage of the total thermal energy localized within the boundary strip is also measured in the sequel, and denoted by Ebrel(t)\mathcal{E}_b^{\mathrm{rel}}(t).

Two-dimensional double Sod test

We consider a two-dimensional extension of the double Sod shock-tube test. The spatial and velocity variables are denoted by x=(x,y)\mathbf{x} = (x,y) and v=(vx,vy)\mathbf{v} = (v_x,v_y), respectively. The spatial domain is Ωx=[0,1.5]2\Omega_x = [0,1.5]^2, with periodic boundary conditions in the xx-direction and reflective boundary conditions in the yy-direction. The initial distribution is given by

f0(x,v)=ρ0(x)2πT0(x)exp⁡(−∣v∣22T0(x)),(69)f_0(\mathbf{x},\mathbf{v}) = \frac{\rho_0(\mathbf{x})}{2\pi T_0(\mathbf{x})} \exp\left(-\frac{\lvert\mathbf{v} \rvert^2}{2T_0(\mathbf{x})}\right), \tag*{(69)}

where the initial density and temperature are defined as

ρ0(x)={0.125,y∈[0,0.5)∪(1,1.5],1,y∈[0.5,1],T0(x)={0.1,y∈[0,0.5)∪(1,1.5],1,y∈[0.5,1].(70)\begin{aligned} \rho_0(\mathbf{x}) &= \begin{cases} 0.125, & y \in[0,0.5) \cup(1,1.5], \\ 1, & y \in[0.5,1], \end{cases}\\ T_0(\mathbf{x}) &= \begin{cases} 0.1, & y \in[0,0.5) \cup(1,1.5], \\ 1, & y \in[0.5,1]. \end{cases} \tag*{(70)} \end{aligned}

Thus, the initial configuration consists of a dense, high-temperature central layer surrounded by two low-density, low-temperature regions. Although the problem is posed in a two-dimensional domain, the initial data are homogeneous in the xx-direction.

The dynamics is evolved up to the final time T=0.5T = 0.5, using the time step Δt=0.01\Delta t = 0.01. The density and the self-consistent electric field are reconstructed on a uniform PIC grid with mx=my=32m_x=m_y=32 nodes, so that Nh=mxmyN_h=m_xm_y. For both the instantaneous and optimal controls, we set

αx=βx=αv=βv=10,γ=10−3,M=50,\alpha_x=\beta_x=\alpha_v=\beta_v=10,\qquad\gamma=10^{-3},\qquad M=50,

with targets y^=0.75\hat{y}=0.75 and v^y=0\hat{v}_y=0. We first consider a reference configuration in which no optimization is performed and the magnetic field is prescribed to be uniform in space and constant in time: Bref(t,x)=0.75B_{\mathrm{ref}}(t,\mathbf{x})=0.75. Figure 1 shows the percentage of the total thermal energy contained in the boundary strip, Ebrel(t)\mathcal{E}^{\mathrm{rel}}_b(t) together with three snapshots of the density profiles. The increase observed during the simulation indicates that an increasing fraction of the thermal energy becomes localized near the upper and lower boundaries.

Density snapshots and relative energy evolution

Figure 1. Two-dimensional double Sod test with the prescribed uniform magnetic field Bref=0.75B_{\mathrm{ref}}=0.75. Snapshots of the uncontrolled dynamics at t=0.2t=0.2, t=0.34t=0.34, and t=0.5t=0.5 and evolution of the percentage of total thermal energy contained in the boundary strip, Ebrel(t)\mathcal{E}^{\mathrm{rel}}_b(t).

We next consider the controlled dynamics, with the control computed using the optimization procedure described in the previous sections and initialized by the instantaneous control. We set Ncx=1N_c^x=1 and consider Ncy∈{5,7,9}N_c^y\in\{5,7,9\}. Consequently, each control cell extends over the entire xx-direction and the reconstructed magnetic field depends only on yy and tt.

To describe the convergence of the optimization algorithm, we introduce the reduced cost functional

J^(b):=J(f[b],b),\widehat{\mathcal{J}}(\mathbf{b}):=\mathcal{J}(\mathbf{f}[\mathbf{b}],\mathbf{b}),

where f[b]\mathbf{f}[\mathbf{b}] denotes the solution of the forward problem associated with the control b\mathbf{b}. Figure 2 reports the value of J^(b(m))\widehat{\mathcal{J}}(\mathbf{b}^{(m)}) and the Euclidean norm of its discrete gradient as functions of the optimization iteration, for the different values of NcyN_c^y. The results indicate that the instantaneous control provides an effective initial guess and is already close to a stationary point of the discrete optimization problem: the initial gradient norm is relatively small and the variation of the cost functional over the iterations is limited. Nevertheless, the optimization produces a non-negligible correction to the instantaneous control and, in most cases, further reduces the fraction of thermal energy localized near the boundaries. This comparison is presented in Figure 3, which shows Ebrel(t)\mathcal{E}_b^{\mathrm{rel}}(t) for the instantaneous and optimal controls. The three panels correspond to Ncy=5N_c^y = 5, Ncy=7N_c^y = 7, and Ncy=9N_c^y = 9, respectively. Overall, the optimal control reduces the concentration of thermal energy in the boundary strip compared with the instantaneous strategy. The first row of Figure 4 shows instead three snapshots of the

Evolution of the reduced cost functional and discrete gradient norm

Figure 2. Two-dimensional double Sod test. Left: evolution of the reduced cost functional J^(b(m))\widehat{\mathcal{J}}(\mathbf{b}^{(m)}) over the optimization iterations. Right: evolution of the discrete gradient norm ∥g(m)∥2\lVert\mathbf{g}^{(m)} \rVert_2. The curves correspond to Ncy∈{5,7,9}N_c^y \in\{5,7,9\}.

Comparison of the percentage of thermal energy in the boundary strip

Figure 3. Two-dimensional double Sod test. Comparison of the percentage of thermal energy contained in the boundary strip, Ebrel(t)\mathcal{E}_b^{\mathrm{rel}}(t), obtained with the instantaneous and optimal controls. From left to right: Ncy=5N_c^y = 5, Ncy=7N_c^y = 7, and Ncy=9N_c^y = 9.

Two-dimensional double Sod test with density snapshots and optimal control coefficients

Figure 4. Two-dimensional double Sod test. First row: snapshots of the controlled dynamics at t=0.2t = 0.2, t=0.34t = 0.34, and t=0.5t = 0.5 for Ncy=9N_c^y = 9. Second row: optimal control coefficients b1,k∗(t)b_{1,k}^*(t). The columns in the second row correspond, from left to right, to Ncy=5N_c^y = 5, Ncy=7N_c^y = 7, and Ncy=9N_c^y = 9.

controlled dynamics at t=0.2t = 0.2, t=0.34t = 0.34, and t=0.5t = 0.5 for Ncy=9N_c^y = 9. Similar qualitative behavior is observed for Ncy=5N_c^y = 5 and Ncy=7N_c^y = 7 and they are not reported. The second row displays the optimal control coefficients b1,k⋆(t)b_{1,k}^{\star}(t), for Ncy∈{5,7,9}N_c^y \in\{5,7,9\}.

Two-stream plasma test

We consider the spatial domain

Ωx=[0,20]×[−1,1],\Omega_x = [0,20] \times[-1,1],

with periodic boundary conditions in the xx-direction and reflective boundary conditions in the yy-direction. The initial distribution is defined by

f0(x,v)=f0+(x,v)χ{y≥0}+f0−(x,v)χ{y<0},(71)f_0(\mathbf{x},\mathbf{v}) = f_0^+(\mathbf{x},\mathbf{v})\chi_{\{y\geq0\}} + f_0^-(\mathbf{x},\mathbf{v})\chi_{\{y<0\}}, \tag*{(71)}

where

f0±(x,v)=ρ0±(y)2πT0exp⁡(−vx2+(vy−uy(x))22T0).(72)f_0^\pm(\mathbf{x},\mathbf{v}) = \frac{\rho_0^\pm(y)}{2\pi T_0}\exp\left(-\frac{v_x^2 + (v_y-u_y(x))^2}{2T_0}\right). \tag*{(72)}

The transverse mean velocity is prescribed as

uy(x)={−3,x<10,3,x≥10,(73)u_y(x) = \begin{cases} -3, & x < 10, \\ 3, & x \geq10, \end{cases} \tag*{(73)}

while the two density profiles are given by

ρ0±(y)=12πσexp⁡(−(y∓cy)22σ2),(74)\rho_0^\pm(y) = \frac{1}{\sqrt{2\pi}\sigma}\exp\left(-\frac{(y \mp c_y)^2}{2\sigma^2}\right), \tag*{(74)}

with

T0=2,cy=0.3,σ=0.1.T_0 = 2,\qquad c_y = 0.3,\qquad\sigma= 0.1.

With this convention, ρ0+\rho_0^+ is centered at y=cyy = c_y and ρ0−\rho_0^- at y=−cyy = -c_y. The regions x<10x < 10 and x≥10x \geq10 are assigned opposite transverse velocities, thus generating counter-propagating particle streams. The dynamics is evolved up to the final time T=0.5T = 0.5, using the time step Δt=0.01\Delta t = 0.01. The density and the electric field are reconstructed on a uniform PIC grid with mx=my=64m_x = m_y = 64 nodes, so that Nh=mxmyN_h = m_xm_y. For both the instantaneous and optimal controls, we set

αx=αv=10,βx=βv=1,γ=10−3,M=50,\alpha_x = \alpha_v = 10,\qquad\beta_x = \beta_v = 1,\qquad\gamma= 10^{-3},\qquad M = 50,

with targets y^=v^y=0\hat{y} = \hat{v}_y = 0. We first consider a reference configuration in which no optimization is performed and the magnetic field is prescribed to be uniform in space and constant in time: Bref(t,x)=0.75B_{\mathrm{ref}}(t,\mathbf{x}) = 0.75. Figure 5 shows three density snapshots at t=0.15t = 0.15, t=0.30t = 0.30, and t=0.50t = 0.50. Owing to the opposite initial transverse velocities, particles are transported toward the upper and lower boundaries and are subsequently reflected. The lower panel on the right reports the corresponding percentage of thermal energy contained in the boundary strip, Ebrel(t)\mathcal{E}^{\mathrm{rel}}_b(t).

Density snapshots and relative thermal energy for the uncontrolled two-stream plasma test

Figure 5. Two-stream plasma test with the prescribed uniform magnetic field Bref=0.75B_{\mathrm{ref}} = 0.75. Density snapshots at t=0.15t = 0.15, t=0.30t = 0.30, and t=0.50t = 0.50 and evolution of the percentage of thermal energy contained in the boundary strip, Ebrel(t)\mathcal{E}^{\mathrm{rel}}_b(t).

We next compute the controlled dynamics. We set Ncx=1N_c^x = 1 and consider Ncy∈{2,4,6}N_c^y \in\{2, 4, 6\}. Hence, the control cells extend over the entire xx-direction and the reconstructed magnetic field depends only on yy and tt. Figure 6 shows the convergence history of the optimization procedure. The two panels report the reduced cost functional J^(b(m))\widehat{J}(\mathbf{b}^{(m)}) and the Euclidean norm of its discrete gradient ∥g(m)∥2\|\mathbf{g}^{(m)}\|_2. As in the previous test, the iterative procedure produces a decrease of the reduced cost functional. Figure 7 shows three density snapshots of the controlled dynamics at t=0.15t = 0.15, t=0.30t = 0.30, and t=0.50t = 0.50 for Ncy=4N_c^y = 4. The control substantially limits transport toward the upper and lower boundaries and retains a larger fraction of the density in the central region. Similar qualitative behavior is observed for Ncy=2N_c^y = 2 and Ncy=6N_c^y = 6. The second row displays the optimal control coefficients over time.

Evolution of the reduced cost functional and discrete gradient norm

Figure 6. Two-stream plasma test. Left: evolution of the reduced cost functional J^(b(m))\widehat{J}(\mathbf{b}^{(m)}). Right: discrete gradient norm ∥g(m)∥2\lVert\mathbf{g}^{(m)} \rVert_2. The curves correspond to Ncx=1N_c^x = 1 and Ncy∈{2,4,6}N_c^y \in\{2,4,6\}.

Density snapshots and optimal control coefficients

Figure 7. Two-stream plasma test. First row: density snapshots of the controlled dynamics at t=0.15t = 0.15, t=0.30t = 0.30, and t=0.50t = 0.50 for Ncy=4N_c^y = 4. Second row: optimal control coefficients b1,k∗(t)b^*_{1,k}(t). The columns in the second row correspond, from left to right, to Ncy=2N_c^y = 2, Ncy=4N_c^y = 4, and Ncy=6N_c^y = 6.

Finally, Figure 8 provides a direct comparison between the instantaneous and optimal controls. The optimized strategy yields a further reduction in the fraction of thermal energy localized near the boundaries.

Relative energy percentage plots comparing instantaneous and optimal controls for $N_c^y=2$, $N_c^y=4$, and $N_c^y=6$

Figure 8. Two-stream plasma test. Comparison of Ebrel(t)\mathcal{E}_b^{\mathrm{rel}}(t) obtained with the instantaneous and optimal controls. From left to right: Ncy=2N_c^y=2, Ncy=4N_c^y=4, and Ncy=6N_c^y=6.

Kelvin–Helmholtz instability

We next consider the development and control of a Kelvin–Helmholtz instability. The spatial domain is

Ωx=[0,36]×[−5,5],\Omega_x = [0,36] \times[-5,5],

with periodic boundary conditions in the xx-direction and reflective boundary conditions in the yy-direction. The initial spatial density is prescribed as

ρ0(x)=1.52πsech⁡(y0.9)[1+ε0cos⁡(3k0x)+ε1sin⁡(k0x)],(75)\rho_0(\mathbf{x})=\frac{1.5}{2\pi}\operatorname{sech}\left(\frac{y}{0.9}\right)\left[1+\varepsilon_0\cos(3k_0x)+\varepsilon_1\sin(k_0x)\right], \tag*{(75)}

where

k0=0.3,ε0=0.5,ε1=0.1.k_0=0.3,\qquad\varepsilon_0=0.5,\qquad\varepsilon_1=0.1.

The initial phase-space distribution is

f0(x,v)=f0+(x,v)χ{y≥0}+f0−(x,v)χ{y<0},(76)f_0(\mathbf{x},\mathbf{v})=f_0^+(\mathbf{x},\mathbf{v})\chi_{\{y\geq0\}}+f_0^-(\mathbf{x},\mathbf{v})\chi_{\{y<0\}}, \tag*{(76)}

with

f0±(x,v)=ρ0(x)2πT0exp⁡(−(vx±ux)2+vy22T0),(77)f_0^\pm(\mathbf{x},\mathbf{v})=\frac{\rho_0(\mathbf{x})}{2\pi T_0}\exp\left(-\frac{(v_x\pm u_x)^2+v_y^2}{2T_0}\right), \tag*{(77)}

where

T0=1,ux=1.T_0=1,\qquad u_x=1.

The dynamics is evolved up to the final time T=5T=5, using Δt=0.05\Delta t=0.05. The density and the electric field are reconstructed on a uniform PIC grid with mx=my=32m_x=m_y=32 nodes, so that Nh=mxmyN_h=m_xm_y. For both control strategies, we set

αx=αv=10,βx=βv=1,γ=10−3,M=50.\alpha_x=\alpha_v=10,\qquad\beta_x=\beta_v=1,\qquad\gamma=10^{-3},\qquad M=50.

with targets y^=v^y=0\hat{y}=\hat{v}_y=0. We first consider the reference dynamics generated by the uniform and time-independent magnetic field Bref(t,x)=0.75B_{\mathrm{ref}}(t,\mathbf{x})=0.75. Figure 9 shows density snapshots at t=2t=2, t=3t=3, and t=4t=4. The deformation of the initially localized layer illustrates the development of the instability. The lower panel on the right reports the corresponding percentage of thermal energy localized within the boundary strip, Ebrel(t)\mathcal{E}_b^{\mathrm{rel}}(t).

Density snapshots at $t=2$, $t=3$, and $t=4$ and relative thermal energy

Figure 9. Kelvin–Helmholtz instability with the prescribed uniform magnetic field Bref=0.75B_{\mathrm{ref}} = 0.75. Density snapshots at t=2t = 2, t=3t = 3, and t=4t = 4 and evolution of the percentage of thermal energy contained in the boundary strip, Ebrel(t)\mathcal{E}^{\mathrm{rel}}_{b}(t).

For the controlled simulations, we set Ncx=1N_c^x=1 and consider Ncy∈{3,5,7}N_c^y\in\{3,5,7\}. Consequently, the control cells extend over the entire xx-direction and the reconstructed magnetic field depends only on yy and tt. Figure 10 reports the convergence history of the optimization procedure. The panels show the reduced cost functional J^(b(m))\widehat{\mathcal{J}}(\mathbf{b}^{(m)}) and the discrete gradient norm ∥g(m)∥2\lVert\mathbf{g}^{(m)}\rVert_2. Figure 11 shows three density snapshots of the controlled dynamics at t=2t=2, t=3t=3, and t=6t=6 for Ncy=5N_c^y=5. Similar qualitative results are obtained for Ncy=3N_c^y=3 and Ncy=7N_c^y=7. The control attenuates the development of the instability and substantially reduces transport toward the upper and lower boundaries. The lower panel show the optimal controls coefficients for Ncy∈{3,5,7}N_c^y \in\{3, 5, 7\}. Finally, Figure 12 provides a direct comparison in terms of relative thermal energy between the instantaneous and optimal controls. As for the previous cases, in most of the cases, the optimized strategy yields a further reduction in the fraction of thermal energy localized near the boundaries.

Evolution of the reduced cost functional and discrete gradient norm

Figure 10. Kelvin–Helmholtz instability. Left: evolution of the reduced cost functional J^(b(m))\widehat{\mathcal{J}}(\mathbf{b}^{(m)}). Right: discrete gradient norm ∥g(m)∥2\lVert\mathbf{g}^{(m)} \rVert_{2}. The curves correspond to Ncx=1N_c^x = 1 and Ncy∈{3,5,7}N_c^y \in\{3, 5, 7\}.

Kelvin–Helmholtz instability density snapshots and optimal control coefficients

Figure 11. Kelvin–Helmholtz instability. Top row: density snapshots of the controlled dynamics at t=2t = 2, t=3t = 3, and t=4t = 4 for Ncy=5N_c^y = 5. Bottom row: optimal control coefficients b1,k∗(t)b_{1,k}^*(t). The columns in the second row correspond, from left to right, to Ncy=3N_c^y = 3, Ncy=5N_c^y = 5, and Ncy=7N_c^y = 7.

Relative energy comparison for instantaneous and optimal controls

Figure 12. Kelvin–Helmholtz instability. Comparison of Eyrel(t)E_y^{\mathrm{rel}}(t) obtained with the instantaneous and optimal controls. From left to right: Ncy=3N_c^y = 3, Ncy=5N_c^y = 5, and Ncy=7N_c^y = 7.

Conclusions

We have developed a full-horizon, open-loop strategy for controlling a 2D–2D Vlasov–Poisson system by means of an externally imposed magnetic field. The field is parameterized by time-dependent coefficients associated with smooth, spatially localized kernels, thereby reducing the infinite-dimensional control problem to the determination of a finite family of control functions. The objective combines the tracking and dispersion of selected position and velocity observables with a penalty on the control effort, while a pointwise bound limits the intensity of the reconstructed magnetic field. Starting from the kinetic optimization problem, we derived the continuous first-order optimality conditions. We then projected this system onto a particle–grid representation and obtained particle adjoint equations. The discretize-then-optimize calculation in Appendix A shows that the same semidiscrete system is recovered when the particle deposition and interpolation operators, the grid quadrature, and the adjoint discrete operators are paired consistently, up to the localization error associated with the particle shape function. The resulting optimization problem is solved by a projected gradient method with Armijo backtracking on a reduced particle ensemble. An instantaneous feedback control, adapted to the same smooth spatial parameterization, provides the initial guess for the full-horizon iteration.

The numerical experiments illustrate the behavior of the method. In the tests considered, the instantaneous control already gives an effective initialization and is often close to a stationary point of the discrete full-horizon problem. Nevertheless, the full-horizon optimization produces a further correction that decreases the reduced objective and generally lowers the normalized thermal-energy diagnostic in the boundary strips.

Several extensions are natural such as the use of collisional or electromagnetic kinetic models, more realistic geometries and coil-generated fields. Moreover, objectives based directly on particle or energy fluxes at material boundaries would bring the framework closer to magnetic confinement applications.

Discretize-then-optimize derivation of the particle–grid optimality system

We provide a discretize-then-optimize derivation of the semidiscrete particle–grid optimality system introduced in § 3. Time remains continuous throughout this appendix. Let

Sℓi(t):=Sh(xℓ−xi(t)),ℓ=1,…,Nh,i=1,…,N,S_{\ell i}(t) := S_h\left(\mathbf{x}_{\ell}-\mathbf{x}_i(t)\right), \qquad\ell=1,\ldots,N_h,\qquad i=1,\ldots,N,

the density and the electric field interpolated to particle ii are

ρℓN=∑i=1NωiSℓi,−(Ahϕ)ℓ=ρℓN−ρˉ,Eih=−∑ℓ=1Nhηℓ(Ghϕ)ℓSℓi.(78)\rho_{\ell}^{N}=\sum_{i=1}^{N}\omega_i S_{\ell i},\qquad-(A_h\phi)_{\ell}=\rho_{\ell}^{N}-\bar{\rho},\qquad\mathbf{E}_i^h=-\sum_{\ell=1}^{N_h}\eta_{\ell}(G_h\phi)_{\ell}S_{\ell i}. \tag*{(78)}

The particle dynamics is therefore

x˙i=vi,v˙i=Fih:=Eih+vi×Bi,Bi=∑k=1Ngbk(t)Kk(xi).(79)\dot{\mathbf{x}}_i=\mathbf{v}_i,\qquad\dot{\mathbf{v}}_i=\mathbf{F}_i^h:=\mathbf{E}_i^h+\mathbf{v}_i\times\mathbf{B}_i,\qquad\mathbf{B}_i=\sum_{k=1}^{N_g}b_k(t)\mathbf{K}_k(\mathbf{x}_i). \tag*{(79)}

Set

MN:=∑i=1Nωi,mqN(t):=1MN∑i=1Nωiψq(xi(t),vi(t)).\mathcal{M}^{N}:=\sum_{i=1}^{N}\omega_i,\qquad m_q^{N}(t):=\frac{1}{\mathcal{M}^{N}}\sum_{i=1}^{N}\omega_i\psi_q(\mathbf{x}_i(t),\mathbf{v}_i(t)).

For q∈I={x,v}q\in\mathcal{I}=\{\mathbf{x},\mathbf{v}\}, define

DqN(t):=αq2∑i=1Nωi∣ψq(xi,vi)−ψ^q∣2+βq2∑i=1Nωi∣ψq(xi,vi)−mqN∣2.(80)\mathcal{D}_q^{N}(t):=\frac{\alpha_q}{2}\sum_{i=1}^{N}\omega_i\left|\psi_q(\mathbf{x}_i,\mathbf{v}_i)-\widehat{\psi}_q\right|^2+\frac{\beta_q}{2}\sum_{i=1}^{N}\omega_i\left|\psi_q(\mathbf{x}_i,\mathbf{v}_i)-m_q^{N}\right|^2. \tag*{(80)}

and

RdiagN(t):=∑k=1Ng∣bk(t)∣2∑i=1NωiKk(xi(t)).(81)\mathcal{R}_{\mathrm{diag}}^{N}(t):=\sum_{k=1}^{N_g}|b_k(t)|^2\sum_{i=1}^{N}\omega_i\mathcal{K}_k(\mathbf{x}_i(t)). \tag*{(81)}

The cost functional at the particle level becomes

JN(X,V,b):=∫0T[∑q∈TDqN(t)+γ2RdiagN(t)] dt+∑q∈TDqN(T),(82)\mathcal{J}_N(\mathbf{X},\mathbf{V},\mathbf{b}) := \int_0^T \left[\sum_{q\in\mathcal{T}} \mathcal{D}_q^N(t) + \frac{\gamma}{2}\mathcal{R}_{\mathrm{diag}}^N(t)\right]\,dt + \sum_{q\in\mathcal{T}}\mathcal{D}_q^N(T), \tag*{(82)}

with X=(x1,…,xN)\mathbf{X}=(\mathbf{x}_1,\ldots,\mathbf{x}_N) and V=(v1,…,vN)\mathbf{V}=(\mathbf{v}_1,\ldots,\mathbf{v}_N). Let GiN\mathcal{G}_i^N denote the particle contribution

GiN(t):=∑q∈T[αq2∣ψq(xi,vi)−ψ^q∣2+βq2∣ψq(xi,vi)−mqN∣2]\mathcal{G}_i^N(t) := \sum_{q\in\mathcal{T}}\left[\frac{\alpha_q}{2}\left|\psi_q(\mathbf{x}_i,\mathbf{v}_i)-\widehat{\psi}_q\right|^2+\frac{\beta_q}{2}\left|\psi_q(\mathbf{x}_i,\mathbf{v}_i)-m_q^N\right|^2\right]
+γ2∑k=1Ng∣bk(t)∣2Kk(xi).(83)\qquad+\frac{\gamma}{2}\sum_{k=1}^{N_g}\left|\mathbf{b}_k(t)\right|^2\mathcal{K}_k(\mathbf{x}_i). \tag*{(83)}

Thus

∇xiGiN:=1ωi∇xi(∑j=1NωjGjN),∇viGiN:=1ωi∇vi(∑j=1NωjGjN),\nabla_{\mathbf{x}_i}\mathcal{G}_i^N:=\frac{1}{\omega_i}\nabla_{\mathbf{x}_i}\left(\sum_{j=1}^{N}\omega_j\mathcal{G}_j^N\right),\qquad \nabla_{\mathbf{v}_i}\mathcal{G}_i^N:=\frac{1}{\omega_i}\nabla_{\mathbf{v}_i}\left(\sum_{j=1}^{N}\omega_j\mathcal{G}_j^N\right),

and the choices

ψx=y,ψv=vy,ψ^x=y^,ψ^v=v^y\psi_x=y,\qquad\psi_v=v_y,\qquad\widehat{\psi}_x=\widehat{y},\qquad\widehat{\psi}_v=\widehat{v}_y

give precisely the expressions in (35). The particle–grid Lagrangian is

LN:=JN(X,V,b)\mathcal{L}_N:=\mathcal{J}_N(\mathbf{X},\mathbf{V},\mathbf{b})
+∑i=1Nωi∫0T[pi⋅(x˙i−vi)+qi⋅(v˙i−Eih−vi×Bi)] dt\qquad+\sum_{i=1}^{N}\omega_i\int_0^T\left[\mathbf{p}_i\cdot(\dot{\mathbf{x}}_i-\mathbf{v}_i)+\mathbf{q}_i\cdot\left(\dot{\mathbf{v}}_i-\mathbf{E}_i^h-\mathbf{v}_i\times\mathbf{B}_i\right)\right]\,dt
+∫0T⟨λ^,−Ahϕ−ρN+ρ‾h⟩h dt.(84)\qquad+\int_0^T\left\langle\widehat{\boldsymbol{\lambda}},-A_h\phi-\boldsymbol{\rho}^N+\overline{\boldsymbol{\rho}}_h\right\rangle_h\,dt. \tag*{(84)}

Here ρ‾h\overline{\boldsymbol{\rho}}_h denotes the grid vector whose entries are all equal to ρ‾\overline{\rho} and ⟨a,c⟩h:=∑ℓ=1Nhηℓaℓ⋅cℓ\langle\mathbf{a},\mathbf{c}\rangle_h:=\sum_{\ell=1}^{N_h}\eta_\ell\mathbf{a}_\ell\cdot\mathbf{c}_\ell. Stationarity with respect to pi\mathbf{p}_i, qi\mathbf{q}_i, and λ^\widehat{\boldsymbol{\lambda}} recovers (A.1)–(A.2). To compute the variation with respect to the grid potential, define

Qℓ:=∑i=1NωiSℓiqi,λih:=∑ℓ=1Nhηℓλ~ℓSℓi.(85)Q_\ell:=\sum_{i=1}^{N}\omega_iS_{\ell i}\mathbf{q}_i,\qquad \lambda_i^h:=\sum_{\ell=1}^{N_h}\eta_\ell\widetilde{\lambda}_\ell S_{\ell i}. \tag*{(85)}

Using the grid inner product and the adjoint operators Ah∗A_h^* and Gh∗G_h^*, stationarity with respect to ϕ\phi gives

−Ah∗λ^+Gh∗Q=0,(86)-A_h^*\widehat{\boldsymbol{\lambda}}+G_h^*Q=0, \tag*{(86)}

which coincides with (36).

Integrating the terms containing x˙i\dot{\mathbf{x}}_i and v˙i\dot{\mathbf{v}}_i by parts in time, and using that the initial particle states are fixed, stationarity with respect to xi\mathbf{x}_i and vi\mathbf{v}_i yields

−p˙i=(∇xiFih)Tqi−∇xiGiN+∇xiλih,-\dot{\mathbf{p}}_i=\left(\nabla_{\mathbf{x}_i}F_i^h\right)^T\mathbf{q}_i-\nabla_{\mathbf{x}_i}\mathcal{G}_i^N+\nabla_{\mathbf{x}_i}\lambda_i^h,
−q˙i=pi−∇viGiN−qi×Bi,i=1,…,N.(87)-\dot{\mathbf{q}}_i=\mathbf{p}_i-\nabla_{\mathbf{v}_i}\mathcal{G}_i^N-\mathbf{q}_i\times\mathbf{B}_i,\qquad i=1,\ldots,N. \tag*{(87)}

Moreover,

∇xiλih=∑ℓ=1Nhηℓλ~ℓ∇xiSh(xℓ−xi),\nabla_{\mathbf{x}_i}\lambda_i^h=\sum_{\ell=1}^{N_h}\eta_\ell\widetilde{\lambda}_\ell\nabla_{\mathbf{x}_i}S_h(\mathbf{x}_\ell-\mathbf{x}_i),

so that (A.10) is exactly (38). The terminal conditions are

pi(T)=−∇xiGT,iN,qi(T)=−∇viGT,iN,(88)\mathbf{p}_i(T)=-\nabla_{\mathbf{x}_i}G^N_{T,i},\qquad\mathbf{q}_i(T)=-\nabla_{\mathbf{v}_i}G^N_{T,i}, \tag*{(88)}

where GT,iNG^N_{T,i} contains only the terminal tracking and dispersion terms, and no control regularization. Finally, the derivative of (A.7) with respect to bkb_k is

gkN(t)=∑i=1Nωi[γbk(t)+vi(t)×qi(t)]Kk(xi(t)),k=1,…,Ng.(89)g^N_k(t)=\sum_{i=1}^{N}\omega_i\left[\gamma b_k(t)+\mathbf{v}_i(t)\times\mathbf{q}_i(t)\right]K_k(\mathbf{x}_i(t)),\qquad k=1,\ldots,N_g. \tag*{(89)}

At the particle–grid level, we enforce the field bound at the spatial grid nodes and set

Uadh:={b∈L2(0,T;RNg):∣∑k=1Ngbk(t)Kk(xℓ)∣≤Bmax⁡ for a.e. t,ℓ=1,…,Nh}.(90)\mathcal{U}^h_{\mathrm{ad}}:=\left\{\mathbf{b}\in L^2(0,T;\mathbb{R}^{N_g}):\left|\sum_{k=1}^{N_g}b_k(t)K_k(\mathbf{x}_\ell)\right|\le B_{\max}\ \text{for a.e. }t,\quad\ell=1,\ldots,N_h\right\}. \tag*{(90)}

For a bounded control, stationarity is then expressed by the variational inequality

∑k=1Ng∫0T(gkN)∗(t)(ck(t)−bk∗(t)) dt≥0for every c∈Uadh.(91)\sum_{k=1}^{N_g}\int_0^T(g^N_k)^*(t)(c_k(t)-b_k^*(t))\,\mathrm{d}t\ge0\qquad\text{for every }\mathbf{c}\in\mathcal{U}^h_{\mathrm{ad}}. \tag*{(91)}

Thus, the discretize-then-optimize calculation recovers (38), (36), (39), and (41). For the particle objective (A.5), the calculation is exact at the particle–grid semidiscrete level when the same discretization are employed. Its comparison with the projected continuous system additionally involves replacing convolutions against ShS_h by evaluations along the particle trajectories. For a shape function of order rr, these localization discrepancies are of order O(hr)O(h^r); in particular, this accounts for the remainder displayed in (41).

Relation with the fully discrete adjoint

The derivation above is semidiscrete, with time kept continuous. Indeed, in Section 4, the gradient (55) is obtained by discretizing in time the semidiscrete optimality system. We now compare it with the gradient obtained by differentiating the fully discrete state problem. The semi-implicit forward scheme (50) can then be written as

{vin+1=Vin:=(Ain)−1(vin+Δt Eih,n),xin+1=Xin:=xin+Δt vin,(92)\begin{cases} \mathbf{v}_i^{n+1}=\mathcal{V}_i^n:=(A_i^n)^{-1}\left(\mathbf{v}_i^n+\Delta t\,\mathbf{E}_i^{h,n}\right),\\ \mathbf{x}_i^{n+1}=\mathcal{X}_i^n:=\mathbf{x}_i^n+\Delta t\,\mathbf{v}_i^n, \tag*{(92)} \end{cases}

where

J=(01−10),Ain:=I−Δt BinJ.\mathbb{J}=\begin{pmatrix}0&1\\-1&0\end{pmatrix},\qquad A_i^n:=I-\Delta t\,B_i^n\mathbb{J}.

Together with the discrete Poisson problem, this defines the fully discrete state map

(Xn,Vn,bn)⟼(Xn+1,Vn+1).(X^n,V^n,b^n)\longmapsto(X^{n+1},V^{n+1}).

Using the same time quadrature as in Section 4, the corresponding discrete Lagrangian may be written as

LN,h,Δt:=JN,h,Δt+∑n=0Nt−1∑i=1Nsω^i[pi,DTOn+1⋅(xin+1−Xin)+qi,DTOn+1⋅(vin+1−Vin)].(93)\mathcal{L}_{N,h,\Delta t}:=J_{N,h,\Delta t} +\sum_{n=0}^{N_t-1}\sum_{i=1}^{N_s}\widehat{\omega}_i\left[\mathbf{p}_{i,\mathrm{DTO}}^{n+1}\cdot(\mathbf{x}_i^{n+1}-\mathcal{X}_i^n)+\mathbf{q}_{i,\mathrm{DTO}}^{n+1}\cdot(\mathbf{v}_i^{n+1}-\mathcal{V}_i^n)\right]. \tag*{(93)}

Since

Xin=xin+Δt vin,\mathcal{X}_i^n = \mathbf{x}_i^n + \Delta t\,\mathbf{v}_i^n,

we introduce

ζin:=(Ain)−T(qi,DTOn+1+Δt pi,DTOn+1),(94)\boldsymbol{\zeta}_i^n := (A_i^n)^{-T}\left(\mathbf{q}_{i,\mathrm{DTO}}^{n+1}+\Delta t\,\mathbf{p}_{i,\mathrm{DTO}}^{n+1}\right), \tag*{(94)}

which collects the contribution of the position and velocity adjoints through the semi-implicit velocity update. Thus, the discrete adjoint writes as

pi,DTOn=pi,DTOn+1+Δt[(∇xiFih,n,+)Tζin−∇xiGiN,n+Lin],qi,DTOn=ζin−Δt ∇viGiN,n,(95)\begin{aligned} \mathbf{p}_{i,\mathrm{DTO}}^n &= \mathbf{p}_{i,\mathrm{DTO}}^{n+1}+\Delta t\left[\left(\nabla_{\mathbf{x}_i}F_i^{h,n,+}\right)^T\boldsymbol{\zeta}_i^n-\nabla_{\mathbf{x}_i}\mathcal{G}_i^{N,n}+L_i^n\right],\\ \mathbf{q}_{i,\mathrm{DTO}}^n &= \boldsymbol{\zeta}_i^n-\Delta t\,\nabla_{\mathbf{v}_i}\mathcal{G}_i^{N,n}, \tag*{(95)} \end{aligned}

with terminal conditions

pi,DTONt=−∇xiGT,iN,qi,DTONt=−∇viGT,iN,\mathbf{p}_{i,\mathrm{DTO}}^{N_t}=-\nabla_{\mathbf{x}_i}\mathcal{G}_{T,i}^N,\qquad\mathbf{q}_{i,\mathrm{DTO}}^{N_t}=-\nabla_{\mathbf{v}_i}\mathcal{G}_{T,i}^N,

and where

Fih,n,+:=Eih,n+vin+1×Bin.F_i^{h,n,+}:=\mathbf{E}_i^{h,n}+\mathbf{v}_i^{n+1}\times\mathbf{B}_i^n.

Here LinL_i^n denotes the contribution associated with the dependence of the electric field on the particle density, obtained from the same discrete Poisson-adjoint relation derived in (A.10) and evaluated at time tnt^n. Finally, stationarity with respect to bknb_k^n gives, for the diagonal regularization (12),

gk,nDTO=∑i=1Nsω^i[γbkn+vin+1×ζin]Kk(xin).(96)g_{k,n}^{\mathrm{DTO}}=\sum_{i=1}^{N_s}\widehat{\omega}_i\left[\gamma b_k^n+\mathbf{v}_i^{n+1}\times\boldsymbol{\zeta}_i^n\right]\mathcal{K}_k(\mathbf{x}_i^n). \tag*{(96)}

We observe that since

(Ain)−T=I−Δt BinJ+O(Δt2),vin+1=vin+O(Δt),(A_i^n)^{-T}=I-\Delta t\,B_i^nJ+\mathcal{O}(\Delta t^2),\qquad\mathbf{v}_i^{n+1}=\mathbf{v}_i^n+\mathcal{O}(\Delta t),

and the adjoint discretizations in (A.17) and (A.19) are such that

ζin=qin+O(Δt),vin+1×ζin=vin×qin+O(Δt).\boldsymbol{\zeta}_i^n=\mathbf{q}_i^n+\mathcal{O}(\Delta t),\qquad\mathbf{v}_i^{n+1}\times\boldsymbol{\zeta}_i^n=\mathbf{v}_i^n\times\mathbf{q}_i^n+\mathcal{O}(\Delta t).

It follows that the gradient used in § 4 is a first-order consistent approximation of the fully discrete gradient, indeed we have

gk,nDTO=gk,n+O(Δt),(97)g_{k,n}^{\mathrm{DTO}}=g_{k,n}+\mathcal{O}(\Delta t), \tag*{(97)}

where gk,ng_{k,n} is the gradient defined in (55).

Computation of the instantaneous control

In the following, we will briefly describe how to compute the baseline instantaneous feedback control B0b(t,x)B_0^b(t,\mathbf{x}), referring to [2, 3] for a complete description of the methodology. Accordingly, the initial guess for the open-loop control Bb(t,x)B_b(t,\mathbf{x}) in the optimization procedure is chosen to be this instantaneous feedback law.

We consider the following minimization problem at the particle level:

min⁡b0∈UadJ(b0)=∫tntn+1[GiN(t)+γ2∑k=1Ng∣bk0(t)∣2Kk(xi)] dt,s.t.  xref‘(3.14)‘,(98)\min_{\mathbf{b}^0\in\mathcal{U}_{ad}}\mathcal{J}(\mathbf{b}^0)=\int_{t^n}^{t^{n+1}}\left[\mathcal{G}_i^N(t)+\frac{\gamma}{2}\sum_{k=1}^{N_g}\left|b_k^0(t)\right|^2\mathcal{K}_k(\mathbf{x}_i)\right]\,\mathrm{d}t,\qquad\text{s.t. }\ {xref}`(3.14)`, \tag*{(98)}

with GiN(t)\mathcal{G}^{N}_{i}(t) and Uad\mathcal{U}_{\mathrm{ad}} defined as in (A.6) and (A.13) respectively, and γ>0\gamma> 0 a penalization coefficient.

We introduce an explicit first order discretization of the characteristic curves in (31) over the time interval [0,T][0,T], with step size Δt>0\Delta t > 0 as follows:

{xin+1=xin+Δtvin+1,vin+1=vin+Δt(Ein+vin×Bi0,n),(99)\begin{cases} \mathbf{x}_{i}^{n+1} = \mathbf{x}_{i}^{n} + \Delta t \mathbf{v}_{i}^{n+1},\\ \mathbf{v}_{i}^{n+1} = \mathbf{v}_{i}^{n} + \Delta t(\mathbf{E}_{i}^{n} + \mathbf{v}_{i}^{n} \times B_{i}^{0,n}), \tag*{(99)} \end{cases}

for any i=1,…,Ni = 1,\ldots,N, and where

Bi0,n=∑k=1Ngbk0,nKk(xi),(100)B_{i}^{0,n} = \sum_{k=1}^{N_g} b_{k}^{0,n}\mathcal{K}_{k}(\mathbf{x}_{i}), \tag*{(100)}

with K\mathcal{K} denoting the spatial basis functions defined as in (60). Following [2, 3], we introduce the scaling

αxΔt=αxΔt,βxΔt=βxΔt,γΔt=γΔt,(101)\alpha_{x}^{\Delta t} = \frac{\alpha_{x}}{\Delta t}, \qquad \beta_{x}^{\Delta t} = \frac{\beta_{x}}{\Delta t}, \qquad \gamma\Delta t = \gamma\Delta t, \tag*{(101)}

while αv\alpha_{v} and βv\beta_{v} are kept fixed. Since the yy-component of the velocity update satisfies

vy,in+1=vy,in+ΔtEy,in−Δtvx,inBi0,n,v_{y,i}^{n+1} = v_{y,i}^{n} + \Delta t E_{y,i}^{n} - \Delta t v_{x,i}^{n} B_{i}^{0,n},

we have

∂vy,in+1∂bj0,n=−Δt vx,inKj(xin),∂yin+1∂bj0,n=−Δt2vx,inKj(xin).\frac{\partial v_{y,i}^{n+1}}{\partial b_{j}^{0,n}} = -\Delta t\,v_{x,i}^{n}\mathcal{K}_{j}(\mathbf{x}_{i}^{n}), \qquad \frac{\partial y_{i}^{n+1}}{\partial b_{j}^{0,n}} = -\Delta t^{2}v_{x,i}^{n}\mathcal{K}_{j}(\mathbf{x}_{i}^{n}).

Using (B.4) and considering the limit Δt→0\Delta t \to0 with tn=nΔt→tt^{n} = n\Delta t \to t, the first-order optimality conditions yield the instantaneous feedback law

H(t)b0(t)=F(t),(102)H(t)\mathbf{b}^{0}(t) = F(t), \tag*{(102)}

where H(t)H(t) is a diagonal matrix based on the assumption (11), and we have

Hjj(t)=γ∑i=1NωiKj(xi(t)),Fj(t)=∑i=1Nωivx,i(t)Ri0(t)Kj(xi(t)),(103)H_{jj}(t) = \gamma\sum_{i=1}^{N}\omega_{i}\mathcal{K}_{j}(\mathbf{x}_{i}(t)), \qquad F_{j}(t) = \sum_{i=1}^{N}\omega_{i}v_{x,i}(t)\mathcal{R}_{i}^{0}(t)\mathcal{K}_{j}(\mathbf{x}_{i}(t)), \tag*{(103)}

with

Ri0(t)=αx(yi(t)−y^)+βx(yi(t)−y‾N(t))+αv(vy,i(t)−v^y)+βv(vy,i(t)−v‾yN(t)).(104)\mathcal{R}_{i}^{0}(t) = \alpha_{x}\left(y_{i}(t)-\widehat{y}\right) + \beta_{x}\left(y_{i}(t)-\overline{y}^{N}(t)\right) + \alpha_{v}\left(v_{y,i}(t)-\widehat{v}_{y}\right) + \beta_{v}\left(v_{y,i}(t)-\overline{v}_{y}^{N}(t)\right). \tag*{(104)}

Equation (B.5) provides the unconstrained instantaneous control, if the constraint in Uadh\mathcal{U}_{\mathrm{ad}}^{h} is active, the corresponding constrained quadratic problem is solved instead. Once determined through this first-order framework, the instantaneous control coefficients b0(t)\mathbf{b}^{0}(t) are applied within the second-order scheme to advance the particle trajectories.

Acknowledgement

This work has been written within the activities of GNCS and GNFM groups of INdAM (Italian National Institute of High Mathematics). The research of FF and LP was supported by the Italian Ministry of University and Research (MUR) through the Fondo Italiano per la Scienza 2022–2023 (Bando FIS 2), under Decree No. 1236 of 01/08/2023, within the FIS2023-01334 Advanced Grant “Tackling Complexity: Advanced Numerical Approaches for Multiscale Systems with Uncertainties” (ADAMUS). The work of LP was supported by the Royal Society Wolfson Fellowship “Uncertainty quantification, data-driven simulations and learning of multiscale complex systems governed by PDEs”. GA and GD acknowledge the support of the PRIN 2026 - 2026PYFEZ9, “Emergent Structures in Kinetic Theory: from Gases to Plasmas” funded with the contribution of MUR.

References

References

  1. [1]G. Albi, Y.-P. Choi, M. Fornasier, and D. Kalise. Mean field control hierarchy. Applied Mathematics & Optimization, 76(1):93–135, 2017.arxiv.org/abs/1608.01728
  2. [2]G. Albi, G. Dimarco, F. Ferrarese, and L. Pareschi. Instantaneous control strategies for magnetically confined fusion plasma. Journal of Computational Physics, 527:113804, 2025.arxiv.org/abs/2403.16254
  3. [3]G. Albi, G. Dimarco, F. Ferrarese, and L. Pareschi. Robust feedback control of collisional plasma dynamics in presence of uncertainties. Journal of Computational Physics, 546:114512, 2026.arxiv.org/abs/2505.19992
  4. [4]R. Bailo, M. Bongini, J. A. Carrillo, and D. Kalise. Optimal consensus control of the Cucker-Smale model. IFAC-PapersOnLine, 51(13):1–6, 2018.DOI
  5. [5]J. Bartsch, P. Knopf, S. Scheurer, and J. Weber. Controlling a Vlasov-Poisson plasma by a particle-in-cell method based on a Monte Carlo framework. SIAM Journal on Control and Optimization, 62(4):1977–2011, 2024.
  6. [6]J. A. Bittencourt. Fundamentals of plasma physics. Springer Science & Business Media, 2013.
  7. [7]J. Blum. Numerical simulation and optimal control in plasma physics. Wiley/Gauthier-Villars Series in Modern Applied Mathematics. John Wiley & Sons, Ltd., Chichester; Gauthier-Villars, Montrouge, 1989.
  8. [8]R. Caflisch, C. Wang, G. Dimarco, B. Cohen, and A. Dimit s. A hybrid method for accelerated simulation of Coulomb collisions in a plasma. Multiscale Modeling & Simulation, 7(2):865–887, 2008.
  9. [9]S. Caprino, G. Cavallaro, and C. Marchioro. Time evolution of a Vlasov-Poisson plasma with magnetic confinement. Kinetic & Related Models, 5(4):729–742, 2012.DOI
  10. [10]E. Chacon-Golcher, S. A. Hirstoaga, and M. Lutz. Optimization of Particle-In-Cell simulations for Vlasov-Poisson system with strong magnetic field. ESAIM: Proceedings and Surveys, 53:177–190, 2016.
  11. [11]P. Chen, R. Jorge, Q. Li, and Y. Yue. Control of a Uniformly Magnetized Plasma with External Electric Fields. Journal of Statistical Physics, 193(40), 2026.arxiv.org/abs/2509.07988
  12. [12]C.-Z. Cheng and G. Knorr. The integration of the Vlasov equation in configuration space. Journal of Computational Physics, 22(3):330–351, 1976.
  13. [13]A. R. Conn, N. I. Gould, and P. L. Toint. Trust Region Methods, SIAM. MPS, 2000.DOI
  14. [14]G.-H. Cottet and P.-A. Raviart. Particle methods for the one-dimensional Vlasov-Poisson equations. SIAM Journal on Numerical Analysis, 21(1):52–76, 1984.DOI
  15. [15]A. Crestetto, N. Crouseilles, and M. Lemou. Kinetic/fluid micro-macro numerical schemes for Vlasov-Poisson-BGK equation using particles. Kinetic and Related Models, 5(4):787–816, 2012.DOI
  16. [16]N. Crouseilles, G. Dimarco, and M.-H. Vignal. Multiscale schemes for the BGK-Vlasov-Poisson system in the quasi-neutral and fluid limits. Stability analysis and first order schemes. Multiscale Model. Simul., 14(1):65–95, 2016.DOI
  17. [17]N. Crouseilles, M. Mehrenberger, and E. Sonnendrücker. Conservative semi-Lagrangian schemes for Vlasov equations. Journal of Computational Physics, 229(6):1927–1953, 2010.DOI
  18. [18]N. Crouseilles, T. Respaud, and E. Sonnendrücker. A forward semi-Lagrangian method for the numerical solution of the Vlasov equation. Comput. Phys. Comm., 180(10):1730–1745, 2009.DOI
  19. [19]P. Degond. Asymptotic-preserving schemes for fluid models of plasmas. In Numerical models for fusion, volume 39/40 of Panor. Synthèses, pages 1–90. Soc. Math. France, Paris, 2013.
  20. [20]P. Degond and F. Deluzet. Asymptotic-preserving methods and multiscale models for plasma physics. Journal of Computational Physics, 336:429–457, 2017.DOI
  21. [21]J. Degrave, F. Felici, J. Buchli, M. Neunert, B. Tracey, F. Carpanese, T. Ewalds, R. Hafner, A. Abdolmaleki, D. de las Casas, et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602(7897):414–419, 2022.DOI
  22. [22]F. Deluzet, G. Fubiani, L. Garrigues, C. Guillet, and J. Narski. Efficient parallelization for 3d-3v sparse grid particle-in-cell: shared memory architectures. Journal of Computational Physics, 480:112022, 2023.DOI
  23. [23]G. Dimarco, Q. Li, L. Pareschi, and B. Yan. Numerical methods for plasma physics in collisional regimes. Journal of Plasma Physics, 81(1):305810106, 2015.DOI
  24. [24]G. Dimarco, L. Mieussens, and V. Rispoli. An asymptotic preserving automatic domain decomposition method for the Vlasov–Poisson–BGK system with applications to plasmas. Journal of Computational Physics, 274:122–139, 2014.DOI
  25. [25]G. Dimarco and L. Pareschi. Numerical methods for kinetic equations. Acta Numerica, 23:369–520, 2014.DOI
  26. [26]L. Einkemmer, Q. Li, C. Mouhot, and Y. Yue. Control of instability in a Vlasov-Poisson system through an external electric field. Journal of Computational Physics, 530:113904, 2025.arxiv.org/abs/2407.15008
  27. [27]L. Einkemmer, Q. Li, L. Wang, and Y. Yang. Suppressing instability in a Vlasov–Poisson system by an external electric field through constrained optimization. Journal of Computational Physics, 498:112662, 2024.arxiv.org/abs/2305.17994
  28. [28]A. Fasoli, S. Brunner, W. A. Cooper, J. P. Graves, P. Ricci, O. Sauter, and L. Villard. Computational challenges in magnetic-confinement fusion physics. Nature Physics, 12(5):411–423, 2016.DOI
  29. [29]F. Ferrarese. Control strategies for magnetized plasma: a polar coordinates framework. arXiv preprint arXiv:2511.10214, 2025.arxiv.org/abs/2511.10214
  30. [30]F. Filbet and L. M. Rodrigues. Asymptotically stable particle-in-cell methods for the Vlasov–Poisson system with a strong external magnetic field. SIAM Journal on Numerical Analysis, 54(2):1120–1146, 2016.
  31. [31]F. Filbet and L. M. Rodrigues. Asymptotically preserving particle-in-cell methods for inhomogeneous strongly magnetized plasmas. SIAM Journal on Numerical Analysis, 55(5):2416–2443, 2017.DOI
  32. [32]F. Filbet and E. Sonnendrücker. Comparison of Eulerian Vlasov solvers. Computer Physics Communications, 150(3):247–266, 2003.
  33. [33]F. Filbet, E. Sonnendrücker, and P. Bertrand. Conservative numerical schemes for the Vlasov equation. Journal of Computational Physics, 172(1):166–187, 2001.DOI
  34. [34]F. Filbet and C. Yang. Numerical Simulations to the Vlasov-Poisson System with a Strong Magnetic Field. arXiv:1805.10888 / hal-01800438, 2018.arxiv.org/abs/1805.10888
  35. [35]J. P. Freidberg. Plasma physics and fusion energy. Cambridge university press, 2008.DOI
  36. [36]P. R. Garabedian. Computational mathematics and physics of fusion reactors. Proceedings of the National Academy of Sciences, 100(24):13741–13745, 2003.pubmed.ncbi.nlm.nih.gov/14614129
  37. [37]V. Grandgirard and Y. Sarazin. Gyrokinetic simulations of magnetic fusion plasmas. Numerical models for fusion, 39:91–176, 2013.
  38. [38]A. Gu, Y. He, and Y. Sun. Hamiltonian particle-in-cell methods for Vlasov–Poisson equations. Journal of Computational Physics, 467:111472, 2022.DOI
  39. [39]E. Hairer and C. Lubich. Energy behaviour of the Boris method for charged-particle dynamics. BIT Numerical Mathematics, 58(4):969–979, 2018.
  40. [40]J. Hesthaven, C. Pagliantini, and N. Ripamonti. Adaptive symplectic model order reduction of parametric particle-based Vlasov–Poisson equation. Mathematics of Computation, 93(347):1153–1202, 2024.arxiv.org/abs/2201.05555
  41. [41]C. T. Kelley. Iterative methods for optimization. SIAM, 1999.DOI
  42. [42]P. Knopf and J. Weber. Optimal control of a Vlasov–Poisson plasma by fixed magnetic field coils. Applied Mathematics & Optimization, 81:961–988, 2020.
  43. [43]J. Lu, L. Wang, and J. Calder. Controlling instability in the Vlasov-Poisson system through moment-based optimization. arXiv preprint arXiv:2508.18412, 2025.arxiv.org/abs/2508.18412
  44. [44]J. Lu, L. Wang, and J. Calder. Dynamical feedback control with operator learning for the vlasov-poisson system. arXiv preprint arXiv:2509.23063, 2025.arxiv.org/abs/2509.23063
  45. [45]A. Medaglia, L. Pareschi, and M. Zanella. A DSMC-PIC coupling method for the Vlasov–Maxwell–Landau system. Journal of Computational Physics, page 115160, 2026.
  46. [46]S. Ravindran. Real-time computational algorithm for optimal control of an MHD flow system. SIAM Journal on Scientific Computing, 26(4):1369–1388, 2005.DOI
  47. [47]J. A. Rossmanith and D. C. Seal. A positivity-preserving high-order semi-Lagrangian discontinuous Galerkin scheme for the Vlasov-Poisson equations. Journal of Computational Physics, 230(16):6203–6232, 2011.DOI
  48. [48]E. Sonnendrücker, J. Roche, P. Bertrand, and A. Ghizzo. The semi-Lagrangian method for the numerical resolution of the Vlasov equation. Journal of Computational Physics, 149(2):201–220, 1999.DOI
  49. [49]J. Weber. Optimal control of the two-dimensional Vlasov-Maxwell system. ESAIM: Control, Optimisation and Calculus of Variations, 27:S19, 2021.DOI
  50. [50]J. Wesson and D. J. Campbell. Tokamaks, volume 149. Oxford university press, 2011.
  51. [51]C. Zhang and I. M. Gamba. A conservative scheme for Vlasov Poisson Landau modeling collisional plasmas. Journal of Computational Physics, 340:470–497, 2017.DOI
  52. [52]Y. Zhu and S. Jin. The Vlasov–Poisson–Fokker–Planck system with uncertainty and a one-dimensional asymptotic preserving method. Multiscale Modeling & Simulation, 15(4):1502–1529, 2017.DOI

Paper details

vlnc:W_k3mpmf7k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents