Spatially Homogeneous Landau Equation

• Spatially Homogeneous Landau Equation is a kinetic PDE that models particle interactions in velocity space, incorporating conservation laws and entropy dissipation.
• It employs spectral analysis, harmonic decomposition, and probabilistic methods to establish global weak solutions and ultra-analytic regularization.
• The model is versatile across hard, Maxwellian, and soft regimes, enabling gradient-flow formulations, numerical approximations, and stochastic particle representations.

The spatially homogeneous Landau equation is the velocity-space evolution equation

$\partial_t f(t,v)=Q(f,f)(v),$

for a one-particle distribution function $f=f(t,v)\ge 0$, where the collision operator acts only in the velocity variable $v\in \mathbb R^3$. It is a variation of the Boltzmann equation in the grazing collision limit, and its interaction kernel is

$$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$

with hard potentials for $\gamma>0$, Maxwell molecules for $\gamma=0$, and soft potentials for $\gamma<0$. The modern theory combines weak formulations, linearization around Maxwellians, spectral decompositions, entropy dissipation, probabilistic particle limits, and numerical approximation; in particular, for Maxwellian molecules the Cauchy problem with Shubin-class data admits a global weak solution and a Gelfand–Shilov smoothing effect based on an “almost diagonal” spectral structure (Jang et al., 13 Jan 2025, Li et al., 2017).

1. Equation, weak structure, and equilibria

In strong form, the spatially homogeneous Landau equation can be written as

$$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$

Equivalent formulations introduce the coefficients

$A[f](v)=a*f,\qquad B[f](v)=b*f,$

so that

$$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$

with

$f=f(t,v)\ge 0$0

In weak form, for $f=f(t,v)\ge 0$1,

$f=f(t,v)\ge 0$2

This weak structure makes the conservation laws transparent: the choices $f=f(t,v)\ge 0$3, $f=f(t,v)\ge 0$4, and $f=f(t,v)\ge 0$5 yield conservation of mass, momentum, and energy, while the Boltzmann entropy

$f=f(t,v)\ge 0$6

is dissipated along the flow (Jang et al., 13 Jan 2025, Carrapatoso, 2013).

After Galilean-energy normalization, the equilibrium is the Maxwellian

$f=f(t,v)\ge 0$7

or, more generally,

$f=f(t,v)\ge 0$8

Accordingly, the homogeneous Landau equation is both a nonlinear diffusion equation and an entropy-dissipating relaxation dynamics toward Maxwellian equilibrium. This dual interpretation underlies spectral, variational, and probabilistic approaches (Carrapatoso, 2013, Carrillo et al., 2020).

2. Linearization, harmonic analysis, and the Maxwellian-molecule spectral theory

A standard close-to-equilibrium parametrization writes

$f=f(t,v)\ge 0$9

so that the fluctuation $v\in \mathbb R^3$0 solves

$v\in \mathbb R^3$1

or, in the notation of perturbative hard/soft-potential works,

$v\in \mathbb R^3$2

For Maxwellian molecules, the linearized Landau operator

$v\in \mathbb R^3$3

is self-adjoint and nonnegative on $v\in \mathbb R^3$4, and

$v\in \mathbb R^3$5

The harmonic oscillator

$v\in \mathbb R^3$6

provides the natural spectral scale. In the Hermite basis $v\in \mathbb R^3$7, $v\in \mathbb R^3$8, and the Shubin space

$v\in \mathbb R^3$9

is the appropriate functional framework. When $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$0, it contains all probability measures, including Dirac masses (Li et al., 2017).

Li and Xu diagonalize $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$1 on the basis $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$2, where $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$3 is built from radial Laguerre functions and spherical harmonics: $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$4 with explicit eigenvalues

$$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$5

The key nonlinear fact is that the matrix elements of $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$6 couple only modes satisfying

$$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$7

so the nonlinear term is “almost diagonal” in this basis. This is the precise mechanism behind the statement that, for Maxwellian molecules, the nonlinear Landau operator is almost linear (Li et al., 2017).

The main existence theorem in this setting states that if $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$8, $$a_{ij}(z)=|z|^{\gamma+2}\Bigl(\delta_{ij}-\frac{z_i z_j}{|z|^2}\Bigr),$$9, and the low-mode projection satisfies

$\gamma>0$0

then the Cauchy problem admits a global weak solution

$\gamma>0$1

Moreover, for any $\gamma>0$2,

$\gamma>0$3

with quantitative bound

$\gamma>0$4

Thus the solution is ultra-analytic and has exponential decay for positive times (Li et al., 2017).

Earlier close-to-equilibrium work had already identified ultra-analytic Gelfand–Shilov smoothing for Maxwellian molecules, showing that for $\gamma>0$5 the fluctuation and its Fourier transform belong to $\gamma>0$6, with Hermite-mode bounds controlled by $\gamma>0$7 (Morimoto et al., 2013). Taken together, these results place the homogeneous Maxwellian-molecule problem in a fully spectral regime where singular initial data, weak solutions, and ultra-analytic regularization can be treated in a common harmonic-oscillator framework.

3. Regularization, analyticity, and relaxation in hard, Maxwellian, and soft regimes

For hard potentials $\gamma>0$8, the long-time theory is organized by the spectral gap of the linearized operator around $\gamma>0$9. Baranger–Mouhot established a constructive spectral gap in $\gamma=0$0, and the semigroup decay can be extended from that Hilbert space to weighted $\gamma=0$1 spaces, with polynomial weights $\gamma=0$2 or stretched-exponential weights $\gamma=0$3. Coupled with Desvillettes–Villani entropy-dissipation estimates, this yields exponential convergence of the global weak solution toward $\gamma=0$4 with the optimal rate $\gamma=0$5, the spectral gap of the linearized operator (Carrapatoso, 2013).

In the perturbative hard-potential regime, Xu and Xu proved analyticity in the time variable: if $\gamma=0$6 and the solution remains sufficiently small in $\gamma=0$7, then for every integer $\gamma=0$8 and $\gamma=0$9,

$\gamma<0$0

The regularizing effect in time is therefore exactly the same as for the heat equation (Xu et al., 2022). Li and Xu obtained the corresponding analytic Gelfand–Shilov effect in velocity for hard potentials: if the initial perturbation is small in $\gamma<0$1, then for every $\gamma<0$2,

$\gamma<0$3

and the analytic radius grows linearly in time, $\gamma<0$4, again paralleling the heat semigroup (Li et al., 2022).

The soft-potential theory $\gamma<0$5 is more delicate because the diffusion degenerates at large velocity, but in the perturbative framework around the global Maxwellian one still has instantaneous analyticity in time and Gelfand–Shilov smoothing in velocity. Under a small exponentially weighted $\gamma<0$6 assumption on the initial fluctuation, the solution satisfies

$\gamma<0$7

and for some $\gamma<0$8,

$\gamma<0$9

hence $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$0 for every $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$1 (Cao et al., 18 Feb 2025).

Axisymmetric measure-valued theory adds a geometric refinement. For hard potentials, any axisymmetric $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$2, $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$3, generates an axisymmetric measure-valued weak solution

$$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$4

and if the initial datum is not a single Dirac mass then $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$5 is real-analytic in $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$6 for every $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$7. In the soft-potential and Maxwellian-molecule cases, there are no solutions whose support is contained in a fixed line for a positive time interval unless the solution is a single Dirac mass (Jang et al., 13 Jan 2025).

These results collectively show that the homogeneous Landau flow regularizes much more strongly than a bare weak-solution theory might suggest. The exact form of the regularization depends on the interaction regime, but analyticity, Gelfand–Shilov smoothing, and exponential relaxation all emerge as structurally stable features.

4. Particle systems, stochastic representations, and propagation of chaos

A large part of the modern homogeneous Landau theory concerns the derivation of the PDE from many-particle systems. For Maxwellian molecules, a conservative Landau master equation can be obtained as the grazing-collision limit of Kac’s Boltzmann master process. The associated $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$8-particle dynamics preserves momentum and energy, and the resulting theory yields quantitative propagation of chaos uniformly in time. In particular, for fixed $$Q(f,f)(v)=\nabla_v\cdot\int_{\mathbb R^3}a(v-v_*)\,[\,f(v_*)\nabla_v f(v)-f(v)\nabla_{v_*}f(v_*)\,]\,dv_*.$$9,

$A[f](v)=a*f,\qquad B[f](v)=b*f,$0

with $A[f](v)=a*f,\qquad B[f](v)=b*f,$1 as $A[f](v)=a*f,\qquad B[f](v)=b*f,$2, and the entropic chaos relation

$A[f](v)=a*f,\qquad B[f](v)=b*f,$3

holds uniformly in time. The same work also proves uniform trend to equilibrium for the $A[f](v)=a*f,\qquad B[f](v)=b*f,$4-particle system (Carrapatoso, 2012).

A complementary probabilistic representation for Maxwellian molecules factors the Landau dynamics into the product of two elementary processes: a Brownian motion $A[f](v)=a*f,\qquad B[f](v)=b*f,$5 on the rotation group $A[f](v)=a*f,\qquad B[f](v)=b*f,$6, and, conditionally on $A[f](v)=a*f,\qquad B[f](v)=b*f,$7, a Gaussian process in $A[f](v)=a*f,\qquad B[f](v)=b*f,$8. The representation

$A[f](v)=a*f,\qquad B[f](v)=b*f,$9

leads to sharp multiscale upper and lower bounds on the transition density. In the nondegenerate case the density exhibits a radial cost and a tangential cost; in the degenerate one-dimensional case the small-time decay is mixed exponential–Gaussian (Delarue et al., 2014).

Recent derivations sharpen the mean-field picture. For a stochastic interacting particle system for Maxwellian molecules, the normalized relative entropy between the $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$0-particle law $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$1 and the tensorized Landau law $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$2 satisfies

$$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$3

after combining a dissipation identity, pointwise logarithmic gradient and Hessian bounds on $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$4, and a law of large numbers at particle level. By Csiszár–Kullback–Pinsker, this gives quantitative propagation of chaos in $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$5 for fixed marginals (Carrillo et al., 2024).

For Coulomb and soft potentials, Kac’s program has also been extended to the Landau equation. A conservative $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$6-particle system obtained through the grazing limit of Kac’s walk yields propagation of chaos, convergence in Wasserstein distance, entropic convergence, and strong $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$7 convergence of marginals toward the unique bounded smooth solution of the Landau PDE. The proof relies on a duality approach adapted to the singular Landau operator and on estimates for weighted Fisher information and second-order Fisher information (Feng et al., 17 Jun 2025).

An even more collision-level realization is given by the microcanonical binary-collision process, a reversible pure-jump Markov process on the Boltzmann sphere. It preserves total momentum and kinetic energy exactly, approximates the Landau operator through small conservative rotations of relative velocities, and yields propagation of chaos in the joint mean-field/grazing limit $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$8 for the full interaction range $$Q(f,f)=\partial_i[(a_{ij}*f)\partial_j f-(b_i*f)f] =(a_{ij}*f)\partial_i\partial_j f-(c*f)f,$$9. The same framework also derives the $f=f(t,v)\ge 0$00-particle Landau master equation as a fixed-$f=f(t,v)\ge 0$01 grazing-collision limit (Du, 12 Nov 2025).

5. Entropy, gradient-flow structure, and Coulomb stability

For soft potentials $f=f(t,v)\ge 0$02, the homogeneous Landau equation admits a gradient-flow interpretation. The equation can be written in continuity form

$f=f(t,v)\ge 0$03

with velocity field

$f=f(t,v)\ge 0$04

The Boltzmann entropy

$f=f(t,v)\ge 0$05

has dissipation

$f=f(t,v)\ge 0$06

and a tailored dynamic metric $f=f(t,v)\ge 0$07 makes the Landau flow the gradient flow of $f=f(t,v)\ge 0$08. In that metric one has the energy-dissipation inequality

$f=f(t,v)\ge 0$09

and a JKO minimizing-movement scheme constructs solutions to a regularized Landau equation (Carrillo et al., 2020).

In the Coulomb case, Desvillettes, He, and Jiang introduced the monotone functional

$f=f(t,v)\ge 0$10

where $f=f(t,v)\ge 0$11. They proved

$f=f(t,v)\ge 0$12

so $f=f(t,v)\ge 0$13 is non-increasing. This monotonicity clarifies the competition between dissipation and nonlinearity, yields a stable regime implying global smoothness, shows eventual return to that regime even for larger data, and constrains possible blow-up scenarios (Desvillettes et al., 2020).

A distinct Coulomb stability principle is formulated in relative entropy. For strong solutions $f=f(t,v)\ge 0$14 and $f=f(t,v)\ge 0$15 with common mass and natural moment, entropy, Fisher-information, and Sobolev assumptions, one has

$f=f(t,v)\ge 0$16

hence

$f=f(t,v)\ge 0$17

The proof uses the continuity-form identity for relative entropy, the coercive negative Fisher-dissipation term generated by $f=f(t,v)\ge 0$18, and cross-term bounds derived from coercivity, Pinsker-type inequalities, and weighted moment/Fisher estimates. The same computation yields an a posteriori error bound for deterministic score-based transport solvers, where the weighted score-matching loss controls the relative-entropy error (Ilin, 16 Oct 2025).

These approaches are complementary rather than competing. The gradient-flow picture emphasizes variational structure, the monotonicity formula isolates a Lyapunov balance specific to Coulomb interactions, and the relative-entropy framework provides quantitative stability and weak–strong uniqueness in a natural information-theoretic metric.

6. Regularized models, numerical approximation, and current directions

A prominent regularized model replaces the singular log-gradient in the Landau flux by a mollified quantity. With a smooth compactly supported mollifier $f=f(t,v)\ge 0$19, the regularized equation becomes

$f=f(t,v)\ge 0$20

or equivalently

$f=f(t,v)\ge 0$21

For $f=f(t,v)\ge 0$22, this equation has a weak well-posedness theory under compact support and, in the very soft regime, an additional $f=f(t,v)\ge 0$23 assumption. Its equally weighted empirical measure

$f=f(t,v)\ge 0$24

solves an interacting ODE system, and under well-prepared initial data one has

$f=f(t,v)\ge 0$25

The original numerical study reported conservation laws, entropy dissipation, and convergence of order roughly $f=f(t,v)\ge 0$26 in Wasserstein metrics when $f=f(t,v)\ge 0$27 (Carrillo et al., 2022).

For Coulomb interactions, a Fourier–Galerkin spectral method on a truncated cube $f=f(t,v)\ge 0$28 yields an explicit $f=f(t,v)\ge 0$29-error estimate of the form

$f=f(t,v)\ge 0$30

The first term is the truncation error, the second the spectral projection error. The analysis supplies explicit conditions on $f=f(t,v)\ge 0$31 and $f=f(t,v)\ge 0$32 to achieve a prescribed accuracy over a fixed time interval, and the simulations recover spectral convergence in the Maxwell case and stable long-time behavior in the Coulomb case (Filbet et al., 11 Sep 2025).

A neural alternative is the physics-informed neural particle method. It uses a Lagrangian interacting-particle representation, a score network $f=f(t,v)\ge 0$33, and a trajectory network $f=f(t,v)\ge 0$34, with continuous-time residual

$f=f(t,v)\ge 0$35

enforced along particle paths rather than through time stepping. The theoretical analysis decomposes the trajectory error into score approximation error, empirical particle approximation error, and physics residual, derives an $f=f(t,v)\ge 0$36 error bound for kernel density reconstruction, and shows Monte Carlo concentration of the empirical score-matching loss. Numerical tests on two- and three-dimensional BKW solutions and reference-free configurations show stable transport, preservation of macroscopic invariants, and competitive or improved accuracy with significantly fewer particles (Kim et al., 11 Mar 2026).

Several current directions are explicitly identified across the recent literature. Natural extensions of the spectral Maxwellian-molecule theory include non-Maxwellian potentials, spatially inhomogeneous cases, and coupling with Boltzmann kernels without cutoff (Li et al., 2017). Open problems in the axisymmetric theory include uniqueness for $f=f(t,v)\ge 0$37 data without further symmetry, extension to the Landau–Poisson system, and quantitative control of the analyticity radius $f=f(t,v)\ge 0$38 as $f=f(t,v)\ge 0$39 and $f=f(t,v)\ge 0$40 (Jang et al., 13 Jan 2025). In the particle derivation program, quantitative chaos rates for Coulomb Landau, uniqueness for the limiting Landau hierarchy, and extension to the inhomogeneous equation remain open (Feng et al., 17 Jun 2025). These problems suggest that the homogeneous Landau equation is no longer studied only as a model kinetic PDE, but also as a testing ground for singular mean-field limits, entropy methods, spectral regularization, and structure-preserving computation.