What is the long time behavior of discrete velocity kinetic equations?

Exponential decay to one constant state

The question is whether a gas whose particles may only travel at a fixed speed along a fixed set of directions eventually settles down, and how fast. The paper answers yes for the nonlinear equations it studies: the densities of particles moving in each direction converge to the same positive constant m∞, and the squared L2 distance to that constant shrinks at least like Λe^(−2λt). Here m∞ is fixed by conservation of mass, equal to one half of the initial integral of u0+v0 in one dimension and one sixth of the summed initial integral over the six densities in three dimensions.

Two conditions shape that answer, and they should be read together with it. The domain is the torus, meaning periodic boundary conditions, and the result is stated for weak solutions that are bounded in L∞. The decay constants Λ and λ are not free parameters: they depend on the speed c, on the L∞ size M of the initial data, on m∞, and, in the singular case, on the positive lower bound δ of the initial data.

Two species, one transport term, one collision term

In one dimension the unknowns are u(t,x) and v(t,x), the densities of particles moving at speed c in the positive and negative x-direction. Each density is transported at speed c, and the two are coupled by a term proportional to k(u,v,x)(v−u) and its mirror image. The function k is the interaction rate: it sets how strongly particles of the two directions exchange identity at a given point and state. Setting k=1 gives the Goldstein-Taylor model, which can be converted into the telegraph equation; setting k=u+v gives the Carleman equation, describing collisions in which two particles of one type become two of the other.

The macroscopic variables make the structure visible. The mass density ρ=u+v satisfies a conservation law with no source term, while the momentum j=u−v obeys an equation in which the collision term appears as −2kj. So transport moves mass and momentum around without dissipating them, and only the interaction term damps the momentum. This split is why the problem is not a straightforward relaxation: dissipation acts in the velocity variable, while the conservative transport term must be handled separately. The paper's stated motivation is to build a hypocoercivity method for the nonlinear case that does not assume the initial data are already close to equilibrium.

Why plain entropy decay is not enough

For any convex function φ, multiplying the two equations by φ′(u) and φ′(v), integrating over the torus, and adding gives a time derivative equal to minus the integral of k(u,v,x)(u−v)(φ′(u)−φ′(v)). Convexity makes this non-positive, so the integral of φ(u)+φ(v) never increases. The right-hand side vanishes exactly when u=v everywhere, which the paper calls a local equilibrium; a local equilibrium that solves the equation is a pair of equal positive constants, the global equilibrium.

This gives monotonicity but not a rate. The paper says explicitly that such functionals are not enough to obtain exponential convergence with explicit rates. The gap is that a non-increasing quantity can stall arbitrarily slowly, and the dissipation controls only the difference u−v, not the full L2 distance to m∞. The constructive step is therefore to take one specific convex function, Boltzmann's entropy, and modify it.

Modifying Boltzmann entropy into a Lyapunov functional

The entropy used here is H[u,v], the integral of (u/m∞)log(u/m∞)−u/m∞+1 times m∞, plus the same expression in v, defined for non-negative bounded u and v with m∞>0. The paper first proves that H is equivalent to the squared L2 distance: it lies between 1/(2M) and 1/m∞ times the integral of (u−m∞)²+(v−m∞)², where M is the larger of the two L∞ norms. Equivalence matters because a decay statement for H then transfers to the L2 norm the theorems are stated in.

The entropy itself dissipates at a rate proportional to the integral of k(u−v)(log u−log v), which is non-negative. To turn this into a Gronwall inequality, the paper adds suitable terms to the entropy so that the resulting functional is decreasing and its derivative is bounded by a negative constant times the functional itself. The constants in that inequality are computed from δ, M, m∞ and c, which is what makes the rate explicit and constructive rather than merely existential. The same construction is then carried out for the three-dimensional system, where six densities interact and the bookkeeping is heavier but the logic is the same.

Singular rates require strictly positive initial data

The paper separates interaction rates into three types. Type 1 rates are bounded on bounded density ranges; type 2 rates blow up at zero density but are bounded away from it; type 3 rates are bounded above and below on any range of densities bounded away from zero. For the power family k=(u+v)^α, type 1 corresponds to α≥0, type 2 to α<0, and type 3 to every real α.

The first main theorem covers type 3 rates, which may be singular, and it requires initial data bounded below by some δ>0. That lower bound is not cosmetic: if k can blow up at u=v=0, the equation needs densities kept away from zero for the estimates to close. The theorem applies, for instance, to k=k1(x)(u+v)^α for any real α with k1 bounded and bounded away from zero. The three-dimensional analogue, Theorem 2.6, carries the same δ>0 requirement and the same structure of constants.

Degenerate rates allow merely non-negative data

The second main theorem removes the lower bound on the initial data, at a price. It assumes k is of type 1 and satisfies k≥k1(x)(u+v)^α for some α in [0,1], with k1 bounded and bounded away from zero. Under that growth assumption k may vanish where u=v=0, so it can be degenerate, but it cannot be singular, and the initial data only need to be non-negative and bounded, with m∞>0. The conclusion is the same form of exponential L2 decay, with constants Λα and λα depending on M, m∞, c and the essential infimum of k1.

This is where the two named models sit. The Goldstein-Taylor equation corresponds to α=0 and the Carleman equation to α=1, so both fall inside the degenerate case and are covered without requiring strictly positive initial data. The three-dimensional version, Theorem 2.7, replaces (u+v)^α by the analogous power of the sum of all six densities and reaches the same conclusion.

What the explicit rate does and does not fix

The value of the result is the combination of generality and explicitness. Earlier treatments of the large time behavior of these equations, as the paper describes them, covered the Goldstein-Taylor and Carleman cases; here a single Lyapunov construction covers a wide class of rates in both one and three dimensions, and the constants are computable from the data rather than left unspecified. The paper also notes that in the linear case, where k=k1(x) is bounded and bounded away from zero, one only needs initial data in L2 and can obtain rates and constants independent of the initial data, which the nonlinear theorems cannot claim.

The scope is equally specific. The estimates are for periodic boundary conditions, which the paper calls a technical convenience it expects could be removed, and for weak solutions bounded in L∞. The rates depend on the L∞ norms of the initial data, so they are not uniform over all data of a given mass. The paper presents whole-space domains, other boundary conditions, higher dimensions and other discrete velocity models such as the discrete Boltzmann equation as extensions rather than results. The practical reading is therefore a threshold between the two regimes: if the interaction rate can be singular at zero density, strictly positive initial data is the condition to check; if it is merely degenerate with growth exponent α in [0,1], non-negative bounded data with positive mass suffices, and that second regime is the one that contains Goldstein-Taylor and Carleman.

Sources used in this answer