The heat hidden
by an average.

Two particle populations can have the same density, average velocity and kinetic energy while carrying energy in different directions. A simple calculation shows why the quantities retained by a model must be chosen with its intended prediction in view.

A fluid model replaces a detailed distribution of particle velocities with a shorter description. That compression makes large calculations possible. It also creates a scientific obligation: identify which distinctions survive in the reduced state, and which predictions depend on information left outside it.

My published research on the geometry of the moment manifold studies statistical distinguishability within a specified relativistic kinetic-theory model. A separate, elementary example brings the choice of representation into view before any distance is calculated. It concerns the energy carried across a surface by particles that have no net bulk motion.

Match the averages and change the transport

Consider two invented, one-dimensional populations of identical classical particles. They have the same number density, n, and particle mass, m. Choose a positive reference speed, v₀, small enough for nonrelativistic mechanics. The proportions below describe the local velocity distribution, rather than a record of a laboratory experiment.

Population A: half the particles move right at v₀, and half move left at the same speed. Its average velocity is zero; its average squared velocity is v₀².

Population B: one fifth move right at 2v₀, and four fifths move left at v₀/2. The weighted average velocity is again zero: (1/5) × 2 − (4/5) × (1/2) = 0, in units of v₀. The average squared velocity is also unchanged: (1/5) × 4 + (4/5) × (1/4) = 1, in units of v₀².

Both populations therefore have zero momentum density and the same kinetic-energy density, nmv₀²/2. A local description retaining only number, momentum and energy assigns them identical values.

Energy flux asks a different calculation. Each particle carries kinetic energy proportional to its speed squared, while its signed velocity determines the direction and rate of transport across a surface. In this one-dimensional rest frame, the heat flux is q = (nm/2) × the average of v³.

For A, the two contributions cancel and q = 0. For B, the dimensionless third moment is (1/5) × 8 − (4/5) × (1/8) = 3/2. Its heat flux is consequently q = 3nmv₀³/4, directed to the right. The smaller group of faster right-moving particles carries more energy across the surface than the larger group returning slowly.

The example separates stored energy from its transport. It establishes a difference in local flux, not an automatic difference in local temperature change. For that next question, the spatial variation of the flux, boundary conditions and collision dynamics must enter the calculation. A spatially uniform flux has no divergence.

A closure supplies the missing relation

Taking moments of a kinetic equation produces a hierarchy: the evolution of a retained quantity generally involves a flux containing higher moments. A closure supplies the relation needed to continue with a finite description. C. David Levermore’s moment-closure hierarchy makes this construction systematic; its first member uses Maxwellian distributions and yields the Euler equations.

In the continuous one-dimensional example, maximising entropy over all real velocities subject only to normalisation, zero mean and the fixed second moment selects the same Gaussian for A and B. That reconstructed Gaussian has zero third moment. It reproduces their common inputs exactly, while missing B’s nonzero heat flux. Here the reconstruction is allowed to use all real velocities; it is not restricted to the discrete speeds in either invented population.

Matching the quantities supplied to a reconstruction is an essential consistency check. Testing an unsupplied quantity provides additional evidence about its usefulness. These two checks have different jobs. A perfect match to density and energy cannot settle whether a model accurately carries heat.

More information changes the reconstruction

Jonathan Ng, Ammar Hakim and Amitava Bhattacharjee examined this issue in a public study of electron distributions in magnetic reconnection. They reconstructed distributions from kinetic-simulation moments. At the reconnection x-point, their 21-moment maximum-entropy model captured a triangular shape that the 10- and 14-moment reconstructions missed, while finer structure remained unresolved. The comparison demonstrates a gain tied to particular retained information and a particular test.

Adding information also imposes mathematical requirements. In an exponential-polynomial family on all real velocities, a nonzero cubic term cannot be the highest term: the exponential grows without bound in one tail. A heat-flux constraint therefore needs a suitable integrable construction, with existence checked for the supplied moments. In the models discussed by Ng and colleagues, a fourth-order moment accompanies the heat-flux information. Merely adding a third-moment entry to a state vector does not finish the closure.

Give the geometric ruler the right state space

The connection to my Physical Review C paper is a question about representation. Its Fisher–Rao geometry is constructed within the DNMR 14-moment truncation for conformal Bjorken flow in the relaxation-time approximation; finite-shear quantities belong to the stated maximum-entropy completion. The classical two-population example above is separate from that relativistic calculation.

If two distributions are mapped to the same reduced state, a distance calculated solely between those reduced states cannot distinguish them. A geometric ruler measures differences present in its chosen representation. Its interpretation therefore starts with the moment map and the completion, as well as the metric.

A useful follow-on evaluation would specify a target observable before choosing the reduced model, reconstruct states using a declared set of moments, and then test a flux or distribution feature that was not used as a fitting constraint. Repeating that comparison with richer admissible moment sets would show what additional information buys, alongside its computational cost. This is a proposed evaluation, not a new simulation result.

The purpose of reduction is to retain enough information for the task at hand. The two populations make that requirement concrete: a model can preserve exactly how much kinetic energy is present while losing the direction in which that energy is being carried.