## 4.2 Relativistic SPH

Besides finite volume schemes, another completely different method is widely used in astrophysics for integrating the hydrodynamic equations. This method is Smoothed Particle Hydrodynamics, or SPH for short [100, 63, 121]. The fundamental idea of SPH is to represent a fluid by a Monte Carlo sampling of its mass elements. The motion and thermodynamics of these mass elements is then followed as they move under the influence of the hydrodynamics equations. Because of its Lagrangian nature there is no need within SPH for explicit integration of the continuity equation, but in some implementations of SPH this is done nevertheless for certain reasons. As both the equation of motion of the fluid and the energy equation involve continuous properties of the fluid and their derivatives, it is necessary to estimate these quantities from the positions, velocities and internal energies of the fluid elements, which can be thought of as particles moving with the flow. This is done by treating the particle positions as a finite set of interpolating points where the continuous fluid variables and their gradients are estimated by an appropriately weighted average over neighboring particles. Hence, SPH is a free-Lagrange method, i.e., spatial gradients are evaluated without the use of a computational grid.

A comprehensive discussion of SPH can be found in the reviews of Hernquist & Katz [76], Benz [12] and Monaghan [120, 121]. The non-relativistic SPH equations are briefly discussed in Section  9.5 . The capabilities and limits of SPH are explored, e.g., in [169, 172], and the stability of the SPH algorithm is investigated in [170].

The SPH equations for special relativistic flows have been first formulated by Monaghan [120]. For such flows the SPH equations given in Section  9.5 can be taken over except that each SPH particle a carries baryons instead of mass  [120, 30]. Hence, the rest mass of particle a is given by , where is the baryon rest mass (if the fluid is made of baryons). Transforming the notation used in [30] to ours, the continuity equation, the momentum and the total energy equations for particle a are given by (unit of velocity is c)

and

respectively. Here, the summation is over all particles other than particle a, and d / dt denotes the Lagrangian time derivative.

is the baryon number density,

the momentum per particle, and

the total energy per particle (all measured in the laboratory frame). The momentum density , the energy density (measured in units of the rest mass energy density), and the specific enthalpy h are defined in Section  2.1 . and are the SPH dissipation terms, and denotes the gradient of the kernel (see Section  9.5 for more details).

Special relativistic flow problems have been simulated with SPH by [90, 80, 102, 104, 30, 164]. Extensions of SPH capable of treating general relativistic flows have been considered by [80, 89, 164]. Concerning relativistic SPH codes the artificial viscosity is the most critical issue. It is required to handle shock waves properly, and ideally it should be predicted by a relativistic kinetic theory for the fluid. However, unlike its Newtonian analogue, the relativistic theory has not yet been developed to the degree required to achieve this. For Newtonian SPH Lattanzio et al. [94] have shown that in high Mach number flows a viscosity quadratic in the velocity divergence is necessary. They proposed a form of the artificial viscosity such that the viscous pressure could be simply added to the fluid pressure in the equation of motion and the energy equation. Because this simple form of the artificial viscosity has known limitations, they also proposed a more sophisticated form of the artificial viscosity terms, which leads to a modified equation of motion. This artificial viscosity works much better, but it cannot be generalized to the relativistic case in a consistent way. Utilizing an equation for the specific internal energy both Mann [102] and Laguna et al. [89] use such an inconsistent formulation. Their artificial viscosity term is not included into the expression of the specific relativistic enthalpy. In a second approach, Mann [102] allows for a time-dependent smoothing length and SPH particle mass, and further proposed a SPH variant based on the total energy equation. Lahy [90] and Siegler & Riffert [164] use a consistent artificial viscosity pressure added to the fluid pressure. Siegler & Riffert [164] have also formulated the hydrodynamic equations in conservation form.

Monaghan [122] incorporates concepts from Riemann solvers into SPH. For this reason he also proposes to use a total energy equation in SPH simulations instead of the commonly used internal energy equation, which would involve time derivatives of the Lorentz factor in the relativistic case. Chow & Monaghan [30] have extended this concept and have proposed an SPH algorithm, which gives good results when simulating an ultra-relativistic gas. In both cases the intention was not to introduce Riemann solvers into the SPH algorithm, but to use them as a guide to improve the artificial viscosity required in SPH.

In Roe's Riemann solver [155], as well as in its relativistic variant proposed by Eulerdink [49, 50] (see Section  3.4), the numerical flux is computed by solving a locally linear system and depends on both the eigenvalues and (left and right) eigenvectors of the Jacobian matrix associated to the fluxes and on the jumps in the conserved physical variables (see Eqs. (26) and (27)). Monaghan [122] realized that an appropriate form of the dissipative terms and for the interaction between particles a and b can be obtained by treating the particles as the equivalent of left and right states taken with reference to the line joining the particles. The quantity corresponding to the eigenvalues (wave propagation speeds) is an appropriate signal velocity (see below), and that equivalent to the jump across characteristics is a jump in the relevant physical variable. For the artificial viscosity tensor, , Monaghan [122] assumes that the jump in velocity across characteristics can be replaced by the velocity difference between a and b along the line joining them.

With these considerations in mind Chow & Monaghan [30] proposed for in the relativistic case the form

when particles a and b are approaching, and otherwise. Here K = 0.5 is a dimensionless parameter, which is chosen to have the same value as in the non-relativistic case [122]. is the average baryon number density, which has to be present in (53), because the pressure terms in the summation of (90) have an extra density in the denominator arising from the SPH interpolation. Furthermore,

is the unit vector from b to a, and

where

Using instead of (see Eq. (51)) the modified momentum , which involves the line of sight velocity , guarantees that the viscous dissipation is positive definite [30].

The dissipation term in the energy equation is derived in a similar way and is given by [30]

if a and b are approaching, and otherwise. involves the energy , which is identical to (see Eq. (52)) except that W is replaced by .

To determine the signal velocity Chow & Monaghan [30] (and Monaghan [122] in the non-relativistic case) start from the (local) eigenvalues, and hence the wave velocities and v of one-dimensional relativistic hydrodynamic flows. Again considering particles a and b as the left and right states of a Riemann problem with respect to motions along the line joining the particles, the appropriate signal velocity is the speed of approach (as seen in the computing frame) of the signal sent from a towards b and that from b to a . This is the natural speed for the sharing of physical quantities, because when information about the two states meets it is time to construct a new state. This speed of approach should be used when determining the size of the time step by the Courant condition (for further details see [30]).

Chow & Monaghan [30] have demonstrated the performance of their Riemann problem guided relativistic SPH algorithm by calculating several shock tube problems involving ultra-relativistic speeds up to v = 0.9999. The algorithm gives good results, but finite volume schemes based on Riemann solvers give more accurate results and can handle even larger speeds (see Section  6).

 Numerical Hydrodynamics in Special Relativity Jose Maria Martí and Ewald Müller http://www.livingreviews.org/lrr-1999-3 © Max-Planck-Gesellschaft. ISSN 1433-8351 Problems/Comments to livrev@aei-potsdam.mpg.de