In this paper the problem of consistency of smoothed particle hydrodynamics (SPH) is solved. A novel error analysis is developed in $n$-dimensional space using the Poisson summation formula, which enables the treatment of the kernel and particle approximation errors in combined fashion. New consistency integral relations are derived for the particle approximation which correspond to the cosine Fourier transform of the classically known consistency conditions for the kernel approximation. The functional dependence of the error bounds on the SPH interpolation parameters, namely the smoothing length $h$ and the number of particles within the kernel support ${\cal{N}}$ is demonstrated explicitly from which consistency conditions are seen to follow naturally. As ${\cal{N}}\to\infty$, the particle approximation converges to the kernel approximation independently of $h$ provided that the particle mass scales with $h$ as $m\propto h^{\beta}$, with $\beta >n$. This implies that as $h\to 0$, the joint limit $m\to 0$, ${\cal{N}}\to\infty$, and $N\to\infty$ is necessary for complete convergence to the continuum, where $N$ is the total number of particles. The analysis also reveals the presence of a dominant error term of the form $(\ln {\cal{N}})^{n}/{\cal{N}}$, which tends asymptotically to $1/{\cal{N}}$ when ${\cal{N}}\gg 1$, as it has long been conjectured based on the similarity between the SPH and the quasi-Monte Carlo estimates.

The problem of consistency of smoothed particle hydrodynamics (SPH) has demanded considerable attention in the past few years due to the ever increasing number of applications of the method in many areas of science and engineering. A loss of consistency leads to an inevitable loss of approximation accuracy. In this paper, we revisit the issue of SPH kernel and particle consistency and demonstrate that SPH has a limiting second-order convergence rate. Numerical experiments with suitably chosen test functions validate this conclusion. In particular, we find that when using the root mean square error as a model evaluation statistics, well-known corrective SPH schemes, which were thought to converge to second, or even higher order, are actually first-order accurate, or at best close to second order. We also find that observing the joint limit when $N\to\infty$, $h\to 0$, and $n\to\infty$, as was recently proposed by Zhu et al., where $N$ is the total number of particles, $h$ is the smoothing length, and $n$ is the number of neighbor particles, standard SPH restores full $C^{0}$ particle consistency for both the estimates of the function and its derivatives and becomes insensitive to particle disorder.