We follow the definitions and notation of
\cite{Dehnen2012}. Motivation for all the material given here can be
found in their paper.

\subsection{General Definitions}

The desirable properties of an SPH kernels $W(\vec{x},h)$ are:
\begin{enumerate}
\item $W(\vec{x},h)$ should be isotropic in $\vec{x}$.
\item $W(\vec{x},h)$ should be positive and decrease monotonically.
\item $W(\vec{x},h)$ should be twice differentiable.
\item $W(\vec{x},h)$ should have a finite support and be cheap to
  compute.
\end{enumerate}
As a consequence, the smoothing kernels used in SPH can
hence be written (in 3D) as

\begin{equation}
 W(\vec{x},h) \equiv \frac{1}{H^3}f\left(\frac{|\vec{x}|}{H}\right),
\end{equation}
where $H=\gamma h$ is defined below and $f(u)$ is a dimensionless
function, usually a low-order polynomial, such that $f(u \geq 1) = 0$
and normalised such that

\begin{equation}
  \int f(|\vec{u}|){\rm d}^3u = 1.
\end{equation}
$H$ is the kernel's support radius and is used as the ``smoothing
length'' in the Gadget code( {i.e.} $H=h$). This definition is,
however, not very physical and makes comparison of kernels at a
\emph{fixed resolution} difficult. A more sensible definition of the
smoothing length, related to the Taylor expansion of the
re-constructed density field is given in terms of the kernel's
standard deviation

\begin{equation}
  \sigma^2 \equiv \frac{1}{3}\int \vec{u}^2 W(\vec{u},h) {\rm d}^3u.
  \label{eq:sph:sigma}
\end{equation}
The smoothing length is then:
\begin{equation}
  h\equiv2\sigma.
    \label{eq:sph:h}
\end{equation}
Each kernel, {\it i.e.} defintion of $f(u)$, will have a different
ratio $\gamma = H/h$. So for a \emph{fixed resolution} $h$, one will
have different kernel support sizes, $H$, and a different number of
neighbours, $N_{\rm ngb}$ to interact with. One would typically choose
$h$ for a simulation as a multiple $\eta$ of the mean-interparticle
separation:

\begin{equation}
  h = \eta \langle x \rangle = \eta \left(\frac{m}{\rho}\right)^{1/3},
\end{equation}
where $\rho$ is the local density of the fluid and $m$ the SPH
particle mass. 
The (weighted) number of neighbours within the kernel support is a
useful quantity to use in implementations of SPH. It is defined (in
3D) as:

\begin{equation}
  N_{\rm ngb} \equiv \frac{4}{3}\pi \left(\frac{H}{h}\eta\right)^3.
\end{equation}
Once the fixed ratio $\gamma= H/h$ is known (via equations
\ref{eq:sph:sigma} and \ref{eq:sph:h}) for a given kernel, the number
of neighbours only depends on the resolution parameter $\eta$.  For
the usual cubic spline kernel (see below), setting the simulation
resolution to $\eta=1.2348$ yields the commonly used value $N_{\rm
  ngb} = 48$.

\subsection{Kernels available in \swift}

The \swift kernels are split into two categories, the B-splines
($M_{4,5,6}$) and the Wendland kernels ($C2$, $C4$ and $C6$). In all
cases we impose $f(u>1) = 0$.\\
The spline kernels are defined as:

\begin{align}
  f(u) &= C M_n(u), \nonumber\\
  M_n(u) &\equiv \frac{1}{2\pi}
  \int_{-\infty}^{\infty}
  \left(\frac{\sin\left(k/n\right)}{k/n}\right)^n\cos\left(ku\right){\rm
  d}k \nonumber,
\end{align}
whilst the Wendland kernels are constructed from:

\begin{align}
  f(u) &= C \Psi_{i,j}(u), \nonumber\\
  \Psi_{i,j}(u) &\equiv
  \mathcal{I}^k\left[\max\left(1-u,0\right)\right], \nonumber\\
  \mathcal{I}[f](u) &\equiv \int_u^\infty f\left(k\right)k{\rm d}k. \nonumber
\end{align}

\subsubsection{Cubic spline ($M_4$) kernel}

In 3D, we have $C=\frac{16}{\pi}$ and $\gamma=H/h = 1.825742$.\\
The kernel function $f(u)$ reads:

\begin{equation}
  M_4(u) = \left\lbrace\begin{array}{rcl}
  3u^3 - 3u^2 + \frac{1}{2} & \mbox{if} & u<\frac{1}{2}\\
  -u^3 + 3u^2 - 3u + 1 & \mbox{if} & u \geq \frac{1}{2}
  \end{array}
  \right.
    \nonumber
\end{equation}


\subsubsection{Quartic spline ($M_5$) kernel}

In 3D, we have $C=\frac{15625}{512\pi}$ and $\gamma=H/h = 2.018932$.\\
The kernel function $f(u)$ reads:

\begin{align}
  M_5(u) &=     \nonumber\\
  &\left\lbrace\begin{array}{rcl}
  6u^4 - \frac{12}{5}u^2 + \frac{46}{125} & \mbox{if} & u < \frac{1}{5} \\
  -4u^4 + 8u^3  - \frac{24}{5}u^2 + \frac{8}{25}u + \frac{44}{125} &  \mbox{if} &  \frac{1}{5} \leq u < \frac{3}{5}\\
  u^4 - 4u^3 + 6u^2 - 4u + 1 &  \mbox{if} &  \frac{3}{5} \leq u \\
  \end{array}
  \right.
  \nonumber
\end{align}


\subsubsection{Quintic spline ($M_6$) kernel}

In 3D, we have $C=\frac{2187}{40\pi}$ and $\gamma=H/h = 2.195775$.\\
The kernel function $f(u)$ reads:

\begin{align}
  M_6(u) &=     \nonumber\\
  &\left\lbrace\begin{array}{rcl}
  -10u^5 + 10u^4 - \frac{20}{9}u^2 + \frac{22}{81} & \mbox{if} & u < \frac{1}{3} \\
  5u^5 - 15u^4 + \frac{50}{3}u^3 - \frac{70}{9}u^2 + \frac{25}{27}u + \frac{17}{81} &  \mbox{if} &  \frac{1}{3} \leq u < \frac{2}{3}\\
  -1u^5 + 5u^4 - 10u^3 + 10u^2 - 5u + 1. & \mbox{if} & u \geq \frac{2}{3}
  \end{array}
  \right.
      \nonumber
\end{align}


\subsubsection{Wendland C2 kernel}

In 3D, we have $C=\frac{21}{2\pi}$ and $\gamma=H/h = 1.936492$.\\
The kernel function $f(u)$ reads:

\begin{align}
  \Psi_{i,j}(u) &= 4u^5 - 15u^4 + 20u^3 - 10u^2 + 1.
    \nonumber
\end{align}


\subsubsection{Wendland C4 kernel}

In 3D, we have $C=\frac{495}{32\pi}$ and $\gamma=H/h = 2.207940$.\\
The kernel function $f(u)$ reads:

\begin{align}
  \Psi_{i,j}(u) &= \frac{35}{3}u^8 - 64u^7 + 140u^6     \nonumber\\
  & - \frac{448}{3}u^5 + 70u^4 - \frac{28}{3}u^2 + 1
    \nonumber
\end{align}


\subsubsection{Wendland C6 kernel}

In 3D, we have $C=\frac{1365}{64\pi}$ and $\gamma=H/h = 2.449490$.\\
The kernel function $f(u)$ reads:

\begin{align}
  \Psi_{i,j}(u) &= 32u^{11} - 231u^{10} + 704u^9 - 1155u^8     \nonumber\\
  & + 1056u^7 - 462u^6 + 66u^4 - 11u^2 + 1
    \nonumber
\end{align}


\subsubsection{Summary}

All kernels available in \swift are shown on Fig.~\ref{fig:sph:kernels}.

\begin{figure}
\includegraphics[width=\columnwidth]{kernels.pdf}
\caption{The kernel functions available in \swift for a mean
  inter-particle separation $\langle x\rangle=1.5$ and a resolution
  $\eta=1.2348$ shown in linear (top) and log (bottom) units to
  highlight their differences. The corresponding kernel support radii
  $H$ (shown by arrows) and number of neighours $N_{\rm ngb}$ are
  indicated on the figure. A Gaussian kernel with the same smoothing
  length is shown for comparison. Note that all these kernels have
  the \emph{same resolution} despite having vastly different number of
  neighbours.}
\label{fig:sph:kernels}
\end{figure}

\begin{figure}
\includegraphics[width=\columnwidth]{kernel_derivatives.pdf}
\caption{The first and second derivatives of the kernel functions
  available in \swift for a mean inter-particle separation $\langle
  x\rangle=1.5$ and a resolution $\eta=1.2348$.  A Gaussian kernel
  with the same smoothing length is shown for comparison.}
\label{fig:sph:kernel_derivatives}
\end{figure}


\subsection{Kernel Derivatives}

The derivatives of the kernel function have relatively simple
expressions and are shown on Fig.~\ref{fig:sph:kernel_derivatives}:

\begin{align}
 \vec\nabla_x W(\vec{x},h) &= \frac{1}{h^4}f'\left(\frac{|\vec{x}|}{h}\right) \frac{\vec{x}}{|\vec{x}|}, \\
 \frac{\partial W(\vec{x},h)}{\partial h} &=- \frac{1}{h^4}\left[3f\left(\frac{|\vec{x}|}{h}\right) + 
\frac{|\vec{x}|}{h}f'\left(\frac{|\vec{x}|}{h}\right)\right].
\end{align}
Note that for all the kernels listed above, $f'(0) = 0$.