diff --git a/theory/Multipoles/fmm_summary.tex b/theory/Multipoles/fmm_summary.tex index d3e0c8320345a74c6b89da2c6207042c8a6eef42..fdcb66cfe185e3e78f2ed79827f7caff3ce9e82f 100644 --- a/theory/Multipoles/fmm_summary.tex +++ b/theory/Multipoles/fmm_summary.tex @@ -221,8 +221,7 @@ The main remaining question is to decide when two cells are far enough from each others that the truncated Taylor expansion used as approximation for the potential (eq. \ref{eq:fmm:expansion}) is accurate enough. The criterion used to make that decision is called the \emph{multipole - acceptance criterion} (MAC). - + acceptance criterion} (MAC). \\ We know that (\ref{eq:fmm:expansion}) is converging towards the correct answer provided $1>|\mathbf{r}_a + \mathbf{r}_b| / |\mathbf{R}|$. This is hence the most basic (and always necessary) MAC that can be designed. If @@ -239,8 +238,7 @@ This lets users have a second handle on the accuracy on the gravity calculation besides the much more involved change in the expansion order $p$ of the FMM method. Typical values for the opening angle are in the range $[0.3, 0.7]$, with the cost of the simulation growing as $\theta_{\rm - cr}$ decreases. - + cr}$ decreases. \\ This method has the drawback of using a uniform criterion across the entire simulation volume and time evolution, which means that the chosen value of $\theta_{\rm cr}$ could be too small in some regions (leading to too many @@ -250,8 +248,7 @@ uses a more adaptive criterion to decide when the multipole approximation can be used. This is based on the error analysis of FMM by \cite{Dehnen2014} and is summarised below for completeness. The key idea is to exploit the additional information about the distribution of particles -that is encoded in the higher-order multipole terms. - +that is encoded in the higher-order multipole terms.\\ We start by defining the scalar quantity $\mathsf{P}_{A,n}$, the \emph{power} of the multipole of order $n$ of the particles in cell $A$, via @@ -309,15 +306,23 @@ the previous time-step\footnote{On the first time-step of a simulation this acceleration in a given cell can be computed at the same time as the P2M and M2M kernels are evaluated in the tree construction phase. The second condition in (\ref{eq:fmm:mac}) is necessary to ensure the convergence of the -Taylor expansion. - +Taylor expansion.\\ +One important difference between this criterion and the purely geometric +one (\ref{eq:fmm:angle}) is that it is not symmetric in $A \leftrightarrow +B$. This implies that there are cases where a multipole in cell $A$ can be +used to compute the field tensors in cell $B$ but the multipole in $B$ +cannot be used to compute the $\mathsf{F}$ values of cell $A$. This affects +the tree walk by breaking the symmetry and potentially leading to cells of +different sizes interacting. \\ In the specific case of the M2P kernel, we have $\rho_B = 0$, which simplifies some of the expressions above. In this case, at order $p$, we get: -\begin{align} - E_{AB} &= \frac{\mathsf{P}_{A,p}}{M_A |\mathbf{R}|^p}, \nonumber \\ - \tilde{E}_{AB} &= 8E_{AB} \nonumber -\end{align} -This leads to the following MAC for the M2P kernel: +\begin{equation} + E_{AB} = \frac{\mathsf{P}_{A,p}}{M_A |\mathbf{R}|^p}, \qquad + \tilde{E}_{AB} = 8E_{AB} \nonumber +\end{equation} +Note that, in this case, only the power term of the order of the +scheme appears; not a sum over the lower-order ones. This leads to the +following MAC for the M2P kernel: \begin{equation} 8\frac{\mathsf{P}_{A,p}}{|\mathbf{R}|^{p+2}} < \epsilon \min_{b\in B}\left(|\mathbf{a}_b|\right) \quad \rm{and} \quad \frac{\rho_A}