19. Appendix

19.1. The MMM family of algorithms

19.1.1. Introduction

In the MMM family of algorithms for the electrostatic interaction, a convergence factor approach to tackle the conditionally convergent Coulomb sum is used (even the authors of the original MMM method have no idea what this acronym stands for). Instead of defining the summation order, one multiplies each summand by a continuous factor \(c(\beta,r_{ij},n_{klm})\) such that the sum is absolutely convergent for \(\beta>0\), but \(c(0,.,.)=1\). The energy is then defined as the limit \(\beta\rightarrow 0\) of the sum, i.e. \(\beta\) is an artificial convergence parameter. For a convergence factor of \(e^{-\beta n_{klm}^2}\) the limit is the same as the spherical limit, and one can derive the classical Ewald method quite conveniently through this approach [Smi81]. To derive the formulas for MMM, one has to use a different convergence factor, namely \(e^{-\beta|r_{ij}+n_{klm}|}\), which defines the alternative energy

\[\tilde{E}=\,\frac{1}{2}\lim_{\beta\rightarrow 0}\sum_{k,l,m}{\sum_{i,j=1}^N}' \frac{q_i q_je^{-\beta|p_{ij} + n_{klm}|}} {|p_{ij} + n_{klm}|} =:\,\frac{1}{2}\lim_{\beta\rightarrow 0}\sum_{i,j=1}^N q_iq_j\phi_\beta(x_{ij}, y_{ij},z_{ij}).\]

\(\phi_\beta\) is given by \(\phi_\beta(x,y,z)=\,\tilde\phi_\beta(x,y,z) + \frac{e^{-\beta r}}{r}\) for \((x,y,z)\neq 0\) and \(\phi_\beta(0,0,0)=\,\tilde\phi_\beta(0,0,0)\), where

\[\tilde\phi_\beta(x,y,z)=\,\sum_{(k,l,m)\neq 0} \frac{e^{-\beta r_{klm}}}{r_{klm}}.\]

The limit \(\tilde{E}\) exists, but differs for three dimensionally periodic systems by some multiple of the square of the dipole moment from the spherical limit as obtained by the Ewald summation [Smi81]. From the physical point of view the Coulomb interaction is replaced by a screened Coulomb interaction with screening length \(1/\beta\). \(\tilde{E}\) is then the energy in the limit of infinite screening length. But because of the conditional convergence of the electrostatic sum, this is not necessarily the same as the energy of an unscreened system. Since the difference to the Ewald methods only depends on the dipole moment of the system, the correction can be calculated easily in linear time and can be ignored with respect to accuracy as well as to computation time.

For one or two dimensionally systems, however, \(\tilde{E}=E\), the convergence factor approach equals the spherical summation limit of the Ewald sum, and MMM1D and MMM2D do not require a dipole correction.

Starting from this convergence factor approach, Strebel constructed a method of computational order \(O(N\log N)\), which is called MMM [Str99]. The favorable scaling is obtained, very much like in the Ewald case, by technical tricks in the calculation of the far formula. The far formula has a product decomposition and can be evaluated hierarchically similarly to the fast multipole methods.

For particles sufficiently separated in the z-axis one can Fourier transform the potential along both x and y. We obtain the far formula as

\[\phi(x,y,z) =\, u_x u_y\sum_{p,q\neq 0} \frac{e^{2\pi f_{pq}z} + e^{2\pi f_{pq}(\lambda_z-z)}}{f_{pq} \left(e^{2\pi f_{pq}\lambda_z} - 1\right)} e^{2\pi i u_y q y}e^{2\pi i u_x p x} + 2\pi u_x u_y\left(u_z z^2 - z + \frac{\lambda_z}{6}\right).\]

where \(\lambda_{x,y,z}\) are the box dimensions, \(f_{pq} =\, \sqrt{(u_x p)^2 + (u_y q)^2},\quad f_p =\, u_x p,\quad f_q =\, u_x q\), \(\omega_p=2\pi u_x p\) and \(\omega_q=2\pi u_y q\). The advantage of this formula is that it allows for a product decomposition into components of the particles. For example

\[e^{2\pi f_{pq}z}=e^{2\pi f_{pq}(z_i-z_j)}=e^{2\pi f_{pq}z_i}e^{-2\pi f_{pq}z_j}\]

etc. Therefore one just has to calculate the sum over all these exponentials on the left side and on the right side and multiply them together, which can be done in \(O(N)\) computation time. As can be seen easily, the convergence of the series is excellent as long as z is sufficiently large. By symmetry one can choose the coordinate with the largest distance as z to optimize the convergence. Similar to the Lekner sum, we need a different formula if all coordinates are small, i.e. for particles close to each other. For sufficiently small \(u_y\rho\) and \(u_xx\) we obtain the near formula as

\[\begin{split}\begin{array}{rl} \tilde\phi(x,y,z)=\, & 2 u_x u_y\sum\limits_{p,q>0} \frac{\cosh(2\pi f_{pq}z)}{f_{pq} \left(e^{2\pi f_{pq}\lambda_z} - 1\right)} e^{2\pi i u_y q y}e^{2\pi i u_x p x} +\\ & 4u_x\sum\limits_{l,p>0}\left(K_0(2\pi u_x p\rho_l) + K_N(2\pi u_x p\rho_{-l})\right)cos(2\pi u_x p x) -\\ & 2u_x\sum\limits_{n\ge 1}\frac{b_{2n}}{2n(2n)!}\Re\bigl((2\pi u_y (z+iy))^{2n}\bigr) +\\ & u_x\sum\limits_{n\ge 0}\left(\begin{array}{c}-\frac{1}{2}\\ n\end{array}\right)\frac{\left( \psi^{(2n)}(1 + u_x x) + \psi^{(2n)}(1 - u_x x)\right)}{(2n)!}\rho^{2n} -\\ & 2\log(4\pi). \end{array}\end{split}\]

Note that this time we calculate \(\tilde{\phi}\) instead of \(\phi\), i.e. we omit the contribution of the primary simulation box. This is very convenient as it includes the case of self energy and makes \(\tilde{\phi}\) a smooth function. To obtain \(\phi\) one has to add the \(1/r\) contribution of the primary box. The self energy is given by

\[\tilde\phi(0,0,0)=\, 2 u_x u_y\sum\limits_{p,q>0} \frac{1}{f_{pq} \left(e^{2\pi f_{pq}\lambda_z} - 1\right)}+ 8u_x\sum\limits_{l,p>0}K_N(2\pi u_x\lambda_y p l) + 2 u_x\psi^{(0)}(1) - 2\log(4\pi).\]

Both the near and far formula are derived using the same convergence factor approach, and consequently the same singularity in \(\beta\) is obtained. This is important since otherwise the charge neutrality argument does not hold.

To obtain the \(O(N\log N)\) scaling, some algorithm tricks are needed, which are not used in MMM1D, MMM2D or ELC and are therefore not discussed here. For details, see [Str99].

19.1.2. MMM2D theory

In the case of periodicity only in the x and y directions, the far formula looks like

\[\begin{split}\begin{array}{rl} \phi(x,y,z) = \, & 4 u_x u_y\sum_{p,q>0} \frac{e^{-2\pi f_{pq}|z|}} {f_{pq}} \cos(\omega_p x)\cos(\omega_q y) +\\ & 2 u_x u_y\left(\sum_{q>0} \frac{e^{-2\pi f_q|z|}}{f_q} \cos(\omega_q y) + \sum_{p>0} \frac{e^{-2\pi f_p|z|}}{f_p} \cos(\omega_p x)\right) -\\ & 2\pi u_x u_y |z| \end{array}\end{split}\]

,

and the near formula is

\[\begin{split}\begin{array}{rl} \tilde\phi(x,y,z)=\, & 4u_x\sum_{l,p>0}\left(K_0(\omega_p\rho_l) + K_0(\omega_p\rho_{-l})\right)\cos(\omega_p x) -\\ & 2u_x\sum_{n\ge 1}\frac{b_{2n}}{2n(2n)!} \Re\bigl((2\pi u_y (z+iy))^{2n}\bigr)\,+\, \sum_{k=1}^{N_\psi-1}\left(\frac{1}{r_{k}} + \frac{1}{r_{-k}}\right) -\\ & u_x\sum_{n\ge 0}\left(\begin{array}{c}-\frac{1}{2}\\n\end{array}\right)\frac{\left( \psi^{(2n)}(N_\psi + u_x x) + \psi^{(2n)}(N_\psi - u_x x)\right)}{(2n)!}(u_x\rho)^{2n} -\\ & 2u_x\log\left(4\pi\frac{u_y}{u_x}\right). \end{array}\end{split}\]

As said before, the energy obtained from these potentials is equal to the electrostatic energy obtained by the spherical summation limit. The deeper reason for this is that in some sense the electrostatic sum is absolutely convergent [AH02b].

The near formula is used for particles with a small distance along the z axis, for all other particles the far formula is used. Below is shown, that the far formula can be evaluated much more efficiently, however, its convergence breaks down for small z distance. To efficiently implement MMM2D, the layered cell system is required, which splits up the system in equally sized gaps along the z axis. The interaction of all particles in a layer S with all particles in the layers S-1,S,S+1 is calculated using the near formula, for the particles in layers \(1,\dots,S-2\), and in layers \(S+2,\dots,N\), the far formula is used.

The implementation of the near formula is relatively straight forward and can be treated as any short ranged force is treated using the link cell algorithm, here in the layered variant. The special functions in the formula are somewhat demanding, but for the polygamma functions Taylor series can be achieved, which are implemented in mmm-common.h. The Bessel functions are calculated using a Chebychev series.

The treatment of the far formula is algorithmically more complicated. For a particle i in layer \(S_i\), the formula can product decomposed, as in

\[\begin{split}\begin{array}{rl} \sum_{j\in I_S, S < S_i - 1} q_iq_j\frac{e^{-2\pi f_{pq}|z_i-z_j|}}{f_{pq}} \cos(\omega_p (x_i - x_j))\cos(\omega_q (y_i - y_j)) = \\ q_i\frac{e^{-2\pi f_{pq}z_i}}{f_{pq}} \cos(\omega_p x_i)\cos(\omega_q y_i) \sum_{j\in I_S, S < S_i - 1}q_je^{2\pi f_{pq}z_j} \cos(\omega_p x_j)\cos(\omega_q y_j) + \\ q_i\frac{e^{-2\pi f_{pq}z_i}}{f_{pq}} \cos(\omega_p x_i)\sin(\omega_q y_i) \sum_{j\in I_S, S < S_i - 1}q_je^{2\pi f_{pq}z_j} \cos(\omega_p x_j)\sin(\omega_q y_j) + \\ q_i\frac{e^{-2\pi f_{pq}z_i}}{f_{pq}} \sin(\omega_p x_i)\cos(\omega_q y_i) \sum_{j\in I_S, S < S_i - 1}q_je^{2\pi f_{pq}z_j} \sin(\omega_p x_j)\cos(\omega_q y_j) + \\ q_i\frac{e^{-2\pi f_{pq}z_i}}{f_{pq}} \sin(\omega_p x_i)\sin(\omega_q y_i) \sum_{j\in I_S, S < S_i - 1}q_je^{2\pi f_{pq}z_j} \sin(\omega_p x_j)\sin(\omega_q y_j). \end{array}\end{split}\]

This representation has the advantage, that the contributions of the two particles are decoupled. For all particles j only the eight terms

\[\xi^{(\pm,s/c,s/c)}_j= q_je^{\pm 2\pi f_{pq}z_j} \sin/\cos(\omega_p x_j)\sin/\cos(\omega_q y_j)\]

are needed. The upper index describes the sign of the exponential term and whether sine or cosine is used for \(x_j\) and \(y_j\) in the obvious way. These terms can be used for all expressions on the right hand side of the product decomposition. Moreover it is easy to see from the addition theorem for the sine function that these terms also can be used to calculate the force information up to simple prefactors that depend only on p and q.

Every processor starts with the calculation of the terms \(\xi^{(\pm,s/c,s/c)}_j\) and adds them up in each layer, so that one obtains

\[\Xi^{(\pm,s/c,s/c)}_s= \sum_{j\in S_s}\xi^{(\pm,s/c,s/c)}_j.\]

Now we calculate

\[\Xi^{(l,s/c,s/c)}_s=\sum_{t < s - 1}\Xi^{(+,s/c,s/c)}_t\]

and

\[\Xi^{(h,s/c,s/c)}_s=\sum_{t > s + 1}\Xi^{(-,s/c,s/c)}_t,\]

which are needed for the evaluation of the product decomposition. While the bottom processor can calculate \(\Xi^{(l,s/c,s/c)}_s\) directly, the other processors are dependent on its results. Therefore the bottom processor starts with the calculation of its \(\Xi^{(l,s/c,s/c)}_s\) and sends up \(\Xi^{(l,s/c,s/c)}_s\) and \(\Xi^{(+,s/c,s/c)}_s\) of its top layer s to the next processor dealing with the layers above. Simultaneously the top processor starts with the calculation of the \(\Xi^{(h,s/c,s/c)}_s\) and sends them down. After the communicated has been completed, every processor can use the \(\Xi^{(l/h,s/c,s/c)}_j\) and the \(\xi^{(\pm,s/c,s/c)}_j\) to calculate the force rsp. energy contributions for its particles.

In pseudo code, the far formula algorithm looks like:

  1. for each layer \(s=1,\ldots,S\)

    1. \(\Xi^{(\pm,s/c,s/c)}_s=0\)

    2. for each particle \(j\) in layer \(s\)

      1. calculate \(\xi^{(\pm,s/c,s/c)}_j\)

      2. \(\Xi^{(\pm,s/c,s/c)}_s += \xi^{(\pm,s/c,s/c)}_j\)

  2. \(\Xi^{(l,s/c,s/c)}_3=\Xi^{(+,s/c,s/c)}_1\)

  3. for each layer \(s=4,\ldots,S\)

    1. \[\Xi^{(l,s/c,s/c)}_s=\Xi^{(l,s/c,s/c)}_{s-1} +\ \Xi^{(+,s/c,s/c)}_{s-2}\]
  4. \(\Xi^{(l,s/c,s/c)}_{S-2}=\Xi^{(-,s/c,s/c)}_S\)

  5. for each layer \(s=(S-3),...,1\)

    1. \[\Xi^{(l,s/c,s/c)}_s=\Xi^{(l,s/c,s/c)}_{s+1} +\ \Xi^{(-,s/c,s/c)}_{s+2}\]
  6. for each layer \(s=1,...,S\)

    1. for each particle \(j\) in layer \(s\)

      1. calculate particle interaction from \(\xi^{(+,s/c,s/c)}_j\Xi^{(l,s/c,s/c)}_s\) and \(\xi^{(-,s/c,s/c)}_j\Xi^{(h,s/c,s/c)}_s\)

For further details, see [AH02b][AH02a][AdJH02b][AdJH02a].

19.1.2.1. Dielectric contrast

A dielectric contrast at the lower and/or upper simulation box boundary can be included comparatively easy by using image charges. Apart from the images of the lowest and topmost layer, the image charges are far enough to be treated by the far formula, and can be included as starting points in the calculation of the \(\Xi\) terms. The remaining particles from the lowest and topmost layer are treated by direct summation of the near formula.

This means, that in addition to the algorithm above, one has to only a few things: during the calculation of the particle and cell blocks \(\xi\) and \(\Xi\), one additionally calculates the contributions of the image charges and puts them either in a separate array or, for the boundary layers, into two extra \(\xi\) cell blocks outside the simulation box. The entries in the separate array are then added up over all processors and stored in the \(\Xi\)-terms of the lowest/topmost layer. This are all modifications necessary for the far formula part. In addition to the far formula part, there is an additional loop over the particles at the boundary to directly calculate their interactions with their images. For details, refer to [TAH07].

19.1.3. MMM1D theory

In one dimensionally periodic systems with z being the periodic coordinate, the far formula looks like

\[\begin{split}\begin{array}{rl} \phi(\rho,z) &=\, 4 u_z\sum_{p\neq 0} K_0(\omega\rho)\cos(\omega z) - 2u_z\log(\frac{\rho}{2\lambda_z}) - 2u_z\gamma\\ F_\rho(\rho,z) &=\, 8\pi u_z^2\sum_{p\neq 0} p K_1(\omega\rho)\cos(\omega z) + \frac{2 u_z}{\rho}\\ F_z(\rho,z) &=\, 8\pi u_z^2 \sum_{p\neq 0} pK_0(\omega\rho)\sin(\omega z), \end{array}\end{split}\]

the near formula is

\[\begin{split}\begin{array}{rl} \tilde{\phi}(\rho,z) &=\, -u_z\sum_{n\ge 0} \left(\begin{array}{c}-\frac{1}{2}\\n\end{array}\right) \frac{\left(\psi^{(2n)}(N_\psi + u_z z) + \psi^{(2n)}(N_\psi - u_z z)\right)}{(2n)!}(u_z\rho)^{2n} - 2u_z\gamma + \\ &\phantom{=\,++} \sum_{k=1}^{N_\psi-1}\left(\frac{1}{r_k}+\frac{1}{r_{-k}}\right)\\ \tilde{F}_\rho(\rho,z) &=\, -u_z^3 \sum_{n\ge 0} \left(\begin{array}{c}-\frac{1}{2}\\n\end{array}\right) \frac{\left(\psi^{(2n)}(N_\psi + u_z z) + \psi^{(2n)}(N_\psi - u_z z)\right)}{(2n)!}(u_z\rho)^{2n-1} + \\ &\phantom{=\,++} \sum_{k=1}^{N_\psi-1}\left(\frac{\rho}{r_k^3}+\frac{\rho}{r_{-k}^3}\right) \\ \tilde{F}_z(\rho,z) &=\, -u_z^2 \sum_{n\ge 0} \left(\begin{array}{c}-\frac{1}{2}\\n\end{array}\right) \frac{\left(\psi^{(2n + 1)}(N_\psi + u_z z) + \psi^{(2n + 1)}(N_\psi - u_z z)\right)}{(2n)!}(u_z\rho)^{2n} + \\ &\phantom{=\,++} \sum_{k=1}^{N_\psi-1}\left(\frac{z+k\lambda_z}{r_k^3}+\frac{z-k\lambda_z}{r_{-k}^3}\right), \end{array}\end{split}\]

where \(\rho\) denotes the xy-distance of the particles. As for the two dimensional periodic case, the obtained energy is equal to the one dimensional Ewald sum. Algorithmically, MMM1D is uninteresting, since neither the near nor far formula allow a product decomposition or similar tricks. MMM1D has to be implemented as a simple NxN loop. However, the formulas can be evaluated efficiently, so that MMM1D can still be used reasonably for up to 400 particles on a single processor [AH05].

19.1.4. ELC theory

The ELC method differs from the other MMM algorithms in that it is not an algorithm for the calculation of the electrostatic interaction, but rather represents a correction term which allows to use any method for three-dimensionally periodic systems with spherical summation order for two-dimensional periodicity. The basic idea is to expand the two dimensional slab system of height h in the non-periodic z-coordinate to a system with periodicity in all three dimensions, with a period of \(\lambda_z>h\), which leaves an empty gap of height \(\delta=\lambda_z - h\) above the particles in the simulation box.

Since the electrostatic potential is only finite if the total system is charge neutral, the additional image layers (those layers above or below the original slab system) are charge neutral, too. Now let us consider the n-th image layer which has an offset of \(n\lambda_z\) to the original layer. If \(n\lambda_z\) is large enough, each particle of charge \(q\_j\) at position \((x_j,y_j,z_j+n\lambda_z)\) and its replicas in the xy-plane can be viewed as constituting a homogeneous charged sheet of charge density \(\sigma_j = \frac{q_j}{\lambda_x\lambda_y}\). The potential of such a charged sheet at distance z is \(2\pi \sigma_j |z|\). Now we consider the contribution from a pair of image layers located at \(\pm n\lambda_z\), n>0 to the energy of a charge q_i at position \((x_i,y_i,z_i)\) in the central layer. Since \(|z_j - z_i| < n\lambda_z\), we have \(|z_j - z_i + n\lambda_z| = n\lambda_z + z_j - z_i\) and \(|z_j - z_i - n\lambda_z|= n\lambda_z - z_j + z_i\), and hence the interaction energy from those two image layers with the charge \(q_i\) vanishes by charge neutrality:

\[2\pi q_i \sum_{j=1}^N \sigma_j(|z_j - z_i + n\lambda_z| + |z_j - z_i - n\lambda_z|) = 4\pi q_i n\lambda_z \sum_{j=1}^N \sigma_j = 0.\]

The only errors occurring are those coming from the approximation of assuming homogeneously charged, infinite sheets instead of discrete charges. This assumption should become better when increasing the distance \(n\lambda_z\) from the central layer.

However, in a naive implementation, even large gap sizes will result in large errors. This is due to the order of summation for the standard Ewald sum, which is spherical, while the above approach orders the cells in layers, called slab-wise summation. Smith has shown that by adding to the Ewald energy the term

\[E_c=2\pi M_z^2 - \frac{2\pi M^2}{3},\]

where M is the total dipole moment, one obtains the result of a slab-wise summation instead of the spherical limit [Smi81]. Although this is a major change in the summation order, the difference is a very simple term. In fact, Smith shows that changes of the summation order always result in a difference that depends only on the total dipole moment.

Using the far formula of MMM2D, one can calculate the contributions of the additional layers up to arbitrarily precision, even for small gap sizes. This method is called electrostatic layer correction, ELC. The advantage of this approach is that for the image layers, z is necessarily large enough, so that all interactions can be represented using the product decomposition. This allows for an order N evaluation of the ELC term.

The electrostatic layer correction term is given by

\[E_{lc}=\sum_{i,j=1}^Nq_iq_j\psi(p_i-p_j),\]

where

\[\begin{split}\begin{array}{rl} \psi(x,y,z)=&4u_xu_y\sum_{p,q>0}\frac{\cosh(2\pi f_{pq}z)}{f_{pq}(e^{2\pi f_{pq}\lambda_z} - 1)} \cos(\omega_p x)\cos(\omega_q y) + \\ &2u_xu_y\sum_{p>0}\frac{\cosh(2\pi f_p z)}{f_p(e^{2\pi f_p\lambda_z} - 1)}\cos(\omega_p x)+\\ &2u_xu_y\sum_{q>0}\frac{\cosh(2\pi f_q z)}{f_q(e^{2\pi f_q\lambda_z} - 1)}\cos(\omega_q y). \end{array}\end{split}\]

The implementation is very similar to MMM2d, except that the separation between slices close by, and above and below is not necessary.

19.1.5. Errors

Common to all algorithms of the MMM family is that accuracy is cheap with respect to computation time. More precisely, the maximal pairwise error, i.e. the maximal error of the \(\psi\) expression, decreases exponentially with the cutoffs. In turn, the computation time grows logarithmically with the accuracy. This is quite in contrast to the Ewald methods, for which decreasing the error bound can lead to excessive computation time. For example, P3M cannot reach precisions above \(10^{-5}\) in general. The precise form of the error estimates is of little importance here, for details see [AdJH02b][AdJH02a].

One important aspect is that the error estimates are also exponential in the non-periodic coordinate. Since the number of close by and far away particles is different for particles near the border and in the center of the system, the error distribution is highly non-homogeneous. This is unproblematic as long as the maximal error is really much smaller than the thermal energy. However, one cannot interpret the error simply as an additional error source.

Error distribution of the ELC method.

Error distribution of the ELC method.

Figure [fig:ELC-error] shows the error distribution of the ELC method for a gap size of \(10\%\) of the total system height. For MMM2D and MMM1D the error distribution is less homogeneous, however, also here it is always better to have some extra precision, especially since it is computationally cheap.