HDR/Manuscript/Chapter4/chapter4.1.tex
2021-05-20 15:13:48 +02:00

92 lines
10 KiB
TeX

%****************************************************************
\section{History}
%****************************************************************
It is well known that highly-accurate wave functions require the fulfilment (or near-fulfilment) of the electron-electron cusp conditions \cite{Kato51, Kato57, Pack66, Morgan93, Myers91, Tew08, QuasiExact09, ExSpherium10, QR12, Kurokawa13, Kurokawa14, Gruneis17}.
For correlated wave functions expanded in terms of products of one-electron Gaussian basis functions, the energy converges as $\order{L^{-3}}$, where $L$ is the maximum angular momentum of the basis set \cite{Kutzelnigg85}.
This slow convergence can be tracked down to the inadequacy of these products to properly model the Coulomb correlation hole \cite{Hattig12, Kong12}.
In the late 20's, Hylleraas solved this issue for the helium atom by introducing explicitly the interelectronic distance $\ree = \abs{\br_1 - \br_2}$ as an additional two-electron basis function \cite{Hylleraas28, Hylleraas29}.
As Kutzelnigg later showed, this leads to a prominent improvement of the energy convergence from $\order{L^{-3}}$ to $\order{L^{-7}}$ \cite{Kutzelnigg85}.
Around the same time, Slater, while studying the Rydberg series of helium \cite{Slater28}, suggested a new correlation factor known nowadays as a Slater geminal:
\begin{equation}
S_{12} = \exp(-\lambda_{12}\, \ree ).
\end{equation}
Unfortunately, the increase in mathematical complexity brought by $\ree$ or $S_{12}$ has been found to be rapidly computationally overwhelming.\footnote{Note that, although Slater was the first to propose such correlation factor, he suggested to set $\lambda_{12} = -1/2$ in order to ensure that the wave function fulfils Kato's electron-electron cusp condition. However, Hartree and Ingman \cite{Hartree33} found this correlation factor physically unreasonable due to its behaviour at large $\ree$, and suggested that a correlation factor of the form $1 - c \exp(-\lambda_{12}\, \ree)$ (with $\lambda_{12} >0$) would be more appropriate. We refer the interested reader to the review of Hattig et al.~\cite{Hattig12} for a detailed historical overview.}
In 1960, Boys \cite{Boys60} and Singer \cite{Singer60} independently proposed to use the Gaussian geminal (GG) correlation factor
\begin{equation}
\label{eq:gg}
G_{12} = \exp(-\lambda_{12}\, \ree^2 ),
\end{equation}
as \cite{Boys60}
\begin{quote}
``\textit{there are explicit formulas for all of the necessary many-dimensional integrals}".
\end{quote}
Interestingly, in the same article, a visionary Boys argued that, even if GGs do not fulfil the electron-electron cusp conditions, they could be used to fit $S_{12}$.
During the following years, variational calculations involving GGs flourished, giving birth to various methods, such as the exponentially-correlated Gaussian method \cite{RychBook, Bubin2005, Bubin2009, Szalewicz2010}.
However, this method was restricted to fairly small systems as it requires the optimisation of a large number of non-linear parameters.
In the 70's, the first MP2 calculations including GGs appeared thanks to the work by Pan and King \cite{Pan70, Pan72}, Adamowicz and Sadlej \cite{Adamowicz77a, Adamowicz77b, Adamowicz78}, and later Szalewicz \textit{et al.} \cite{Szalewicz82, Szalewicz83}.
Even if these methods represented a substantial step forward in terms of computational burden, they still require the optimization of a large number of non-linear parameters.
In 1985, Kutzelnigg derived a first form of the MP2-R12 equations using $\ree$ as a correlation factor \cite{Kutzelnigg85}.
Kutzelnigg's idea, which was more formally described together with Klopper in 1987 \cite{Klopper87}, dredged up an old problem: in addition to two-electron integrals (traditional ones and new ones), three-electron and four-electron integrals were required.
At that time, the only way to evaluate them would have been via an expensive one- or two-dimensional Gauss-Legendre quadrature \cite{Preiskorn85, Clementi89}.
Additionally, citing Kutzelnigg and Klopper \cite{Kutzelnigg91},
\begin{quote}
\textit{``even if fast procedures for the evaluation of these integrals were available, one would have to face the problem of the large number of these integrals; while that of two-electron integrals is $\sim N^{4}$, there are $\sim N^{6}$ three-electron and $\sim N^{8}$ four-electron integrals.
The storing and manipulating of these integrals could be handled only for extremely small basis sets.''}
\end{quote}
Undoubtedly, in the late 80's, the two-electron integrals technology was still in development \cite{MD78, PH78, King76, Dupuis76, Rys83, OS1, HGP, Gill94b}.
Nowadays, though still challenging, these integrals could be computed much more effectively via judicious recursive schemes, designing the quadrature only to the fundamental integrals \cite{3ERI1}.
Another important remark is that the actual number of \textit{significant} (i.e.~greater than a given threshold) three- and four-electron integrals in a large system, is, at worst, $\order{N^{3}}$ or $\order{N^{4}}$.
These kinds of scaling are achievable, for example, by exploiting robust density fitting \cite{Womack14} or upper bound-based screening methods, as discussed below \cite{3ERI2}.
Nevertheless, the success of the R12 method was triggered by the decision to avoid the computation of these three- and four-electron integrals through the use of the \textit{resolution of the identity} (RI) approximation \cite{Kutzelnigg91, Hattig12, Werner07}.
In this way, three- and four-electron integrals are approximated as linear combinations of products of two-electron integrals.
Several key developments and improvements of the original MP2-R12 approach have been proposed in the last decade \cite{MOLPRO2011, Hattig12, Werner07, Kong12}.
Of course, the accuracy of the RI approximation relies entirely on the assumption that the auxiliary basis set is sufficiently large, i.e.~$N_\text{RI} \gg N$, where $N$ and $N_\text{RI}$ are the number of basis functions in the primary and auxiliary basis sets, respectively.
The use of RI as method of choice does not seem definitive to us.
In fact, eschewing the RI approximation would offer at least two advantages:
i) smaller one-electron basis as the larger auxiliary basis set would not be required anymore;
ii) the three- and four-electron integrals would be computed exactly.
Moreover, one could avoid the commutator rearrangements involved in the computation of integrals over the kinetic energy operator \cite{Rohse93}.
In 1996, Persson and Taylor killed two birds with one stone.
Using a pre-optimized GG expansion fitting $\ree$
\begin{equation}
\label{eq:r12fit}
\ree \approx \sum_{k} a_{k} \qty[1-\exp(-\lambda_{k} \ree^{2}) ],
\end{equation}
they avoided the non-linear optimisation, and eschewed the RI approximation thanks to the analytical integrability of three- and four-electron integrals over GGs \cite{Persson96}.
They were able to show that a six- or ten-term fit introduces a $0.5$ mE$_\text{h}$ or $20$ $\mu$E$_\text{h}$ error, respectively \cite{Persson96}.
Unfortunately, further improvements were unsuccessful due to the failure of $\ree$ in modelling the correct behaviour of the wave function for intermediate and large $\ree$ \cite{Tew05, Tew06}.
In fact, Ten-no showed that $S_{12}$ is near-optimal at describing the correlation hole, and that a 10-term GG fit of $S_{12}$ yields very similar results.
This suggests that, albeit not catching the cusp per se, the Coulomb correlation hole can be accurately represented by GGs \cite{Tenno04a, Tenno07, May04, May05}.
Methods for evaluating many-electron integrals involving GGs have already been developed.
As mentioned previously, Persson and Taylor \cite{Persson97} derived recurrence relations based on Hermite Gaussians, analogously to the work of McMurchie and Davidson for two-electron integrals \cite{MD78}.
These recurrence relations were implemented by Dahle \cite{DahleThesis, Dahle2007, Dahle2008}.
Saito and Suzuki \cite{Saito01} also proposed an approach based on the work by Obara and Saika \cite{OS1, OS2}.
More recently, a general formulation using Rys polynomials \cite{King76, Dupuis76, Rys83} was published by Komornicki and King \cite{Komornicki11}.
Even if limited to the three-centre case, it is worth mentioning that May has also developed recurrence relations for two types of three-electron integrals \cite{MayThesis}.
These recurrence relations were implemented by Womack using automatically-generated code \cite{WomackThesis}.
Recently, we have developed recurrence relations for three- and four-electron integrals for generic correlation factors \cite{3ERI1, 4ERI1}.
All these recursive schemes have variable computational cost depending on the degree of contraction of the integral class to be computed.
Unsurprisingly, these algorithms are, to a large extent, complementary, and none of the algorithms has proven optimal under all circumstances.
A major limitation of all these approaches is that they do not include any integral screening.
Indeed, a remarkable consequence of the short-range nature of the Slater and Gaussian correlation factors is that, even if formally scaling as $\order{N^{6}}$ and $\order{N^{8}}$, there are only $\order{N^{2}}$ \textit{significant} (i.e.~greater than a given threshold) three- and four-electron integrals in a large system \cite{3ERI1, 3ERI2}.
Therefore, it is paramount to devise rigorous upper bounds to avoid computing the large number of negligible integrals.
This chapter is organised as follows.
First, we discuss Gaussian basis functions, many-electron integrals and the structure of the three- and four-electron operators considered here.
In the next three sections, we introduce the main ingredients for the efficient computation of three- and four-electron integrals involving GGs: i) fundamental integrals (FIs), ii) upper bounds (UBs), and iii) recurrence relations (RRs).
Finally, we give an overall view of our algorithm which is an extension of the late-contraction path of PRISM, also known as the Head-Gordon-Pople (HGP) algorithm (see Refs.~\cite{HGP, Gill94b} and references therein).
Note that the HGP algorithm corresponds to the Obara-Saika scheme where one has introduced additional RRs and provides a precise description of how these RRs can be used in tandem to good effects.