# DLPNO-MP2: Domain-Based Local Pair Natural Orbital MP2¶

Code author: Zach Glick

Section author: Zach Glick

Module: Keywords, PSI Variables, DLPNOMP2

## Introduction¶

The steep polynomial scaling (in both time and memory) of post-HF dynamic correlation methods prohibits calculations on large systems, even for efficient codes like PSI4‘s DF-MP2. This poor scaling is in part due to the use of canonical HF orbitals, which are entirely delocalized across the molecule. Canonical orbitals are commonly used because of mathematical convenience. Another possible choice is localized orbitals. Any two orbitals localized to separate regions of a molecule can be treated as non-interacting to a good approximation. Thus, when working with localized orbitals, the number of interacting orbital pairs (and triples, quadruples, etc.) scales linearly with system size. If carefully implemented, programs that exploit this sparsity can be made to scale linearly (or else with lower order than their canonical counterparts) at the cost of of modest, controllable errors. This is the defining insight of DLPNO-MP2 and all related “local correlation” methods.

The DLPNO-MP2 code is a linear-scaling alternative to the DF-MP2 code, and is intended for use with large systems for which DF-MP2 is intractable. When running DLPNO-MP2 with default settings, approximately 99.9% of the DF-MP2 correlation energy is recovered. The general outline of the method is as follows:

1. Localize the active occupied MOs (with the Foster-Boys method)

2. Construct projected AOs (PAOs) from the virtual MOs

3. Calculate three-index integrals in the (sparse) LMO/PAO basis

4. Perform local density fitting to construct (sparse) exchange integrals

5. Transform local virtuals from PAOs to pair natural orbitals (PNOs), and truncate

6. Solve the iterative local MP2 equations in the LMO/PNO basis

An example input file is:

molecule h2o {
0 1
O
H 1 1.0
H 1 1.0 2 104.5
symmetry c1
}

set basis cc-pvdz
set scf_type df
set freeze_core True
set pno_convergence normal

energy('dlpno-mp2')


The main difference between this input and a DF-MP2 input is the energy('dlpno-mp2') call to energy(). The only other addition is the PNO_CONVERGENCE keyword, which determines the accuracy of the local approximations underlying the DLPNO-MP2 method. Note that the water molecule in this example is not large enough for DLPNO-MP2 to be of any benefit relative to DF-MP2.

The theory of the DLPNO-MP2 method and practical recommendations for using the code are presented below.

## Theory¶

See DF-MP2 for background on the theory of (non-local) density-fitted MP2. PSI4‘s DLPNO-MP2 implementation is based on the manuscript by Pinski et al. [Pinski:2015:034108].

In DLPNO-MP2, as in all local MP2 methods, the second-order MBPT energy is determined variationally via the Hylleraas functional [Hylleraas:1930:209]:

(1)$E^{(2)} = 2 \langle \Psi_{0}^{(0)} | \hat{H} - E_{0}^{(0)} | \Psi_{0}^{(1)} \rangle - \langle \Psi_{0}^{(1)} | \hat{H}^{(0)} - E_{0}^{(0)} | \Psi_{0}^{(1)} \rangle = \min_{| \Psi_{0}^{(1)} \rangle}.$

Determining the optimal $$| \Psi_{0}^{(1)} \rangle$$ entails iteratively minimizing the following residual [Pulay:1986:357]:

(2)$R_{ij}^{ab} = (ia|jb) + (\epsilon_a + \epsilon_b - f_{ii} - f_{jj})t_{ij}^{ab} - \sum_{k \ne j} f_{ik} \sum_{c,d} S_{ac}t_{kj}^{c,d}S_{db} - \sum_{k \ne i} f_{kj} \sum_{cd} S_{ac}t_{ik}^{cd}S_{db}$

where i, j, and k are (not necessarily canonical) occupied orbitals, a, b, c, and d are virtual orbitals, $$f_{ij}$$ are fock matrix elements, $$S_{ab}$$ are orbital overlaps, and finally $$t_{ij}^{ab}$$ are the MP2 amplitudes to be solved for. Virtual orbitals may be different for each pair of occupied orbitals. For a given occupied orbital pair ij, all virtuals are orthogonal and canonical, but virtuals belonging to different pair domains may not be orthogonal.

The following expression is used to evaluate the energy of a given set of amplitudes:

(3)$\begin{split}E^{(2)} &= \sum_{i,j} e_{ij}, \\ e_{ij} &= \sum_{a, b}((ia|jb) + R_{ij}^{ab})(2t_{ij}^{ab} - t_{ij}^{ba}).\end{split}$

The error in $$E^{(2)}$$ scales quadratically with the error in the amplitudes.

No local approximations have been made so far, and this iterative approach can be used to exactly determine $$E^{(2)}$$ with $${\cal O}(N^5)$$ cost. In DLPNO-MP2, the first local approximation is to screen distant, non-interacting orbital pairs ij. Orbital pairs are screened if below both an overlap criteria:

(4)$DOI_{ij} \equiv \sqrt{\int d\mathbf{r} | \chi_{i}(\mathbf{r}) | ^{2} | \chi_{j}(\mathbf{r}) | ^{2}},$

and a pair energy estimate:

(5)$e_{ij}^{approx} = -\frac{4}{R^{6}} \sum_{a_{i} \in [i],b_{j} \in [j]} \frac{ (2 \langle i | \mathbf{r} | a_{i} \rangle \langle j | \mathbf{r} | b_{j} \rangle)^{2}}{\epsilon_{a_{i}} + \epsilon_{b_{j}} - f_{ii} - f_{jj}},$

in which small domains of virtual orbitals are used for each local MO. As a result, an asymptotically linear number of ij pairs enter the local MP2 equations, and the approximate pair energy of neglected pairs is added to the final energy.

The second major local approximation in DLPNO-MP2 is the truncation of the virtual space. Initially, exchange integrals are calculated in the LMO/PAO basis using the standard density fitting approach:

(6)$(ia|jb) = \sum_{K,L \in [ij]} (ia|K)[\mathbf{J}^{-1}]_{KL}(L|jb)$

This is done with linear scaling effort by exploiting the locality of the LMOs, PAOs, and auxiliary basis functions. Solving the iterative local MP2 equations in the LMO/PAO basis requires large PAO domains to achieve reasonable accuracy. Instead, the virtual space is transformed into the much more compact pair natural orbital representation. The (approximate) PNOs diagonalize the virtual-virtual block of the (approximate) MP2 density matrix:

(7)$D_{ij}^{ab} = \frac{1}{1 + \delta_{ij}}[\tilde{t}_{ij}^{\dagger}t_{ij} + \tilde{t}_{ij}t_{ij}^{\dagger}]^{ab}$

which is constructed from semicanonical amplitudes:

(8)$\begin{split}t_{ij}^{ab} &= - \frac{(iajb)}{\epsilon_{a} + \epsilon_{b} - \epsilon_{i} - \epsilon_{j}}, \\ \tilde{t}_{ij}^{ab} &= 2t_{ij}^{ab} - t_{ij}^{ba}.\end{split}$

PNOs with small occupation numbers are discarded, and the local MP2 equations are solved in the LMO/PNO basis.

## Recommendations¶

Some practical notes on running the code:

• DLPNO-MP2 is not a drop-in replacement for DF-MP2. Instead, it should be used for large calculations that cannot reasonably be performed with DF-MP2. The crossover point between DF-MP2 and DLPNO-MP2 depends on details of both the calculation and the hardware, but can be as low as 2,000 basis functions.

• The accuracy of DLPNO-MP2 (relative to DF-MP2) can be controlled with the PNO_CONVERGENCE keyword according to recommendation by Liakos et al. [Liakos:2015:1525]. For non-covalent interactions TIGHT is highly recommended.

• The greater the spatial sparsity of a molecular system, the smaller the pair domains and consequently the faster the calculation. DLPNO-MP2 is much faster for linear alkanes than for globular proteins, all else constant.

• Similar to molecular sparsity, the sparsity of the orbital basis affects runtime. Diffuse functions increase the size of the pair domains and therefore lead to longer calculations.

• All aspects of DLPNO-MP2 run in core; no disk is required. As a result, the code exhibits very good intra-node parallelism, and benefits from many threads. The amount of memory needed scales linearly with system size.

• DLPNO-MP2 is not symmetry aware. This should not be a concern for large systems in which symmetry is seldom present.

• As with DF-MP2, freezing core orbitals (by setting FREEZE_CORE to True) is recommended for efficiency. In DLPNO methods, this is also recommended for accuracy, since core excitations are known to exhibit greater errors relative to valence excitations.

• At the moment, the DLPNO-MP2 code is only compatible with with RHF references.