MP²A: MP2 Predictor & Multi-Precision Arithmetic

• MP²A is an ambiguous designation that encompasses both a MP2 accuracy predictor (MAP) for weak interactions and a Fortran-based multiple-precision arithmetic system.
• MAP uses the Hartree–Fock adiabatic connection and SPL interpolation to gauge MP2 reliability by quantifying the linearity of the interaction energy curve.
• The MP context employs ANSI Fortran subroutines for high-precision arithmetic, facilitating machine-independent evaluation of elementary and special functions.

Searching arXiv for the cited works and the term to ground the article. arXiv search query: (Vuckovic et al., 2020) MP2 accuracy predictor weak interactions “MP$^2$A” is not presented in the cited arXiv sources as the formal name of a single standardized method or software system. In the materials associated with this label, two distinct contexts appear. One is MAP, the MP2 accuracy predictor for weak interactions, which estimates the reliability of second-order Møller–Plesset perturbation theory from the structure of the Hartree–Fock adiabatic connection (Vuckovic et al., 2020). The other is MP, Brent’s portable Fortran package for multiple-precision floating-point arithmetic and special functions (Brent, 2010). This suggests that “MP$^2$A” is best understood as an ambiguous designation rather than a canonical technical term.

1. Nomenclature and scope

The most direct chemically relevant interpretation is MAP, introduced as the MP2 accuracy predictor. MAP is described as a simple, physically motivated diagnostic for how reliable second-order Møller–Plesset perturbation theory (MP2) is for weak interactions, especially noncovalent binding energies (Vuckovic et al., 2020). Its purpose is not to replace MP2, but to indicate when the MP2 approximation is likely to be trustworthy and when it should be treated with caution.

A separate and unrelated interpretation arises from Brent’s MP package. That work does not describe a separate package called “MP$^2$A”; rather, it describes MP itself, a package of ANSI Standard Fortran subroutines for multiple-precision floating-point arithmetic and the evaluation of elementary and special functions (Brent, 2010). In this reading, the notation “MP$^2$A” can only be treated as an interpretive label associated with MP and its extended arithmetic capabilities, not as the formal name of a distinct package.

2. Adiabatic-connection basis of the MP2 accuracy predictor

MAP is built from the Hartree–Fock (HF) adiabatic connection (AC). In that framework one defines a coupling-constant dependent Hamiltonian

$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$

where $\hat{T}$ is kinetic energy, $\hat{V}_{\rm ext}$ is the external potential, $\hat{V}_{ee}$ is the electron-electron repulsion, and $\hat{J}$, $\hat{K}$ are the Hartree and exchange operators computed from the physical HF reference and held fixed along the path (Vuckovic et al., 2020). The corresponding correlation energy is obtained from

$^2$0

with HF-AC integrand

$^2$1

Expanding the integrand around $^2$2 gives the perturbation series, and truncation at second order yields

$^2$3

In this sense, MP2 is exactly the linear approximation to the exact HF AC curve near $^2$4 (Vuckovic et al., 2020). The central physical interpretation follows directly: if the exact HF AC curve is close to a straight line between $^2$5 and $^2$6, then MP2 should be good; if the exact AC curve has strong curvature, MP2 will be less reliable.

The paper also discusses an earlier exact indicator based on comparing the endpoint $^2$7 with the initial slope $^2$8. For a perfectly linear AC curve, the indicator equals $^2$9, and MP2 is exact; convexity or concavity shifts the value away from $^2$0, thereby quantifying nonlinearity. The helium isoelectronic series illustrates this idea: as nuclear charge $^2$1 increases, the AC curve becomes more linear and the indicator tends to $^2$2, consistent with the increasing accuracy of MP2 in the large-$^2$3 limit (Vuckovic et al., 2020).

3. Interpolated construction of MAP

The exact indicator is not practical for routine use because it requires the fully interacting wavefunction $^2$4, through $^2$5. MAP is obtained by replacing the exact endpoint with an interpolated estimate along the HF adiabatic connection using the Seidl–Perdew–Levy (SPL) form

$^2$6

with input set

$^2$7

For the HF case, $^2$8 (Vuckovic et al., 2020).

For a complex $^2$9 composed of fragments $^2$0, the interaction AC curve is defined as

$^2$1

The interpolated interaction curve is then

$^2$2

The construction works cleanly for systems whose fragments have non-degenerate ground states, because under that condition the interaction AC vanishes at infinite separation (Vuckovic et al., 2020).

The practical predictor is defined through

$^2$3

In operational terms, MAP is a dimensionless measure of how far the HF AC interaction curve deviates from the MP2 straight-line picture, with the practical endpoint supplied by interpolation rather than by the exact wavefunction (Vuckovic et al., 2020).

The interpolation requires a large-$^2$4 input. For that purpose the paper uses the point-charge-plus-continuum (PC) model,

$^2$5

with

$^2$6

This is a semilocal approximation evaluated on the HF density, and the authors note that the “bare” PC model works best in this context (Vuckovic et al., 2020).

4. Computational characteristics, tested systems, and observed behavior

A notable property of MAP is that it comes at negligible extra cost beyond the MP2 calculation itself (Vuckovic et al., 2020). The required ingredients are the usual MP2 correlation energy, the HF density $^2$7, its gradient $^2$8, and the PC-model estimate $^2$9. This makes the predictor suitable as a screening tool for MP2 reliability in weakly bound complexes.

The method is illustrated on the helium isoelectronic series—H$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$0, He, Be$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$1, Ne$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$2—and on noncovalent dimers including the benzene dimer and acetic acid dimer (Vuckovic et al., 2020). Its main practical assessment is on the S22 and S66 noncovalent benchmark sets. The reported trend is monotonic in the intended sense: as MAP decreases toward 0, the AC curve becomes more linear and MP2 becomes more accurate.

The benchmark observations are specific. When MAP < 0.20, the relative MP2 errors are always below 25% in the reported tests; around MAP $$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$3, MP2 errors can become very large, up to about 80% (Vuckovic et al., 2020). The large-MAP cases are often stacking complexes, which are known failure cases for MP2, whereas hydrogen-bonded systems tend to have smaller MAP values and better MP2 performance.

For the S66 subsets, mean absolute errors and mean absolute relative errors are reported separately for hydrogen-bonded, mixed/others, and dispersion-dominated complexes. The averaged MAP increases in the order

$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$4

and MP2 accuracy decreases correspondingly (Vuckovic et al., 2020). This aligns with the established observation that MP2 is less reliable for dispersion-driven and stacked systems than for many hydrogen-bonded complexes.

5. Applicability, limits, and interpretive status

MAP is intended for interaction energies / weak interactions, not for arbitrary strongly correlated or multireference problems (Vuckovic et al., 2020). The predictor is expected to be useful when the system is a weakly interacting complex, the fragments have non-degenerate ground states, the interaction can be reasonably described by the HF AC plus interpolation, and the MP2 error is mainly governed by the curvature of the AC curve. Under those conditions, a small MAP functions as a compact warning signal of near-linearity and probable MP2 reliability, whereas a large MAP signals pronounced curvature and probable unreliability.

The stated limitations are explicit. The current predictor is restricted to systems dissociating into fragments with non-degenerate ground states. The SPL interpolation is not exact for the HF AC curve itself, but is used as a practical surrogate. For fragments with degeneracy or near-degeneracy, a different treatment is needed, and future work is mentioned on small covalently bonded diatomics to extend the approach (Vuckovic et al., 2020).

The interpretive status of MAP is therefore diagnostic rather than foundational. It does not alter the formal definition of MP2; instead, it uses AC geometry to anticipate when the MP2 approximation is likely to succeed or fail. A plausible implication is that MAP is most informative in exactly those regimes where weak interaction energies are inexpensive enough to compute at MP2 level, yet sufficiently delicate that qualitative misranking or substantial overbinding remains a practical concern.

6. Multiple-precision arithmetic: the separate MP context

In a separate line of work relevant to the designation “MP$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$5A,” MP is a package of ANSI Standard Fortran subroutines for multiple-precision floating-point arithmetic and the evaluation of elementary and special functions (Brent, 2010). It is intended to let a Fortran program work with real numbers at arbitrarily high precision, subject to storage limitations and the chosen number of digits. Brent emphasizes machine independence or near machine independence: the package should run on any machine with an ANSI Fortran (1966) compiler, enough memory, and at least a 16-bit integer wordlength.

The key runtime parameters are the base $$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$6 and the number of digits $$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$7, both user-selectable and dynamically changeable. A multiple-precision number is stored in an integer array in unpacked form as

$$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$8

Word 1 contains the sign $$\hat{H}_{\lambda}^{\rm HF}=\hat{T}+\hat{V}_{\rm ext}+\lambda \hat{V}_{ee}+ \Big(1-\lambda \Big) \Big(\hat{J}+\hat{K} \Big),$$9, $\hat{T}$0, or $\hat{T}$1; word 2 contains the exponent to base $\hat{T}$2; and the remaining words store the normalized fraction with one base-$\hat{T}$3 digit per word (Brent, 2010). A packed format is also supported, with two base-$\hat{T}$4 digits per word, to save storage.

The package supports truncated arithmetic, rounded-to-nearest arithmetic, and directed rounding. The rounding mode is controlled by RNDRL in COMMON /MPCOM/, where 0 means truncate/chop, 1 means round to nearest with ties to even, 2 means round down, and 3 means round up (Brent, 2010). For rounded arithmetic, the basic operations produce correctly rounded $\hat{T}$5-digit base-$\hat{T}$6 results; for truncated arithmetic, the guide states a relative error on the order of $\hat{T}$7.

Global control is organized through the labeled common block

$\hat{T}$8

which stores parameters such as [BASE](https://www.emergentmind.com/topics/benchmarking-autonomy-in-scientific-experiments-base-scale), NUMDIG, MAXEXP, LUN, workspace pointers, and controls for precision, formatting, rounding, and underflow (Brent, 2010). Exponents range from $\hat{T}$9 to $\hat{V}_{\rm ext}$0, so representable numbers lie roughly between $\hat{V}_{\rm ext}$1 and almost $\hat{V}_{\rm ext}$2. Underflow is handled by MPUNFL, which usually sets the result to zero, while overflow calls MPOVFL and terminates execution with an error. The package also records the minimum and maximum exponents encountered in a run through MNEXPN and MXEXPN.

The technical restrictions are explicit: $\hat{V}_{\rm ext}$3, $\hat{V}_{\rm ext}$4, $\hat{V}_{\rm ext}$5, $\hat{V}_{\rm ext}$6 must fit in the machine’s largest integer, and $\hat{V}_{\rm ext}$7 is recommended for practical decimal accuracy (Brent, 2010). MP avoids dependence on standard library functions such as SIN, EXP, or ALOG, because one goal is to use MP to test those routines rather than depend on them. It also avoids reliance on REAL and DOUBLE PRECISION except in a few conversion routines.

Usability is extended by the Augment precompiler, which lets Fortran code declare variables as MULTIPLE or MULTIPAK and then write expressions such as

$\hat{V}_{\rm ext}$8

instead of manually invoking routines like MPADDI, MPEXP, MPMUL, MPDIV, and MPADD (Brent, 2010). The package also provides a broad function library, including arithmetic, powers and roots, elementary functions, special functions, constants, input/output, conversions, and comparisons. Representative routines include MPADD, MPMUL, MPDIV, MPEXP, MPLN, MPPI, MPBESJ, MPERF, MPGAM, and MPZETA. Algorithmic details are given for many routines; for example, MPEXP and MPLN have time $\hat{V}_{\rm ext}$9, whereas MPMUL uses the straightforward $\hat{V}_{ee}$0 multiplication algorithm with guard digits (Brent, 2010).

Taken together, these two contexts show that “MP$\hat{V}_{ee}$1A” does not denote a single unified formal object in the cited literature. In chemical electronic-structure theory it aligns most naturally with MAP, a diagnostic for MP2 reliability derived from HF adiabatic-connection curvature (Vuckovic et al., 2020). In numerical computation it may evoke the MP ecosystem for multiple-precision arithmetic, but that work describes MP, not a distinct system formally named “MP$\hat{V}_{ee}$2A” (Brent, 2010).