Configuration interaction extension of AGP for incorporating inter-geminal correlations
Airi Kawasaki, Fei Gao, Gustavo E. Scuseria · 2026
Read Configuration interaction extension of AGP for incorporating inter-geminal correlations by Airi Kawasaki, Fei Gao, Gustavo E. Scuseria (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-14115.txt.
Introduction
Kawasaki, Gao, and Scuseria extend a geminal power framework for quantum chemistry by adding correlations between electron pairs. To keep the new wave functions workable, they rewrite them as combinations of simpler geminal states whose overlaps and energy terms follow standard machinery. Tests on a model lattice and on water and nitrogen molecules show steady high accuracy, beating the simpler combination most clearly for larger systems and strongly correlated regimes.
Catalog record
| CREATOR | Airi Kawasaki, Fei Gao, Gustavo E. Scuseria |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43521 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.14115
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Configuration interaction extension of AGP for incorporating inter-geminal correlations Airi Kawasaki, Fei Gao, Gustavo E. Scuseria Abstract In this paper, we develop a class of antisymmetrized geminal power configuration interaction (AGP-CI) wave functions that extend the AGP framework by incorporating inter-geminal correlations through a CI expansion. To make these wavefunctions computationally tractable, we evaluate them by rewriting the AGP-CI ansatz as a linear combination of AGPs (LC-AGP), for which overlaps and Hamiltonian matrix elements can be computed with standard AGP machinery. Motivated by border-rank decompositions, we further reorganize this ansatz into a compact linear combination of AGPs depending on a small deformation parameter τ \tau , which controls how closely the truncated expansion approximates the full AGP-CI state. Benchmark applications to the Hubbard model and to the small molecules H 2 O and N 2 demonstrate that the proposed wavefunctions achieve consistently high accuracy and outperform the LC-AGP, particularly for systems with more electrons and in strongly correlated regimes. I Introduction The accurate description of electron correlation remains a central challenge in quantum chemistry 35 ; 25 . Beyond the independent particle approximation to incorporate electron correlation, various computational methods have been developed, including many-body perturbation theory 16 , coupled-cluster theory (CC) 2 and configuration interaction (CI) 29 approaches. However, these approaches are fundamentally constructed upon the one-body electron approximation, and their performance deteriorates as the electron correlation becomes stronger. For precise calculations in systems exhibiting significant electron correlation, that is, strongly correlated electron systems, methods based on geminals 30 ; 5 ; 37 , which are constructed from correlated electron pairs, have recently attracted considerable attention. A geminal can be defined as \displaystyle\hat{F}=\sum_{ab}^{2m}F_{ab}\hat{c}^{{\dagger}}_{a}\hat{c}^{{\dagger}}_{b}. (1) where 2m is the number of spin orbitals and F is an antisymmetric matrix. The simplest wavefunction constructed from this geminal is the antisymmetrized geminal power (AGP) wavefunction 5 ; 3 , which can be written as follows. \displaystyle\ket{\Psi_{\mathrm{AGP}}}=\hat{F}^{n}\ket{-}\equiv\ket{F} (2) where 2n is the number of electrons and \ket{-} denotes the vacuum state. For notational simplicity, we omit the normalization factor 1/n! . The AGP wavefunction constructed from identical geminal products can capture correlations within electron pairs, but correlations between pairs are treated only at the mean-field level. A simple strategy to include inter-geminal correlations is to construct a wavefunction as a linear combination of AGP wavefunctions using distinct geminals 39 ; 40 ; 7 ; 6 , referred to as LC-AGP, which can be written as follows. \displaystyle\ket{\Psi_{\mathrm{LC-AGP}}} \displaystyle= \displaystyle\sum_{k}^{K}\left(\hat{F}^{(k)}\right)^{n}\ket{-}=\sum_{k}^{K}\ket{F^{(k)}} (3) Here, k denotes the index of the geminal type and K is the total number of geminal types (the number of AGP terms) used. The coefficients associated with each term in the LC-AGP wavefunction are omitted, as they can be absorbed into the matrix F . The overlaps of the AGP and LC-AGP wavefunctions are calculated using the following Onishi formula 24 . \displaystyle\braket{F^{(\lambda)}|F^{(\mu)}}=\left.\exp\left(\frac{1}{2}\mathrm{tr}\left[\ln(1+F^{(\mu)}F^{(\lambda)T}x)\right]\right)\right|_{x^{n}} (4) Here, x is an auxiliary variable, and |_{x^{n}} denotes extraction of the coefficient of x^{n} from the generating function. The matrix elements of the Hamiltonian and the gradients required for the variational optimization can be expressed in an analogous way using the Onishi formula. Although AGP and LC-AGP can both be treated variationally using the Onishi formula, the practical limitations of LC-AGP are twofold. First, maintaining accuracy for larger systems or in strongly correlated regimes may require a rapidly increasing number of AGP components 39 . Second, while the linear coefficients of an LC-AGP expansion are easily determined once the constituent AGPs are fixed, full variational optimization of the geminals entering each AGP is considerably more difficult, resembling a nonorthogonal multiconfigurational optimization and being prone to numerical difficulties. Another representative approach to incorporate correlations between geminals is the antisymmetric product of geminals (APG) wavefunction 17 ; 22 ; 15 , which is constructed as a product of distinct geminals. For a system with 2n electrons, the APG wave function is given by \displaystyle\ket{\Psi_{\mathrm{APG}}} \displaystyle= \displaystyle\hat{F}^{(1)}\cdots\hat{F}^{(n)}\ket{-} (5) Since the APG wavefunction is constructed from a product of distinct geminals, the Onishi formula (Eq. ( 4 )) cannot be applied straightforwardly. Therefore, we employ the classical Fischer formula 8 , which expresses the product of n commuting geminals as a signed linear combination of n -th powers of their linear combinations. \displaystyle\hat{F}^{(1)}\hat{F}^{(2)}\cdots\hat{F}^{(n)}=\frac{1}{2^{n-1}n!}\sum_{i_{2}=0}^{1}\cdots\sum_{i_{n}=0}^{1}(-1)^{i_{2}+\cdots+i_{n}}\left(\hat{F}^{(1)}+(-1)^{i_{2}}\hat{F}^{(2)}+\cdots+(-1)^{i_{n}}\hat{F}^{(n)}\right)^{n} (6) In tensor language, this is a Waring (symmetric-rank) decomposition of the corresponding polynomial. This decomposition allows the APG wavefunction to be rewritten in LC-AGP form, making it possible to employ the Onishi formula (Eq. ( 4 )). APG has been reported to achieve higher accuracy compared to LC-AGP when the same number of geminal types is employed 15 . However, as can be seen in Eq. ( 6 ), the number of LC-AGP terms generated by the Fischer decomposition grows exponentially with the number of electrons, leading to a prohibitive increase in computational cost. This APG example illustrates the broader difficulty addressed in the present work. For general geminal-based wave functions, efficient deterministic evaluation is considerably less straightforward than for AGP or LC-AGP. While direct formulas are known in some cases, they are generally tractable only for special structured families, such as AGP, antisymmetrized product of strongly orthogonal geminals (APSG) 32 ; 1 ; 10 ; 38 ; 36 ; 12 ; 21 ; 31 ; 33 ; 34 ; 14 ; 11 ; 23 , or Richardson–Gaudin-type states 27 ; 26 ; 28 ; 13 ; in more general settings, reducing the target state to an LC-AGP form remains one of the few broadly practical ways to exploit existing AGP overlap and Hamiltonian machinery efficiently. Motivated by this viewpoint, in this work we introduce a class of antisymmetrized geminal power configuration interaction (AGP-CI) wavefunctions, formulated in a manner analogous to a CI expansion around an AGP reference. These wave functions incorporate inter-geminal correlations through one or more independent pair-creation operators while retaining a polynomial structure in commuting geminal operators. In this sense, AGP-CI is similar in spirit to APG, but it is designed so that its exact Waring decomposition into LC-AGP grows only polynomially with the number of electrons, rather than exponentially as in the APG case. Thus, AGP-CI provides a more expressive geminal ansatz than AGP while remaining much more manageable than a direct APG-to-LC-AGP expansion. We then take a further step and exploit border-rank approximations of the AGP-CI polynomial. Introducing a small deformation parameter \tau , we reorganize the AGP-CI wave function into an even more compact LC-AGP representation, which we denote AGP-CI \tau . The role of the border-rank construction is therefore practical rather than purely formal. On the one hand, AGP-CI is the physically motivated target ansatz, whereas LC-AGP is the computational representation in which overlaps and observables can be evaluated efficiently; the border-rank approximation therefore provides a compressed bridge between the two. On the other hand, by reducing the number of AGP components that must be handled explicitly, it offers a more compact variational form for directly optimizing AGP-CI \tau and can also reduce the cost of subsequent correlation treatments based on the optimized AGP-CI \tau state. In addition, because the border-rank representation embeds AGP-CI into the LC-AGP manifold in a structured way, it can also serve as a physically informed initial guess for subsequent LC-AGP optimization, allowing one to explore