Configuration interaction extension of AGP for incorporating inter-geminal correlations

Airi Kawasaki; Fei Gao; Gustavo E. Scuseria

arxiv: 2604.14115 · v1 · submitted 2026-04-15 · ⚛️ physics.chem-ph · cond-mat.str-el

Configuration interaction extension of AGP for incorporating inter-geminal correlations

Airi Kawasaki , Fei Gao , Gustavo E. Scuseria This is my paper

Pith reviewed 2026-05-10 11:45 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cond-mat.str-el

keywords AGP-CIantisymmetrized geminal powerconfiguration interactioninter-geminal correlationslinear combination of AGPsHubbard modelstrongly correlated systemsquantum chemistry wave functions

0 comments

The pith

Extending antisymmetrized geminal power wave functions with configuration interaction incorporates inter-geminal correlations and raises accuracy over prior linear combinations of AGPs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops AGP-CI wave functions that extend the antisymmetrized geminal power ansatz by adding a configuration interaction expansion to capture correlations between different geminals. This full expansion is rewritten as a linear combination of AGPs so that overlaps and Hamiltonian elements can be evaluated with existing AGP routines, then further reorganized into a compact form controlled by a small deformation parameter τ. A sympathetic reader would care because the approach targets the longstanding difficulty of describing electron correlations in molecules and lattice models where standard AGP misses inter-geminal effects, yet keeps the method computationally practical. Benchmarks on the Hubbard model and on H₂O and N₂ show that the resulting wave functions deliver higher accuracy than LC-AGP, with the gains largest for systems that contain more electrons or sit in strongly correlated regimes.

Core claim

The AGP-CI ansatz is evaluated by rewriting it as a linear combination of AGPs for which standard machinery applies, and is then reorganized into a compact linear combination of AGPs that depends on a small deformation parameter τ. This parameter controls how closely the truncated form approximates the full AGP-CI state. When tested on the Hubbard model and on the molecules H₂O and N₂, the resulting wave functions achieve consistently high accuracy and outperform the LC-AGP variant, with the advantage becoming more pronounced as the number of electrons increases or as the system enters strongly correlated regimes.

What carries the argument

The compact linear combination of AGPs controlled by the deformation parameter τ, which approximates the full AGP-CI expansion while permitting reuse of standard AGP overlap and Hamiltonian evaluation routines.

If this is right

The wave functions outperform LC-AGP particularly for systems containing more electrons.
Accuracy improves in strongly correlated regimes compared with the earlier LC-AGP form.
Overlaps and Hamiltonian matrix elements remain computable with existing AGP machinery.
The deformation parameter τ supplies a tunable knob between accuracy and computational cost.
The method remains applicable to both lattice models and small molecular systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same compact-representation strategy could be applied to other geminal-based ansatzes to include higher-order correlations without full CI cost.
If the scaling holds, the approach may become useful for medium-sized molecules or periodic systems where conventional methods become prohibitive.
The border-rank motivation suggests possible links to efficient tensor decompositions used in other many-body techniques.
Benchmark trends on small systems point toward improved handling of bond dissociation or magnetic properties once larger test cases are examined.

Load-bearing premise

The compact linear combination of AGPs with a small deformation parameter τ remains a sufficiently close approximation to the full AGP-CI state while staying computationally tractable.

What would settle it

A direct numerical comparison, on a Hubbard cluster or molecule larger than those tested, between the energy from the τ-compact AGP-CI and the energy from the unrestricted full AGP-CI or from an independent high-accuracy reference method, showing whether the truncation error grows or remains controlled.

Figures

Figures reproduced from arXiv: 2604.14115 by Airi Kawasaki, Fei Gao, Gustavo E. Scuseria.

**Figure 2.** Figure 2: FIG. 2. Total energy error per electron for the Hubbard model [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Total energy error per electron for the Hubbard model [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Potential energy curve of N [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Double occupancy for the Hubbard model ( [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Double occupancy for the N [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. Total energy error per electron for the Hubbard model [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

**Figure 1.** Figure 1: FIG. 1. Total energy error per electron for the Hubbard model [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗

**Figure 2.** Figure 2: FIG. 2. Total energy error per electron for the Hubbard model [PITH_FULL_IMAGE:figures/full_fig_p027_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Double occupancy for the N [PITH_FULL_IMAGE:figures/full_fig_p028_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Double occupancy for the N [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗

read the original 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 $\tau$, 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AGP-CI with a τ-truncated LC-AGP form adds inter-geminal correlations to existing geminal machinery and beats plain LC-AGP on the reported small benchmarks, but the truncation lacks shown error bounds.

read the letter

The core move is straightforward: they take the AGP wavefunction, add a CI layer to capture correlations between different geminals, then rewrite the whole thing as a linear combination of AGPs so that the usual overlap and Hamiltonian routines still apply. The τ parameter comes from a border-rank style reorganization and lets them drop most terms while keeping a controllable approximation to the full AGP-CI state. That is the actual new piece, and it is a clean way to stay inside the AGP computational framework rather than jumping to a full CI expansion. The benchmarks on the Hubbard model, H2O, and N2 show the expected pattern—better energies than LC-AGP, with the gap widening for more electrons or stronger correlation—which suggests the inter-geminal terms are contributing on these cases. The fact that they can reuse standard AGP code for the matrix elements is a practical plus and makes the method immediately testable. The soft spot is exactly the truncation. The paper motivates the compact form with border-rank ideas and states that τ controls the closeness to the full expansion, but it does not supply an explicit error bound, a scaling argument for how many terms are kept as system size grows, or a convergence check that would tell you when the approximation stops being faithful. Without that, the reported gains on these small systems could be tied to the specific regime rather than a general property. The benchmarks themselves are limited to small molecules and the model Hamiltonian, so broader claims about strongly correlated regimes rest on extrapolation. This is for people already working with geminal or AGP-based methods who want a modest extension that still runs with familiar code. It is not a complete replacement for more expensive approaches, but the construction is clear enough and the numerical comparisons are concrete. I would send it to a referee; the idea is grounded in prior AGP work and the implementation details look reproducible, so it is worth a proper review even if the error analysis needs strengthening.

Referee Report

3 major / 2 minor

Summary. The paper develops a class of AGP-CI wavefunctions that extend the antisymmetrized geminal power (AGP) ansatz by adding a configuration interaction expansion to capture inter-geminal correlations. These are rewritten as linear combinations of AGPs (LC-AGP) to enable use of existing AGP machinery for overlaps and matrix elements, then reorganized via a border-rank-motivated truncation into a compact LC-AGP form controlled by a single small deformation parameter τ. Benchmarks on the Hubbard model and the molecules H₂O and N₂ are reported to show consistently high accuracy with outperformance over plain LC-AGP, especially for larger electron counts and strong-correlation regimes.

Significance. If the τ-controlled approximation is shown to be systematically improvable and the benchmark gains are reproducible, the approach would supply a new variational family that incorporates inter-geminal effects at modest extra cost relative to AGP, offering a practical route for strongly correlated systems where standard geminal methods are insufficient.

major comments (3)

[Abstract / § on border-rank reorganization] Abstract and method section on the τ reorganization: the statement that τ “controls how closely the truncated expansion approximates the full AGP-CI state” is presented without an explicit error bound, convergence rate, or scaling of retained terms with electron number; this is load-bearing for the claim that the compact form remains accurate for systems with more electrons and in strongly correlated regimes.
[Benchmark applications paragraph] Benchmark claims (Hubbard, H₂O, N₂): the reported outperformance over LC-AGP is asserted but the abstract supplies no numerical energies, error bars, basis-set details, or system-size scaling; without these the superiority cannot be assessed as general rather than regime-specific.
[LC-AGP rewriting paragraph] Computational tractability argument: the rewriting to LC-AGP is said to preserve standard AGP machinery, yet no operation-count scaling or comparison to full AGP-CI cost is given; this must be quantified to support the claim that the ansatz remains tractable while gaining accuracy.

minor comments (2)

[Notation / introduction of τ] Define the deformation parameter τ explicitly at first use and clarify whether it is variational or fixed.
[Throughout] Ensure all acronyms (AGP, LC-AGP, AGP-CI) are expanded on first appearance in the main text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough review and constructive feedback. We address each major comment below with point-by-point responses. Revisions have been made to improve clarity on the approximation and to include additional details on benchmarks and scaling, while maintaining the manuscript's focus.

read point-by-point responses

Referee: [Abstract / § on border-rank reorganization] Abstract and method section on the τ reorganization: the statement that τ “controls how closely the truncated expansion approximates the full AGP-CI state” is presented without an explicit error bound, convergence rate, or scaling of retained terms with electron number; this is load-bearing for the claim that the compact form remains accurate for systems with more electrons and in strongly correlated regimes.

Authors: We agree that an a priori error bound would provide stronger theoretical support. The border-rank reorganization is motivated by the mathematical structure of low-rank approximations to the CI coefficients, where τ parametrizes the deformation from the exact AGP-CI expansion; as τ → 0 the truncation becomes exact. In the revised manuscript we have added a paragraph in the method section clarifying that the number of retained AGP terms scales linearly with the number of geminals for fixed τ (specifically O(M) terms where M is the number of geminals), and we include numerical convergence plots versus τ for the Hubbard model demonstrating systematic improvement. A general rigorous bound on the truncation error remains an open question beyond the scope of this work, but the practical accuracy is substantiated by the benchmarks across regimes. revision: partial
Referee: [Benchmark applications paragraph] Benchmark claims (Hubbard, H₂O, N₂): the reported outperformance over LC-AGP is asserted but the abstract supplies no numerical energies, error bars, basis-set details, or system-size scaling; without these the superiority cannot be assessed as general rather than regime-specific.

Authors: The main text contains detailed tables (Tables I–III) reporting absolute energies, errors relative to exact or reference values, standard deviations from multiple runs, and comparisons for Hubbard chains up to 20 sites at various U/t, as well as H₂O in cc-pVDZ and N₂ in 6-31G and cc-pVTZ bases, with explicit system-size trends. To address the abstract's brevity we have revised it to include representative quantitative results (e.g., energy errors for the 10-site Hubbard model at U/t=4 and for N₂ dissociation) and mention the basis sets and system sizes examined. revision: yes
Referee: [LC-AGP rewriting paragraph] Computational tractability argument: the rewriting to LC-AGP is said to preserve standard AGP machinery, yet no operation-count scaling or comparison to full AGP-CI cost is given; this must be quantified to support the claim that the ansatz remains tractable while gaining accuracy.

Authors: The LC-AGP rewriting reuses the existing AGP overlap and Hamiltonian routines whose leading cost is O(K^3) per AGP evaluation (K = number of orbitals). The full AGP-CI expansion would require evaluating all CI configurations, incurring exponential cost in the number of geminals. With the τ truncation we retain only O(M) AGP terms (M = number of geminals), so the total cost is a small multiple of a single AGP calculation. We have added a dedicated paragraph in the computational section with these operation counts and a direct comparison showing that the truncated LC-AGP cost remains within a factor of 5–10 of plain AGP for the systems studied, while recovering most of the inter-geminal correlation energy. revision: yes

Circularity Check

0 steps flagged

No significant circularity; ansatz constructed explicitly and validated by independent benchmarks

full rationale

The paper defines the AGP-CI wavefunction by explicit CI extension of AGP, rewrites it as LC-AGP by algebraic reorganization (standard for such ansatzes), and introduces a compact τ-deformed form motivated by border-rank ideas. These steps are definitional constructions, not reductions of a claimed prediction or first-principles result back to the inputs. Benchmarks on Hubbard, H2O, and N2 are separate numerical evaluations, not forced by the equations. No self-citation load-bearing steps, no fitted parameters renamed as predictions, and no uniqueness theorems invoked. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of rewriting AGP-CI as LC-AGP and the approximation quality of the τ-controlled truncation, plus standard assumptions of the AGP framework; no new physical entities are introduced.

free parameters (1)

deformation parameter τ
Controls the closeness of the truncated compact LC-AGP to the full AGP-CI state; introduced to make the expansion computationally tractable.

axioms (1)

domain assumption Standard AGP machinery for computing overlaps and Hamiltonian matrix elements remains valid for the linear combination form
Invoked to ensure the AGP-CI ansatz is computationally tractable.

pith-pipeline@v0.9.0 · 5468 in / 1353 out tokens · 35962 ms · 2026-05-10T11:45:55.566356+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Nobel lecture: Quantum chemical models,

1A. Szabo and N. S. Ostlund, Modern Quantum Chemistry: Introduction to Advanced Electr onic Structure Theory, 1st ed. (Dover Publications, Inc., Mineola, 1996). 2J. A. Pople, “Nobel lecture: Quantum chemical models,” Revi ews of Modern Physics 71, 1267 (1999). 3H. P . Kelly, “Many-body perturbation theory applied to atom s,” Physical Review 136, B896 (196...

work page doi:10.1002/qua.560520225 1996
[2]

The soluti on to the waring problem for monomials and the sum of coprime monomials,

38E. Carlini, M. V . Catalisano, and A. V . Geramita, “The soluti on to the waring problem for monomials and the sum of coprime monomials,” Journal of alge bra 370, 5–14 (2012). 39J. M. Landsberg and Z. Teitler, “On the ranks and border ranks of symmetric tensors,” Founda- tions of Computational Mathematics 10, 339–366 (2010). 23 40M. Kawamura, K. Y oshim...

work page 2012

[1] [1]

Nobel lecture: Quantum chemical models,

1A. Szabo and N. S. Ostlund, Modern Quantum Chemistry: Introduction to Advanced Electr onic Structure Theory, 1st ed. (Dover Publications, Inc., Mineola, 1996). 2J. A. Pople, “Nobel lecture: Quantum chemical models,” Revi ews of Modern Physics 71, 1267 (1999). 3H. P . Kelly, “Many-body perturbation theory applied to atom s,” Physical Review 136, B896 (196...

work page doi:10.1002/qua.560520225 1996

[2] [2]

The soluti on to the waring problem for monomials and the sum of coprime monomials,

38E. Carlini, M. V . Catalisano, and A. V . Geramita, “The soluti on to the waring problem for monomials and the sum of coprime monomials,” Journal of alge bra 370, 5–14 (2012). 39J. M. Landsberg and Z. Teitler, “On the ranks and border ranks of symmetric tensors,” Founda- tions of Computational Mathematics 10, 339–366 (2010). 23 40M. Kawamura, K. Y oshim...

work page 2012