Low-Rank APG in Geminal Wavefunction Theory

• The paper introduces low-rank APG as a geminal-based wavefunction method that restricts each geminal's pairing matrix to dominant modes for more efficient electron correlation modeling.
• It employs structured approaches such as Takagi factorization and block-diagonal representations to transform #P-hard permanent evaluations into determinant-based computations, reducing complexity.
• Benchmark studies on molecules like H2O and N2 show that low-rank APG recovers substantial correlation energy, validating its use for strong-correlation systems with manageable computational cost.

Searching arXiv for the cited papers to ground the article in the current literature. Low-rank antisymmetric product of geminals (APG) is a constrained form of the general APG wavefunction in which each geminal’s pairing matrix is restricted to a small number of dominant pairing modes. Within geminal theory, APG is the most general pair-product ansatz: an $N$-electron state is built from $N/2$ two-electron building blocks and antisymmetrized automatically through the fermionic algebra. The low-rank variant retains this pair-product structure while truncating the internal rank of the geminal coefficient matrices, thereby reducing the parameter count and computational burden relative to full APG. In modern electronic-structure theory, this places low-rank APG at the intersection of strong-correlation modeling, seniority-based methods, and compact variational wavefunction design (Gaikwad et al., 17 Apr 2026).

1. Position within geminal wavefunction theory

Geminal wavefunctions were introduced in the late 1950s as compact representations of electron pairing, and they have long been associated with static or strong correlation, especially in near-degenerate regimes and bond breaking. The contemporary classification emphasized in recent reviews places the general antisymmetrized product of geminals (APG), its seniority-zero specialization APIG/DOCI, strongly orthogonal products APSG/GVB-PP, and the single-geminal antisymmetrized geminal power (AGP) within one overarching framework of pair-based ansätze (Gaikwad et al., 17 Apr 2026).

The distinction between these families is structural. AGP repeats one collective geminal, APSG imposes strong orthogonality by assigning different geminals to disjoint orbital subspaces, and APG allows $N/2$ distinct geminals without those restrictions. In the seniority-zero language used for paired-electron systems, a convenient APIG/APG form is

$$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$

with $N=2N_P$ and $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$. AGP is recovered when all geminals are identical, while APSG is recovered when geminals act on mutually orthogonal orbital subspaces (Johnson et al., 2022).

Low-rank APG is not a separate many-electron symmetry class but a restricted parametrization of APG. Its purpose is to preserve the physically transparent pairing picture of APG while making optimization and matrix-element evaluation tractable. This suggests a hierarchy: AGP may be viewed as an extreme one-geminal limit, APSG as a block-structured low-rank limit, and modern low-rank APG as an intermediate regime designed to capture inter-geminal correlation more flexibly than APSG without incurring the full combinatorial burden of unconstrained APG (Gaikwad et al., 17 Apr 2026).

2. Formal definition and meaning of “rank”

In second quantization, a general pair operator can be written as

$$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$

and the corresponding pair-product wavefunction as

$$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$

A conceptually equivalent compact form is

$$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$

where each geminal has its own pairing matrix $g^{(k)}$ (Gaikwad et al., 17 Apr 2026).

For low-rank APG, “rank” refers to the rank of the geminal coefficient matrix after a suitable factorization, not to the CP or tensor-product rank of the full antisymmetric many-electron tensor. In the review formulation, one keeps only the dominant singular or eigenmodes of each pairing matrix. For complex symmetric singlet-pairing matrices, a Takagi factorization is used:

$N/2$0

For general or antisymmetric matrices, singular value decomposition or skew-symmetric canonical forms are used analogously, with truncation implemented by discarding the small singular or eigenvalues (Gaikwad et al., 17 Apr 2026).

This point is central. Low-rank APG reduces the dimensionality of each geminal’s internal pairing space, whereas the antisymmetric-tensor literature studies low-rank approximations of the full $N/2$1-electron tensor. Those two notions of rank are not identical. In APG, a low-rank geminal may still generate a highly nontrivial antisymmetric $N/2$2-electron state after antisymmetrization; conversely, lower bounds on CP or tensor-product rank address exact antisymmetry in a different representation class (Wang et al., 10 Jan 2025).

A related operator-theoretic perspective comes from geminal-basis formulations in which geminal creation operators obey a composite-boson Lie algebra rather than exact bosonic commutation relations. In that formalism, APG is

$N/2$3

and stationarity can be expressed through a geminal Brillouin condition. The nontrivial commutator structure implies that geminal rotations generate both primary and secondary rotations, a feature that becomes important in variational optimization of low-rank parametrizations (Sørensen, 2019).

3. Structured low-rank realizations

Modern low-rank APG work consists largely of imposing additional algebraic structure on the geminal matrices so that amplitudes, overlaps, and reduced density matrices become accessible at polynomial cost. Several forms highlighted in recent literature are summarized below.

Model	Defining constraint	Stated consequence
AP1roG / pCCD	Each geminal anchored to one occupied reference pair with virtual pair excitations	Tractable seniority-zero ansatz; formally equivalent to pCCD
APr2G	Reciprocal rank-two structure $N/2$4	Permanents become determinant ratios by Borchardt’s theorem
2D-block geminals	Block-diagonal $N/2$5 geminal matrices with PI2O/EPI2O constraints	Polynomial complexity with mixed seniority
Low-rank APG via spectral truncation	Retain only dominant eigenvalue blocks of each geminal matrix	Compact hierarchy of APG ansätze

AP1roG is the best-known chemically motivated rank-restricted realization. Its geminals have the form

$N/2$6

and the resulting wavefunction is formally equivalent to pair-coupled-cluster doubles,

$N/2$7

In the review’s formulation, the one-reference-orbital constraint acts as an effective rank restriction per geminal and yields a seniority-zero description closely approximating DOCI (Gaikwad et al., 17 Apr 2026).

The 2D-block geminal construction of Cassam-Chenaï and co-workers relaxes both strong orthogonality and seniority-zero. Here each geminal matrix is block diagonal with $N/2$8 blocks chosen from sets including $N/2$9, $N/2$0, $N/2$1, and $N/2$2, together with blockwise PI2O/EPI2O constraints. In the reported proof-of-principle calculations, this sharply reduces the number of terms contributing to overlaps and reduced density matrices while preserving a bona fide antisymmetrized electronic wavefunction (Cassam-Chenaï et al., 2022).

A different route is provided by Waring decomposition and AGP-CI. In that approach, a product of distinct geminals is rewritten as a finite sum of AGP states; for a monomial product of $N/2$3 geminals, Fischer’s formula gives $N/2$4 AGP terms. This gives an efficient variational reformulation of APG in an AGP basis, but the exact AGP-CI space still grows exponentially with electron number. The same work then introduces explicit low-rank APG through Schur or canonical decomposition of skew-symmetric geminal matrices and reports that only a few non-zero eigenvalues, up to half the number of electrons, are typically significant in the optimized geminals (Kawasaki et al., 2024).

4. Evaluation, scaling, and optimization

The principal obstacle in APG is combinatorial. In a Slater or occupation-number basis, APG and APIG amplitudes involve matrix permanents, and direct evaluation is therefore #P-hard. This intractability is the main reason low-rank and structured parametrizations are introduced in the first place (Gaikwad et al., 17 Apr 2026).

Several structured reductions are now available. In RG-type or Cauchy-type parametrizations, scalar products, form factors, and reduced density matrices can be written as determinants or determinant ratios rather than permanents. The bivariational APG framework of Van Raemdonck and collaborators uses dual pair and pair-hole representations and the asymmetric energy functional

$N/2$5

Within this construction, Borchardt-like identities and determinant formulas render the evaluation of $N/2$6, $N/2$7, and $N/2$8 polynomial for the structured off-shell RG class, although the functional is strictly variational only when primal and dual representations are consistent; without that consistency, it is not bounded from below (Johnson et al., 2022).

The review situates these strategies within a broader algorithmic landscape. APr2G uses determinant ratios with projected Schrödinger equations. AP1roG/pCCD attains near mean-field scaling $N/2$9 and an overall cost of about $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$0 when orbital optimization is included. 2D-block geminals reduce overlap and matrix-element counts through block traces. Graphical APG maps overlaps to weighted bipartite matchings. AGP-based correlators exploit elementary symmetric polynomials for reduced density matrices, and Pfaffian ansätze permit compact stochastic evaluation in QMC (Gaikwad et al., 17 Apr 2026).

The most explicit low-rank APG orbital optimization presently described in the literature is the direct Givens rotation (DGR) method. In that formulation, each antisymmetric geminal matrix is transformed to canonical $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$1 blocks, only the $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$2 largest blocks are retained, and all geminals share a common orbital rotation $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$3. The orbital unitary is then parameterized as a product of Givens rotations, and analytic gradients are obtained by error backpropagation through the rotation sequence. The central orbital-gradient seed is an $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$4-matrix built from one- and two-body expectation values in the canonical pair basis, and the full sweep over rotation angles avoids numerical differentiation of $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$5 (Kawasaki et al., 15 Aug 2025).

A useful limiting case follows immediately from the shared-$$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$6 low-rank form. For rank $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$7,

$$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$8

which is, up to normalization, a single Slater determinant. In that formulation, electron correlation appears only for rank $$\ket{\mathrm{APIG}}=\prod_{\alpha=1}^{N_P}\Bigg(\sum_{i=1}^{N_{\mathrm{orb}}} g_\alpha^i\,S_i^+\Bigg)\ket{\theta},$$9 (Kawasaki et al., 15 Aug 2025).

5. Reported performance and applications

The empirical literature portrays low-rank APG as a compact strong-correlation model whose quality depends strongly on orbital optimization and the chosen structural constraint. The mini-review reports that low-rank APG benchmarks on $N=2N_P$0 and $N=2N_P$1 recover a substantial fraction of correlation energy and yield potential-energy curves superior to AGP or UHF at compact ranks (Gaikwad et al., 17 Apr 2026).

The most detailed recent small-molecule benchmark is the DGR study on $N=2N_P$2 and $N=2N_P$3 in STO-6G. For $N=2N_P$4 at equilibrium geometry, the reported total energies are: Exact $N=2N_P$5 Ha, APG $N=2N_P$6 Ha, Rank-5 APG $N=2N_P$7 Ha $N=2N_P$8, Rank-4 APG $N=2N_P$9 Ha $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$0, Rank-3 APG $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$1 Ha $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$2, Rank-2 APG $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$3 Ha $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$4, AGP $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$5 Ha $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$6, and UHF $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$7 Ha $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$8. The same work reports that rank-2 APG tracks the exact potential-energy curve closely across O–H stretching and that the convergence time for rank-2 APG was reduced from more than one week in the earlier $S_i^+=a^\dagger_{i\uparrow}a^\dagger_{i\downarrow}$9 scheme to less than ten minutes with DGR (Kawasaki et al., 15 Aug 2025).

The earlier Schur-decomposition study on Hubbard rings and small molecules reached closely related conclusions. For the half-filled 6-site Hubbard model at $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$0, the reported energies are Exact $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$1, APG $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$2, AGP-CI $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$3, AGP $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$4, and HF $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$5. In the same work, low-rank APG for $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$6 in STO-6G recovers approximately $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$7 of the APG correlation energy at rank 2 and $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$8 at rank 3, while rank-2(a) in the 6-site Hubbard model recovers approximately $$\hat{G}_p^\dagger=\sum_{q<r} C^p_{qr}\,a_q^\dagger a_r^\dagger,$$9 of the APG correlation energy for $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$0 and approximately $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$1 for $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$2 (Kawasaki et al., 2024).

The 2D-block geminal model gives another benchmark point on the low-rank spectrum. For second-row diatomics at equilibrium, using spin-restricted 3-type blocks, the reported total energies improve over APSG while remaining computationally affordable. Examples include $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$3/cc-pVDZ with $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$4 versus $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$5 for APSG, $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$6/cc-pVDZ with $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$7 versus $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$8, $$\Psi=\prod_{p=1}^{N/2}\hat{G}_p^\dagger\,|0\rangle.$$9/cc-pVDZ with $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$0 versus $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$1, and $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$2/cc-pVTZ with $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$3 versus $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$4 (Cassam-Chenaï et al., 2022).

A closely related but more specialized low-rank form is AP1roG. It is not a general low-rank APG, yet it is the canonical example of how severe structural restriction can produce chemically useful compactness. AP1roG-LCC calculations on $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$5, $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$6, $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$7, and the symmetric dissociation of $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$8 show that an LCC correction built on the AP1roG reference is more robust and reliable than perturbative corrections in those tests, with singles playing a particularly important role for the $$|{\rm APG}\rangle=\mathcal{A}\,\prod_{k=1}^{N/2}\left(\sum_{ij} g^{(k)}_{ij}\,a_i^\dagger a_j^\dagger\right)|0\rangle,$$9 chain (Boguslawski et al., 2015).

6. Limitations, conceptual issues, and open directions

Low-rank APG inherits the main limitations of pair-product wavefunctions. The review emphasizes that, although low-rank truncation captures dominant static correlation compactly, interpair dynamic correlation is still missing unless the ansatz is augmented. Reported remedies include perturbative corrections, ERPA fluctuation–dissipation corrections, explicit F12 corrections, CC-like hybrids, and Jastrow correlators. Size consistency is ansatz-dependent: AGP alone is not size-consistent, APSG is size-extensive and variational but suppresses intergeminal correlation, and AP1roG becomes size-consistent when orbitals are fully optimized (Gaikwad et al., 17 Apr 2026).

A second limitation is optimization landscape complexity. In the bivariational APG formulation, the unconstrained asymmetric functional is not bounded from below, and practical duality constraints are overdetermined and must be enforced on a restricted determinant set. In the geminal-basis formulation, composite-boson commutators imply that nominal single-geminal rotations induce secondary rotations as well, so the low-rank variational manifold is not as simple as a naive orbital-rotation picture might suggest (Johnson et al., 2022).

A third issue is that the term “low-rank” can be misleading unless its representation dependence is made explicit. In APG, rank refers to the number of retained pairing modes in each geminal matrix. By contrast, for exact antisymmetric tensor-product-function representations in a fixed finite-dimensional function class, the minimum number of product terms grows at least as

$g^{(k)}$0

so exact antisymmetry is fundamentally incompatible with low CP/TPF rank in that setting (Wang et al., 10 Jan 2025). This does not invalidate low-rank APG; rather, it shows that geminal-matrix rank reduction and full antisymmetric tensor-product rank reduction are different compression problems.

A separate tensor-oriented literature reinforces that point by constructing low-rank approximations directly in the antisymmetric tensor space through antisymmetrized outer products and structure-preserving alternating least squares. For order-3 tensors, the format

$g^{(k)}$1

uses only three vectors while preserving antisymmetry. This suggests a possible bridge between APG and antisymmetric-tensor compression, although the cited results are formulated for tensor approximation rather than as a geminal electronic-structure method (Begovic et al., 2022).

The open problems identified in the current APG review are therefore structural rather than merely numerical: size consistency for more general APG constraints, robust orbital optimization in the presence of multiple local minima, systematic recovery of intergeminal dynamic correlation at polynomial cost, extension of low-rank APG to open shells with spin projection, scalable matrix-element evaluation beyond permanents, and the possibility of exploiting geminal structure in quantum algorithms such as AGP preparation, seniority-zero pUCCD, and Jastrow-type circuit factorizations (Gaikwad et al., 17 Apr 2026).