One-Body Reduced Density-Matrix Functional

• 1RDMF is a variational many-body framework using the one-body reduced density matrix to capture kinetic, coherence, and static-correlation effects.
• It employs constrained searches and nonlocal potentials in both finite-temperature canonical and grand-canonical settings to ensure unique v-representability.
• Recent computational advances in 1RDMF enable accurate treatment of static correlation, molecular dissociation, and Mott transitions in strongly correlated systems.

Searching arXiv for recent and foundational papers on one-body reduced density-matrix functional theory, including canonical/grand-canonical finite-temperature formulations, excited-state extensions, computational scaling, and recent strongly correlated applications. One-body reduced density-matrix functional theory, usually abbreviated 1RDMFT or 1RDMF, is a variational many-body framework in which the basic variable is the one-body reduced density matrix $\gamma$ rather than the density alone. In second quantization, $$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$, while in coordinate representation $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$; its eigenvectors are the natural orbitals and its eigenvalues are the natural occupation numbers, with $0 \le n_i \le 1$ for fermions and $\sum_i n_i = N$. The density is the diagonal, $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$, but the off-diagonal structure carries kinetic, coherence, and static-correlation information that is absent from a density-only description. In finite basis sets, a universal functional of $\gamma$ can be defined by constrained search over many-body density operators yielding the prescribed 1RDM, and both finite-temperature thermodynamic potentials and zero-temperature energies can be minimized directly in 1RDM space (Giesbertz et al., 2017, Baldsiefen et al., 2012, Sutter et al., 2022).

1. Basic variable, representability, and constrained search

The 1RDM interpolates between the full many-body state and the density. For an $N$-electron wavefunction $\Psi$, it is obtained by tracing out $N-1$ coordinates, and in the natural-orbital representation it takes the spectral form

$$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$0

This representation is central because the occupation numbers directly encode departures from idempotency and therefore multi-reference character, bond breaking, and related static-correlation effects (Wetherell et al., 2020).

The universal-functional construction follows the Gilbert–Levy–Valone logic. In the canonical finite-temperature setting, the universal 1RDM functional is

$$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$1

and the Helmholtz free energy is obtained from

$$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$2

In the grand-canonical setting, the analogous constrained-search construction yields a universal grand-potential functional $$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$3, and equilibrium is obtained by minimizing $$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$4 over the ensemble-$$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$5-representable domain (Sutter et al., 2022, Baldsiefen et al., 2012).

For finite basis sets, ensemble representability is particularly clean. The admissible 1RDMs are exactly the positive semidefinite Hermitian matrices, with the additional Pauli constraint $$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$6 for fermions. In this sense, the finite-basis formulation is mathematically tighter than the continuum case, and it is also the regime in which standard finite-basis DFT effectively becomes a 1RDM functional theory because locality and nonlocality of the one-body potential are no longer distinguishable at the level of basis-matrix elements (Giesbertz et al., 2017).

2. Potential mappings, nonlocal Kohn–Sham structure, and uniqueness

A defining structural feature of 1RDMF is that its conjugate external variable is a general nonlocal one-body potential rather than a strictly local scalar potential. At finite temperature in the grand-canonical ensemble, equilibrium properties are uniquely determined by the equilibrium 1RDM for local or nonlocal external potentials, and the theory establishes the chain

$$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$7

The corresponding universal grand-potential functional can be decomposed as

$$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$8

with explicit kinetic, Hartree, exchange, and entropy terms, plus a correlation functional $$\gamma_{ij}[\hat{\rho}] = \mathrm{Tr}\{\hat{\rho}\,\hat a_j^\dagger \hat a_i\}$$9 (Baldsiefen et al., 2012).

Finite temperature is decisive for the Kohn–Sham construction. In the grand-canonical formulation, all equilibrium occupation numbers satisfy $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$0, so a noninteracting Hamiltonian can always be chosen to reproduce the interacting 1RDM by inverting the Fermi–Dirac relation,

$$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$1

This yields a nonlocal Kohn–Sham potential that reproduces the exact equilibrium 1RDM. The corresponding canonical finite-temperature theory proves the same essential point in fixed-$$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$2 language: all 1RDMs with a purely fractional occupation spectrum are uniquely $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$3-representable up to a constant, and the universal functional is differentiable on that interior domain (Baldsiefen et al., 2012, Sutter et al., 2022).

By contrast, the zero-temperature situation is subtler. In linear response, the retarded 1RDM response function is invertible only up to a nontrivial kernel. For nondegenerate ground states, the kernel of the static response coincides with the full nonuniqueness class of the nonlocal potential in ground-state 1RDMF, including constants, continuous-symmetry generators, operators acting purely within fully occupied or fully unoccupied natural-orbital blocks, and special two-electron structures tied to degenerate occupation numbers (Giesbertz, 2015). For mixed thermodynamic ensembles, this kernel simplifies drastically: in the retarded response problem, only perturbations commuting with the unperturbed Hamiltonian remain in the kernel, so the finite-temperature mapping is much less encumbered by the zero-temperature “pathological” directions (Giesbertz, 2016).

3. Functional constructions and formal extensions

Several distinct routes to explicit or approximate 1RDM functionals have emerged. One route starts from finite-temperature MBPT. By restricting the Klein or Luttinger–Ward grand-potential functionals $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$4 to the subset of noninteracting Green’s functions $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$5 that reproduce a prescribed 1RDM, one obtains approximate 1RDM functionals derived from diagrammatic perturbation theory. In the zero-temperature limit this yields an energy functional $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$6; in a model hydrogen molecule described by an extended Hubbard Hamiltonian, the Luttinger–Ward version in the GW approximation performs the best and is able to reproduce energies close to the GW energy corresponding to the stationary point (Giesbertz et al., 2018).

A second route explicitly hybridizes DFT and 1RDMFT. “Density Functional Theory transformed into a One-electron Reduced Density Matrix Functional Theory” introduces a modified Kohn–Sham energy

$$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$7

or equivalently

$$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$8

where idempotency is relaxed, a quadratic 1RDM correction is added to the density-functional exchange-correlation backbone, and the resulting optimization is carried out as a semidefinite program with $$\gamma(x,x') = \mathrm{tr}\{\hat D\,\hat\psi^\dagger(x')\,\hat\psi(x)\}$$9 scaling (Gibney et al., 2022).

A third extension targets neutral excitations. The $0 \le n_i \le 1$0-ensemble formulation generalizes ground-state RDMFT by introducing a weight vector $0 \le n_i \le 1$1 and a $0 \le n_i \le 1$2-dependent universal functional

$0 \le n_i \le 1$3

This formalism yields the first explicit excited-state 1RDM functionals for the symmetric Hubbard dimer and for the homogeneous Bose gas in the Bogoliubov regime, and it makes the domain geometry itself part of the functional through generalized exclusion principles and spectral polytopes (Liebert et al., 2022).

A concise comparison of these constructions is useful.

Formulation	Basic thermodynamic object	Distinctive feature
Finite-$0 \le n_i \le 1$4 grand-canonical 1RDMFT	$0 \le n_i \le 1$5	Exact 1RDM–potential mapping for nonlocal $0 \le n_i \le 1$6
Finite-$0 \le n_i \le 1$7 canonical 1RDMFT	$0 \le n_i \le 1$8, $0 \le n_i \le 1$9	Unique $\sum_i n_i = N$0-representability for purely fractional occupations
MBPT-derived 1RDMF	$\sum_i n_i = N$1, $\sum_i n_i = N$2	Klein/Luttinger–Ward restriction to $\sum_i n_i = N$3
DFT-transformed 1RDMFT	$\sum_i n_i = N$4	Non-idempotent KS extension with quadratic 1RDM correction
$\sum_i n_i = N$5-ensemble 1RDMFT	$\sum_i n_i = N$6, $\sum_i n_i = N$7	Excited-state ensembles and generalized exclusion constraints

These developments collectively indicate that 1RDMF is not a single approximation family but a variational architecture that can be populated by perturbative, density-functional, ensemble, or explicitly constrained-search constructions.

4. Computational formulations and minimization technology

The formal appeal of 1RDMF depends heavily on whether its functionals and derivatives can be evaluated without prohibitive orbital-space tensor algebra. A major advance is the recognition that most approximate 1RDM functionals are separable. For such functionals, the four-index transformation of two-electron integrals to the natural-orbital basis can be avoided entirely by evaluating Coulomb-type and exchange-type contractions directly in the AO basis. For separable functionals without diagonal corrections, AO-direct evaluation eliminates the $\sum_i n_i = N$8 transformation bottleneck and yields observed scaling $\sum_i n_i = N$9 to $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$0; with diagonal corrections, Schwarz screening reduces the practical scaling to around $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$1 (Giesbertz, 2016).

In the canonical finite-temperature setting, a separate computational bottleneck is the inversion from NOONs to the effective orbital energies of the noninteracting maximum-entropy reference ensemble. Bosonic and fermionic Sinkhorn algorithms solve precisely this inversion problem and thereby provide $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$2, $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$3, and noninteracting approximations to the interaction energy. The bosonic algorithm is particularly robust; the fermionic one is effective but more sensitive to near-degenerate occupations. The same formalism highlights a direct connection to entropically regularized multi-marginal optimal transport (Kooi, 2022).

Parametrization matters as much as raw scaling. When the 1RDM is written as

$\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$4

the convexity of the exact energy functional $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$5 is generally lost in the variables $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$6. Degeneracies in the occupation numbers then generate additional critical points, and these are classified as saddle points rather than minima. The specific orbital parametrization further reshapes the landscape: Cayley and Householder parametrizations do not create extra critical points, whereas Givens rotations and the exponential can, although these are not of practical concern in standard small-step update schemes. This provides an explicit rationale for the empirical success of second-order minimization procedures in modern 1RDMF implementations (Cartier et al., 31 Jan 2025).

5. Static correlation, dissociation, and Mott physics

The most persistent motivation for 1RDMF has been the treatment of static or strong correlation in regimes where idempotent Kohn–Sham pictures fail. The DFT-transformed semidefinite-programming formulation gives a clear molecular example. Using a scalar weight matrix $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$7 fitted to $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$8, the 1RDMFT dissociation curve in a 6-31G basis recovers the CASSCF dissociation energy to within $\rho(\mathbf r)=\gamma(\mathbf r,\mathbf r)$9 kcal/mol and the ACSE dissociation energy to within $\gamma$0 kcal/mol. On the MR-MGN-BE17 subset of 11 diatomics, the same fitted parameter reduces the severe B3LYP overestimation of dissociation energies from MUE $\gamma$1 kcal/mol to $\gamma$2 kcal/mol against $\gamma$3 CASSCF, and from $\gamma$4 to $\gamma$5–$\gamma$6 kcal/mol against ACSE, depending on the fitting protocol. For the linear $\gamma$7 chain, the same framework reproduces the FCI dissociation curve with an error of only $\gamma$8 kcal/mol and restores the correct insulating limit of the off-diagonal 1RDM, whereas B3LYP retains an erroneous metallic character (Gibney et al., 2022).

A broader condensed-matter test is the recent explicit construction of a 1RDMF for the multi-orbital Hubbard model in the thermodynamic limit using the $\gamma$9 ansatz of variational discrete action theory. In that setting, the interaction functional is obtained by actually executing the constrained search within the variational manifold, and the resulting 1RDMF reproduces the underlying VDAT solution, including Mott and Hund physics. The central structural result is that non-analytic behavior emerges in the 1RDMF at fixed integer filling and gives rise to the Mott transition; moreover, a nonzero Hund exchange drives the continuous Mott transition to become first-order (Cheng et al., 18 Aug 2025).

These two strands are complementary rather than redundant. Molecular dissociation benchmarks show that fractional occupations and non-idempotent optimization repair the classic single-determinant failure mode. The Hubbard-model construction shows that the same basic variable can also encode bona fide quantum phase transitions through non-analyticities of the functional itself. A plausible implication is that 1RDMF’s most distinctive strength is not merely “fractional occupations,” but the possibility of building functionals whose singularity structure tracks physically correct correlation regimes.

6. Spin structure, generalized constraints, and emerging directions

One recurring misconception is that open-shell 1RDMFT can always be formulated with a single spin-independent spatial natural-orbital set and spin-dependent occupations. Exact three-electron studies show otherwise. For doublet systems, except for maximally polarized states, the natural orbitals of spin eigenstates are generally spin dependent, and the spatial parts of the up- and down-spin natural orbitals form two different sets. The paper introduces a quantitative measure of this spin dependence,

$N$0

and shows that it ranges from $N$1 to $N$2 across representative three-electron systems. The same analysis makes clear that generalized Pauli constraints, occupation-number ordering, and quasi-pinning have direct implications for 1RDMFT minimization schemes and for the sparse CI structure compatible with a given exact 1RDM (Theophilou et al., 2017).

Another emerging direction is data-driven structure discovery. For exactly solvable one-dimensional model systems, PCA and convolutional autoencoders show that physically realized 1RDMs occupy a highly constrained low-dimensional manifold. In a 62-point grid representation, PCA compresses the 1RDM data to 327 components without loss, while a CAE with an $N$3 architecture reconstructs the 1RDM with a mean absolute error of $N$4 a.u. per matrix element. A PCA-based linear functional $N$5 already captures about $N$6 of the total off-diagonal contribution, and denoising autoencoders learn the nonlinear correction significantly more effectively when the baseline is a physics-informed linear model or UHF rather than a purely noninteracting 1RDM (Wetherell et al., 2020).

Several open issues remain explicit in the literature. Zero-temperature $N$7-representability and nonuniqueness are still more delicate than their finite-temperature counterparts; the general static degenerate case remains incomplete in the response-based analysis (Giesbertz, 2015). Excited-state functionals appear naturally piecewise because of level crossings and the “curse of universality” (Liebert et al., 2022). Canonical and grand-canonical noninteracting references differ in practically significant ways for strong correlation and bond breaking (Kooi, 2022). These points suggest that future progress is likely to come not from a single universal closed form, but from functionals that combine exact domain geometry, nonlocal Kohn–Sham structure, constrained-search information, and computationally controlled approximations tailored to the correlation regime.