Completing the renormalization-group approach for electronic structure: coarse-graining of bound-state atoms, core-electron integration, and low-temperature auxiliary-field sampling without sign-problem bias (White-Wilkins-Wilson, 1986)

In 1986 White, Wilkins, and Wilson proposed computing the electronic structure of atoms and molecules by combining a lattice functional-integral formulation in space and imaginary time, renormalization-group (RG) blocking to coarsen the lattice, perturbative integration of core electrons, and Langevin-equation Monte Carlo over auxiliary fields for the valence electrons. They demonstrated exact blocking for noninteracting particles in a one-dimensional well and Langevin sampling of a three-site interacting model, and explicitly left open five issues: core-electron integration in the interacting problem; the loss of locality and causality of the renormalized action when the coarse-graining scale reaches bound-state energies; singular configurations with det A = 0 that corrupt low-temperature expectation values; the untested scaling of the unbiased estimator; and extension to spin-1/2 electrons and multi-dimensional systems. The original blocking strategy was abandoned after the failure of real-space RG for interacting fermions was diagnosed and superseded by the density matrix renormalization group, while the auxiliary-field line evolved into constrained-path and phase-free auxiliary-field QMC, which made realistic low-temperature calculations practical with pseudopotentials but retain an uncontrolled constraint; exact bias-free fermion sampling is NP-hard in general. The precise surviving open core is a coarse-grained, core-integrated, low-temperature auxiliary-field treatment of realistic electronic structure with quantified errors.

Background

Calculating the correlated electronic structure of atoms and molecules is a central problem of computational physics. Around 1985 each standard tool had a known failure mode: configuration interaction (CI), which expands the many-electron wavefunction in Slater determinants (antisymmetrized products of one-electron orbitals), converges too slowly because the number of determinants grows combinatorially with electron number; density-functional theory (DFT), which replaces the interacting problem by a self-consistent one-particle problem with an approximate exchange-correlation functional, offered no systematic route to improve the functional and performed poorly for systems such as negative ions and excitation energies; and Green's-function Monte Carlo, which samples the many-electron state stochastically, was dominated by the large energy fluctuations associated with the core (inner-shell) electrons of large atoms.

In this context White, Wilkins, and Wilson (1986) proposed a strategy combining four elements. First, an exact functional-integral representation of the single-particle Green's function - the amplitude for an electron added at one point of space and imaginary time (the formal variable $\tau = it$ that turns quantum evolution into diffusion and underlies finite-temperature partition functions) to be removed at another - on a discrete lattice in space and imaginary time, written with anticommuting Grassmann variables that encode Fermi statistics and a quadratic action $S=\psi A\psi$ whose matrix $A$ encodes the finite-difference Hamiltonian. Second, renormalization-group (RG) transformations in the spirit of Wilson's numerical renormalization group for the Kondo problem (the screening of a magnetic impurity by conduction electrons): neighboring spatial sites are 'blocked' and alternate imaginary-time slices 'decimated' (integrated out), reducing the number of degrees of freedom while preserving the coarse-grained Green's function. Third, perturbative treatment of core electrons, whose dynamics is dominated by the static screened nuclear potential, so that only the valence electrons require nonperturbative treatment. Fourth, Monte Carlo sampling of the interacting valence problem: the two-body Coulomb repulsion is rewritten exactly as an average over a fluctuating one-body auxiliary field $\phi$ (a Hubbard-Stratonovich transformation), the Grassmann integrals are done exactly, giving a bosonic weight $s(\phi)$ containing $-\ln\det A(\phi)$, and $\phi$ is sampled with the Langevin equation, a stochastic differential equation whose stationary distribution is $e^{-s}$. A constant shift $V\to V-\lambda I$ of the discretized interaction makes $s(\phi)$ real, and a variance-reduction trick from lattice-field-theory Langevin work (Batrouni et al. 1985) replaces the matrix inverse $A^{-1}$ by an unbiased statistical estimator, empirically reducing the cost from order $N^{2}$ to roughly linear in the number of spatial sites $N$.

What the 1986 paper demonstrated was deliberately modest: an exact RG calculation for noninteracting spinless particles in a one-dimensional potential well $U(x)=-V\,\mathrm{sech}^{2}\,x$, where for moderate coarse-graining the renormalized action stays short-ranged and causal (its matrix elements $A_{bc}(\Delta l)$ vanish for negative imaginary-time separations $\Delta l<0$), and an energy-dependent effective Hamiltonian $H_s(E)$ reproduces the bound-state eigenvalues to high accuracy; plus a Langevin simulation of a three-site interacting model reaching temperatures only as low as $t/3$ or $u/6$. The paper itself flagged five unresolved issues. (i) Core-electron integration was only sketched: expand the action quadratically about its stationary point $\partial S/\partial\phi=0$ - whose stationary value reproduces the Hartree approximation, each electron moving in the average field of the others - and treat the remainder perturbatively; this was not implemented, analyzed, or demonstrated. (ii) When the coarse-graining scale $a$ grows so that the energy scale $-1/s$ ($s$ being the imaginary-time lattice spacing, with $r=s/a^{2}$ the coarse-graining ratio) becomes comparable to low bound-state energies, the renormalized action stops decaying exponentially, becomes long-ranged, and can acquire non-causal components $A_{bc}(\Delta l)\neq 0$ for $\Delta l<0$; the boundary of this breakdown as a function of $r$ and the potential strength, and remedies for it, were left open. (iii) Although the $V\to V-\lambda I$ shift renders the auxiliary-field action real, the weight contains $-\ln\det A(\phi)$, which is singular at field configurations with $\det A(\phi)=0$; at low temperatures these singularities entered the sampled region and produced incorrect expectation values in the authors' preliminary calculations. (iv) The unbiased-estimator speedup was validated only empirically on the three-site model. (v) Only spinless electrons in one dimension were treated (spin-1/2 requires twice as many auxiliary fields), and no atom or molecule in two or three dimensions was attempted.

The later history of these questions is intertwined with the origins of two major methods. White and Noack (1992) showed that standard real-space RG blocking of a Hamiltonian fails even for a single particle in a box, because a block held at fixed boundary conditions cannot represent the environment of its neighbors; this diagnosis led directly to White's density matrix renormalization group (DMRG, 1992), which truncates each block's Hilbert space using the reduced density matrix of an enlarged system rather than the block's own low-lying eigenstates, and which became the method of choice for one-dimensional correlated systems and, from White and Martin (1999) on, a standard tool of ab initio quantum chemistry. In parallel, the auxiliary-field line evolved from Langevin sampling into random walks in the space of Slater determinants: the constrained-path method (Zhang, Carlson, and Gubernatis 1995) and the phase-free method (Zhang and Krakauer 2003) suppress the exponential signal-to-noise decay caused by sign-changing or complex sampling weights - the fermion sign problem - by constraining the walk with a trial wavefunction, at the price of an uncontrolled approximation; Troyer and Wiese (2005) proved that the general sign problem is NP-hard (at least as hard as the hardest problems in the NP complexity class), so no exact general remedy is expected. The most direct later continuation of the RG-for-atoms line is Murthy and Kais (1998), who applied a finite-size numerical RG to low-lying electronic excitations of carbon and silicon atoms.

Problem Statement

Complete - or rigorously delimit - the renormalization-group approach to electronic structure of White, Wilkins, and Wilson (1986), whose single common objective is quantitatively accurate, systematically improvable calculation of the electronic structure of atoms and molecules by coarse-grained functional-integral sampling. Within one coherent methodology for the interacting problem, an answer must resolve:

1. Coarse-graining of bound-state systems. Map the boundary, in the coarse-graining ratio $r=s/a^{2}$ and the potential strength, beyond which the renormalized action of a system with bound states stops decaying exponentially in space and imaginary time and develops long-ranged or non-causal components ($A_{bc}(\Delta l)\neq 0$ for $\Delta l<0$), and provide remedies or corrective truncations that keep the blocked action local and causal with quantified error.

2. Core-electron integration. Implement the integration of core degrees of freedom in the interacting functional-integral formulation - the quadratic expansion of the action about its stationary point (the Hartree approximation) plus controlled higher corrections - with an error analysis and a demonstration on realistic atoms, so valence-electron properties can be computed without explicitly simulating core states.

3. Low-temperature auxiliary-field sampling. Eliminate or control the incorrect expectation values caused by singular configurations $\det A(\phi)=0$ of the real auxiliary-field action at low temperatures - exactly where possible, otherwise with a rigorously quantified bias - for realistic Coulomb systems.

4. Validation. Demonstrate the stability, accuracy, and computational scaling of the whole (including the unbiased-estimator modification) at lattice sizes and temperatures adequate for realistic Coulomb systems, with spin-1/2 electrons in more than one spatial dimension.

These deliverables are jointly needed for the common objective: (1) and (2) are the coarse-graining and core-valence separation that make the interacting lattice problem tractable, (3) is the unbiased sampling of what remains, and (4) assembles and validates (1)-(3). Successor methods (DMRG, phase-free auxiliary-field QMC) that meet parts of the overall objective by abandoning the framework do not by themselves answer this question, though they bound what any answer must achieve. Likewise, the rigorously quantified bias allowed in item 3 must be a bound on the error of the auxiliary-field expectation values caused by the $\det A(\phi)=0$ singularities within this framework, not an uncontrolled approximation.

References

1. S. R. White, J. W. Wilkins, and K. G. Wilson, "Renormalization-Group Approach for Electronic Structure", Phys. Rev. Lett. 56, 412-415 (1986). DOI: 10.1103/PhysRevLett.56.412, https://doi.org/10.1103/PhysRevLett.56.412 (source paper)
2. K. G. Wilson, "The renormalization group: Critical phenomena and the Kondo problem", Rev. Mod. Phys. 47, 773-840 (1975). DOI: 10.1103/RevModPhys.47.773, https://doi.org/10.1103/RevModPhys.47.773
3. G. G. Batrouni, G. R. Katz, A. S. Kronfeld, G. P. Lepage, B. Svetitsky, and K. G. Wilson, "Langevin simulations of lattice field theories", Phys. Rev. D 32, 2736-2747 (1985). DOI: 10.1103/PhysRevD.32.2736, https://doi.org/10.1103/PhysRevD.32.2736
4. S. R. White and R. M. Noack, "Real-Space Quantum Renormalization Groups", Phys. Rev. Lett. 68, 3487 (1992). DOI: 10.1103/PhysRevLett.68.3487, https://doi.org/10.1103/PhysRevLett.68.3487
5. S. R. White, "Density Matrix Formulation for Quantum Renormalization Groups", Phys. Rev. Lett. 69, 2863-2866 (1992). DOI: 10.1103/PhysRevLett.69.2863, https://doi.org/10.1103/PhysRevLett.69.2863
6. S. R. White, "Real Space Numerical Renormalization Groups for Hubbard Models", in Computational Approaches in Condensed-Matter Physics (eds. Miyashita, Imada, Takayama), Springer Proceedings in Physics vol. 70, pp. 97-104 (1992). DOI: 10.1007/978-3-642-84821-6_15, https://link.springer.com/chapter/10.1007/978-3-642-84821-6_15
7. U. Schollwöck, "The density-matrix renormalization group", Rev. Mod. Phys. 77, 259-315 (2005). DOI: 10.1103/RevModPhys.77.259, https://arxiv.org/abs/cond-mat/0409292
8. S. R. White and R. L. Martin, "Ab initio quantum chemistry using the density matrix renormalization group", J. Chem. Phys. 110, 4127-4130 (1999). DOI: 10.1063/1.478295, https://doi.org/10.1063/1.478295
9. S. Zhang, J. Carlson, and J. E. Gubernatis, "Constrained Path Quantum Monte Carlo Method for Fermion Ground States", Phys. Rev. Lett. 74, 3652-3655 (1995). DOI: 10.1103/PhysRevLett.74.3652, https://arxiv.org/abs/cond-mat/9503055
10. S. Zhang and H. Krakauer, "Quantum Monte Carlo Method using Phase-Free Random Walks with Slater Determinants", Phys. Rev. Lett. 90, 136401 (2003). DOI: 10.1103/PhysRevLett.90.136401, https://doi.org/10.1103/PhysRevLett.90.136401
11. M. Troyer and U.-J. Wiese, "Computational Complexity and Fundamental Limitations to Fermionic Quantum Monte Carlo Simulations", Phys. Rev. Lett. 94, 170201 (2005). DOI: 10.1103/PhysRevLett.94.170201, https://arxiv.org/abs/cond-mat/0408370
12. G. Murthy and S. Kais, "Renormalization group approach for electronic excitations in atoms", Chem. Phys. Lett. 290, 199-204 (1998). DOI: 10.1016/S0009-2614(98)00499-0, https://doi.org/10.1016/S0009-2614(98)00499-0
13. M. Motta and S. Zhang, "Ab initio computations of molecular systems by the auxiliary-field quantum Monte Carlo method", WIREs Comput. Mol. Sci. 8, e1364 (2018). DOI: 10.1002/wcms.1364, https://arxiv.org/abs/1711.02242