Full Configuration-Interaction Quantum Monte Carlo (FCIQMC)

• FCIQMC is a stochastic projector method that samples the full configuration-interaction wave function via a population of signed walkers, providing exact ground-state energies in the large-walker limit.
• The algorithm dynamically controls walker populations through spawning, death, and annihilation processes, efficiently navigating vast Hilbert spaces in strongly correlated systems.
• Enhancements like initiator adaptation, adaptive-shift feedback, and unbiased reduced density matrix sampling extend FCIQMC to excited states, symmetry adaptations, and basis-set corrections.

Full Configuration-Interaction Quantum Monte Carlo (FCIQMC) is a stochastic projector method for solving the ground-state Schrödinger equation in a finite basis by sampling the full configuration-interaction (FCI) wave function in determinant space. In the usual formulation, it is the direct quantum Monte Carlo analogue of full configuration interaction: the exact ground state is sought over the space of all Slater determinants, but the coefficient vector is not stored explicitly. Instead, the wave function is represented by a population of signed walkers whose long-time average samples the FCI amplitudes, so that the method is exact within the chosen basis in the large-walker limit and can reach Hilbert spaces far beyond deterministic diagonalization [(Shepherd et al., 2014); (Kersten et al., 2016)].

1. Projector formulation and walker dynamics

The basic FCIQMC ansatz is the FCI expansion

$|\Psi\rangle = \sum_I C_I |D_I\rangle,$

with imaginary-time projection used to isolate the lowest-energy state in the chosen symmetry sector. In discrete form, one common propagation equation is

$$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$

while an equivalent short-time projector form is

$$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$

Here the shift $S$ controls normalization and, after equilibration, tracks the ground-state energy (Kersten et al., 2016, Blunt et al., 2015).

Operationally, FCIQMC realizes this propagation through spawning, death or cloning, and annihilation. Walkers on determinant $D_j$ attempt to spawn onto connected determinants according to off-diagonal Hamiltonian matrix elements; diagonal terms generate removal or replication events; and walkers of opposite sign on the same determinant are canceled. The annihilation step is crucial because it suppresses uncontrolled noise and enables the correct fermionic sign structure to emerge. Projected estimators, typically onto a reference determinant or a compact trial space, provide low-variance energy estimates, while shift averages offer a complementary population-control estimator [(Shepherd et al., 2014); (Hu et al., 6 Jul 2026)].

This stochastic representation changes the storage model rather than the target problem. Deterministic FCI requires explicit construction and diagonalization of the full Hamiltonian matrix, whereas FCIQMC samples the same full-space solution without explicit matrix storage. This makes the method particularly attractive in strongly correlated and multireference settings, where determinant spaces are combinatorially large but the physically relevant amplitudes can still be discovered by projector dynamics (Kersten et al., 2016).

2. Sign structure, annihilation plateaus, and representation regimes

FCIQMC does not remove the fermion sign problem; it reformulates it. The central difficulty is that the signs of the exact CI coefficients are not known in advance. Without annihilation, the walker dynamics tend toward the positive-definite solution of a modified Hamiltonian,

$$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$

which is the determinant-space analogue of a bosonic problem with no sign structure. The sign problem in FCIQMC is therefore the problem of maintaining the correct fermionic sign pattern under stochastic propagation (Shepherd et al., 2014).

A central diagnostic is the annihilation plateau, defined as the minimum walker population needed for an exact-on-average representation of the ground-state vector. In the one-dimensional $k$-space Hubbard model, the plateau was analyzed through

$N_{\text{plateau}}=\beta N_{\text{det}}^\gamma,$

revealing distinct scaling regimes. In a conventional linear regime, observed at larger $U$ and sufficiently large systems, $\gamma=1$ with $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$0. In a sub-linear regime, especially at smaller $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$1 and moderate system size, $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$2; the smallest observed exponent was about $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$3 for $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$4. For $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$5, the return to $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$6 appeared around $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$7 (Shepherd et al., 2014).

These results established that the exact wave function can sometimes be represented much more cheaply than the Hilbert-space size would suggest, despite being a linear combination of determinants. The corresponding sign-problem landscape is representation-dependent rather than universal. In the same work, a 70-site half-filled Hubbard model with a space of $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$8 determinants was solved in 250 core hours, with reported energy $$n_J(\tau+\Delta\tau)=n_J(\tau)-\Delta\tau\sum_I (H_{JI}-E_S\delta_{IJ})\,n_I(\tau),$$9, whereas exact diagonalization was estimated to require about $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$0 core hours. This was described as the largest space sampled in an unbiased fashion at that time (Shepherd et al., 2014).

A common misconception is that FCIQMC is “sign-problem free” once annihilation is present. The literature instead treats annihilation as the mechanism that controls sign instability, not as a guarantee that sign-related cost disappears. The plateau, sub-linear regimes, and basis dependence show that the severity of the sign problem must be characterized empirically for each Hamiltonian representation (Shepherd et al., 2014).

3. Population control, initiator bias, and feedback schemes

Because the walker population would otherwise grow or decay exponentially, FCIQMC relies on feedback through the shift. In its present form, the algorithm can be cast as a discrete-time Markov chain, and this perspective reveals a population-control bias analogous to that in diffusion Monte Carlo. In both the two-determinant model and realistic atomic calculations, the bias decreases approximately as $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$1, where $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$2 is the average population. A post-processing reweighting scheme based on the shift history was shown to be effective for removing this bias, and smaller damping parameters reduce the bias prefactor in the projected-energy estimator (Vigor et al., 2014).

A later refinement replaced the original two-stage control protocol with a target-seeking feedback law,

$$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$3

which drives the walker number toward a prescribed target $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$4. In a scalar approximation, the logarithm of the population behaves as a damped harmonic oscillator, with critical damping at

$$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$5

The fixed point is $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$6, and linear stability of the discrete map requires $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$7. This modification substantially suppresses overshoots, improves predictability of memory use, and leaves the standard error of the shift estimator and the population-control bias unaffected within error bars. The same work introduced a population growth witness $$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$8 that exposes annihilation plateaus even when population control hides them in the raw walker count (Yang et al., 2020).

The other major finite-population approximation is the initiator adaptation, i-FCIQMC. Here, spawning from low-population determinants to previously unoccupied determinants is suppressed, stabilizing the dynamics at the cost of a systematic bias that vanishes as the walker number increases. The large-walker limit remains

$$\hat{P}=\mathbb{1}-\Delta\tau(\hat{H}-S\mathbb{1}).$$9

and the initiator method was later interpreted as expanding an already existing reduced-scaling regime rather than creating a fundamentally new sparsity mechanism [(Shepherd et al., 2014); (Hu et al., 21 Jun 2026)].

Adaptive-shift corrections were introduced to reduce this initiator undersampling bias by assigning determinant-dependent local shifts. In the offset adaptive-shift formulation,

$S$0

with $S$1 corresponding to the full adaptive shift. This offset weakens the correction in a controlled way, does not alter the exact large-walker limit, and often improves finite-walker convergence substantially. A practical rule of thumb reported in the benchmarks is that an offset around half the correlation energy is often near optimal (Ghanem et al., 2020).

4. Reduced density matrices, pure estimators, and basis-set corrections

FCIQMC was initially used primarily for energies, but its formal scope broadened once unbiased reduced density matrix (RDM) sampling became practical. One- and two-body RDMs,

$S$2

can be accumulated on the fly during walker evolution. Because quadratic observables are biased if estimated from a single stochastic realization, replica sampling with two statistically independent simulations is required for unbiased RDMs. This made possible not only RDM-based energy diagnostics but also explicitly correlated corrections, orbital optimization, excited-state properties, and transition moments [(Booth et al., 2012); (Blunt et al., 2017)].

Distributed sparse storage then removed the memory bottleneck for large-basis RDM calculations. Excited-state RDMs were combined with the Gram–Schmidt excited-state algorithm, and transition density matrices enabled transition dipole moments and oscillator strengths. For LiH, BH, and MgO, energies displayed small initiator error while dipole moments were markedly more sensitive, reinforcing the point that property convergence is often slower than energy convergence (Blunt et al., 2017).

A major practical limitation of determinant-space methods is basis-set incompleteness error. FCIQMC is exact in a given one-particle basis, but determinant expansions are ill-suited to the electron–electron cusp, so basis convergence is slow. Two universal explicitly correlated remedies were developed in the FCIQMC framework. The first, FCIQMC-$S$3, is a diagonalize-then-perturb scheme that applies an a posteriori internally contracted perturbative correction using FCIQMC-sampled density matrices. The second, canonical transcorrelation, is a perturb-then-diagonalize scheme that transforms the Hamiltonian before the FCIQMC simulation. On the G1 set, both reduced the median absolute atomization-energy error in aug-cc-pVDZ from about $S$4 with plain FCIQMC to about $S$5 with CT-FCIQMC and about $S$6 with FCIQMC-$S$7; for total energies per correlated electron, the average absolute error was reduced from $S$8 to $S$9 and $D_j$0, respectively (Kersten et al., 2016).

The same theme appears in the earlier FCIQMC-F12 work on carbon dimer. There, the non-parallelity error in cc-pVDZ was reported to drop from $D_j$1 to $D_j$2, and the improvement was summarized as equivalent to increasing the quality of the one-electron basis by two cardinal numbers (Booth et al., 2012). Beyond energies, a general analytical nuclear-force formulation later made FCIQMC usable as a force engine with frozen cores and arbitrary orbital choices, enabling construction of a full-dimensional ground-state potential energy surface of water by Gaussian-process regression from FCIQMC energies and forces (Jiang et al., 2022).

5. Excited states, symmetry adaptation, and generalized configuration spaces

Although FCIQMC is a ground-state projector method by construction, excited-state extensions were developed by running several simulations simultaneously and orthogonalizing higher states against lower ones at every iteration. The propagator for state $D_j$3 becomes

$D_j$4

with

$D_j$5

This Gram–Schmidt step is computationally inexpensive, requires no trial wave functions or space partitioning for the propagation itself, and targets several of the lowest states within the same symmetry sector. On the carbon dimer, the method produced low-energy potential curves in cc-pVDZ, cc-pVTZ, and cc-pVQZ, with agreement to about $D_j$6 against high-accuracy DMRG in the cc-pVQZ benchmarks (Blunt et al., 2015).

Spin adaptation pushed the symmetry resolution further. In the GUGA-FCIQMC formulation, the determinant basis is replaced by configuration state functions constructed through the Graphical Unitary Group Approach, so the sampled wave function is an eigenfunction of $D_j$7. This explicitly resolves different spin sectors, permits direct treatment of low- and intermediate-spin states, and accelerates convergence when spin gaps are small. Illustrative applications included the $D_j$8 spin gap of the cobalt atom and the $D_j$9, $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$0, $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$1, and $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$2 spin gaps of stretched N$$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$3 (Dobrautz et al., 2019).

The FCIQMC framework has also been generalized beyond purely fermionic determinant spaces. For coupled electron–boson Hamiltonians, the configuration space becomes a product of fermionic configurations and bosonic occupation states, and the method can sample without a priori truncation of boson occupation. In a sign-problem-free Hubbard–Holstein model, unbiased energies extrapolated to the thermodynamic limit were obtained by combining importance sampling and reweighting, while general electron–boson auxiliary models in embedding calculations were sampled successfully despite a formal sign problem, including faithful reconstruction of converged reduced density matrices (Anderson et al., 2022).

6. Benchmark domains, limitations, and methodological significance

FCIQMC first established its large-scale relevance in periodic electronic structure. For the 54-electron homogeneous electron gas at $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$4 and $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$5 a.u., the method sampled Hilbert spaces from about $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$6 to $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$7 Slater determinants and produced finite-basis and extrapolated energies competitive with, and in some cases lower than, state-of-the-art backflow diffusion Monte Carlo results. The extrapolated complete-basis estimates reported were $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$8 and $$\tilde H_{ij}= H_{ij}\delta_{ij} - (1-\delta_{ij})|H_{ij}|,$$9 a.u./electron at $k$0 and $k$1 a.u., respectively (Shepherd et al., 2011).

Subsequent applications broadened the domain substantially. Quasi-Newton preconditioning accelerated convergence of the instantaneous projected energy by over an order of magnitude in uniform-electron-gas tests, with only $k$2 extra computational cost, and enabled highly accurate Cr$k$3 calculations in an all-electron basis with Hilbert-space size about $k$4 (Neufeld et al., 2019). Semi-stochastic propagation, in which a selected determinant subspace is propagated exactly while the rest remains stochastic, produced efficiency gains that could exceed $k$5 in favorable Hubbard cases and proved especially effective for replica-sampled RDMs and post-FCIQMC corrections (Blunt et al., 2015).

In molecular benchmark work, adaptive-shift semi-stochastic FCIQMC provided near-exact basis-set-limited descriptions of 13 low-lying states of ScO, TiO, and VO, including states with significant multi-configurational character, and was used to assess coupled-cluster and density-functional approximations (Jiang et al., 2021). In nuclear structure, FCIQMC was introduced as a full-configuration-space solver for chiral EFT Hamiltonians and used to compute ground-state energies and charge radii of $k$6He, $k$7Be, $k$8C, and $k$9O with sub-percent-level many-body uncertainties; for $N_{\text{plateau}}=\beta N_{\text{det}}^\gamma,$0O at $N_{\text{plateau}}=\beta N_{\text{det}}^\gamma,$1, the FCI dimension was reported as about $N_{\text{plateau}}=\beta N_{\text{det}}^\gamma,$2 (Hu et al., 21 Jun 2026). A detailed follow-up validated projected, trial, and RDM-based pure estimators against deterministic FCI in small spaces, analyzed finite-walker extrapolations in larger spaces, and demonstrated low-lying excited states of $N_{\text{plateau}}=\beta N_{\text{det}}^\gamma,$3Li by multi-state orthogonalized propagation (Hu et al., 6 Jul 2026).

The method’s limitations are correspondingly well defined. FCIQMC is exact within a chosen basis, not at the complete-basis limit; basis-set incompleteness can remain the dominant error source unless explicitly correlated corrections or basis extrapolations are used [(Booth et al., 2012); (Kersten et al., 2016)]. Population-control bias exists and must be controlled or reweighted (Vigor et al., 2014). The initiator approximation introduces finite-walker bias, and pure estimators, such as radii in nuclear applications or dipole moments in molecular applications, often converge more slowly than projected energies (Blunt et al., 2017, Hu et al., 6 Jul 2026). Most fundamentally, the sign problem persists, but its severity depends on representation, Hamiltonian structure, and system size, so the practical goal is not merely to “solve” the sign problem in the abstract but to characterize and enlarge regimes where the exact state can be represented efficiently (Shepherd et al., 2014).

Across these developments, FCIQMC emerged as a stochastic exact-diagonalization methodology whose distinctive strengths are exact-on-average full-space sampling, a walker representation that can exploit sparsity not evident from Hilbert-space size alone, and a flexible algorithmic ecosystem spanning initiator corrections, semi-stochastic propagation, symmetry adaptation, explicitly correlated corrections, force evaluation, and generalized fermion–boson or nuclear many-body applications. The recurring development target in the literature is to expand the sub-linear or otherwise favorable representation regimes while retaining exact-on-average accuracy (Shepherd et al., 2014).