Fermion Sign Problem

• Fermion sign problem is the challenge in simulating fermionic systems due to cancellation between large positive and negative contributions in the partition function.
• It arises from antisymmetrization effects such as permutation parity and sign-indefinite determinants in world-line and auxiliary-field approaches.
• Various strategies, including sign-free reformulations, attenuation methods, and extrapolative techniques, are used to mitigate its computational impact.

The fermion sign problem (FSP) is the computational obstruction that arises when antisymmetrization of identical-fermion contributions makes the partition function, or any estimator derived from it, a small residual difference of large positive and negative terms. In world-line and coordinate-space path integrals, the sign originates from permutation parity; in auxiliary-field formulations, it appears after integrating out fermions as sign-indefinite determinants; in projector methods, it appears as an exponential loss of efficiency with increasing projection time and particle number. Across these representations, the central symptom is an exponentially decaying average sign, which converts unbiased Monte Carlo sampling into an exponentially hard cancellation problem (Dornheim, 2019, Iazzi et al., 2014).

1. Formal definition and standard statistical formulation

A canonical coordinate-space representation makes the origin of the FSP explicit:

$$Z = \frac{1}{N!} \sum_{\sigma\in S_N} \mathrm{sgn}(\sigma)\int d\mathbf{R}\,\langle \mathbf{R}|e^{-\beta \hat H}|\hat\pi_\sigma \mathbf{R}\rangle .$$

Here the antisymmetry of fermionic exchange inserts alternating signs into the partition function. In the language used for path-integral Monte Carlo (PIMC), the full partition function is rewritten as

$Z=\int d\mathbf{X}\, W(\mathbf{X}),$

but for fermions the weight $W(\mathbf{X})$ is not positive definite. The standard workaround is to sample the modified partition function

$Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$

and recover observables by reweighting,

$$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$

The average sign,

$S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$

is therefore the quantitative control parameter of the problem in canonical PIMC (Dornheim, 2019).

The same structure appears in auxiliary-field QMC. One samples absolute weights and computes

$$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$

with

$$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$

Once $\langle s\rangle$ becomes exponentially small in inverse temperature and volume, fixed-precision estimates require exponentially increasing effort (Iazzi et al., 2014).

A basic but important distinction is that the FSP is not identical to the “sign problem” of complex oscillatory path integrals. One strand of the literature separates an oscillatory complex-action problem from the fermionic cancellation problem caused by antisymmetrization of real-valued density-matrix elements. In that terminology, the fermionic problem is fundamentally combinatorial and exchange-driven rather than merely phase-oscillatory (Filinov et al., 2020).

2. Exchange topology, geometric phase, and representation dependence

In the world-line picture, the sign is tied directly to exchange topology. Odd fermionic exchanges generate negative weights, and the configuration can be visualized topologically as a Möbius-strip-like sector rather than a cylinder. This already indicates that the sign is not a local numerical artifact but a property of the global topology of paths (Iazzi et al., 2014).

In auxiliary-field QMC, the same issue becomes less transparent because the fermions are integrated out and the sign is encoded in determinants. A key conceptual result is that the negative sign of an auxiliary-field configuration can be interpreted as a geometric phase: specifically, an imaginary-time analogue of the Aharonov–Anandan phase, reducing to a Berry phase in the adiabatic limit. For real auxiliary Hamiltonians, this phase is quantized to $0$ or $Z=\int d\mathbf{X}\, W(\mathbf{X}),$0, so a negative sign expresses a topological obstruction to choosing globally smooth, periodic eigenvectors in imaginary time. This perspective implies that the sign problem is not removed by taking smaller time steps, using higher precision, or smoothing the auxiliary field; its origin is topological rather than discretization-based (Iazzi et al., 2014).

The severity of the FSP is also representation dependent. In continuum projector QMC analyzed through the Shadow Wave Function formalism, the efficiency obeys

$Z=\int d\mathbf{X}\, W(\mathbf{X}),$1

showing exponential degradation with particle number $Z=\int d\mathbf{X}\, W(\mathbf{X}),$2, effective projection length $Z=\int d\mathbf{X}\, W(\mathbf{X}),$3, and density $Z=\int d\mathbf{X}\, W(\mathbf{X}),$4. The same analysis connects stronger orbital localization with weaker sign cancellations, indicating that localization suppresses the sign problem by reducing effective permutation fluctuations (Calcavecchia et al., 2016).

A related representation-dependent mechanism appears in full configuration interaction quantum Monte Carlo (FCIQMC). There the sign problem is not formulated as a path-integral weight problem but as the competition between the physical signed population $Z=\int d\mathbf{X}\, W(\mathbf{X}),$5 and an unphysical in-phase population $Z=\int d\mathbf{X}\, W(\mathbf{X}),$6. Without annihilation, the latter can dominate. Efficient annihilation of opposite-sign walkers is therefore the mechanism that restores convergence to the fermionic ground state, and the characteristic population plateau becomes a practical measure of sign-problem severity (Spencer et al., 2011).

3. Exact sign-free sectors and positivity theorems

A substantial body of work does not attenuate the FSP but eliminates it exactly for restricted Hamiltonian classes by exploiting algebraic structure, symmetry, or nontrivial reformulation.

One prominent example is the fermion bag approach for half-filled, spin-polarized fermions on bipartite lattices. For the repulsive $Z=\int d\mathbf{X}\, W(\mathbf{X}),$7 model,

$Z=\int d\mathbf{X}\, W(\mathbf{X}),$8

the free hopping matrix obeys

$Z=\int d\mathbf{X}\, W(\mathbf{X}),$9

After summing over diagonal particle-hole variables, the partition function becomes

$W(\mathbf{X})$0

and the determinant is nonnegative because $W(\mathbf{X})$1 is real antisymmetric and $W(\mathbf{X})$2. This yields a sign-problem-free formulation for repulsive half-filled bipartite systems, with extensions to broader half-filled interactions and to attractive odd-flavor Hubbard models such as $W(\mathbf{X})$3 (Huffman et al., 2013).

A more general algebraic framework is provided by Majorana positivity. In auxiliary-field QMC, two sufficient conditions were proved for nonnegative configuration weights: Majorana reflection positivity and Majorana Kramers positivity. Reflection positivity requires bilinear kernels of the form

$W(\mathbf{X})$4

with Hermitian $W(\mathbf{X})$5 semidefinite. Kramers positivity requires matrices $W(\mathbf{X})$6 and $W(\mathbf{X})$7 satisfying

$W(\mathbf{X})$8

together with $W(\mathbf{X})$9, $Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$0, $Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$1, and $Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$2. These criteria unify many previously known sign-free AFQMC models and extend them to new cases, including repulsive spinless fermions without particle-hole symmetry and interacting topological insulators with spin-flip terms (Wei et al., 2016).

Exact elimination can also arise from constrained world-line geometry. In a frustrated quantum impurity model of three spin-$Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$3 chains coupled through a triangular impurity, a sequence of mappings leads to an effectively one-dimensional fermion model in which sign changes come only from self-crossing world lines. By pairing fermion world lines on half the system, all negative processes occur in matched pairs along closed loops, so global signs cancel and the original frustrated spin model becomes sign-free in a composite-state basis (Hann et al., 2016).

Not all fermion-bag reorganizations remove the problem completely. In strongly coupled lattice QED with one Wilson fermion, the bag formulation reveals that “complex bags” carry nearly symmetric positive and negative weight distributions and drive a severe sign problem, whereas “simple bags” are mostly positive. In the special limit $Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$4, however, all bag weights become nonnegative (Chandrasekharan et al., 2010).

These exact results establish an important principle: sign freedom is often not absent because fermions are benign, but because an alternative representation makes positivity manifest for a sharply delimited class of models.

4. Attenuation, auxiliary systems, and extrapolative strategies

Outside such exact sectors, most progress is based on attenuation rather than elimination.

One line of attack replaces explicit antisymmetrization by an effective phase-space exchange pseudopotential in a Wigner-function formulation. In the pair approximation, identical-fermion exchange is encoded by a logarithmic repulsive pseudopotential depending on coordinates, momenta, spin, and degeneracy,

$Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$5

This converts Pauli blocking into a positive effective interaction in phase space. For the ideal degenerate Fermi gas, the method reproduces momentum distributions in good agreement with analytical Fermi-Dirac results over a substantial degeneracy range (Larkin et al., 2017). Closely related Wigner/pseudopotential formulations explicitly present this as a way to avoid explicit permutation sums and thereby mitigate the fermionic sign problem (Filinov et al., 2020).

Another attenuation strategy adds an auxiliary repulsion to improve the average sign and then extrapolates back to the physical Hamiltonian. For

$Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$6

with a short-range dipolar perturbation

$Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$7

the modified system has a much larger sign. In 2D and 3D quantum dots, this approach yielded speed-ups exceeding $Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$8 in favorable cases while retaining relative accuracy of $Z'=\int d\mathbf{X}\, |W(\mathbf{X})|$9. The same work introduced thermodynamic integration,

$$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$0

as a more reliable low-temperature correction than direct energy extrapolation alone (Dornheim et al., 2020).

A conceptually different extrapolative approach introduces a statistics parameter $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$1 interpolating between bosons, distinguishable particles, and fermions:

$$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$2

Here $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$3 corresponds to bosons, $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$4 to distinguishable particles, and $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$5 to fermions. Because the path-integral weights are sign-problem-free for $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$6, one can sample only that side with PIMD, fit observables such as the energy as functions of $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$7, and extrapolate to $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$8. The method is empirically reliable when $$\langle \hat A\rangle=\frac{\langle \hat A \hat S\rangle'}{\langle \hat S\rangle'}, \qquad S(\mathbf{X})=\frac{W(\mathbf{X})}{|W(\mathbf{X})|}=\pm 1.$$9 is monotonic and has no inflection point on $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$0; in that regime it predicts fermionic energies for systems as large as $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$1, beyond the reach of direct sign reweighting. The same work reported the approximate relation

$S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$2

for high-temperature noninteracting particles and for strongly repulsive interacting particles at low temperature, reflecting near-linearity in $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$3 (Yunuo et al., 2022).

Recent auxiliary-system variants go further by replacing each short-time fermionic determinant by its absolute value, defining a pseudo-fermion propagator that is sign-free by construction. Fermionic energies are then inferred by an $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$4-dependent correction calibrated in the noninteracting limit. Quantum-dot benchmarks from $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$5 to $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$6 in 2D harmonic confinement show excellent agreement with established methods in the reported cases (Xiong et al., 13 Aug 2025).

5. Lee–Yang zeros, analytic continuation, and low-temperature obstruction

The interpolation parameter $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$7 also exposes the analytic structure behind extrapolation-based methods. If the partition function is treated as a polynomial in $S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$8,

$S=\frac{Z}{Z'}=e^{-\beta N(f-f')},$9

then its zeros in the complex $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$0-plane are Lee–Yang zeros, and thermodynamic quantities become nonanalytic at those points (He et al., 30 Jul 2025).

At $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$1, one analysis finds an especially rigid zero structure:

$$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$2

Thus the fermionic endpoint $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$3 is itself a zero of the partition function in the zero-temperature limit. This directly obstructs analytic continuation from the sign-problem-free interval $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$4 to fermions whenever the continuation path intersects, or is strongly distorted by, these zeros (He et al., 30 Jul 2025).

A finite-temperature solvable benchmark sharpens this picture. For noninteracting particles on a one-dimensional ring, the zero emerging from $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$5 remains exponentially close to $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$6 at low temperature,

$$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$7

and dominates the sign factor

$$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$8

At low and moderate temperature, the zeros lie on the negative real axis; at higher temperature they move into complex-conjugate pairs, opening a smoother analytic corridor on the real $$\langle O\rangle=\frac{\langle O s\rangle'}{\langle s\rangle'},$$9 axis. This explains why direct $$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$0-extrapolation and contour-based implicit fitting can fail even at high polynomial order in the low-temperature regime but become reasonable again once the relevant zeros move away from the real axis (He et al., 5 Jun 2026).

From this viewpoint, empirical extrapolation criteria such as monotonicity and absence of inflection in $$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$1 are practical diagnostics of an underlying analytic constraint. A plausible implication is that many successful continuation-based FSP methods operate not because the fermionic endpoint is benign, but because the relevant Lee–Yang zeros remain far enough from the sampled domain for a given temperature window (Yunuo et al., 2022, He et al., 5 Jun 2026).

6. The sign as physical observable, critical correlator, and structural diagnostic

An important recent development is the reinterpretation of the sign itself as a source of physical information rather than only a measure of numerical failure.

In auxiliary-field QMC, the spin-resolved average sign

$$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$2

has been shown to obey finite-size scaling near phase transitions:

$$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$3

This produces universal crossings and data collapse in the vicinity of quantum critical points and Kosterlitz–Thouless transitions. In the SU(2) honeycomb Hubbard model, the square-lattice ionic Hubbard model, and the attractive Hubbard model, the sign behaves as a “minimal correlator” of criticality, even in situations where the total sign is symmetry-protected and remains unity (Mondaini et al., 2022).

The same idea has been used in the two-dimensional attractive Hubbard model with spin imbalance. There the total sign and spin-resolved signs distinguish vacuum, equal-density, partially polarized, and fully polarized regimes, and the onset of the partially polarized phase—where FFLO-like physics is expected—coincides with departures of $$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$4 from unity. The reported scaling is consistent with a free-fermion universality class with $$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$5 and $$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$6, and a Berezinskii–Kosterlitz–Thouless analysis of the sign yields indirect evidence for a robust polarized superfluid phase in 2D (Yi et al., 2024).

A distinct but related strategy is to preempt the sign problem by extracting universal information before the average sign collapses. Short-time imaginary-time relaxation near a quantum critical point already exhibits universal scaling,

$$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$7

while the sign remains much larger than in equilibrium ground-state projection. In the reported examples, average signs around $$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$8–$$\langle s\rangle=\frac{Z}{Z'}=\exp(-\beta V\Delta f).$$9 at short times replaced equilibrium values of $\langle s\rangle$0–$\langle s\rangle$1, enabling numerically exact determination of critical points and exponents in sign-problematic Dirac fermion models, including an $\langle s\rangle$2 staggered-flux Hubbard model with a reported $\langle s\rangle$3-antiferromagnetic critical point in a new universality class (Yu et al., 2024).

The sign structure also appears in geometric diagnostics of many-body states. For backflow-modified Slater determinants, nodal hypersurfaces can become fractal, with correlation integral

$\langle s\rangle$4

over a scale window $\langle s\rangle$5. Yet the second Rényi entropy $\langle s\rangle$6 remains largely insensitive to the detailed fractal geometry of the sign structure: in 2D it exhibits a crossover from volume scaling $\langle s\rangle$7 in the critical backflow regime to the Fermi-liquid form $\langle s\rangle$8 on larger scales. This suggests that entanglement entropy detects the presence of a dense sign structure more readily than its fine nodal organization (Kaplis et al., 2016).

The contemporary view of the FSP is therefore stratified rather than monolithic. Exact positivity results exist, but only for highly structured sectors; most general-purpose methods attenuate, extrapolate, or bypass the problem in restricted regimes; and a parallel conceptual literature treats the sign as a topological invariant, an analytic obstruction, or even a critical observable. The persistent misconception is that these are interchangeable notions of “solution.” The literature instead supports a sharper distinction between exact sign-free reformulations, controlled regime-dependent mitigations, and physical diagnostics extracted from the very structure that makes fermionic sampling difficult (Iazzi et al., 2014, Dornheim et al., 2020).