Sign Problem Preview

A sign problem occurs when a chosen stochastic representation assigns negative or complex weights to configurations that one would like to sample as probabilities. For the finite thermal systems considered here, the partition function remains positive and well defined. The obstruction is that large positive and negative, or differently phased, contributions cancel before producing the physical answer.

Suppose an exact representation gives

$$Z = \sum_C W(C),$$

with $W(C)\in\mathbb C$. Define

$$\begin{aligned} Z_{\mathrm{abs}} &= \sum_C |W(C)|, \\ \phi(C) &= \begin{cases} W(C)/|W(C)|, & W(C)\neq0, \\ 0, & W(C)=0. \end{cases} \end{aligned}$$

Assume $\beta>0$ and $0<Z_{\mathrm{abs}}<\infty$. The absolute-weight ensemble is then a valid probability distribution,

$$q_{\mathrm{abs}}(C) = \frac{|W(C)|}{Z_{\mathrm{abs}}},$$

but the desired partition function is recovered only after the phase is restored:

$$\left\langle\phi\right\rangle_{\mathrm{abs}} = \frac{Z}{Z_{\mathrm{abs}}}.$$

When this average sign or phase is exponentially small, ordinary sampling spends nearly all its effort estimating contributions that cancel.

Required background. Quantum Monte Carlo explains configuration weights and sampling estimators.

Signed, complex, and vanishing weights

Three related cases should be distinguished.

A sign problem has real weights with

$$\phi(C) = \operatorname{sgn}W(C) \in \{-1,+1\}$$

for nonzero configurations.

A phase problem has genuinely complex weights,

$$\phi(C) = e^{i\theta(C)}.$$

A zero-overlap or rare-event problem can occur even with nonnegative weights when the sampling ensemble rarely visits configurations important to the target observable. It may produce similarly bad statistics, but it is not automatically a sign problem.

Negative and complex weights are not negative or complex probabilities. They are amplitudes, determinants, expansion coefficients, or effective-action factors whose sum yields a physical quantity. Sampling $|W|$ changes the theory being sampled; reweighting by $\phi$ is what restores the original one.

The exact reweighting identity

Let a configuration estimator $A(C)$ represent an observable. Then

$$\begin{aligned} \langle A\rangle &= \frac{ \sum_C W(C)A(C) }{ \sum_C W(C) } \\ &= \frac{ \left\langle \phi A \right\rangle_{\mathrm{abs}} }{ \left\langle \phi \right\rangle_{\mathrm{abs}} }. \end{aligned}$$

This identity is exact whenever the sums exist and $Z\neq0$. It does not say that the ratio is statistically well conditioned.

A common and usually advantageous estimator uses the same sampled configurations for numerator and denominator,

$$\begin{aligned} \widehat X &= \frac{1}{N} \sum_{t=1}^{N} \phi_t A_t, \\ \widehat Y &= \frac{1}{N} \sum_{t=1}^{N} \phi_t, \\ \widehat A &= \frac{\widehat X}{\widehat Y}. \end{aligned}$$

Their covariance matters when samples are shared. Independent simulations are also valid, but then their separate uncertainties must be propagated. Reporting a small error bar for $\widehat X$ while treating $\widehat Y$ as exact can be dramatically misleading.

A two-configuration cancellation

Take

$$W(C_1) = 1, \qquad W(C_2) = -(1-\varepsilon),$$

with $0<\varepsilon\ll1$. Then

$$\begin{aligned} Z &= \varepsilon, \\ Z_{\mathrm{abs}} &= 2-\varepsilon, \\ \langle s\rangle_{\mathrm{abs}} &= \frac{\varepsilon}{2-\varepsilon}. \end{aligned}$$

Both sampled contributions are order one, while the desired residual is order $\varepsilon$. The difficulty is cancellation, not a small integrand.

Average sign as a free-energy ratio

For a thermal partition function of a Hermitian Hamiltonian, $Z>0$. Define

$$\begin{aligned} F &= -\frac{1}{\beta}\ln Z, \\ F_{\mathrm{abs}} &= -\frac{1}{\beta}\ln Z_{\mathrm{abs}}. \end{aligned}$$

Because $Z_{\mathrm{abs}}\ge Z$,

$$F \ge F_{\mathrm{abs}}.$$

The average phase is exactly

$$\left\langle\phi\right\rangle_{\mathrm{abs}} = \exp \left[ -\beta \left( F-F_{\mathrm{abs}} \right) \right].$$

For a volume $V$, write

$$F-F_{\mathrm{abs}} = V\Delta f_V.$$

Then

$$\left\langle\phi\right\rangle_{\mathrm{abs}} = e^{-\beta V\Delta f_V}.$$

If $\Delta f_V\to\Delta f>0$ as $V\to\infty$, the average phase falls exponentially in the space-time volume $\beta V$. This is not merely an empirical fitting form: $\Delta f_V$ is the finite-volume free-energy-density difference between the physical and absolute-weight ensembles.

The absolute ensemble is often called phase quenched, even in the real-sign case. It is generally a different statistical theory. Its dominant configurations need not resemble the configurations that survive after cancellation.

Why the sampling cost becomes exponential

For independent samples of $\phi$, let

$$\mu_\phi = \langle\phi\rangle_{\mathrm{abs}}.$$

Since $|\phi|=1$ on nonzero-weight configurations,

$$\mathbb E_{\mathrm{abs}} \left[ \left| \overline\phi-\mu_\phi \right|^2 \right] = \frac{ 1-|\mu_\phi|^2 }{ N }.$$

The relative root-mean-square uncertainty is therefore

$$\frac{ \sqrt{ \mathbb E \left| \overline\phi-\mu_\phi \right|^2 } }{ |\mu_\phi| } = \sqrt{ \frac{ 1-|\mu_\phi|^2 }{ N|\mu_\phi|^2 } }.$$

To resolve the denominator to relative precision $\epsilon$ requires approximately

$$N_{\mathrm{eff}} \gtrsim \frac{ 1-|\mu_\phi|^2 }{ \epsilon^2|\mu_\phi|^2 }.$$

When $|\mu_\phi|\ll1$,

$$N_{\mathrm{eff}} \gtrsim \epsilon^{-2} e^{2\beta V\Delta f_V}.$$

For a stationary correlated chain with a summable covariance and an appropriate central-limit regime,

$$N_{\mathrm{eff}} \approx \frac{ N_{\mathrm{meas}} }{ 2\tau_{\mathrm{int},\phi} },$$

where for a complex phase the covariance uses $\operatorname{Re}\langle(\phi_t-\mu_\phi)^*(\phi_{t+k}-\mu_\phi)\rangle$ and is normalized by $1-|\mu_\phi|^2$. Autocorrelation modifies the sampling burden; it does not remove the small mean. These estimates concern resolving the denominator. A special observable, such as $A=1$, can have an exact ratio because its numerator and denominator cancel sample by sample.

Set $x=\beta V\Delta f$ and $\mu=e^{-x}$. The solid line is $-\ln\mu=x$; the dashed curve is the exact independent-sample estimate $\ln(\epsilon^2N)=\ln(e^{2x}-1)$ for relative RMS error $\epsilon$. Its slope approaches two at large $x$.

Ratio uncertainty

Set

$$X = \phi A, \qquad Y = \phi, \qquad R = \frac{\mu_X}{\mu_Y}.$$

To first order,

$$\widehat R-R \approx \frac{ (\overline X-\mu_X) -R(\overline Y-\mu_Y) }{ \mu_Y }.$$

When the denominator is well resolved, the leading large-sample mean-square error for independent samples is

$$\mathbb E \left| \widehat R-R \right|^2 \approx \frac{ \mathbb E_{\mathrm{abs}} \left| X-RY \right|^2 }{ N|\mu_Y|^2 }.$$

The same small denominator appears. For correlated chains, the numerator is replaced by an integrated autocovariance. Blocking, jackknife, or bootstrap analysis must preserve the joint numerator–denominator fluctuations.

The ratio estimator also has finite-sample bias. When $\widehat Y$ frequently approaches or crosses zero, a Gaussian error bar around $\widehat A$ is not trustworthy.

Where the cancellations come from

The sign problem is not synonymous with fermions, and fermions do not imply a sign problem in every formulation. Several mechanisms recur.

Fermion exchange in worldlines

For identical particles in a coordinate representation,

$$Z_N^{(\pm)} = \frac{1}{N!} \sum_P (\pm1)^P \int dR\, \rho(R,PR;\beta).$$

Bosonic exchange sectors add. Fermionic sectors carry the permutation parity $(-1)^P$. At low temperature, long exchange cycles can make even and odd sectors individually large while their difference remains small.

This is one origin of the fermion sign problem in path-integral Monte Carlo. It is not the only origin, and integrating out fermions can move the same cancellation into determinants rather than remove it.

Fermion determinants

After an auxiliary-field transformation, a common weight is

$$W[s] = W_{\mathrm B}[s]\, \det M_\uparrow[s]\, \det M_\downarrow[s].$$

The Grassmann integral producing each determinant is exact. The sampling problem begins when the determinant product is negative or complex across the ordinary auxiliary-field configurations.

At particle-hole-symmetric half filling, the repulsive one-band Hubbard Model on a bipartite lattice has a standard determinant-QMC formulation in which the spin determinants are related so that their product is nonnegative. Doping, frustrating hopping, magnetic flux, spin imbalance, or a different decoupling can destroy that relation.

Frustrated spins

Consider the transverse exchange of an antiferromagnetic Heisenberg model,

$$\begin{aligned} H_\perp &= \frac{J}{2} \sum_{\langle ij\rangle} \left( S_i^+S_j^- +S_i^-S_j^+ \right), \\ J &> 0. \end{aligned}$$

A local $\pi$ rotation about $z$ acts as

$$S_i^\pm \longmapsto \eta_i S_i^\pm, \qquad \eta_i\in\{-1,+1\}.$$

Every transverse bond becomes nonpositive if

$$\eta_i\eta_j = -1$$

on every antiferromagnetic edge. A bipartite graph permits this two-coloring. Around a triangle, however,

$$\begin{aligned} (\eta_1\eta_2) (\eta_2\eta_3) (\eta_3\eta_1) &= +1, \\ (-1)^3 &= -1. \end{aligned}$$

The requirements are inconsistent. The odd cycle carries a gauge-invariant obstruction to this simple sign rotation. This is how geometric frustration can generate signs even in a spin model with no mobile fermions.

Flux, Berry phases, and oscillatory actions

Complex hopping amplitudes may be written

$$t_{ij} = |t_{ij}|e^{iA_{ij}}.$$

Local phase redefinitions move $A_{ij}$ between bonds, but the loop phase

$$\Phi_\gamma = \sum_{(ij)\in\gamma} A_{ij} \pmod{2\pi}$$

is gauge invariant. A nontrivial flux can prevent all matrix elements from being made simultaneously real and nonpositive.

Berry-phase terms, finite-density fermion determinants, and real-time factors $e^{iS/\hbar}$ provide further complex-weight examples. Their detailed analytic structures differ, so one should not assume that a technique useful for one applies to all.

Stoquasticity and its limits

In a specified orthonormal basis $\{|x\rangle\}$, a real Hamiltonian is stoquastic if

$$\langle x|H|y\rangle \le 0 \qquad (x\neq y).$$

For a finite matrix, this ensures that $e^{-\tau H}$ is entrywise nonnegative for every $\tau\ge0$, not merely to first order in a time step. Choose $c\ge\max_x H_{xx}$; then $B=cI-H$ is entrywise nonnegative, and

$$e^{-\tau H}=e^{-c\tau}\sum_{n=0}^{\infty}\frac{\tau^n B^n}{n!}$$

has nonnegative entries. This gives a direct route to nonnegative time-sliced or series weights. Efficient sampling is a separate question.

Three qualifications matter.

1. Stoquasticity is basis dependent.
2. Stoquasticity is sufficient for important QMC constructions, not a universal equivalence between Hamiltonians and all possible algorithms.
3. A sign-free representation can still be computationally difficult because of critical slowing down, topological freezing, poor overlap, ill-conditioned matrices, or hard-to-measure observables.

The quantum annealing discussion applies this distinction to annealing Hamiltonians.

Basis and representation dependence

The spectrum of a Hamiltonian does not determine the sign problem by itself. For example,

$$H = t\sigma^x, \qquad t>0,$$

has positive off-diagonal elements in the $\sigma^z$ basis. With $U=\sigma^z$,

$$UHU^\dagger = -t\sigma^x,$$

which is stoquastic in that basis. The physics is unchanged only if states and observables are transformed with the Hamiltonian.

This toy cure is local and obvious. In a many-body problem, a useful transformation must also be:

• computable without solving the original problem;
• compatible with locality or another tractable structure;
• accompanied by tractable transformed observables;
• effective over the parameter region of interest.

A formally sign-free basis that makes every operator exponentially nonlocal is not an efficient computational cure.

Representation dependence extends beyond basis rotations. Trotter decompositions, operator groupings, Hubbard–Stratonovich channels, fermion bags, loop decompositions, and partial analytic sums can assign different signs to their intermediate configurations while producing the same exact $Z$.

The average sign is therefore not a physical observable of $H$. It is a diagnostic of a specified representation and sampling measure.

Important sign-free structures

A sign-free construction establishes positivity for every configuration in its stated domain.

Structure	Positivity mechanism	Scope warning
unfrustrated bipartite spin model	a Marshall-type sublattice rotation makes transverse matrix elements nonpositive	additional frustrating bonds or phases can break the construction
spin-balanced attractive Hubbard formulation	the two real spin determinants coincide, giving a square	imbalance or altered couplings need separate analysis
repulsive bipartite Hubbard model at symmetric half filling	particle-hole symmetry relates the spin determinants	doping, frustrating hopping, flux, or another decoupling may restore signs
antiunitary determinant symmetry	eigenvalues or determinants occur in complex-conjugate pairs	the antiunitary action and its square must satisfy the theorem’s hypotheses
Majorana or reflection positivity	the fermionic weight factorizes or obeys a positivity cone	only specified interaction channels and decompositions qualify
meron or cluster regrouping	sign-canceling configurations are summed analytically before sampling	exact only for model and observable classes admitting the required cluster structure

Conjugate determinant pairing

A particularly transparent sufficient condition is

$$M_\downarrow[s] = M_\uparrow[s]^*.$$

Then

$$\begin{aligned} \det M_\downarrow[s] &= \left( \det M_\uparrow[s] \right)^*, \\ \det M_\uparrow[s]\, \det M_\downarrow[s] &= \left| \det M_\uparrow[s] \right|^2 \\ &\ge 0. \end{aligned}$$

More general antiunitary symmetries can enforce analogous conjugate pairing without a literal spin-up/spin-down factorization. The symmetry condition is formulation specific; time-reversal invariance of the physical Hamiltonian by itself is not enough unless it acts appropriately on every configuration matrix.

What counts as a cure?

The phrase “solve the sign problem” is too broad unless the method and guarantees are stated.

Exact transformations

An exact basis change, duality, auxiliary-field choice, or analytic regrouping can produce nonnegative weights while preserving the desired observable. It removes cancellation for that representation. An efficient solution additionally requires tractable transformed weights, observables, and sampling; diagonalizing the original Hamiltonian first is formally sign free but does not solve its computational difficulty.

Meron-cluster and fermion-bag ideas can cancel families of signed configurations before stochastic sampling. They are powerful precisely because they exploit special algebraic structure; they are not black-box cures for arbitrary Hamiltonians.

Variance mitigation

A change of representation may increase the average sign without making it one. Better proposals, correlated sampling, control variates, and contour deformations may reduce variance. These can extend the reachable regime, but residual exponential scaling must still be measured.

Reweighting by $W/|W|$ is an exact identity, not a cure. It exposes the cancellation in a denominator.

Biased approximations

Fixed-node, constrained-path, phaseless, or related restrictions replace the uncontrolled cancellation with a condition informed by a trial state or path constraint. They can be extremely useful, but the constraint introduces bias unless an exactness theorem applies.

The result should be described as a constrained approximation and tested against released-constraint calculations, exact small systems, variational bounds when available, and trial-state dependence.

Analytic continuation from sign-free regions

Simulating at imaginary chemical potential, altered couplings, or another sign-free parameter and extrapolating to the target is a model-dependent inference problem. Its reliability depends on analyticity, truncation control, and proximity to singularities. A good fit does not by itself establish an exact cure.

The complexity result, stated carefully

Troyer and Wiese constructed a class of quantum systems for which a generic polynomial-time solution of the associated fermion sign problem would yield a polynomial-time solution of an NP-hard problem. For bounded-error randomized sampling, the relevant consequence is an efficient randomized algorithm for the encoded hard family (often phrased as NP being contained in BPP). Such a generic algorithm is not known and is not generally expected. The theorem is a worst-case obstruction, not a claim about the runtime at every physical parameter point.

This result does not prove that:

• every fermionic model is NP-hard;
• every negative weight is fatal at accessible sizes;
• every nonstoquastic Hamiltonian is hard;
• no special symmetry, duality, or regrouping can remove a particular sign problem;
• quantum Monte Carlo is inferior to every other method on sign-problem instances.

Related work shows that even finding a curing local basis transformation can be NP-complete for specified transformation classes. Again, this is a family-level worst-case statement. It does not replace model-specific analysis.

The useful conclusion is restrained: search energetically for structure in the problem at hand, but do not expect a representation-independent black-box algorithm that removes all sign problems efficiently.

Diagnosing severity in practice

Measure both the average phase and its uncertainty over the sizes and temperatures of interest.

State the ensemble

Report the exact configuration weight $W(C)$, the sampled weight $|W(C)|$ or other proposal measure, and the phase factor used in reweighting. “We used QMC” is not enough.

Measure scaling, not one point

The average phase should be measured over a grid of:

• inverse temperatures $\beta$;
• spatial sizes and aspect ratios;
• densities, couplings, fluxes, and frustration parameters;
• boundary conditions;
• bases or decoupling channels being compared.

If exponential behavior is claimed, analyze

$$-\frac{ 1 }{ \beta V } \ln \left| \langle\phi\rangle_{\mathrm{abs}} \right|.$$

This quantity estimates $\Delta f_V$ only while the average phase is statistically resolved. Taking a logarithm of noise after the mean is consistent with zero creates a biased, meaningless plateau.

Resolve the denominator

An average sign should be quoted with an autocorrelation-aware uncertainty. A useful calculation requires more than a nonzero point estimate:

$$\frac{ |\widehat\mu_\phi| }{ \operatorname{SE}(\widehat\mu_\phi) } \gg 1.$$

The required significance depends on the observable ratio and the desired precision. Five standard errors in the denominator do not guarantee a precise nonlinear observable.

Preserve covariance

Store numerator and denominator in the same blocks. Apply jackknife or bootstrap resampling to the complete ratio, not to separately summarized error bars.

Benchmark the signed theory

At small size or high temperature, compare with exact diagonalization, direct enumeration, a series expansion, or a known sign-free limit. Agreement of the phase-quenched theory with itself is not a validation of the physical theory.

Test a claimed sign-free theorem

If symmetry predicts $W(C)\ge0$ configuration by configuration, inspect actual generated weights. A negative value then signals numerical roundoff, an inconsistent convention, or an implementation error rather than an unavoidable physical sign problem.

Exercises

Exercise 2: Free-energy cost

A phase-quenched simulation has

$$\beta V\Delta f_V = 5,$$

and $\tau_{\mathrm{int},\phi}=3$. Estimate the measurements needed to resolve the average phase to $10\%$ relative precision, using the small-phase approximation.

Solution

The average phase is

$$|\mu_\phi| = e^{-5}.$$

With $\epsilon=0.1$,

$$\begin{aligned} N_{\mathrm{eff}} &\gtrsim \epsilon^{-2} e^{10} \\ &\approx 100 \times 2.20\times10^4 \\ &\approx 2.2\times10^6. \end{aligned}$$

Autocorrelation gives

$$\begin{aligned} N_{\mathrm{meas}} &\approx 2\tau_{\mathrm{int},\phi} N_{\mathrm{eff}} \\ &\gtrsim 6 \times 2.2\times10^6 \\ &\approx 1.3\times10^7. \end{aligned}$$

This only resolves the denominator. A noisy observable numerator can require more samples.

Exercise 4: Determinant pairing

Suppose an auxiliary-field formulation obeys

$$M_2[s] = U M_1[s]^* U^{-1}$$

for every configuration $s$, with invertible $U$. Show that

$$\det M_1[s]\, \det M_2[s] \ge 0.$$

Solution

Similarity leaves the determinant unchanged:

$$\begin{aligned} \det M_2 &= \det \left( U M_1^* U^{-1} \right) \\ &= \det M_1^* \\ &= (\det M_1)^*. \end{aligned}$$

Hence

$$\begin{aligned} \det M_1\det M_2 &= |\det M_1|^2 \\ &\ge 0. \end{aligned}$$

The proof must hold configuration by configuration. Pairing only after averaging would not define a nonnegative sampling weight.

Exercise 5: Covariance in the reweighted ratio

Let $\widehat R=\overline X/\overline Y$ and $R=\mu_X/\mu_Y$. Derive the first-order fluctuation of $\widehat R$ and explain why strongly correlated numerator and denominator noise can reduce the ratio variance.

Solution

Write

$$\overline X = \mu_X+\delta X, \qquad \overline Y = \mu_Y+\delta Y.$$

Expanding the reciprocal,

$$\frac{1}{\mu_Y+\delta Y} \approx \frac{1}{\mu_Y} \left( 1-\frac{\delta Y}{\mu_Y} \right).$$

Therefore

$$\begin{aligned} \widehat R &\approx \frac{\mu_X}{\mu_Y} + \frac{\delta X}{\mu_Y} - \frac{\mu_X\delta Y}{\mu_Y^2} \\ &= R + \frac{ \delta X-R\delta Y }{ \mu_Y }. \end{aligned}$$

The variance depends on the combination $\delta X-R\delta Y$, which contains

$$-2\operatorname{Re} \left[ R^* \operatorname{Cov}(\delta X,\delta Y) \right].$$

Common fluctuations can cancel in the ratio. Separate error propagation discards this covariance and can overestimate or underestimate the uncertainty.

Exercise 6: Read the complexity claim

Classify each statement as supported or unsupported by the generic NP-hardness result.

1. No sign-problem-free formulation exists for any interacting fermion model.
2. A universal polynomial-time cure for the constructed family would have unlikely complexity consequences.
3. A particular doped Hubbard cluster is impossible to simulate accurately.
4. Special symmetries can remove signs without contradicting the worst-case theorem.

Solution

Statement 2 is supported: it captures the family-level worst-case result.

Statement 4 is also supported. A worst-case hardness theorem permits easy subclasses, special parameter points, and model-specific transformations.

Statements 1 and 3 are unsupported. Sign-free interacting fermion formulations are known, and a worst-case theorem does not decide the practical cost of one finite instance.

References

• D. M. Ceperley, “Path Integrals in the Theory of Condensed Helium,” Reviews of Modern Physics 67, 279–355 (1995). doi:10.1103/RevModPhys.67.279
• J. E. Hirsch, “Two-Dimensional Hubbard Model: Numerical Simulation Study,” Physical Review B 31, 4403–4419 (1985). doi:10.1103/PhysRevB.31.4403
• V. I. Iglovikov, E. Khatami, and R. T. Scalettar, “Geometry Dependence of the Sign Problem in Quantum Monte Carlo Simulations,” Physical Review B 92, 045110 (2015). doi:10.1103/PhysRevB.92.045110
• E. Y. Loh Jr., J. E. Gubernatis, R. T. Scalettar, S. R. White, D. J. Scalapino, and R. L. Sugar, “Sign Problem in the Numerical Simulation of Many-Electron Systems,” Physical Review B 41, 9301–9307 (1990). doi:10.1103/PhysRevB.41.9301
• Z.-X. Li, Y.-F. Jiang, and H. Yao, “Solving the Fermion Sign Problem in Quantum Monte Carlo Simulations by Majorana Representation,” Physical Review B 91, 241117(R) (2015). doi:10.1103/PhysRevB.91.241117
• Z.-X. Li, Y.-F. Jiang, and H. Yao, “Majorana-Time-Reversal Symmetries: A Fundamental Principle for Sign-Problem-Free Quantum Monte Carlo Simulations,” Physical Review Letters 117, 267002 (2016). doi:10.1103/PhysRevLett.117.267002
• M. Marvian, D. A. Lidar, and I. Hen, “On the Computational Complexity of Curing Non-Stoquastic Hamiltonians,” Nature Communications 10, 1571 (2019). doi:10.1038/s41467-019-09501-6
• P. Henelius and A. W. Sandvik, “Sign Problem in Monte Carlo Simulations of Frustrated Quantum Spin Systems,” Physical Review B 62, 1102–1113 (2000). doi:10.1103/PhysRevB.62.1102
• G. G. Batrouni and R. T. Scalettar, “Anomalous Decouplings and the Fermion Sign Problem,” Physical Review B 42, 2282–2289 (1990). doi:10.1103/PhysRevB.42.2282
• N. Hatano and M. Suzuki, “Representation Basis in Quantum Monte Carlo Calculations and the Negative-Sign Problem,” Physics Letters A 163, 246–249 (1992). doi:10.1016/0375-9601(92)91006-D
• L. Wang, Y.-H. Liu, M. Iazzi, M. Troyer, and G. Harcos, “Split Orthogonal Group: A Guiding Principle for Sign-Problem-Free Fermionic Simulations,” Physical Review Letters 115, 250601 (2015). doi:10.1103/PhysRevLett.115.250601
• Z.-C. Wei, C. Wu, Y. Li, S. Zhang, and T. Xiang, “Majorana Positivity and the Fermion Sign Problem of Quantum Monte Carlo Simulations,” Physical Review Letters 116, 250601 (2016). doi:10.1103/PhysRevLett.116.250601
• S. Chandrasekharan and U.-J. Wiese, “Meron-Cluster Solution of Fermion Sign Problems,” Physical Review Letters 83, 3116–3119 (1999). doi:10.1103/PhysRevLett.83.3116
• M. Troyer and U.-J. Wiese, “Computational Complexity and Fundamental Limitations to Fermionic Quantum Monte Carlo Simulations,” Physical Review Letters 94, 170201 (2005). doi:10.1103/PhysRevLett.94.170201
• Z.-X. Li and H. Yao, “Sign-Problem-Free Fermionic Quantum Monte Carlo: Developments and Applications,” Annual Review of Condensed Matter Physics 10, 337–356 (2019). doi:10.1146/annurev-conmatphys-033117-054307
• C. Wu and S.-C. Zhang, “Sufficient Condition for Absence of the Sign Problem in the Fermionic Quantum Monte Carlo Algorithm,” Physical Review B 71, 155115 (2005). doi:10.1103/PhysRevB.71.155115