Finite Density and Sign Problems

Finite density changes a Euclidean lattice measure before it changes any Monte Carlo algorithm: a chemical potential weights sectors by conserved charge, generally making the fermion determinant complex. This chapter separates that algebraic fact from three different computational obstacles—phase cancellation, poor overlap, and method-specific bias—and then asks what each proposed method must prove before its output supports a physical claim. There is no representation-independent severity number and no universal cure; there are exact reformulations and controlled windows for particular theories, observables, parameter ranges, and regulators.

The chapter uses determinant symmetries from fermion determinants, Pfaffians, and measure positivity and the thermal path-integral language of Euclidean correlators and Schwinger functions. For the computational sections, estimators, covariance, and resampling reviews ratio-estimator diagnostics, while branches, sheets, continuation, and monodromy reviews analytic continuation.

Chapter convention card. All identities are first stated at finite lattice spacing and finite volume. Real chemical potential is $μ$, $β_T=1/T$, and $V_s$ is spatial volume; $β_0$ is reserved for the coupling in the one-angle fixture. Every method must preserve the same action, charge normalization, boundary conditions, and normalized observable before its results are compared.

Enter finite density through the measure

For a conserved charge $Q$,

$$Z(\beta,\mu)=\operatorname{Tr}e^{-\beta(H-\mu Q)} =\sum_Q e^{\beta\mu Q}Z_Q(\beta).$$

On a Euclidean lattice, the fugacity is implemented by asymmetric temporal hopping factors $e^{\pm a\mu q}$ Hasenfratz and Karsch 1983. The resulting Dirac operator obeys

$$D(\mu)^\dagger=\gamma_5D(-\mu^*)\gamma_5, \qquad \det D(\mu)^*=\det D(-\mu^*).$$

Thus a real baryon chemical potential does not generally leave a real determinant, whereas imaginary chemical potential or paired opposite chemical potentials can restore a nonnegative measure under additional flavor assumptions. The chemical-potential page derives these statements and fixes the forward/backward-link convention.

For a complex weight $w(x)=|w(x)|e^{i\phi(x)}$, phase reweighting gives

$$\langle O\rangle =\frac{\langle Oe^{i\phi}\rangle_{\rm pq}} {\langle e^{i\phi}\rangle_{\rm pq}}, \qquad \langle e^{i\phi}\rangle_{\rm pq} =\frac{Z}{Z_{\rm pq}}=e^{-\beta V_s\Delta f}.$$

The final equality defines the full–phase-quenched free-energy-density difference $\Delta f$ at finite spatial volume $V_s$. A small denominator creates exponentially poor signal-to-noise for direct phase reweighting, but it does not by itself locate the rare configurations that dominate a particular observable. Nor does a stable numerical trajectory prove that a complexified stochastic process has the desired stationary expectation. The anatomy page makes those distinctions quantitative.

Choose a method by its correctness contract

The nine routes form a decision sequence rather than a ranking.

1. Use chemical potential on the Euclidean lattice to identify the charge, temporal-link prescription, determinant symmetry, and sign-free exceptions.
2. Use anatomy and severity to measure average phase, relative variance, overlap, volume scaling, and observable dependence separately.
3. Use basis dependence and complexity to decide whether a change of variables removes negative weights without introducing a nonlocal action, hard observable, or exponentially costly transformation. The standard NP-hardness result is a worst-case reduction with explicit hypotheses, not a statement about every instance Troyer and Wiese 2005.
4. In a controlled analytic neighborhood, compare reweighting, Taylor expansion, and imaginary density. Their shared endpoint is the same target observable, but their failure diagnostics differ; the original imaginary-density construction explicitly restricted continuation to an analytic interval de Forcrand and Philipsen 2002.
5. When an exact expansion exists, test dual variables, worldlines, fluxes, and tensors for positivity, complete constraints, boundary sectors, and an exact observable map.
6. Treat complex Langevin as correct only when the integration-by-parts argument has controlled tails and singularities, the process is ergodic, and exact or sign-free benchmarks agree.
7. Treat Lefschetz thimbles and holomorphic flow as contour identities only after homology, all contributing saddle sectors, Jacobians, residual phases, and multimodal sampling are controlled.
8. Test canonical, fugacity, and density-of-states reconstructions for normalization, Fourier resolution, dynamic range, omitted tails, and cancellation-amplified uncertainty.
9. Apply the cross-method reliability standards before combining results. Agreement is informative only when the methods do not share the same hidden approximation and the comparison lies inside each method’s demonstrated domain.

One exact fixture, several failure modes

Every method can be checked on the finite integral

$$Z_1(\mu)=\int_{-\pi}^{\pi}d\theta\,e^{\beta_0\cos\theta} \bigl(1+he^{\mu+i\theta}\bigr) \bigl(1+he^{-\mu-i\theta}\bigr),$$

with $\beta_0=1$ and $h=0.2$. Fourier orthogonality gives the analytic answer

$$Z_1(\mu)=2\pi\left[(1+h^2)I_0(\beta_0) +2h\cosh\mu\,I_1(\beta_0)\right].$$

Independent-copy volumes $Z_V=Z_1^V$, with $V=1,4,16,64$, isolate exponential phase severity without introducing a thermodynamic interaction. Real $\mu=0,0.5,1,1.5$ probes complex weights; imaginary $\mu=i\mu_I$ at $\mu_I=0,\pi/4,\pi/2,3\pi/4$ probes analytic continuation and Fourier reconstruction. The model is an exact numerical fixture, not continuum QCD and not evidence that a method scales to an interacting field theory.

Reliability ceiling

A finite-density result should report the target observable and regulator, the sampled or transformed measure, a severity and overlap diagnostic, the method’s correctness conditions, an exact or sign-free benchmark, a deliberate failure witness, volume and discretization behavior, and every residual extrapolation or truncation. A single successful parameter point establishes only that point. A smooth curve, small optimizer residual, stationary Langevin history, or stable flowed ensemble is not a substitute for correctness.

The two chapter figures summarize phase severity and the method-specific failure branches. The semantic comparison table on the reliability page is the compact record to use when deciding how far a claim may extend. Dated assessments of algorithms at the research frontier belong on the corresponding Research pages; this overview states durable conditions rather than a current winner.

Learning outcomes

After completing the chapter, you should be able to:

• Given a finite-density lattice measure, compute its determinant symmetry, average-phase estimator, leading sample-cost scaling, and at least one independent overlap diagnostic.
• Given a proposed result from any method in the chapter, write a method-specific validation matrix containing an exact fixture, a negative control, volume and regulator tests, residual uncertainties, and a defensible claim boundary.

Review the chapter

A complex-Langevin estimate and a canonical reconstruction agree for the density at $\mu=1.5$. The Langevin drift has a power-law tail, while the canonical calculation used four Fourier points for independent-copy volume $V=16$. What conclusion is licensed?

Solution

Neither method is validated at that point. The Langevin tail violates the decay needed for its correctness argument, and four Fourier points cannot resolve the $2V+1=33$ canonical modes without an additional exact sparsity theorem. Their agreement could be coincidental or share a bias. The defensible result stops at the largest connected parameter region where both methods pass their own diagnostics and an exact or sign-free comparison.

Chapter handoff

The next chapter, Hamiltonian lattice field theory, replaces Euclidean importance sampling by regulated Hamiltonians, Hilbert spaces, transfer matrices, and real-time evolution. The sign problem can change form under that move, but truncation, state-preparation, and dynamical-cost questions replace rather than erase the need for controlled inference.

References

• de Forcrand, Philippe, and Owe Philipsen. “The QCD Phase Diagram for Small Densities from Imaginary Chemical Potential.” Nuclear Physics B 642 (2002): 290–306. doi:10.1016/S0550-3213(02)00626-0.
• Hasenfratz, Peter, and Frithjof Karsch. “Chemical Potential on the Lattice.” Physics Letters B 125 (1983): 308–310. doi:10.1016/0370-2693(83)91290-X.
• Troyer, Matthias, and Uwe-Jens Wiese. “Computational Complexity and Fundamental Limitations to Fermionic Quantum Monte Carlo Simulations.” Physical Review Letters 94 (2005): 170201. doi:10.1103/PhysRevLett.94.170201.