The Blume-Capel Model as a $\phi^6$ Field Theory: A Monte Carlo and Flow Model Study

I. INTRODUCTION

• The spin-1 Blume–Capel model has continuous and first-order transition lines meeting at a tricritical point, while scalar ϕ6 theory describes these regimes through quartic and sixth-order couplings.A negative quartic coupling permits a first-order transition stabilized by the sixth-order term; in two dimensions, scalar ϕ6 studies confirm tricritical-Ising behavior.Source: §I · p. 1
• Normalizing flows map an easy-to-sample distribution through invertible transformations to represent correlated field distributions and generate independent configurations.Such configurations can be used directly with importance weights or as Metropolis proposals, reducing autocorrelation associated with local or molecular-dynamics-based Markov chains.Source: §I · p. 1
• First-order transitions challenge flow models because competing phases create multimodal distributions separated by exponentially suppressed interfacial regions, raising mode-collapse and rare-event-sampling concerns.Earlier scalar-field flow benchmarks had mainly focused on ϕ4 theories near continuous critical points.Source: §I · p. 1
• A two-dimensional ϕ6 model interpolates between a soft scalar field and the three-state Blume–Capel structure, enabling study of phase-boundary and tricritical-point evolution as fields become more discrete.The introduction also identifies an entropic T ln 2 shift arising from unequal vacancy-well and occupied-well curvatures in the soft-to-discrete connection.Source: §I · p. 1
• The study compares flow samplers with cluster and extended-ensemble Monte Carlo across continuous, tricritical, and first-order regimes in one two-dimensional ϕ6 phase diagram.The sampling challenges range from critical slowing down to tricritical fluctuations and phase coexistence with free-energy barriers.Source: §I · p. 1

II. SOFT-SPIN ISING AND BLUME–CAPEL MODELS

• The ϕ4 analogue remains continuous as it interpolates toward the hard-spin Ising limit, making it a control case.It separates effects of relaxing discrete spin amplitudes from the tricritical and first-order behavior introduced by ϕ6 theory.Source: §II · p. 2
• The section uses the ϕ4 model to connect the continuous-field critical temperature clearly to its Ising progenitor..Source: §II · p. 2
• The comparison is designed to isolate effects associated simply with softening the local spin amplitude.The transition remains continuous throughout the interpolation to the hard-spin Ising limit.Source: §II · p. 2

A. The Soft-Spin Ising Model: ϕ4

• The continuous model uses a scalar field on an L × L square lattice, coupled between near-neighbor sites by exchange J>0, with λ controlling a potential minimized at ϕ=±1.The continuous field ranges from negative to positive infinity.Source: A. The Soft-Spin Ising Model: ϕ4 · p. 2
• Reporting Tc/⟨ϕ²⟩ removes the trivial critical-temperature scaling associated with changing the spin length.In mean-field theory on the square lattice, the normalized temperature is 4J, independent of λ.Source: A. The Soft-Spin Ising Model: ϕ4 · p. 2
• Binder-parameter crossings across lattice sizes L=8, 12, and 16 are used to determine the continuous-spin model’s critical temperature.The representative figure reports Tc/J=3.641 for λ/J=0.5.Source: A. The Soft-Spin Ising Model: ϕ4 · p. 2, Figure 1 · p. 2
• The normalized critical temperature is steadily suppressed as λ decreases, reflecting increased entropy of the continuous ϕ field after removing the trivial spin-length effect.The potential softens as λ decreases.Source: A. The Soft-Spin Ising Model: ϕ4 · p. 2

B. The Soft-Spin Blume–Capel Model: ϕ6

• The Blume–Capel model adds vacancies S_i = 0 to Ising spins S_i = ±1, and its continuous counterpart uses a ϕ6 potential favoring ϕ_i = 0, ±1.The parameter α controls that preference, while Vδ connects the continuous model to the discrete one.Source: B. The Soft-Spin Blume–Capel Model: ϕ6 · pp. 2–3
• The discrete model’s transition changes from second order to first order at the tricritical point (T_t/J, ∆_t/J) = (0.609, 1.966).In the hard-well limit, the corresponding δ_t/J is approximately 1.544.Source: B. The Soft-Spin Blume–Capel Model: ϕ6 · p. 3
• The soft model approaches the discrete Blume–Capel model with ∆ = δ + T ln 2, because the vacancy well has lower curvature than the occupied-spin wells.The resulting parameter-free hard-well relation is δ = ∆ − T ln 2.Source: B. The Soft-Spin Blume–Capel Model: ϕ6 · p. 3
• Figure 3 shows that increasing δ favors vacancies, while increasing α confines the field toward 0 and ±1; the softer curvature at ϕ = 0 blurs the discrete-model mapping.The vacancy-well curvature is 2α, compared with 8α at ϕ = ±1.Source: Figure 3 · p. 3, B. The Soft-Spin Blume–Capel Model: ϕ6 · p. 3

A. ‘Advanced’ Markov Chain Monte Carlo

• The simulations combine single-site Metropolis–Hastings proposals from a three-component Gaussian mixture with Wolff updates of embedded Ising signs.Updates are vectorized across each lattice sublattice.Source: A. ‘Advanced’ Markov Chain Monte Carlo · pp. 3–4
• Wolff updates become less effective near tricriticality and cannot overcome the phase-coexistence barrier near the first-order transition.The cluster update’s dynamical critical exponent near tricriticality degrades to the local-update value.Source: A. ‘Advanced’ Markov Chain Monte Carlo · p. 4
• Replica exchange swaps in δ support sampling, while umbrella sampling and weighted-histogram analysis provide a cross-check near the tricritical point from the first-order side.The umbrella-sampling histograms are reweighted to locate equal-weight peaks for the finite-size critical field.Source: A. ‘Advanced’ Markov Chain Monte Carlo · p. 4
• The finite-size tricritical point is located from zero skewness of the density histogram, followed by Binder-cumulant scaling along that curve.Fits use the second thermal exponent y_g = 4/5 of the tricritical Ising universality class.Source: A. ‘Advanced’ Markov Chain Monte Carlo · p. 4
• The critical δ/J at low temperatures is independent of T/J, and the ordered phase extends to higher T as α decreases.The phase boundaries for several α values and the Blume–Capel case are shown in Figure 4.Source: A. ‘Advanced’ Markov Chain Monte Carlo · pp. 4–5, Figure 4 · p. 4
• As α changes by three orders of magnitude, ∆_t/J varies by only a few percent, while T_t/J increases by almost a factor of two as spins harden.The tricritical point converges toward the discrete Blume–Capel limit as α → ∞.Source: Figure 5 · p. 4, A. ‘Advanced’ Markov Chain Monte Carlo · p. 5

B. Flow Model

• The model maps a tractable base density to the ϕ6 target with invertible coupling layers, spline conditioners, and a coarse-to-fine architecture.Its symmetries are imposed during training, which minimizes reverse KL divergence.Source: B. Flow Model · p. 5
• Flow pairs give consistent first-order transition estimates, while MCMC umbrella-sampler fluctuations produce a less precise fit.The analysis fits the finite-size transition location after reweighting in δ.Source: B. Flow Model · p. 5, Figure 7 · p. 6
• On the second-order side, flow and MCMC estimates of the critical temperature show remarkably good agreement.The comparison uses the same system sizes and fitting approach for both methods.Source: B. Flow Model · p. 6
• Tricritical flow analysis is limited by compute time to α = 8, 16, 32 and sizes L = 12, 16, 20.For these data, the study fits the density Binder cumulant along the reweighted zero-skewness curve, sharing U*ρ and r across the three α values.Source: B. Flow Model · p. 6

CONCLUSIONS

• The study computes finite-α phase diagrams and tracks how the tricritical point moves as the soft-spin model approaches the spin-1 limit.The soft-spin construction makes the connection to Blume–Capel physics explicit through the vacancy-potential mapping and well curvatures.Source: CONCLUSIONS · p. 6
• The flow benchmark tests how performance evolves from Ising criticality through tricriticality to first-order transitions.The conclusions frame this as extending prior first-order flow studies to behavior across a tricritical point.Source: CONCLUSIONS · p. 6

Appendix A: Mean Field Theory

• The ϕ4 mean-field critical condition is an instability of the zero-field state set by the zero-field spin variance.Expanding the self-consistency equation about m = 0 yields the criterion that fixes Tc.Source: §1 · p. 10
• Tc/⟨ϕ²⟩Tc = 4J for the square-lattice ϕ4 model at every λ, although the unnormalized Tc varies with λ.The mean-field Tc diverges as λ → 0 and approaches 4J as λ → ∞, with a shallow intermediate minimum also seen in Monte Carlo.Source: §1 · p. 10, Figure A1 · p. 11
• The soft-spin Blume–Capel model has Tc/⟨ϕ²⟩Tc = 4J on the square lattice whenever its transition is second order.The normalized critical temperature depends on α only through the spin length.Source: §2 · p. 11
• (Tt/J, δt/J) = (1.333, 0.924) for the discrete mean-field point after the T ln 2 shift, matching the α = 2048 prediction.The soft-spin tricritical positions are tabulated across α; the table caption identifies them as mean-field predictions.Source: §2 · p. 11, Table A1 · p. 11

1. Base Density

• The base density is a three-component multidimensional Gaussian mixture to encode the field theory’s three-well modality.The ordered component’s mean is initialized by minimizing reverse KL against the action; its resulting spectrum is replaced by a free-field form.Source: §1 · p. 11
• Lattice symmetries constrain the base covariance spectra through Fourier-space diagonal parameterization over unique D4 momentum orbits.Unless otherwise specified, the mixture weight πc is fixed at 1/2.Source: §1 · p. 11

2. Coupling Layers

• Each coupling layer transforms one lattice sublattice conditioned on the other, making the update invertible.Alternating the checkerboard roles in paired layers updates all sites; a sequence of blocks forms the full flow.Source: §2 · p. 12
• Rational quadratic splines parameterized by a convolutional network implement the coupling transformations.Adding coupling blocks permits progressively richer transformations of the base distribution.Source: §2 · p. 12

3. Multiscale Architecture

• An orthonormal Haar transform decomposes each 2 × 2 block into one block-mean and three detail coefficients.The L/2 × L/2 block-means pass through coupling layers first.Source: §3 · p. 12
• The detail coefficients are conditionally rescaled and shifted before an inverse Haar transform recombines them with the transformed block-means.The conditional transform uses a standard RealNVP-style affine transform.Source: §3 · p. 12
• Unlike up-sampling from a coarse lattice with added noise, the model begins from an L × L base whose 2 × 2 sites already have some correlation.Further coupling layers refine the full-resolution field after the inverse Haar transform.Source: §3 · p. 12

• The flow is trained by minimizing reverse KL to the target density, with the target partition function dropped because it is parameter-independent.Source: §4 · p. 12
• Minimizing this loss is equivalent to minimizing the free energy of the flow-defined distribution.Source: §4 · p. 12
• Reverse-KL training avoids target configurations but can collapse onto only one mode, including one symmetry-broken or coexisting phase.Source: §4 · p. 12

• The target symmetries are Z2, square-lattice D4, and translation; checkerboard splitting can break translation symmetry between sublattices.Source: §5 · p. 13, Figure B1 · p. 13
• For Z2 symmetry, conditioner outputs are symmetrized by reversing spline parameters so the active-branch transformation is odd.Source: §5 · p. 13
• D4-equivariant kernels store one value per orbit, using three numbers per kernel per channel and roughly halving iteration time versus manual kernel averaging.Source: §5 · p. 13
• For coupling-only models, a one-site translation mixture restores translation invariance lost through the checkerboard split.Source: §5 · p. 13
• Multiscale blocking leaves invariance under T′ ⋊ D2, so full symmetrization averages over the surviving translations and point-group subgroup.Source: §5 · pp. 13–14
• Symmetrized densities are used for evaluation and reweighting, not training, because most quality gains came at lower training cost.Source: §5 · p. 14

6. Mixtures of Walled Flows

• At first-order coexistence, density ρ is multimodal, so two separately trained flows are restricted to the ordered and disordered phases with one-sided harmonic walls.The disordered-side stiffness is kd = 50 and the ordered-side stiffness scales as ko = 9N.Source: §6 · p. 14
• Near tricriticality, where density peaks merge, a third flow models intermediate densities using boundaries ρl = 0.25 and ρr = 0.65.Source: §6 · p. 14
• Independently trained flows are combined only at evaluation: first-order pairs use equal weights, while triples use weights 0.45, 0.45, and 0.1.Source: §6 · p. 14
• Unlike partitioning flows [24], this method uses one-sided harmonic walls rather than exponential walls and leaves more sophisticated histogram stitching for future work.Source: §6 · p. 14

7. Optimization

• Optimization uses natural-gradient descent with a trust region and a path-gradient estimator; the information metric is built from per-sample log-density scores.Source: §7 · p. 14
• The damped metric system is solved with the Woodbury identity, and the update size is limited by a KL-change radius κ = 3 × 10−4L2.Source: §7 · p. 14

8. Estimating Physical Quantities

• Reweighted flow samples estimate target expectations exactly when the flow’s support contains the target distribution’s support.Source: §8 · pp. 14–15