RFBEGM: Random Field Blume-Emery-Griffiths Model

• RFBEGM is a disordered spin-1 model featuring ternary states that couple magnetic order and occupancy through quenched random fields, crystal fields, and bilinear/biquadratic interactions.
• The model dynamically incorporates the 0-state, leading to rich metastability, non-abelian relaxation, and distinct avalanche behaviors compared to the RFIM.
• RFBEGM frameworks extend to sparse associative-memory and random crystal field models, elucidating phase transitions, tricritical points, and disorder-induced critical phenomena.

The Random Field Blume-Emery-Griffiths model (RFBEGM) is a disordered spin-1 extension of the Blume-Emery-Griffiths framework in which the local environment is quenched and spatially heterogeneous, while the microscopic variables remain ternary, typically $s_i\in\{-1,0,+1\}$. In the strict sense represented by the fully connected random-field formulation, the disorder enters through quenched local magnetic fields $h_i$, a crystal-field term $\Delta$, and bilinear and biquadratic couplings that jointly control magnetization and occupancy sectors (E et al., 16 Jul 2025). In a broader literature, RFBEGM-like behavior also includes random crystal-field BEG models on sparse graphs, quenched randomness in local state availability, and random-field-type effective descriptions arising in sparse associative-memory dynamics (Jr. et al., 2023). Across these variants, the defining structural feature is the coupling of magnetic order, quadrupolar or occupancy order, and quenched disorder within a spin-1 state space.

1. Canonical model and conceptual scope

In the fully connected formulation, the RFBEGM Hamiltonian is

$$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$

with $s_i\in\{-1,0,+1\}$, $J>0$, $K$ the biquadratic coupling, $H$ a uniform external field, $h_i$ quenched Gaussian random fields of mean $0$ and variance $h_i$0, and $h_i$1 a crystal field controlling the energetic cost of the $h_i$2 state (E et al., 16 Jul 2025). The limit $h_i$3 recovers the RFIM, which provides the principal reference point for identifying what is specific to RFBEGM dynamics.

The spin-1 structure introduces a second local sector beyond ordinary magnetization: the occupancy or quadrupolar variable $h_i$4. This makes the model sensitive not only to sign alignment but also to the presence or absence of active spins. The sign sector is governed by bilinear ferromagnetic alignment, while the occupancy sector is governed by the biquadratic interaction and the crystal field. A central consequence is that the $h_i$5-state may become dynamically relevant, which is impossible in the RFIM and underlies much of the distinct RFBEGM phenomenology (E et al., 16 Jul 2025).

Several adjacent models in the literature are not RFBEGM in the strict random-field sense but are directly relevant to RFBEGM-type physics. On random graphs, the BEG Hamiltonian with a site-dependent random crystal field,

$h_i$6

with $h_i$7 drawn from a bimodal distribution, is an explicit random-crystal-field realization of disordered BEG physics (Jr. et al., 2023). A different mean-field quenched-disorder generalization randomizes the local state space itself by making some sites pure Ising and others full BEG sites, which the authors explicitly relate to the broader RFBEGM class because the local thermodynamics becomes site-dependent and frozen in advance (Schreiber et al., 2022). By contrast, hierarchical-lattice and spin-glass variants with bond disorder rather than site disorder are best regarded as analogs rather than direct RFBEGM realizations (Antenucci et al., 2013).

2. Local fields, metastability, and zero-temperature dynamics

For the quasi-statically driven fully connected RFBEGM, zero-temperature dynamics is implemented as single-spin-flip Glauber evolution under infinitesimal increase of the uniform field $h_i$8. The paper derives two local fields,

$h_i$9

and

$\Delta$0

which jointly determine the preferred state at site $\Delta$1 (E et al., 16 Jul 2025).

When $\Delta$2, the stable state is selected by a three-branch rule: $\Delta$3

$\Delta$4

$\Delta$5

When $\Delta$6, the intermediate state becomes irrelevant and the site reduces effectively to Ising-like behavior,

$\Delta$7

This partition makes explicit that the RFBEGM is not merely an RFIM with an extra state: the $\Delta$8-state appears or disappears dynamically depending on the occupancy field $\Delta$9 (E et al., 16 Jul 2025).

The same “two-field” structure reappears in the sparse BEG associative-memory model, where retrieval is governed by a signed local field

$$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$0

and an activity field

$$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$1

The zero-temperature retrieval rule depends jointly on $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$2 and $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$3, so that the local update is determined by a random field term plus an activity-dependent threshold term (Heusel et al., 2017). This is not the same Hamiltonian as the driven RFBEGM, but it exhibits what the paper describes as random-field-type sparse neural behavior: the spin sign and the occupancy variable are acted on by distinct quenched couplings.

A further common feature across RFBEGM-like models is metastability. Under slow monotone driving, the system relaxes through metastable states separated by avalanches (E et al., 16 Jul 2025). On random graphs, the replica-symmetric cavity description similarly replaces a single mean field by a distribution $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$4 of effective magnetic and quadrupolar fields, with $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$5 the local magnetic field and $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$6 the local crystal-field-like field (Jr. et al., 2023). This suggests a unifying interpretation: RFBEGM-type systems are most naturally described by coupled random effective fields acting separately on $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$7 and $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$8.

3. No-passing, non-abelianity, and avalanche structure

A central dynamical issue in the driven RFBEGM is the status of the no-passing rule (NPP). In the RFIM, if two configurations satisfy a sitewise ordering $$\mathcal{H}[s] = -\frac{J}{2N}\left(\sum_i s_i\right)^2 -\frac{K}{2N}\left(\sum_i s_i^2\right)^2 -\sum_i (H+h_i)s_i +\Delta \sum_i s_i^2,$$9, monotone driving preserves that ordering. This implies deterministic relaxation, uniqueness of the final metastable state from a given initial condition, and abelian avalanches (E et al., 16 Jul 2025). The RFBEGM generically violates this structure.

The violation is tied to the coexistence of the $s_i\in\{-1,0,+1\}$0-state and the biquadratic interaction. The paper identifies the regime

$s_i\in\{-1,0,+1\}$1

more specifically

$s_i\in\{-1,0,+1\}$2

as the frustrated regime in which no-passing is broken (E et al., 16 Jul 2025). In this regime, transitions such as

$s_i\in\{-1,0,+1\}$3

can occur while $s_i\in\{-1,0,+1\}$4 is increased, contradicting the monotone ordering logic familiar from the RFIM. The mechanism is that, for repulsive biquadratic coupling $s_i\in\{-1,0,+1\}$5, the change induced in $s_i\in\{-1,0,+1\}$6 by a flip can outweigh the monotone contribution carried by $s_i\in\{-1,0,+1\}$7. The resulting dynamics is non-abelian: the final metastable state can depend on update order.

The paper organizes the $s_i\in\{-1,0,+1\}$8 plane into several regimes. When $s_i\in\{-1,0,+1\}$9 or $J>0$0, the system behaves essentially like the RFIM and NPP holds. For $J>0$1 with $J>0$2, NPP can still be violated, but the dynamics is effectively decoupled into $J>0$3 and then $J>0$4 transitions, without the same frustrated signatures. The most distinctive regime is the frustrated sector $J>0$5, where the repulsive biquadratic term competes directly with the bilinear ferromagnetic coupling (E et al., 16 Jul 2025).

The most explicit dynamical diagnostic found in this regime is a discontinuity in the distribution of field increments between avalanches. If $J>0$6 is the minimum field increment required to trigger the next instability after an avalanche, then in the frustrated no-passing-violation regime the distribution develops a gap at

$J>0$7

By contrast, in the NPP or RFIM-like regime, $J>0$8 is flat near small increments and decays smoothly. In all regimes the scaling law

$J>0$9

holds, so the essential distinction is not the $K$0 scaling itself but the finite discontinuity in the scaled distribution (E et al., 16 Jul 2025).

Avalanche statistics reinforce this picture. In the frustrated regime, the fraction of avalanches of size $K$1 is significantly enhanced, consistent with the idea that frustration blocks collective rearrangements and fragments avalanches into isolated flips. At the same time, the integrated avalanche size distribution at critical disorder remains RFIM-like,

$K$2

so the microscopic non-abelianity and the waiting-increment gap do not alter the mean-field critical avalanche exponent reported in the paper (E et al., 16 Jul 2025).

4. Sparse associative-memory realization and RFBEGM-like local fields

A distinct but structurally related realization appears in the sparse BEG associative-memory model. Here the network has $K$3 neurons with ternary states

$K$4

and stores sparse ternary patterns $K$5 with

$K$6

The storage load is

$K$7

The model employs two quenched couplings,

$K$8

which generate the signed field $K$9 and the activity field $H$0 (Heusel et al., 2017).

At zero temperature, the original asynchronous retrieval rule is

$H$1

while the sparse-adjusted rule is

$H$2

with $H$3. The paper’s interpretation is that the original BEG rule is too permissive when activity is of order $H$4, because inactive neurons can be spuriously activated by cross-talk noise. Subtracting $H$5 suppresses these false activations (Heusel et al., 2017).

The fixed-point criterion is

$H$6

For the unmodified rule, the paper proves instability at sparse load: $H$7 For the thresholded rule, the main theorem states that if

$H$8

where $H$9 is the unique root of

$h_i$0

then

$h_i$1

The bound is sharp, and $h_i$2 is the maximal admissible threshold. At $h_i$3, the maximal admissible load is

$h_i$4

so the model stores on the order of

$h_i$5

patterns with exact fixed-point retrieval (Heusel et al., 2017).

The relevance to RFBEGM is structural rather than literal. Retrieval is governed by a random sign field and a random activity field, the state space is ternary, and the sparse regime requires an explicit threshold matched to $h_i$6. This suggests that RFBEGM-like behavior can emerge whenever a spin-1 system is controlled by coupled quenched fields acting on $h_i$7 and $h_i$8, even outside equilibrium statistical mechanics.

5. Random crystal fields, quenched heterogeneity, and renormalization-group formulations

On finite-connectivity random graphs, disordered BEG physics is treated by replica symmetry through an order-parameter function $h_i$9 and, more transparently, through a distribution $0$0 of effective local fields (Jr. et al., 2023). The random crystal field is bimodal,

$0$1

so a fraction $0$2 of sites have the crystal field effectively switched off. The self-consistent equation for $0$3 takes the Poisson-cavity form

$0$4

with auxiliary functions $0$5, $0$6, and $0$7 defined in terms of neighboring cavity fields (Jr. et al., 2023).

Within this formalism, the basic observables are the magnetization

$0$8

and occupancy

$0$9

The numerical solution uses a population dynamics algorithm with $h_i$00 fields, about $h_i$01 iterations, and averaging over $h_i$02 runs per point (Jr. et al., 2023). The resulting phase diagrams show that average connectivity $h_i$03 changes the topology of the phase diagram and interpolates between sparse-network behavior and the fully connected mean-field limit.

The phases reported include FM, FM$h_i$04, FM$h_i$05, PM, PM$h_i$06, and PM$h_i$07. Random crystal-field disorder can split the single ferromagnetic phase of the clean system into high- and low-magnetization branches, and for sufficiently large $h_i$08 can likewise split the paramagnetic sector. The paper reports reentrance, tricritical points, critical points, and critical end points, with moderate disorder being particularly sensitive to connectivity (Jr. et al., 2023). Since the disorder couples to $h_i$09 rather than the sign of $h_i$10, this realization is best described as a random-crystal-field BEG model, but it sits naturally within the broader RFBEGM family.

A more structural variant randomizes the local state space rather than the local field. In the mean-field quenched-random hybrid model, some sites are “strong” and admit only Ising values $h_i$11, while others are “weak” and admit full BEG values $h_i$12. The Hamiltonian is

$h_i$13

with $h_i$14, and the concentration $h_i$15 of strong sites is the disorder parameter (Schreiber et al., 2022). This is not a standard random-field model, but the randomness is again quenched, local, and thermodynamically consequential. The paper finds different phase portraits in concentration-temperature space in the canonical and microcanonical ensembles, including different tricritical points at $h_i$16: $h_i$17 in the canonical ensemble and

$h_i$18

in the microcanonical ensemble (Schreiber et al., 2022).

Position-space renormalization on hierarchical lattices provides a different perspective on disordered BEG-like systems. For the disorder-generalized BEG Hamiltonian,

$h_i$19

the RG generates random effective field-like terms even when the initial disorder is placed only in $h_i$20 (Antenucci et al., 2013). The paper concludes that hierarchical lattices can capture some qualitative features, including spin-glass phases and some reentrance, but fail to reproduce the first-order transition present in mean-field theory and finite-dimensional simulations. This is directly relevant to RFBEGM interpretation because it separates what may be generic to spin-1 quenched disorder from what is sensitive to lattice topology and RG representation.

6. Criticality, field-like perturbations, and ensemble structure

The BEG framework is unusually sensitive to perturbations that act as effective fields. On an fcc lattice, the dipole-quadrupole interaction

$h_i$21

breaks the symmetry between $h_i$22 and $h_i$23 in a way analogous to an external magnetic field (Özkan et al., 2010). At criticality, the paper postulates

$h_i$24

and finds $h_i$25, in good agreement with the universal field exponent $h_i$26. This result concerns a clean BEG model rather than RFBEGM, but it clarifies why field-like perturbations are especially consequential in spin-1 systems: they act simultaneously on symmetry breaking and occupancy balance.

Mean-field BEG models also display multicritical structures that remain relevant as baselines for RFBEGM. In the infinite-range model with negative biquadratic coupling,

$h_i$27

the canonical phase diagram has a fourth-order critical point at

$h_i$28

while the microcanonical ensemble has a corresponding point at different parameter values,

$h_i$29

The continuous transition line coincides in the two ensembles, but the first-order structure and multicritical topology do not (Prasad et al., 2019). This ensemble inequivalence is not an RFBEGM result in itself, yet it shows that even before quenched random fields are introduced, the BEG phase structure can depend sensitively on whether one probes free-energy or entropy maximization.

At the level of fluctuations, the mean-field BEG model exhibits a hierarchy of scaling laws controlled by the minimum type of the Landau function $h_i$30. In the single-phase region the scaled total spin is Gaussian with rate $h_i$31; on the second-order curve it has a quartic non-Gaussian limit with rate $h_i$32; and at the tricritical point it has a sextic non-Gaussian limit with rate $h_i$33 (Eichelsbacher et al., 2013). When the parameters approach criticality along sequences $h_i$34, the same limiting distribution can appear with different convergence rates, a phenomenon the paper calls an “additional phase transition” in rates. A plausible implication is that disorder in RFBEGM-like systems may further enrich not only the phase topology but also the fluctuation structure around critical sets, although that extension is not developed in the cited work.

Taken together, these results establish RFBEGM as a family of spin-1 disordered systems in which magnetization, occupancy, and quenched heterogeneity are inseparable. In the strict random-field setting, the model provides a mean-field laboratory for frustration-induced no-passing violation, non-abelian relaxation, and avalanche-gap statistics (E et al., 16 Jul 2025). In broader RFBEGM-like realizations, random crystal fields, random local state constraints, and activity-coupled effective fields generate phase splitting, reentrance, and ensemble-dependent criticality (Jr. et al., 2023). The common lesson is that the $h_i$35 sector is not an auxiliary embellishment of Ising behavior; it is the source of a distinct class of disordered dynamics and thermodynamics.