Composite Potential Fields in Theory and Practice

• Composite potential fields are effective potentials assembled from simpler constituents, offering flexible modeling across diverse disciplines.
• They enable tractable approximations in Bayesian inference, induce reentrant localization in quantum systems, and speed up dynamic motion planning in robotics.
• Extensions in field theory utilize composite constructions for symmetry breaking, gauge-invariant reformulations, and the emergence of operator-based effective variables.

Composite potential fields are composite constructions in which an effective field, potential, or potential-like object is assembled from simpler constituents rather than specified as a single primitive quantity. Across the cited literature, this phrase does not denote one universal formalism. Instead, it refers to several technically distinct operations: products of tractable local Gibbs factors in Bayesian inference for Gibbs random fields, additive superposition of periodic and quasi-periodic onsite modulations in localization theory, voxelwise minimum composition of signed-distance fields for dynamic motion planning, and linear combinations of semantic maps for decision-making in stealth-game AI (Friel, 2012, Wei, 13 Apr 2025, Finean et al., 2020, Xu et al., 25 Aug 2025). Related field-theoretic work extends the same compositional logic to scalar, gauge, and operator-valued fields, where composite structure is used to encode symmetry breaking, gauge-invariant reformulations, or enlarged dynamical operators (Zheltukhin, 2022, Fournel et al., 2012, Reshetnyak, 2014, Weng, 11 Jan 2025, Alonso-Izquierdo et al., 29 Dec 2025, Chkareuli, 2021).

1. Statistical composite fields in Gibbs random fields

In spatial statistics, composite potential fields arise as approximations to an intractable global Gibbs random field by products of tractable local factors. For binary variables $y=\{y_1,\dots,y_n\}$ on a graph, the Gibbs random field is written in exponential-family form as

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$

with normalizing constant

$z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$

For the first-order autologistic model,

$$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$

and the sitewise full conditional is

$$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$

The local-interaction interpretation is explicit: the global joint law is generated by local site and pairwise interaction terms, and the Hammersley–Clifford theorem links that conditional structure to the Gibbs form (Friel, 2012).

The composite approximation replaces the intractable likelihood by

$$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$

or, in weighted form,

$$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$

The cited work distinguishes full likelihood, marginal composite likelihood, and conditional composite likelihood, and focuses almost entirely on the conditional case for the autologistic model. The blockwise conditional factors are defined on contiguous $k\times k$ lattice blocks, yielding $CCL_3$, $CCL_4$, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$0, and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$1. Because the blocks overlap, the same pairwise interactions are included multiple times; this is computationally useful but statistically consequential (Friel, 2012).

The empirical picture is sharply defined. All conditional composite likelihoods outperform pseudolikelihood in posterior mean bias on $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$2 and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$3 Ising lattices. At the same time, posterior variances are severely underestimated by all composite likelihood methods: on the $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$4 lattice, they are about an order of magnitude smaller than the true posterior variance, and the same under-dispersion persists on the $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$5 lattice. Larger blocks are much more expensive, yet the differences between block sizes are modest; in the reported runs, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$6 blocks are used exhaustively, while only about $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$7, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$8, and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$9 of possible $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$0, $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$1, and $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$2 blocks are used. A common misconception is that more faithful local blocks automatically produce calibrated Bayesian posteriors; the reported results show that they do not, because overlapping composition overcounts information unless additional weighting or calibration is introduced (Friel, 2012).

2. Additive composite potentials and reentrant localization

In one-dimensional tight-binding systems, a composite potential can mean an additive onsite modulation formed from a periodic commensurate component and a quasi-periodic incommensurate component. The Hamiltonian is

$z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$3

with onsite potential

$z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$4

The periodic term generates commensurate band structure, while the quasi-periodic term tends to localize states in the Aubry–André sense; the $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$5 limit recovers the Aubry–André transition at $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$6 (Wei, 13 Apr 2025).

The notable phenomenon is reentrant localization. For rational $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$7, increasing $z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$8 can induce sequences such as

$z(\theta)=\sum_{y\in Y}\exp(\theta^T s(y)).$9

The 2025 study shows that this is not restricted to the staggered case $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$0; it also appears for $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$1, with different critical amplitudes and different bands. For $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$2, the periodic potential has a 9-site unit cell, the spectrum splits into nine subbands at large $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$3, and reentrant localization appears prominently in the second and seventh bands for $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$4 and $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$5 (Wei, 13 Apr 2025).

The mechanism is symmetry-sensitive rather than amplitude-only. The phase $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$6 of the periodic component is crucial, whereas the quasi-periodic phase $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$7 is largely irrelevant. For $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$8, the reentrant-supporting phases repeat with $$s_0(y)=\sum_{i=1}^n y_i,\qquad s_1(y)=\sum_{j=1}^n\sum_{i\sim j} y_i y_j,$$9, and even a tiny shift from $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$0 to $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$1 can destroy the reentrant windows in certain bands. The cited work identifies translational invariance within the commensurate unit cell and mirror symmetry within each period as jointly necessary; translational invariance alone is not sufficient, and perturbations that break either symmetry destroy reentrance (Wei, 13 Apr 2025).

Localization diagnostics are given by the mean fractal dimension

$$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$2

and the Lyapunov exponent

$$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$3

For the seventh band at $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$4, finite-size scaling yields critical points $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$5, $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$6, and $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$7, with exponents $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$8, $$p(y_i\mid y_{-i},\theta)\propto \exp\left( \theta_0 y_i+\theta_1 y_i(y_{i-m}+y_{i-1}+y_{i+1}+y_{i+m}) \right).$$9, and $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$0, consistent with the Harris criterion. The numerical interpretation given in the paper is paradoxical but precise: increasing $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$1 enhances localization through stronger onsite modulation, yet simultaneously strengthens the mirror-symmetric periodic structure that favors extended states. The reentrant transition is attributed to that competition rather than to monotone disorder-strength effects (Wei, 13 Apr 2025).

3. Min-composed signed-distance fields for dynamic motion planning

In robotic motion planning, composite potential fields take the form of predicted signed-distance fields assembled from object-centric SDF primitives. The environment is decomposed into static obstacles and moving obstacles, a static ESDF $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$2 is computed once, and each moving object $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$3 is assigned an occupancy box $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$4 and a local object SDF $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$5. At a future time $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$6, each object SDF is translated to its predicted pose and merged with the static field by voxelwise minimum: $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$7 The method exploits the claim that an SDF is compositional under $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$8, making it unnecessary to recompute a full workspace ESDF from scratch at every update (Finean et al., 2020).

The predictive component is organized through a tracking-and-propagation loop. From successive occupancy grids, moving objects are isolated, their centroids and velocities are estimated, and a constant-velocity model propagates them forward. The implementation repeatedly updates object SDFs and poses, updates velocities, predicts future object positions at planning times, generates the predicted composite SDF at each time, and updates the motion planner with the resulting field. The framework is not tied to constant velocity; the cited work explicitly states that KLT tracking or an Unscented Kalman Filter can be substituted (Finean et al., 2020).

The principal technical result is a speed–accuracy trade-off tailored to motion optimization. The composite SDF is guaranteed to match the exact SDF up to distance $$CL(y\mid\theta) =\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta),$$9 from obstacle surfaces, which is the active band for collision costs. On a $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$0 grid, composite initialization costs $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$1 ms, full exact computation costs $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$2 ms, but subsequent composite prediction costs only $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$3 ms, corresponding to a $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$4 repeat-prediction speed-up. Across tested workspace sizes, the repeat-prediction speed-up ranges from $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$5 to $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$6, summarized as an $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$7–$$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$8 reduction in time for subsequent predictions. For the $$\prod_{i=1}^C p(y_{A_i}\mid y_{B_i},\theta)^{w_i},\qquad w_i>0.$$9 setup, mean SDF generation rates are reported as $k\times k$0 Hz for composite predictions versus $k\times k$1 Hz for exact recomputation (Finean et al., 2020).

The composite field is integrated with GPMP2, where each obstacle factor in the factor graph can be assigned the SDF associated with its own time index rather than sharing a single static field. In the simulated $k\times k$2-DoF WAM setup, the horizon is $k\times k$3 s, there are $k\times k$4 support states at $k\times k$5 s intervals, dense interpolation uses $k\times k$6 ms with $k\times k$7, $k\times k$8 m, $k\times k$9, and the environment resolution is $CCL_3$0 cm. In the closed-loop Panda-arm demonstration with a Toyota HSR as the moving obstacle, the reported obstacle speed is $CCL_3$1 m/s, the same $CCL_3$2 workspace and $CCL_3$3 s horizon are used in simulation, and the early-iteration update rate is $CCL_3$4 Hz with standard deviation $CCL_3$5 Hz, rising to about $CCL_3$6 Hz later. A common misconception is that the method is an exact replacement for full ESDF generation; the cited paper presents it instead as a predictive approximation that preserves the SDF geometry where the planner needs it most (Finean et al., 2020).

4. Composite potential fields in stealth-game guard AI

In stealth-game AI, Composite Potential Fields (CPF) designates a specific training-free framework in which guard behavior is determined by combining three scalar maps over a grid or graph: Information $CCL_3$7, Confidence $CCL_3$8, and Connectivity $CCL_3$9. The Information field is a negative-valued “where might the player be?” map that decays temporally as

$CCL_4$0

and, upon player detection at node $CCL_4$1, is reseeded by

$CCL_4$2

The reported implementation uses $CCL_4$3, $CCL_4$4, and $CCL_4$5. The Confidence field is a positive anti-revisit signal: $CCL_4$6 with $CCL_4$7 and $CCL_4$8. The Connectivity field is static: $CCL_4$9 so dead ends and chokepoints receive larger values (Xu et al., 25 Aug 2025).

The composite potential is the linear mixture

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$00

The baseline patrol weights are $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$01, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$02, and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$03; the maximum information weight is $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$04; and the weight decay rate is $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$05. Rather than minimizing $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$06 directly, the framework applies local kernel filtering: $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$07 with $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$08 and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$09, and the next move is selected by

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$10

This local averaging is introduced to suppress pointwise noise and reduce oscillatory or myopic behavior (Xu et al., 25 Aug 2025).

The environment abstraction can be either an occupancy grid or a NavMesh partition graph. On a grid, the candidate set is the four orthogonal walkable neighbors. On a triangulated NavMesh graph, candidates are generated using BFS with a lookahead threshold $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$11, stated to be typically around $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$12–$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$13 times the average inter-node distance, with fallback to direct neighbors if necessary. The same field definitions are preserved across both abstractions; only the graph structure and distance metric change (Xu et al., 25 Aug 2025).

The framework differs from classical artificial potential fields in several stated ways: it uses multiple semantic maps rather than a single goal–obstacle decomposition, includes temporal memory through decays in $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$14 and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$15, incorporates topology through $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$16, relies on kernel filtering rather than raw pointwise minima, and adapts weights continuously rather than switching through an explicit finite state machine. The cited evaluation compares PF against Random Walk, FSM, Staleness-based FSM, and Cheat. In the Grid environment, PF attains capture rate $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$17 with capture time $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$18 s; in the Partition environment, PF attains capture rate $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$19 with capture time $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$20 s. For fixed-time patrol, PF yields grid coverage about $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$21 with backtracking $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$22, while in the Partition environment the reported coverage is about $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$23 and backtracking is $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$24 or $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$25 depending on player policy. The comparison with Staleness is particularly important: Staleness may match or exceed total captures in some cases, but it produces much more repetitive motion, with partition backtracking around $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$26. The framework is also extended by modifying the same field-update rules to account for footstep noise, thrown or static decoys, corpse discovery, lighting, weather, and concealment (Xu et al., 25 Aug 2025).

5. Composite scalar potentials and emergent scalar modes

In cosmological field theory, composite structure can mean that a scalar degree of freedom is not elementary but arises from a tensor field in curved space. One construction begins from a massless symmetric tensor $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$27 on a fixed 4D background $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$28, with a generally covariant quartic self-interaction potential

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$29

The trace

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$30

acts as a composite Nambu–Goldstone scalar boson when global Weyl and scale symmetries are spontaneously broken. The nontrivial extremum

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$31

aligns the tensor vacuum with the background metric, leaving the trace as the active scalar mode. Degeneracy removal selects a constant vacuum

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$32

and the cosmological constant on the broken-symmetry vacuum is

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$33

so $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$34 when $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$35 (Zheltukhin, 2022).

In multifield soliton theory, a composite potential is constructed by coupling two previously independent scalar theories on the product target space $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$36. If the original theories admit superpotentials $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$37 and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$38, the composite superpotential is

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$39

and the BPS potential becomes

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$40

The corresponding first-order equations are

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$41

A central structural result is that original kinks survive as boundary kinks when one sector is frozen at one of its vacua, while extra vacua can appear through

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$42

All extra vacua share the same superpotential value,

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$43

so BPS kinks between two extra vacua are forbidden (Alonso-Izquierdo et al., 29 Dec 2025).

These two uses of compositeness are distinct. In the cosmological model, the scalar itself is composite, being the trace of a tensor field. In the kink construction, the potential is composite because the target-space sectors are coupled through a superpotential cross-term. A plausible synthesis is that both works treat composite structure as a way of preserving analytically controlled degrees of freedom while enlarging the vacuum structure and symmetry content (Zheltukhin, 2022, Alonso-Izquierdo et al., 29 Dec 2025).

6. Gauge, geometric, and operator-based composite fields

In gauge theory and differential geometry, composite fields are often built to neutralize gauge redundancy or to reformulate nonlocal insertions as controlled effective variables. A general construction uses a dressing field $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$44 with transformation law $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$45 to form

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$46

which is gauge invariant when $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$47 is a connection. Matter fields transform similarly, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$48. This mechanism is used to reinterpret electroweak symmetry reduction, where the scalar doublet is decomposed as $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$49 and the dressed $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$50-invariant field

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$51

yields the familiar combinations $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$52, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$53, and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$54. It also appears in Cartan-connection gravity, where the dressing field $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$55 produces

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$56

interpreted as a gauge-invariant Christoffel-symbol-like object. The cited authors emphasize that dressing is not itself symmetry breaking but a change of variables that trivializes the gauge action on the composite fields (Fournel et al., 2012).

A separate use of composite fields appears in the Gribov–Zwanziger problem, where the horizon functional is introduced as a composite insertion

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$57

The generating functional becomes

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$58

and the effective action is defined by a double Legendre transform in the ordinary fields and the composite field. The resulting Ward identities acquire explicit composite-field terms, and the central gauge-dependence statement is that, on the mass shell determined by the Yang–Mills fields alone,

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$59

both $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$60 and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$61 do not depend on the gauge choice. This construction treats the horizon functional not as an ad hoc nonlocal modification of the action but as a composite operator insertion with controlled gauge variation (Reshetnyak, 2014).

The octonion-field framework of 2025 pushes the operator idea further by allowing field potential, field strength, and other physical quantities to enter the basic differential operator directly. Besides the field-strength-centered composite operator

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$62

the paper introduces potential-based operators of the form

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$63

leading to equations such as

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$64

It then generalizes to multi-quantity operators involving $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$65, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$66, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$67, $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$68, and $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$69. The explicit claim is that field strength does not occupy a unique central position: field potential can be an active ingredient in the source, momentum, torque, and force equations, and different composite operators generate different field equations (Weng, 11 Jan 2025).

Another line of work treats gauge bosons themselves as constrained composite fields made from fermion currents. In that framework, the defining constraint is

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$70

or, in the non-Abelian case,

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$71

The vector field is introduced as an auxiliary composite boson coupled to fermion currents, and after integrating out the hidden fermions the effective action acquires the standard kinetic term

$$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$72

The central claim is that the starting global symmetry $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$73 turns into a local symmetry $$f(y\mid \theta)\propto \exp\big(\theta^T s(y)\big)=q(y\mid \theta),$$74, while the nonlinear constraint functions as a gauge-fixing condition rather than an observable Lorentz-violating restriction. In this sense, the emergent vector bosons are naturally massless because the constrained composite construction removes the radial mode and protects the remaining modes by the generated gauge invariance (Chkareuli, 2021).

Taken together, these works show that in field theory the composite-field idea serves at least four distinct purposes: neutralizing gauge redundancy by dressing, converting nonlocal modifications into composite insertions with Ward identities, broadening the operator content of field equations, and generating apparently fundamental gauge fields from constrained fermionic composites. The underlying constructions are not interchangeable, but they share the principle that physically relevant variables may emerge more naturally from composite organization than from primitive potentials or connections (Fournel et al., 2012, Reshetnyak, 2014, Weng, 11 Jan 2025, Chkareuli, 2021).