Equi-Centric Model: A Framework of Symmetry and Fairness

• Equi-Centric Model is a framework that incorporates invariant, symmetric, or fairness criteria to structure diverse methods in finance, mRNA modeling, facility location, and geometry.
• Each domain employs distinct mathematical tools—from spectral analysis and group-equivariant mappings to optimization proxies—to enforce the central organizing principle.
• The models yield practical benefits, offering consistent estimators, improved predictions, scalable optimization, and theoretical insights into symmetry-driven design.

The expression Equi-Centric Model does not denote a single standardized formalism across the literature. In current arXiv usage, it refers to several domain-specific constructions that place an equalizing, symmetric, or symmetry-respecting structure at the center of modeling. These include the equi-correlated normal population in random matrix theory and finance, the codon-symmetry-equivariant architecture of Equi-mRNA for mRNA language modeling, an equity-centric facility location framework based on the Kolm–Pollak Equally-Distributed Equivalent, and a geometric formalization of centers as equivariant maps between $G$-spaces (Akama, 2022, Yazdani-Jahromi et al., 20 Aug 2025, Horton et al., 2024, Prieto-Martínez, 2023). A plausible implication is that the phrase functions less as the name of one model than as a family resemblance across methods that encode equality, symmetry, or fairness as the organizing principle.

1. Terminological scope and organizing principle

Domain	Core object	Organizing structure
Random matrix theory and finance	Sample correlation or covariance matrix	Constant pairwise correlation $\rho$
mRNA language modeling	Codon-level autoregressive model	Synonymous-codon symmetry as cyclic subgroups of $SO(2)$
Facility location	Mixed-integer optimization model	Kolm–Pollak EDE with inequality aversion $\epsilon$
Geometry	Map from objects to points	$G$-equivariance and stabilizer fixed points

In the random-matrix setting, the equi-centric construction is the equi-correlated normal population, an $N$-dimensional normal vector with unit variances and common off-diagonal correlation. In the mRNA setting, the term is used for an explicit equivariant model in which synonymous codons are represented by cyclic rotations in latent space. In facility location, the expression describes an equity-centric optimization framework in which the objective is built around a welfare-valid burden metric rather than average distance alone. In geometry, the term refers to the idea that a center should be understood as a $G$-equivariant map, so that symmetries of the object are mirrored by symmetries of the selected point (Akama, 2022, Yazdani-Jahromi et al., 20 Aug 2025, Horton et al., 2024, Prieto-Martínez, 2023).

A common misconception would be to treat these as direct variants of one another. The sources do not support that interpretation. Each formulation is autonomous, with its own state space, objective, and notion of equivariance or equality. What they share is structural: each paper centers its theory on a formally specified invariant, symmetry class, or fairness criterion.

2. Equi-correlated normal populations in random matrix theory and finance

In the random-matrix formulation, the population correlation matrix is

$$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$

where $\rho \in [0,1)$ is the common pairwise correlation. Its population spectrum is explicit: $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$ and

$\rho$0

on the $\rho$1-dimensional subspace orthogonal to $\rho$2. For $\rho$3, the model therefore has one dominant spike and a flat bulk. In the normal case it admits the one-factor representation

$\rho$4

with independent standard normals $\rho$5 and $\rho$6, so the common factor strength is controlled by $\rho$7 (Akama, 2022).

The sample correlation matrix is constructed from centered row vectors and standardized to unit sample variance, with

$\rho$8

in the high-dimensional regime

$\rho$9

The paper proves that for a broad class of factor models, including the equi-correlated case, the empirical spectral distribution of $SO(2)$0 converges almost surely to a Marčenko–Pastur law with index $SO(2)$1 and scale equal to the limiting ratio of specific variance to total variance. In the equi-correlated normal specialization,

$SO(2)$2

Thus the bulk support is asymptotically

$SO(2)$3

The corresponding largest-eigenvalue result is a strong-consistency theorem: $SO(2)$4 This establishes that the dominant sample eigenvalue scales linearly with $SO(2)$5 and consistently estimates the equi-correlation coefficient after normalization by $SO(2)$6. In the covariance case, for $SO(2)$7 and $SO(2)$8, the largest eigenvalue satisfies

$SO(2)$9

with

$\epsilon$0

The limiting law is Gaussian rather than Tracy–Widom because the spike is well separated and diverges with $\epsilon$1.

The paper also discusses a BBP-type phase transition when $\epsilon$2. Since

$\epsilon$3

the classical threshold translates into a condition on $\epsilon$4. If

$\epsilon$5

the spike remains above the MP edge and the largest sample eigenvalue is asymptotically normal around a deterministic shift. If

$\epsilon$6

the spike merges with the edge and the fluctuation regime becomes Tracy–Widom; the borderline case yields a generalized Tracy–Widom law.

In finance, these results formalize the empirical picture of one dominant “market mode” plus an MP-like bulk. The paper proves that the heuristic fit of stock-return correlation eigenvalue histograms by an MP law with scale parameter $\epsilon$7 is justified, because $\epsilon$8. It follows that

$\epsilon$9

is a consistent estimator of market-mode strength, and the fitted law $G$0 supports eigenvalue cleaning or shrinkage procedures. The same section of the source also notes the model’s limitations: normality and equi-correlation are idealizations, while empirical financial data may exhibit heteroskedasticity, time dependence, and multiple factors (Akama, 2022).

3. Equivariant mRNA LLMs

In the mRNA setting, the equi-centric construction is explicitly symmetry-theoretic. Let $G$1 be the set of 64 codons and $G$2 the set of 20 amino acids plus a stop symbol, with surjective genetic-code map $G$3. For each amino acid $G$4, the synonym set is $G$5 with cardinality $G$6. The central observation is that synonymous substitutions preserve amino-acid identity but alter translational speed, mRNA stability, local RNA structure, co-translational folding, and expression through usage bias, GC content, and tRNA abundance. Equi-mRNA models this by assigning each synonym class a cyclic subgroup

$G$7

with rotations at angles $G$8 and codon-to-rotation map $G$9 (Yazdani-Jahromi et al., 20 Aug 2025).

The codon-level encoder is required to satisfy an equivariance constraint: $N$0 Operationally, synonymous codons act by rotations in an amino-acid-specific $N$1-dimensional latent subspace while amino-acid identity and positional context are preserved. The architecture uses codon tokenization aligned to the reading frame, with a GPT-2 or hybrid Mamba–Transformer backbone. Each amino acid $N$2 owns a base embedding vector $N$3 and a learned $N$4-frame $N$5, yielding

$N$6

with $N$7. The source also gives an equivalent canonical $N$8-dimensional form: $N$9 where $G$0.

Three generator variants are described. Fixed $G$1 uses the uniform cyclic angle $G$2. Learned $G$3 parameterizes the generator and imposes soft constraints so that $G$4 remains a homomorphism. Fuzzy $G$5 places a soft distribution over $G$6 angle prototypes per amino acid, with effective angle

$G$7

The architecture can also be extended to block-diagonal $G$8 actions by partitioning $G$9 into multiple $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$0-dimensional subspaces.

Sequence-level aggregation is also symmetry-aware. A canonical invariant pooling is

$$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$1

which removes the rotational component and yields an amino-acid-specific invariant summary. The paper further explores polar pooling, Fourier/DFT pooling, and $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$2 mean pooling via $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$3 of weighted cosine and sine sums.

Training combines standard autoregressive next-token prediction with an auxiliary equivariance loss,

$$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$4

Pretraining is reported on a $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$5M-sequence ablation corpus for $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$6 epochs and on a $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$7M-sequence corpus for $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$8 epochs, with bf16 mixed precision, cosine learning-rate scheduling, max length $$\Sigma_\rho = (1-\rho) I_N + \rho\,\mathbf{1}\mathbf{1}^\top,$$9 codons, and effective batch size $\rho \in [0,1)$0. Ablations were run on $\rho \in [0,1)$1 NVIDIA H200 GPUs; final runs used $\rho \in [0,1)$2 NVIDIA H100 GPUs.

Empirically, the model is reported to deliver up to approximately $\rho \in [0,1)$3 improvements in downstream property prediction relative to vanilla codon embeddings. The details given include: fixed-angle plus equivariance exceeded vanilla on E. coli expression $\rho \in [0,1)$4 and mRFP $\rho \in [0,1)$5 Spearman); learned $\rho \in [0,1)$6 plus Stiefel plus equivariance achieved mRFP Spearman $\rho \in [0,1)$7 and E. coli accuracy $\rho \in [0,1)$8; fuzzy $\rho \in [0,1)$9 was robust on noisy assays, with MLOS approximately $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$0 Spearman and SARS-CoV-2 degradation approximately $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$1. At $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$2M scale, the $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$3M-parameter Equi-mRNA achieved the highest accuracy in $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$4 datasets, often matching or surpassing codon-aware baselines while using approximately $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$5 of HELM.

For generation, autoregressive completion on iCodon prompts with top-$$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$6, nucleus $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$7, and temperatures $$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$8 is evaluated with Fréchet BioDistance,

$$\lambda_1^{\mathrm{pop}} = 1 + (N-1)\rho,\qquad \mathbf{v}_1 \propto \mathbf{1},$$9

The best equivariant fuzzy-$\rho$00 model at $\rho$01M scale achieved FBD approximately $\rho$02 at $\rho$03, approximately $\rho$04 better than vanilla at approximately $\rho$05. On the $\rho$06M corpus, Equi-mRNA $\rho$07M$\rho$08 achieved $\rho$09 and Equi-mRNA $\rho$10M$\rho$11 achieved $\rho$12. Generated suffixes also showed up to approximately $\rho$13 MSE reduction versus vanilla in functional property preservation.

The model is designed to be interpretable. Entropy of learned angles increases almost linearly with GC content, with Pearson $\rho$14 $\rho$15. The codon angle has Spearman correlation $\rho$16 $\rho$17 with normalized human tRNA copy number, so codons decoded by more abundant tRNAs receive systematically smaller angles. The reported limitations are equally explicit: the $\rho$18 construction ignores potential non-cyclic permutation symmetries and codon-pair interactions; codon usage is organism-specific; and the present scope is restricted to coding regions and codonized inputs rather than UTRs, RNA structures, or non-coding regulatory signals (Yazdani-Jahromi et al., 20 Aug 2025).

4. Equity-centric facility location

In facility location, the equi-centric model is not based on symmetry but on an explicitly normative equity functional. The central object is the population-weighted Kolm–Pollak Equally-Distributed Equivalent for burdens such as travel distance: $\rho$19 where $\rho$20 is the population of residential area $\rho$21, $\rho$22, and $\rho$23 with $\rho$24 for burdens and

$\rho$25

The source states that the KP EDE satisfies symmetry, population independence, scale dependence, transfers, mirror property, and separability. Because $\rho$26, the inequality penalty raises the EDE above the mean distance (Horton et al., 2024).

The limiting behavior connects standard facility-location objectives. As $\rho$27 and hence $\rho$28,

$\rho$29

which reproduces p-median behavior. As $\rho$30,

$\rho$31

which yields p-center-like behavior. The paper therefore treats $\rho$32 as a single policy parameter controlling the efficiency–fairness trade-off without introducing a separate mixing weight.

With demand nodes $\rho$33, candidate facilities $\rho$34, demand weights $\rho$35, travel costs $\rho$36, and binary assignment variables $\rho$37, the nonlinear EDE objective is converted into a linear proxy. Since minimizing

$\rho$38

is equivalent to minimizing

$\rho$39

the paper proves that in the unsplittable case

$\rho$40

so the resulting model is a fully linear MILP: $\rho$41 subject to the standard facility budget, assignment, linkage, and optional capacity constraints. Split assignment $\rho$42 is also supported, and the source argues that the same linear proxy more appropriately penalizes poor distances under splitting than the original nonlinear expression.

The framework is further extended with location-specific penalties. If undesirable sites $\rho$43 have penalties $\rho$44 measured in distance units, the added proxy term is

$\rho$45

with

$\rho$46

and tangent-line inequalities

$\rho$47

used to obtain a pure MILP. When all per-site penalties are equal, choosing $\rho$48 yields exact linearization. The source also gives an approximation bound via

$\rho$49

The empirical emphasis is scalability. The model solved all grocery-store instances for the $\rho$50 largest U.S. cities to optimality, with problem sizes up to $\rho$51 binary variables in New York City and with computational times comparable to p-median and substantially faster than p-center. Averaged across the $\rho$52 cities, relative to p-median, mean distance increased by $\rho$53 m for $\rho$54, $\rho$55 m for $\rho$56, and $\rho$57 m for $\rho$58, while maximum distance decreased by $\rho$59 m, $\rho$60 m, and $\rho$61 m, respectively. In polling-location experiments, KP EDE, mean, max, and standard deviation either matched or improved p-median, while p-center sometimes showed worse distributional properties and failed to reach optimality within three hours in one scenario. In a Gwinnett County early-voting case study, penalties reduced the use of less-suitable sites from eight to two with typical penalty approximation errors in meters of $\rho$62 to $\rho$63.

The paper’s practical guidance is correspondingly concrete: choose $\rho$64 for distances, start at $\rho$65, calibrate $\rho$66 by a two-pass procedure, prune extremely long assignments if coefficient ranges become problematic, and report KP EDE together with mean, max, upper quantiles, descriptive inequality measures, and subgroup analyses. The limitations stated in the source are static demand, deterministic travel times, single-mode distances, homogeneous disutility, and numerical difficulties at extreme $\rho$67 (Horton et al., 2024).

5. Centers as equivariant maps in geometry

In geometry, the equi-centric formulation is categorical and group-theoretic. Let $\rho$68 be a group acting on sets or spaces, and let $\rho$69 and $\rho$70 be $\rho$71-spaces. A map

$\rho$72

is $\rho$73-equivariant if

$\rho$74

for all $\rho$75 and $\rho$76. The stabilizer of $\rho$77 is

$\rho$78

and for a subgroup $\rho$79, the fixed-point set is

$\rho$80

A center is then defined to be an equivariant map from a family of geometric objects to the ambient space (Prieto-Martínez, 2023).

This definition subsumes classical examples. For triangles in $\rho$81 under the similarity group, the centroid

$\rho$82

is equivariant; the incenter and circumcenter are likewise equivariant on their natural domains. For finite subsets of $\rho$83, the centroid is

$\rho$84

For nondegenerate conics in the real projective plane, the map sending a conic to its center is equivariant under $\rho$85.

The paper’s main theorem is a general existence statement. If $\rho$86 and $\rho$87 satisfy the compatibility condition

$\rho$88

for every $\rho$89, and if $\rho$90 and $\rho$91, then there exists an equivariant map

$\rho$92

such that $\rho$93. The proof first defines $\rho$94 on the orbit $\rho$95 by

$\rho$96

which is well defined because $\rho$97 is fixed by $\rho$98. If $\rho$99 is not transitive on $SO(2)$00, the map is then extended orbit by orbit using representatives and fixed points chosen via the Axiom of Choice.

A central application is an analogue of Edmonds’s center conjecture for equifacetal simplices. Let $SO(2)$01 be the family of $SO(2)$02-point subsets of $SO(2)$03, with elements interpreted as possibly degenerate $SO(2)$04-simplices, and let $SO(2)$05 under the natural Euclidean action. The paper proves: $SO(2)$06 If $SO(2)$07 is equifacetal, the fixed-point set of its symmetry group is a singleton, so every equivariant center agrees there. If $SO(2)$08 is not equifacetal and is affinely independent, then $SO(2)$09 has more than one point, and the existence theorem allows different equivariant centers to realize different values at $SO(2)$10. In the degenerate case, the paper constructs an explicit weighted centroid

$SO(2)$11

where $SO(2)$12 is the distance from $SO(2)$13 to the centroid of $SO(2)$14.

The theorem has immediate low-dimensional consequences. For an equilateral triangle, all equivariant centers coincide at the unique fixed point. For an isosceles but non-equilateral triangle, the fixed-point set is the symmetry axis, and any prescribed point on that line can be realized as $SO(2)$15 for some equivariant center. For a scalene triangle with trivial stabilizer, any point in the plane can occur as the center value of some equivariant map. The source is equally explicit about limitations: the existence theorem is purely set-theoretic, continuity is not guaranteed, and the continuous version of Edmonds’s original conjecture remains open (Prieto-Martínez, 2023).

6. Comparative perspective, limitations, and recurrent themes

Taken together, these works suggest that Equi-Centric Model has two dominant interpretations. In finance, mRNA modeling, and geometry, it is primarily symmetry-centric: constant pairwise correlation, codon-substitution group actions, or object symmetries determine the admissible structure of spectra, latent states, or center maps. In facility location, it is equity-centric: the model is organized around a welfare functional that makes inequality aversion explicit rather than auxiliary (Akama, 2022, Yazdani-Jahromi et al., 20 Aug 2025, Horton et al., 2024, Prieto-Martínez, 2023).

The methodological contrast is equally sharp. The finance paper is asymptotic and spectral, with almost-sure convergence, central-limit behavior, and BBP-type phase transitions. The mRNA paper is representational and algorithmic, combining group actions, Stiefel-manifold parameterization, auxiliary equivariance loss, and autoregressive sequence modeling. The facility-location paper is optimization-theoretic, converting a nonlinear welfare objective into a scalable MILP through an exact linear proxy and piecewise-linear penalty approximation. The geometry paper is foundational and set-theoretic, emphasizing existence and nonuniqueness of equivariant selections under stabilizer constraints.

The limitations recorded in the sources are also domain-specific rather than incidental. The equi-correlated normal population assumes normality, independence across time, and a one-factor or limited-factor structure. Equi-mRNA models synonymous codon sets as cyclic rotations in $SO(2)$16 and therefore ignores non-cyclic permutation symmetries, codon-pair interactions, and non-coding regulatory context. The equitable facility-location framework assumes static demand and deterministic travel times, and its exponential coefficients can become numerically challenging. The geometric existence theorem does not preserve continuity or measurability and therefore does not settle the continuous center conjecture.

A plausible implication is that the most stable meaning of equi-centric is not attached to any one equation or architecture. Rather, it designates a modeling stance: specify the relevant equivalence structure first, and then require the representation, objective, or selection rule to respect that structure exactly. In one case the structure is the constant correlation coefficient $SO(2)$17; in another it is the cyclic symmetry of synonymous codons; in another it is the inequality-aversion parameter $SO(2)$18 embedded in a Kolm–Pollak welfare criterion; and in another it is the stabilizer fixed-point set $SO(2)$19. Across these otherwise unrelated literatures, the equi-centric move is to make that structure constitutive rather than post hoc.