Stratum: A Multi‐Disciplinary Layered Concept

• Stratum is defined as a layer or subset partitioned by fixed invariants or operational constraints, serving as an analytical unit across diverse fields.
• In mathematical and statistical contexts, stratification clarifies complex objects by enabling cohomology computations, Lyapunov analysis, and precise variance estimation.
• In biology and engineering, strata refer to physical or infrastructural layers that influence transport phenomena, system optimization, and network orchestration.

Stratum denotes a layer, subset, or component in a structured decomposition. In the research literatures surveyed here, the term appears in several technically distinct senses: the basic stratum of Shimura varieties, strata of abelian differentials and moduli spaces, strata of geodesics, survey and experimental-design strata, principal strata in causal survival analysis, the stratum corneum of skin, and architectural layers or named systems such as the Network Intelligence Stratum, the Stratum mining protocol, and the stratum system for agent-centric machine learning (Kret, 2012, Fils, 27 Mar 2025, Hu et al., 2020, Mosaferi, 2019, Sun et al., 13 Mar 2025, Saiko, 2023, Soto et al., 2024, Recabarren et al., 2017, Phani et al., 3 Mar 2026). This suggests a unifying abstraction: a stratum is the level at which a complex object is partitioned by fixed invariants, operational constraints, or layer-specific behavior.

1. Arithmetic and representation-theoretic strata

In arithmetic geometry, a stratum is often a locally closed locus cut out by $p$-adic or group-theoretic invariants. For PEL-type Shimura varieties modulo a prime of good reduction, the reduction is stratified by Newton strata, indexed by isocrystals with $G_{\mathbb{Q}_p}$-structure. The basic stratum is the Newton stratum associated to the unique basic class, characterized by an isocrystal with only one slope; under simplifying hypotheses, its $\ell$-adic cohomology is related to the cohomology of the complex Shimura variety, and explicit formulas for its point counts over finite fields are obtained by truncating Kottwitz’s point-counting formula and applying the trace formula (Kret, 2011).

For certain unitary Shimura varieties associated to division algebras, the cohomology of the basic stratum $B$ is expressed explicitly in terms of automorphic representations of the group in the Shimura datum. Only rigid representations at $p$ contribute, and their contribution is controlled by compact traces of Kottwitz functions. The combinatorics is encoded by weighted sums over non-intersecting strict Dyck paths, with the Lindström–Gessel–Viennot lemma handling the non-intersection constraint. The same analysis yields an explicit dimension formula,

$$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$

and at $q=1$ the number of non-intersecting Dyck paths computes the Euler–Poincaré characteristic of $B$ (Kret, 2012).

A related use appears in the Drinfeld stratification of Deligne–Lusztig varieties and their parahoric analogues for inner forms of $\mathrm{GL}_n$. For a twisted Levi subgroup $L$,

$G_{\mathbb{Q}_p}$0

and the corresponding images $G_{\mathbb{Q}_p}$1 define the stratification. In the inner-form setting, the strata are indexed by divisors $G_{\mathbb{Q}_p}$2, with locally closed pieces

$G_{\mathbb{Q}_p}$3

The unique closed stratum is $G_{\mathbb{Q}_p}$4; its cohomology is concentrated in a single degree for suitable eigenspaces, and the stratum is a maximal variety in the sense of Boyarchenko–Weinstein (Chan et al., 2020).

2. Strata in moduli spaces of differentials and flat bundles

In the moduli space of abelian differentials, a stratum $G_{\mathbb{Q}_p}$5 is specified by zero multiplicities $G_{\mathbb{Q}_p}$6 with $G_{\mathbb{Q}_p}$7. For a closed oriented surface $G_{\mathbb{Q}_p}$8 and finite $G_{\mathbb{Q}_p}$9, the relative period map takes values in $\ell$0. A complete characterization is now known for which classes $\ell$1 arise from an abelian differential in a given connected component $\ell$2 of a prescribed stratum: the necessary and sufficient conditions are a positive volume condition $\ell$3 and, when $\ell$4 is a lattice in $\ell$5, a partition inequality determined by an equivalence relation on $\ell$6 and the orders of zeros (Fils, 27 Mar 2025). The volume functional is

$\ell$7

and for period maps arising from $\ell$8 it equals $\ell$9 (Fils, 27 Mar 2025).

The de Rham moduli space $B$0 admits an oper stratification induced by the $B$1-action on the Hodge moduli space. The closed oper stratum is the unique minimal stratum, with dimension

$B$2

while the open dense stratum of irreducible flat bundles with stable underlying vector bundles is the unique maximal stratum, with dimension

$B$3

A Simpson filtration on a flat bundle produces the limiting complex variation of Hodge structure that determines the stratum (Hu et al., 2020).

For quadratic differentials on Teichmüller space, the principal stratum consists of differentials all of whose zeros are simple. Random walks on the mapping class group with finite first moment, non-elementary support, and semigroup containing a pseudo-Anosov whose invariant Teichmüller geodesic lies in the principal stratum eventually produce pseudo-Anosov mapping classes whose invariant geodesics are in the principal stratum, almost surely (Gadre et al., 2016).

An Outer space analogue replaces zero orders by ideal Whitehead graphs of ageometric fully irreducible outer automorphisms. For a graph $B$4, the corresponding stratum of axes is

$B$5

The principal stratum $B$6 corresponds to the disjoint union of $B$7 triangles. Nearby axes to a dominant-stratum axis lie in the basin of that stratum, but the principal stratum is not open: sequences of nonprincipal axes can converge to a principal axis because folding can identify vertices and remove components of the ideal Whitehead graph (Algom-Kfir et al., 2017).

In low-genus translation-surface theory, the even minimal stratum $B$8 in genus $B$9 is identified as an orbifold classifying space for a central extension of $p$0. Its projectivization is isomorphic as an orbifold to a moduli space of pointed genus-$p$1 curves with semigroup $p$2, and the kernel of the monodromy to $p$3 contains a non-abelian free group of rank $p$4 (Giannini, 2024).

3. Stratification as a tool in dynamics and optimization

In reaction-network theory, the stratum approach divides the positive orthant into regions determined by the ordering of monomials $p$5 associated with complexes. For a permutation $p$6, the corresponding stratum $p$7 records a strict ordering of these ratios. This geometric decomposition is used to analyze global stability of complex balanced mass-action systems: for a boundary face $p$8 adjacent to a stratum, one constructs a linear Lyapunov function $p$9 with $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$0 for $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$1 and $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$2 otherwise, and proves $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$3 along trajectories within that stratum. The method generalizes earlier detailed-balanced results to the broader class of complex balanced systems and clarifies how boundary approach could only occur through repeated stratum switching (Siegel et al., 2010).

In nonlinear semidefinite programming, stratification is used to resolve the nonsmoothness of the KKT system caused by the projection onto $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$4. The matrix space is decomposed by inertia pattern:

$$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$5

yielding a Whitney stratification of $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$6. Lifting this to the primal–dual space makes the KKT map smooth on each lifted stratum. The paper introduces stratum-restricted strong metric regularity and characterizes it by the weak second order condition (W-SOC) and weak strict Robinson constraint qualification (W-SRCQ). Geometrically, W-SRCQ is interpreted via transversality, which yields genericity in the ambient space and stability along strata. Algorithmically, a stratified Gauss–Newton method with normal steps and a correction mechanism is shown to converge globally to directional stationary points; under W-SOC and SRCQ it achieves local quadratic convergence to KKT pairs and eventually identifies the active stratum (Bao et al., 13 Jan 2026).

These works use “stratum” not merely as a partition label but as the domain on which local smoothness, Lyapunov monotonicity, or regularity becomes available. This suggests why stratification is recurrent in problems where the global object is singular or only piecewise regular.

4. Statistical and causal meanings of stratum

In survey sampling, a stratum is a sampling subgroup. Under a one-primary-sampling-unit-per-stratum design, the point estimator is unbiased and efficient, but an unbiased variance estimator does not exist. The classical collapsed-stratum estimator has bias

$$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$7

so the bias is driven by differences in means between collapsed strata. Empirical Bayes and constrained empirical Bayes variance estimators shrink the per-group variance proxies and outperform the classical collapsed estimator in empirical relative mean squared error in simulation (Mosaferi, 2019).

In two-stratum case-control analysis, each stratum is summarized by a $$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$8 table with stratum-specific odds ratio

$$\dim(B) = \sum_{\wp \mid p} \left( \sum_{v \in \wp} \frac{s_v(1-s_v)}{2} + \sum_{j=0}^{s_\wp-1}\left\lceil j\frac{n}{s_\wp}\right\rceil \right),$$9

A commonly proposed two-stage procedure first tests homogeneity of the stratum-specific log-odds ratios and then either pools or reports separate confidence intervals. For large sample sizes, Monte Carlo and large-sample analysis show that this preliminary test has a very harmful effect on coverage probabilities: in the real-data scenario analyzed, the minimum simultaneous coverage probability dropped to about $q=1$0 in simulation and $q=1$1 under large-sample calculation, despite a nominal $q=1$2 target (Kabaila et al., 2011).

In response-surface methodology, a multi-stratum design arises from restrictions on randomization produced by hard-to-set factors. The unit structure may involve nesting, crossing, or both, and the paper develops a general stratum-by-stratum design construction for crossed and nested multi-stratum structures. The model

$q=1$3

or, for quantitative factors, $q=1$4, is paired with compound optimality criteria and projector matrices tailored to blocked or row-by-column structures. Good designs are reported even for large experiments in complex structures (Trinca et al., 2024).

In causal survival analysis with competing risk of death, “principal stratum” denotes the counterfactual subgroup of always survivors at time $q=1$5,

$q=1$6

The principal stratum hazard is the instantaneous risk of the non-fatal event within this subgroup, and the proportional principal stratum hazards model

$q=1$7

defines the principal stratum hazard ratio as a direct treatment effect on the underlying non-fatal event process. Principal stratum membership is identified probabilistically through a shared frailty model, and simulation studies verify the reliability of the estimators (Sun et al., 13 Mar 2025).

5. Physical strata and layered biological transport

In biophysics and skin optics, stratum can denote a literal anatomical layer. The stratum corneum is the topmost layer of the epidermis, composed of dead skin cells and characterized by low water content. Because its water content is about $q=1$8, compared with about $q=1$9 in underlying epidermal tissues, a refractive-index contrast arises via a Gladstone–Dale mixture law,

$B$0

with $B$1. Analytical single-layer and two-layer models predict light confinement in the stratum corneum: for biologically relevant conditions, the light intensity in the stratum corneum is $B$2–$B$3 higher than in the immediately underlying tissue layer, and the effect is most prominent for smaller diffuse reflectance of the underlying tissue (Saiko, 2023).

The same layer is treated as the primary permeation barrier in molecular transport studies. Molecular-dynamics simulations of acetone, 6-methyl-5-hepten-2-one, and water through a model stratum corneum lipid membrane determine position-dependent diffusivities using VACF- and PACF-based estimators and combine them with free-energy profiles through the inhomogeneous solubility–diffusion model,

$B$4

The propagator analysis shows that the VACF approach yields an upper bound and the PACF approach a lower bound for the true diffusivity, producing permeability estimates that differ by about one order of magnitude. The main mechanistic conclusion is that permeation is governed primarily by energetic barriers rather than by molecular mobility (Thomas et al., 16 Oct 2025).

These two studies use “stratum” in the most literal layered sense, but they retain the same analytical role seen in abstract stratification: the outer layer is the locus where transport, reflection, and confinement acquire qualitatively different behavior from the adjacent medium.

6. Architectural and infrastructural strata in networked systems

In 6G architecture, the Network Intelligence Stratum is proposed as a new independent layer designed to integrate AI/ML-powered Network Intelligence instances natively across RAN, Core, Edge, and orchestration domains. Its main components are the Network Intelligence Orchestrator, the NIF Manager, and the NIF-C Manager; its building blocks are NIFs, NIF-Cs, and composite NISs organized around an extended N-MAPE-K loop. The design targets end-to-end automation, closed-loop control, lifecycle management, conflict detection and resolution, knowledge sharing, and integration with Kubernetes, Kubeflow, and Eclipse Zenoh, with validation on the Smart Highway Belgian testbed (Soto et al., 2024).

In large-scale agent-centric ML systems, stratum is the name of a system infrastructure that decouples pipeline execution from agentic reasoning and planning. It fuses batches of pipelines into optimized operator DAGs, performs rule-based rewrites and backend selection, and executes across heterogeneous backends including a Rust-based runtime with multithreading and intermediate-result reuse. Preliminary experiments report up to $B$5 speedup over a sequential baseline and $B$6 over naive multiprocessing (Phani et al., 3 Mar 2026).

In blockchain infrastructure, Stratum is the de-facto mining communication protocol used by Bitcoin pools and miners. Because it uses plaintext JSON-RPC over TCP/IP, passive attacks such as StraTap and ISP Log can infer miner earnings, and the active BiteCoin attack can hijack submitted shares and payouts using the WireGhost TCP-hijacking tool. The proposed Bedrock extension introduces a per-miner mining cookie

$B$7

which is embedded in puzzle computations and prevents attackers from reconstructing or hijacking puzzles. Bedrock is reported to impose a daily overhead of $B$8 on a pool server handling traffic from $B$9 miners (Recabarren et al., 2017).

In these engineering uses, “stratum” denotes either an explicit architectural layer or a named protocol/system. A plausible implication is that the term is attractive where decoupling, modularity, and layer-specific optimization are central design goals.

7. Conceptual synthesis

Across the cited literature, stratum has three recurrent functions. First, it is a classification unit: Newton strata, oper strata, principal strata, and inertia strata isolate objects with fixed invariants (Kret, 2012, Hu et al., 2020, Bao et al., 13 Jan 2026). Second, it is an analysis unit: within a stratum, one obtains formulas for cohomology, tractable path combinatorics, local smoothness of nonsmooth maps, or Lyapunov monotonicity (Kret, 2012, Siegel et al., 2010, Bao et al., 13 Jan 2026). Third, it is an operational layer: experimental strata encode randomization restrictions, principal strata encode counterfactual subpopulations, the stratum corneum encodes barrier transport, and network or systems strata encode orchestration boundaries (Trinca et al., 2024, Sun et al., 13 Mar 2025, Saiko, 2023, Soto et al., 2024).

The term therefore spans pure mathematics, statistics, biology, and systems engineering without losing its core semantic content. In each case, a stratum is the piece on which a heterogeneous whole becomes describable by stable local rules, whether those rules are automorphic trace formulas, period inequalities, mixed-model variance structure, transport coefficients, or API-level orchestration.