Generic Concentration in Middle Degree

• Generic Concentration in Middle Degree is a phenomenon where mod p cohomology of Shimura varieties and deviations in empirical processes concentrate in a unique middle degree or intermediate scale.
• It employs deep structural tools such as BGG resolutions with theta-linkage in modular representation theory and Talagrand’s generic chaining in statistical models.
• This concentration enables precise classification in arithmetic geometry and yields tight, non-asymptotic risk bounds in statistical learning settings with dependent, heavy-tailed data.

The concept of generic concentration in middle degree refers to precise phenomena in both mod $p$ cohomology of Shimura varieties and concentration inequalities for empirical processes, where—under genericity hypotheses—cohomology or deviations localize or concentrate in a single “middle” degree or intermediate scale. This localization is central in the formulation and proof of geometric and statistical results such as the weight part of Serre's conjecture for automorphic Galois representations (Ortiz, 16 Jan 2026), and in obtaining sharp control of suprema of empirical processes with dependent, heavy-tailed data (Amorino et al., 1 Nov 2025). Both contexts exploit deep structural and probabilistic tools: BGG-like resolutions in modular representation theory and Talagrand's generic chaining in empirical process theory, respectively.

1. Concentration Phenomena in Mod $p$ Cohomology

For a compact unitary Shimura variety $Sh$ of signature $(2,1)$ with $G_{\mathbb{Q}_p} \simeq GL_3 \times \mathbb{G}_m$, and for sufficiently generic $p$-restricted Serre weight $\lambda_0$ in the lowest Weyl alcove, consider its affine reflection $\lambda_1$ in the upper alcove. The main theorem establishes that for generic, non-Eisenstein maximal ideal $\mathfrak m$ of the spherical Hecke algebra (as per Hamann–Lee's definition), the mod $p$ de Rham cohomology with coefficients in $L(\lambda_1)$ on a toroidal compactification $Sh^{tor}$ is concentrated in the unique “middle” degree $j = \dim Sh = 2$: $$H^j_{dR}(Sh^{tor},\underline{L(\lambda_1)})_{\mathfrak m} = 0 \quad \text{for}\ j \neq 2.$$ Thus, $$H^\bullet_{dR}(Sh^{tor},\underline{L(\lambda_1)})_{\mathfrak m}$$ is supported only in degree $2$.

This concentration is structurally reflected by an explicit three-step filtration of $H^2_{dR}$ with graded pieces involving cohomology of automorphic bundles and theta-linkage operators, namely: $$\mathrm{coker}\{H^0(\omega(\lambda_0+\eta)) \xrightarrow{\theta_{\lambda_0 \uparrow \lambda_1}} H^0(\omega(\lambda_1+\eta))\}_{\mathfrak m}, \quad H^1(L_M(\tilde \mu_1)^\vee)_{\mathfrak m}, \quad \ker\{H^2(\omega(\tilde \lambda_1)^\vee) \to H^2(\omega(\tilde \lambda_0)^\vee)\}_{\mathfrak m},$$ where $\eta$ denotes the Hodge character twist and $\theta_{\lambda_0\uparrow\lambda_1}$ is the mod $p$ theta-linkage operator (Ortiz, 16 Jan 2026).

2. Precise Notion of “Genericity” and Hecke Eigenstructures

The term generic is rigorously defined at the level of weights and maximal ideals:

• An $\epsilon$-generic weight $\lambda \in X^*(T)$ satisfies

$$N_\gamma p+\epsilon < \langle\lambda, \gamma^\vee\rangle < (N_\gamma+1)p-\epsilon$$

for each root $\gamma \in \Phi$ and some integer $N_\gamma$, avoiding proximity to affine Weyl hyperplanes.

• A maximal ideal $\mathfrak m \subset \mathbb T$ is non-Eisenstein if the associated residual Galois representation $\bar r_{\mathfrak m}$ is irreducible at $p$, and is generic (in the sense of Hamann–Lee) if it avoids a finite list of lower-dimensional loci in the Emerton–Gee stack.

Under these genericity conditions, mod $p$ de Rham cohomology vanishes outside the middle dimension, mirroring previously established vanishing results for generic étale cohomology (Ortiz, 16 Jan 2026). There is thus a bijection between generic weights with nonzero middle-degree de Rham cohomology and Serre weights in the set $W(\mathfrak m)$.

3. Structural Tools: BGG Resolutions and Theta-Linkage Operators

The proof of generic concentration exploits a mod $p$ version of category $\mathcal{O}$, involving $(U\mathfrak{g}, P)$-modules over $\mathbb F_p$ admitting parabolic actions. For $W$ a representation of $P$, the parabolic Verma module is defined as

$\Ver_P(W) = U\mathfrak{g} \otimes_{U\mathfrak{p}} W.$

A key construction is the generalized mod $p$ BGG resolution: $BGG_{L(\lambda)} = \left(\Ver_P(\wedge^\bullet (\mathfrak{g}/\mathfrak{p}) \otimes L(\lambda))\right)_{\chi_\lambda} \simeq L(\lambda)$ for upper alcove weights $\lambda$ ($C_1$ for $GL_3$, $C_1$, $C_2$ for $GSp_4$), yielding filtrations with graded pieces corresponding to affine-Weyl reflections. The de Rham realization proceeds via an exact functor $$\Psi: D^b(\mathcal{O}_{P,\mathbb F_p}) \to D^b(C_{Sh^{tor}/})$$ which recovers de Rham complexes of automorphic bundles and connects Verma modules to differential operators (theta-linkage).

Through a spectral sequence and vanishing theorems for higher coherent cohomology groups—specifically, $H^i(Sh^{tor}, \omega^{sub}(\lambda))=0$ for $i>0$ and $\epsilon$-generic $p$-restricted $\lambda$—together with the injectivity of the theta linkage map on Hecke eigenspaces, all but the middle cohomology are forced to vanish in each graded BGG piece (Ortiz, 16 Jan 2026).

4. Generic Concentration for Empirical Processes: Middle Scale Interpolation

A parallel phenomenon appears in the theory of empirical processes with dependent data. Let $\{Z_t\}_{t=1}^T$ be a stationary $\beta$-mixing sequence with values in a Polish space, and consider a class of functions $g: \mathcal Z \times \Theta \to \mathbb R$. The supremum

$$\sup_{\theta \in \Theta} \left| \frac{1}{T} \sum_{t=1}^T g(Z_t, \theta) - \mathbb E[g(Z_t, \theta)] \right|$$

is controlled by a two-level concentration inequality under two high-level conditions:

• A sub-Weibull increment condition with respect to a pseudo-metric $d_\Theta$;
• A coupling condition via the $\beta$-mixing coefficients.

The main result is a non-asymptotic concentration bound (Theorem 2.1 in (Amorino et al., 1 Nov 2025)) interpolating between sub-Gaussian (local) and sub-Weibull (global) regimes, with middle-scale “hybrid” behavior governed by Talagrand’s $\gamma$-functionals: $$\gamma_\alpha(\Theta, d_\Theta) = \inf_{\{\mathcal A_k\}} \sup_{\theta \in \Theta} \sum_{k=0}^\infty 2^{k/\alpha} \Delta(A_k(\theta)).$$ The resulting tail inequality combines a $\gamma_2$ term for small distances and a $\gamma_\alpha$ term for large distances, “stitched together” level by level in the generic chaining sum to control the supremum even at intermediate (“middle scale”) distances, ensuring no scale-dependent gaps in deviation control (Amorino et al., 1 Nov 2025).

5. Applications: Shimura Varieties and Statistical Learning

In the arithmetic context, concentration in middle degree for mod $p$ de Rham cohomology underlies a clean classification of Serre weights and enables an explicit match with étale-theoretic constructions, supporting a purely geometric approach to the weight part of Serre’s conjecture for groups of higher rank.

For statistical learning, the generic concentration phenomenon translates into sharp non-asymptotic risk bounds for empirical risk minimization in regression models with dependent, possibly heavy-tailed data. For example, for a single-layer neural network model of the form

$g(Z_t, \theta) = (Y_t - f_\theta(X_t))^2,$

with $f_\theta$ a single-hidden-layer perceptron, the bound recovers classical rates like $\sqrt{\log n / n}$ (for fixed $d$) but now with $n \ll T$ reflecting dependence, and maintains tight deviations throughout the intermediate scales (Amorino et al., 1 Nov 2025).

6. Extensions, Explicit Examples, and Conjectural Frameworks

Explicit resolutions are constructed for $GSp_4$ in the mod $p$ context, yielding multi-step filtrations for weights in the first three $p$-restricted alcoves. While full middle-degree concentration remains unproven for weights in $C_2$, these resolutions correctly encode shadow-weight combinatorics anticipated by Breuil–Mézard cycles.

A major conjecture posits an equivalence between the “de Rham–Serre weights” and usual Serre weights for any Hodge-type Shimura variety and any generic, non-Eisenstein $\mathfrak m$: $W_{dR}(\mathfrak m) = W(\mathfrak m).$ Prismatic arguments suggest that

$$\dim H^i_{dR}(Sh^{tor},L(\lambda)) \geq \dim H^i_{et}(Sh_{\bar{\mathbb Q}_p}, \underline{F(\lambda)}),$$

implying $W(\mathfrak m) \subset W_{dR}(\mathfrak m)$; together with generic concentration, the two descriptions agree for lowest-alcove weights, and the BGG-filtered complexes fully describe upper-alcove (shadow) weights in terms of geometry (Ortiz, 16 Jan 2026).

7. Significance and Broader Impact

The generic concentration in middle degree provides a robust bridge between geometric, representation-theoretic, and probabilistic frameworks in arithmetic geometry and statistical learning. In the cohomological setting, it paves the way for a systematic, purely geometric formulation of Serre’s conjecture in higher rank, leveraging coherent cohomology and mod $p$ representation theory. In empirical process theory, it resolves longstanding obstacles in obtaining non-asymptotic, fine-scale control over deviations for high-dimensional, strongly dependent, and heavy-tailed data, integrating generic chaining and coupling techniques for comprehensive concentration across all scales.

References:

(Amorino et al., 1 Nov 2025, Ortiz, 16 Jan 2026)