--- title: State Rank Stratification Overview url: https://www.emergentmind.com/topics/state-rank-stratification type: topic --- # State Rank Stratification Overview State rank stratification denotes a family of constructions in which an ambient set is partitioned into strata indexed by a rank, a rank-derived state, or a rank-dependent reward band. In current literature, the term is used in several non-equivalent but structurally related senses: strategic allocation rules can partition applicants into reward states defined by rank thresholds; state-level policy analysis can induce tiers from aggregated rank scores; algebraic geometry uses rank-labeled strata of secant and moduli spaces; learning theory and efficient sequence models use effective-rank strata of parameter or runtime-state manifolds; and quantum information stratifies density matrices by matrix rank [2109.08240], [2604.00330], [1010.3546], [2502.09540], [2211.11623], [2605.27907]. ## 1. Conceptual scope and recurring formal structure Across these literatures, a recurring structure is present: a rank variable is defined, the underlying space is partitioned into rank-homogeneous subsets, and transitions between subsets are described by thresholds, closure relations, or singular boundaries. The resulting strata may be combinatorial, geometric, statistical, or dynamical. | Domain | Stratified object | Rank/state variable | |---|---|---| | Strategic allocation | Applicants on a rank interval | Reward strata $\psi_k=[c_k,c_{k+1})$ | | Policy evaluation | States or districts | Joint score, UP, or posterior merit | | Algebraic geometry | Secant or ambient varieties | Symmetric rank, border rank, $X$-rank | | Moduli theory | Curves or abelian-variety moduli | $p$-rank or SK-stratum dimension | | Learning dynamics | Function space or recurrent heads | Model rank or effective state rank | | Quantum information | Density-matrix manifold | Matrix rank $r$ | In school-choice preference modeling, rank stratification means allowing the choice process to vary by list position, so that rank-specific parameter vectors $\theta_r$ are regularized across adjacent ranks by a Laplacian penalty; this is explicitly contrasted with rank-homogeneous multinomial logit models [2306.01801]. In ranked set sampling and judgment post-stratification, the sample is partitioned into strata indexed by judgment ranks $R=1,\dots,k$, and inference on the unknown distribution function $F$ exploits the fact that conditional on $R_i=r$, $X_i$ has the distribution of the $r$-th order statistic from a sample of size $k$ [1304.6950]. ## 2. Strategic ranking as state rank stratification The most literal algorithmic use of the notion appears in strategic ranking, where a designer chooses a non-decreasing reward function $\lambda:[0,1]\to[0,1]$ under a capacity constraint $E_{\theta_{\mathrm{post}}}[\lambda(\theta_{\mathrm{post}})]=\rho$, and step-function reward designs partition the rank space into discrete strata or “states” [2109.08240]. For a $K$-level policy, $$ \lambda(\theta)=\sum_{k=0}^{K-1}\ell_k\,1\{\theta\in[c_k,c_{k+1})\}, $$ with $0=c_00, \end{cases} $$ with $\tilde e_{k-1}$ defined by $$ p(\tilde e_{k-1})=p(e_{k-1}(c_k))+\ell_k-\ell_{k-1}. $$ Applicants in band $k$ therefore expend just enough effort to prevent those at the top of band $k-1$ from profitably jumping upward. This produces the phenomenon explicitly described as State Rank Stratification: step thresholds $c_k$ define strata $\psi_k$, equilibrium efforts create sharp margins at thresholds, effort monotonically decreases within a band until it hits the baseline $e_0$, and post-effort score profiles become piecewise flat or monotone. The sharpness of the margins is localized at the $c_k$ thresholds rather than spread smoothly over the population [2109.08240]. The framework also makes explicit the welfare and fairness trade-offs induced by stratification. Applicant welfare is $$ W=\rho-E[p(e)], $$ societal utility is $U_{\mathrm{soc}}=E[v]$, and school utility is $U_{\mathrm{pri}}=E[v\,\lambda(\theta_{\mathrm{post}})]$. Under two-group heterogeneity with environment factors $\psi_A>\psi_B$, the welfare gap $$ G(\theta_{\mathrm{true}})=W^A(\theta_{\mathrm{true}})-W^B(\theta_{\mathrm{true}}) $$ is nonnegative and strictly positive above the disadvantaged threshold under nontrivial cutoffs, while disadvantaged-group access $$ A=E_{\theta_{\mathrm{true}}}\big[\lambda(\theta_{\mathrm{post}}(\theta_{\mathrm{true}},\psi_B))\big] $$ decreases with the cutoff $c$ under convex $f^{-1}$. Randomization softens stratification: among two-level policies, $W(c)$ is non-increasing in $c$, $U_{\mathrm{pri}}(c)$ is non-decreasing in $c$, and $U_{\mathrm{soc}}(c)$ can attain an interior optimum. Pure randomization yields $G(\theta_{\mathrm{true}})\equiv 0$, and lower cutoffs monotonically reduce welfare gap and increase access. ## 3. State-level statistical and policy stratification In applied policy analysis, state rank stratification often refers to the construction of state-level scores, rankings, and tiers from clustered lower-level data. One formulation uses a rank-based information fusion framework derived from the Longitudinal Rank-Sum Test. Counties are pooled across states, transformed outcome-by-outcome into mid-ranks, aggregated within states, optionally averaged across years, and then combined across outcomes into a joint score $$ S_{gi}=\sum_{k=1}^p w_k\,\bar R_{gi}^{(k)}. $$ Group means of these joint scores define an omnibus statistic $$ T=\frac{\bar S^{(1)}-\bar S^{(0)}}{\sqrt{\widehat{\mathrm{Var}(\bar S^{(1)}-\bar S^{(0)})}}}, $$ with asymptotic $N(0,1)$ behavior under cluster-level independence [2604.00330]. The framework then produces complete rankings of states and strata by quantiles, significance, or resampling stability. In the refundable-versus-non-refundable EITC application, 200 independent subsamples yielded consistently positive $T$; with $C=50$, rejection occurred in $95\%$ of repetitions at the $5\%$ level and $100\%$ at the $10\%$ level, while sensitivity over $C=30,40,50$ gave mean $T$ values of $10.4$, $14.0$, and $17.6$, respectively. A distinct state-stratification methodology models district ranks within each state by a discrete generalized beta distribution, $$ f_{(a,b)}(r)=A\frac{(N+1-r)^b}{r^a}, $$ and then quantifies intra-state uncertainty by the entropy $$ S_N(a,b)=-\sum_{r=1}^N f_{(a,b)}(r)\log f_{(a,b)}(r), $$ normalized as $$ UP=100\times \frac{S_N(\hat a,\hat b)}{\log N}. $$ States are treated as first-tier strata, districts are ranked within each state, and the resulting Uncertainty Percentage provides a comparable measure of distributional uniformity across states with different district counts [2102.10308]. In the 2011 Indian data, the highest literacy-rate UP values were reported for Delhi, Kerala, and Uttarakhand, while Chhattisgarh and Odisha had the lowest; for work participation rate, Karnataka, West Bengal, and Meghalaya had the highest UP, and Jammu & Kashmir, Nagaland, and Odisha the lowest. The same study reports that literacy and work-participation rates are distributed independently of the population distributions, even though the numbers of literate and working people correspond linearly to population size. A Bayesian paired-comparison variant orders states or union territories by posterior means of Bradley–Terry merits $\mu_i$, with pairwise counts $X_{ij}\sim\mathrm{Binomial}(K,\pi_{ij})$ and $$ P(i\succ j)=\pi_{ij}=\frac{e^{\mu_i}}{e^{\mu_i}+e^{\mu_j}}, $$ while prior covariance shares information across economically similar units through log per-capita-income differences [2602.18150]. The analysis of NFHS-5 indicators uses full-sample rankings together with PCI-zone stratification into low-, middle-, and high-income subsets, and reports that extreme ranks remain stable under both classical and Bayesian Bradley–Terry approaches, while mid-ranks exhibit minor swaps. A resampling-based robustness formulation is provided by the Stratified Bootstrap Test. There the strata are states, items are variables to be ranked, and the Non-Containment Index $$ \mathrm{NCI}_{g,k}=1-\frac{1}{B}\sum_{b=1}^B \mathbf{1}\!\left[T_{g,k}\subseteq T_{g,k}^{*(b)}\right] $$ measures how often a state’s observed top-$k$ items fail to reappear under bootstrap resampling [2512.15057]. This usage shifts the emphasis from estimating a single latent merit to assessing stability of rank patterns within and across strata. ## 4. Algebraic and tensorial rank stratifications In algebraic geometry, rank stratification is a precise decomposition of a secant variety, a projective ambient space, or a curvilinear locus into subsets indexed by symmetric rank, border rank, or minimal curvilinear label. For secant varieties of Veronese varieties, a curvilinear subscheme $Z$ of degree $t$ carries a partition label $(t_1,\dots,t_s)$ recording the degrees of its connected components, and the associated quasi-stratum $$ \Sigma(t)=\overline{\bigcup_{Z\in \mathrm{Hilb}^c_t(X_{m,d})[t]}\langle Z\rangle} $$ yields a quasi-stratification of the curvilinear locus $\sigma_t(X_{m,d})^\dagger$ [1010.3546]. When $$ 2\le t\le \left\lfloor\frac{d-1}{2}\right\rfloor, $$ the quasi-strata are disjoint, the curvilinear quasi-stratification becomes a true stratification, each $\Sigma(t)$ is irreducible of dimension $$ \dim \Sigma(t)=mt+l(t)-1, $$ and the largest stratum is $\Sigma(2,1,\dots,1)$, identified with the tangential join $\tau(X_{m,d},t)$. For the fourth secant variety of a Veronese variety, the stratification by symmetric rank is completely explicit. If $P\in \sigma_4(X)\setminus \sigma_3(X)$, the possible symmetric ranks depend on $m$ and $d$. For example, when $m\ge 3$ and $d\ge 6$, the possible ranks are $$ 4,\ d-2,\ d,\ d+2,\ 2d-2,\ 2d,\ 3d-2, $$ while for ternary forms with $m=2$ and $d\ge 6$ they are $$ 4,\ d-2,\ d,\ d+2,\ 2d-2 $$ [1005.3465]. These strata are described geometrically by the type of a smoothable Gorenstein scheme of length $4$: four reduced points, schemes supported on a line or conic, two disjoint double points, or a connected curvilinear length-$4$ scheme at one point. A closely related classification holds for degree-$d\ge 9$ homogeneous polynomials of border rank $5$ that depend essentially on at least $5$ variables. Such a polynomial has a unique associated degree-$5$ zero-dimensional scheme $A$, and the symmetric rank depends only on the number of connected components of $A$ [1702.01914]. The only possible ranks are $$ 5,\ d+3,\ 2d+1,\ 3d-1,\ 4d-3. $$ More precisely, $s=1$ connected component gives $4d-3$, $s=2$ gives $3d-1$, $s=3$ gives $2d+1$, $s=4$ gives $d+3$, and $s=5$ gives $5$. The two $s=2$ types, $(3,2)$ and $(4,1)$, and the two $s=3$ types, $(3,1,1)$ and $(2,2,1)$, are geometrically distinct even though they share the same rank. Each irreducible family determined by the degrees $(b_1,\dots,b_s)$ has projective dimension $5m+s-1$. For a linearly normal elliptic curve $X\subset \mathbb{P}^n$, the $X$-rank stratification of $\mathbb{P}^n$ is controlled by the border rank $b_X(P)$. If $$ P\in \sigma_w(X)\setminus \sigma_{w-1}(X),\qquad n\ge 2w+2, $$ then $$ r_X(P)\in \{w,\ n+1-w\}, $$ and both possibilities occur; on the same stratum, the open rank is constant and equals $n+1-w$ [1210.7444]. This yields a two-valued rank decomposition of each sufficiently low secant stratum, with tangential configurations responsible for the complementary rank. ## 5. Moduli-theoretic rank strata In arithmetic and algebro-geometric moduli problems, rank stratification often means stratification by $p$-rank. For genus-$g$ curves admitting a double cover of a fixed elliptic curve $E$ in characteristic $p>2$, the closed $p$-rank strata $$ V_f(\overline{B}_{E,g}) $$ are pure of dimension $$ \dim V_f(\overline{B}_{E,g})=g-2+f-f_E,\qquad f_E\le f\le g-1+f_E, $$ where $f_E\in\{0,1\}$ is the $p$-rank of $E$ [2502.09540]. The strata form a nested sequence $$ V_{f_E}\subset V_{f_E+1}\subset \cdots \subset V_{g-1+f_E}=\overline{B}_{E,g}, $$ and existence is established for all admissible $f$ except the case $(p,g,f)=(3,2,0)$. For the Siegel moduli space with Iwahori level structure, the $p$-rank stratification is refined by the Kottwitz–Rapoport stratification. If $\mathcal{A}_I^{(d)}$ denotes the locus of $p$-rank $d$, then $$ \dim \mathcal{A}_I^{(d)}=\left\lfloor \frac{g^2+d}{2}\right\rfloor, $$ and the closure relations are parity-sensitive: when $g-d$ is even, $$ \overline{\mathcal{A}_I^{(d)}}=\bigcup_{d'\le d}\mathcal{A}_I^{(d')}, $$ whereas when $g-d$ is odd, top-dimensional Kottwitz–Rapoport strata in certain lower $p$-rank loci must be removed from the naive union [1109.5061]. Here rank stratification is not merely a partition by isogeny invariant; it is intertwined with affine Weyl group combinatorics and Bruhat order. A different moduli-theoretic use appears in four-dimensional $\mathcal{N}=2$ SCFTs. The singular locus of a rank-$r$ Coulomb branch carries a canonical special Kähler stratification $$ \mathcal{S}^{(r)}\supset \mathcal{S}^{(r-1)}\supset \cdots \supset \mathcal{S}^{(0)}, $$ whose codimension-$1$ strata are modeled by rank-$1$ elementary slices drawn from the Kodaira and irregular families [2007.00012]. The combinatorial data of these strata are constrained by sum rules relating them to conformal and flavor central charges, so the stratification becomes a classification device rather than a purely local decomposition. ## 6. Dynamical, computational, and information-geometric forms In nonlinear learning theory, rank stratification refers to a decomposition of the model function space by the effective dimension of the tangent function space. For a model $f(\cdot;\theta)$, the model rank at parameter $\theta$ is $$ \mathrm{rank}_{f_\theta}(\theta)=\dim\!\Big(\mathrm{span}\{\partial_{\theta_i}f(\cdot;\theta)\}_{i=1}^M\Big), $$ and the function space decomposes as $$ \mathcal{F}_{f_\theta}=\bigsqcup_r \mathcal{F}_r,\qquad \mathcal{F}_r=\{f^\star:\mathrm{rank}_{f_\theta}(f^\star)=r\} $$ [2211.11623]. The phase-transition theorem states that for analytic models and generic data, linear stability fails when $n<\mathrm{rank}_{f_\theta}(f^\star)$ and holds almost everywhere when $n\ge \mathrm{rank}_{f_\theta}(f^\star)$. This gives a target-dependent sample-complexity threshold, such as $2rd-r^2$ for matrix factorization and $k(d+1)$ for sums of $k$ distinct tanh neurons. In linear attention large language models, State Rank Stratification is an observed runtime bifurcation of attention heads. Each head maintains a recurrent state matrix $S_h(t)$, and its effective rank is measured by $$ \mathrm{Rank}_{\mathrm{eff}}(S(t))=\sum_{i=1}^d \mathbf{1}(\sigma_i>\epsilon \sigma_1),\qquad \epsilon=10^{-4}. $$ Empirically, some heads remain low-rank and oscillate near zero, while others rapidly saturate to a head-specific upper bound [2602.02195]. The state rank satisfies $$ \mathrm{rank}(S(t))\le \min(t,d), $$ and more sharply, $$ \mathrm{rank}(S_h(t))\le \min\{\dim K_h(t),\dim V_h(t)\}\le m_h\le d. $$ Temporal invariance is strong: cosine similarity of headwise nuclear norms is typically above $0.98$ and of headwise effective ranks above $0.97$, while Spearman correlations across widely separated steps exceed $0.90$ in most layers. Ablations indicate that low-rank heads are indispensable for reasoning, whereas high-rank heads are comparatively redundant. The resulting Joint Rank-Norm Pruning score, $$ S_h=\alpha T_h+(1-\alpha)\frac{n_h}{\max_j n_j}, $$ supports a zero-shot pruning strategy that reduces KV-cache memory usage by $38.9\%$ while largely maintaining accuracy. In quantum information, the mixed-state manifold is stratified by matrix rank: $$ S_r=\{\rho\mid \mathrm{rank}(\rho)=r\},\qquad \dim S_r=2Nr-r^2-1. $$ The Bures metric, $$ ds_B^2=\frac12\sum_{i,j}\frac{|\langle i|\,d\rho\,|j\rangle|^2}{\lambda_i+\lambda_j}, $$ is smooth across the pure-state boundary for $N=2$, where the apparent radial divergence in Bloch coordinates is a coordinate artifact and the scalar curvature remains $R=24$ [2605.27907]. For $N\ge 3$, under controlled transverse approaches to a pure state the metric reduces to a cone, $$ ds^2=du^2+u^2 h_{ab}(\theta)\,d\theta^a d\theta^b, $$ with genuine curvature singularities at the tip: a Dirac delta-function curvature for a two-dimensional cone and a power-law divergence $$ R(u)=\frac{R_h-d(d-1)}{u^2} $$ for higher-dimensional cones. Here state rank stratification is literal manifold stratification by density-matrix rank, and the singular geometry at rank-changing points governs geodesic and Lindblad dynamics. Taken together, these usages show that state rank stratification is not a single formalism but a recurrent technical pattern. In each domain, a rank variable induces a hierarchy of strata; the substantive content lies in how the strata are generated, how transitions between them are constrained, and which observables—welfare, utility, entropy, central charges, effective rank, or curvature—remain stable or singular at the boundaries between rank-defined states. Source: https://www.emergentmind.com/topics/state-rank-stratification