Rank-1 Dominance: Concepts & Applications

• Rank-1 dominance is a phenomenon where rank-one perturbations simplify and control complex, high-dimensional structures across diverse mathematical domains.
• In Donaldson–Thomas theory, higher-rank invariants are explicitly derived as polynomials of rank-one invariants, linking DT and Gromov–Witten frameworks.
• Across Coxeter groups, operator theory, and optimization, rank-one structures enable dimension reduction, spectral isolation, and efficient algorithmic strategies.

Rank-1 dominance refers broadly to a class of phenomena—arising in representation theory, algebraic geometry, operator theory, and optimization—in which rank-1 structures or perturbations exert controlling influence over larger or higher-rank objects. This notion crystallizes as explicit reduction formulae, dimension reduction, or spectral dominance in several distinct yet linked domains. Representative settings include the structure of DT invariants on Calabi–Yau 3-folds, dominance orderings in Coxeter root systems, spectral theory under rank-one minorization, and signal recovery in low-rank fine-tuning.

1. Rank-1 Dominance in Donaldson–Thomas Theory

On a smooth projective Calabi–Yau 3-fold $X$ satisfying the Bayer–Macrì–Toda (BMT) Bogomolov–Gieseker tilt-stability conjecture and the MNOP conjecture, every higher-rank generalized Donaldson–Thomas (DT) invariant is universally and explicitly determined as a polynomial in the rank-one invariants. For any numerical class $v=(r,\beta,n)$, Joyce’s invariant $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$ admits the formula

$$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$

where $k$ runs over all positive integers dividing $(r,\beta,n)$ simultaneously and $C_{r,k}$ is a universal combinatorial coefficient determined by the cohomology ring of $X$, its Chern classes, the polarization, and numerical data $\{ (\beta_i, n_i) \}_{i=1}^\ell$. The factor $1/k^2$ encodes the multiple-cover contribution, reflecting automorphism and stabilizer factors for semistable $v=(r,\beta,n)$0-fold direct sums (Feyzbakhsh et al., 2021). This result shows that rank-one DT invariants, once computed, completely control the theory in higher rank. Under the proven cases of the MNOP conjecture, all DT invariants are thus ultimately governed by Gromov–Witten invariants.

2. Rank-1 Dominance in Coxeter Root Systems

Within the geometric representation of Coxeter groups, dominance relations capture subtle structure in root systems. For a Coxeter group $v=(r,\beta,n)$1 of finite rank and simple root basis $v=(r,\beta,n)$2, the dominance relation between roots $v=(r,\beta,n)$3 is given by

$v=(r,\beta,n)$4

where $v=(r,\beta,n)$5 denote positive and negative roots, respectively. The "rank-1 dominance" scenario corresponds to those roots in the set $v=(r,\beta,n)$6, dominating exactly one other positive root (Fu, 2011). A key structural theorem states that

$v=(r,\beta,n)$7

where $v=(r,\beta,n)$8 is the set of elementary roots (those not dominating any other), and $v=(r,\beta,n)$9 is the reflection in $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$0. In finite Coxeter groups, $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$1; in infinite groups of finite rank $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$2 is finite but nonempty. Explicit upper bounds include $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$3. Thus, the entirety of rank-1 dominance in the dominance hierarchy is controlled by elementary roots and their pairwise reflections. Algorithms for enumerating $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$4 and higher $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$5 rely on this combinatorial reduction.

3. Rank-One Dominant Spectral Poles in Operator Theory

For positive bounded operators $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$6 on a real Banach lattice $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$7 admitting a rank-one Doeblin-type minorization, spectral dominance by a single, simple, positive eigenvalue arises through an explicit scalar analytic function. If

$\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$8

then the spectrum $\operatorname{DT}_X(r,\beta,n)\in\mathbb Q$9 has a unique dominant eigenvalue $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$0 (the spectral radius) given by the unique real solution to $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$1, where

$$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$2

This eigenvalue is algebraically simple, and its spectral projection is a rank-one operator expressible as

$$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$3

Such an approach generalizes the Perron–Frobenius theory and applies even outside compact or trace-class regimes, reducing the spectral analysis to the zero set of a scalar function (Chino et al., 12 Feb 2026).

4. Rank-1 Dominance in Low-Rank Adaptation and Zeroth-Order Optimization

In zeroth-order (ZO) optimization for LoRA-adapted neural network layers, topology-aware queries aligned with rank-1 decompositions eliminate high-rank signal collapse. The LoRA parameterization

$$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$4

admits updates by individually querying each rank-1 "atom" $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$5. Standard two-point ZO with naive scaling $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$6 yields finite-difference signal-to-noise ratio (FD-SNR) scaling as $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$7, leading to vanishing signal for large $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$8. AR1-ZO alternates queries over rank-1 atoms while increasing the scale to $$\operatorname{DT}_X(r,\beta,n) = \sum_{\,k\,|\, (r,\beta,n)} \frac{1}{k^2} \sum_{\substack{(\beta_1,n_1),\dots,(\beta_\ell,n_\ell) \ \sum_i (\beta_i, n_i) = (\frac\beta k, \frac n k)}} C_{r,k}\bigl\{ (\beta_i, n_i) \bigr\}_{i=1}^\ell \prod_{i=1}^\ell \operatorname{DT}_X(1,\beta_i, n_i)$$9, which restores rank-independence of the FD-SNR and bounds the remaining rank effect to the amortized coverage cost $k$0 for all atoms. Thus, the high-dimensional ZO optimization reduces—via rank-1 dominance—to a sequence of rank-1 actions, inheriting both minimality and statistical optimality (Chen et al., 19 May 2026).

5. Universal Principles and Comparative Table

Rank-1 dominance recurs across mathematical and algorithmic contexts through dimension reduction, structural control, and spectral isolation. Despite differences in domain, three canonical features appear:

Domain	Rank-1 Structure	Dominance Manifestation
DT Theory	Rank-1 sheaves	All higher-rank DT invariants explicit in rank-1 counts
Coxeter Groups	Reflections of elementary roots	$k$1 roots determined by $k$2
Positive Operators	Rank-one minorization ($k$3)	Spectral pole isolated, rank-1 projection
LoRA-ZO Optimization	Rank-1 atoms in adapter decomposition	Signal optimality achieved by atomwise query

A plausible implication is that rank-1 dominance signals intrinsic block-decomposability, minimal sufficient control, or dominance relations within a broader high-dimensional structure.

6. Consequences and Applications

In DT theory and enumerative geometry, rank-1 dominance enables practical computation of all higher-rank invariants from base data, yielding integrality and vanishing results, and identifies the determining power of Gromov–Witten invariants via MNOP. In representation theory, it organizes the structure of root systems, providing tight control and termination proofs for combinatorial algorithms. In operator theory, it yields closed-form expressions for principal eigenvalues and projections without determinant computations. In machine learning, the AR1-ZO strategy demonstrates strong empirical performance, closing much of the performance gap to first-order fine-tuning with only two forward passes per optimization step.

7. Structural Insights and Outlook

Rank-1 dominance universally reduces complexity by enabling explicit analytical or algorithmic reduction from high-rank, high-dimensional, or weakly-structured objects to rank-1 units. This organizing principle not only supports efficient computation and theoretical clarity but also exposes the minimal building blocks required for control, spectral analysis, or enumeration within complex systems. Its continued study is likely to deepen connections among algebraic, analytic, and algorithmic frameworks.