---
title: 'Locus in Research: Definitions & Applications'
url: https://www.emergentmind.com/topics/locus
type: topic
---

# Locus in Research: Definitions & Applications

“Locus” denotes several distinct technical objects across contemporary research. In geometry and dynamical systems it denotes a set of points or parameters singled out by a defining condition, such as a conjugate locus, a hyperbolic locus, or a locus configuration; in algebraic geometry it denotes closed or locally closed subvarieties in moduli or parameter spaces; in chromatin mechanics it denotes a fixed-size genomic segment; and in several recent systems papers it appears as an acronymic method name, including LoCUS, LOCUS, and Locus [2210.17273] [2603.22171] [2512.22820] [2310.01095].

## 1. Geometric and analytic meanings

In the theory of hyperplane arrangements, a central hyperplane arrangement in \(\mathbb{C}^2\) with multiplicities is a “locus configuration” when it satisfies a family of locus equations on each hyperplane. Greg Muller showed that the first locus equation is exactly a force-balance equation for a charged trigonometric Calogero–Moser system on \(\mathbb{C}^*\), with charges \(q_i=m_i(m_i+1)\). When the particles lie on \(S^1\subset \mathbb{C}^*\), there is a unique equilibrium up to rotation; for coarsely symmetric multiplicity lists this yields explicit real 2D locus configurations and hence new Schrödinger operators \(L=\Delta+u(x)\) with Baker–Akhiezer functions [1012.5287].

In Poncelet geometry, the locus is the traced set of a triangle center over a 1-parameter family of interscribed triangles. For confocal pairs and for an outer ellipse with an inner concentric circular caustic, the paper shows that many such loci admit the normalized parametrization
\[
F(\lambda)=u\lambda+v\lambda^{-1}+w,\qquad \lambda\in T,
\]
which is an ellipse when the center is a fixed affine combination of \(X_2\), \(X_3\), and a stationary center. In the confocal case, the theory gives explicit criteria for degeneration to a segment, criteria for a circular locus, and the statement that the locus turning number is \(\pm 3\); monotonicity fails only in the degenerate case \(|u|=|v|\) [2106.00715].

In convex 3-manifolds, the conjugate locus of a point \(p\) is the set of points \(\gamma(t)\) conjugate to \(p\) along geodesics \(\gamma(t)=\exp_p(tv)\). The cited work develops a Jacobi-field-based coordinate system on \(S(T_pM)\), detects conjugate points by the vanishing of the signed area \([J_\xi,J_\eta](t)\) swept by a 1-parameter family of orthogonal Jacobi fields, and shows that in 3D the first conjugate locus is organized into two sheets whose singularities are cuspidal edges called ribs. On the quadraxial ellipsoid, each sheet has one closed rib and two partial ribs, and the overlaid sheets share a third closed rib at umbilic directions where \(R_1=R_2\) [2210.17273].

## 2. Parameter-space loci in control and cocycle dynamics

In linear systems on Hilbert spaces, the root locus is the family of closed-loop eigenvalue branches for proportional output feedback. For a SISO system \(\Sigma(A,B,C,D)\), the closed-loop generator is
\[
A_k=A-Bk(1+Dk)^{-1}C,
\]
and the closed-loop eigenvalues satisfy \(1+kG(s)=0\), equivalently \(G(s)=-1/k\). Under the paper’s assumptions—\(A\) generating a \(C_0\)-semigroup, discrete spectrum of isolated eigenvalues of finite algebraic multiplicity, and minimality—each branch is well-defined and continuous, branches are simple and non-self-intersecting, and every branch either converges to a transmission zero or escapes to infinity. For collocated self-adjoint generators, poles and zeros are real and interlace; for collocated skew-adjoint generators, poles and zeros lie on the imaginary axis and the closed-loop spectrum moves into the open left half-plane for \(k>0\) [1409.7081].

For locally constant \(\mathrm{SL}(2,\mathbb{R})\)-valued cocycles, the hyperbolic locus \(\mathcal{H}\subset \mathrm{SL}(2,\mathbb{R})^N\) consists of uniformly hyperbolic \(N\)-tuples. In the projective picture, Avila–Bochi–Yoccoz’s criterion says that \(F=(f_1,\dots,f_N)\in \mathrm{PSL}(2,\mathbb{R})^N\) is uniformly hyperbolic if and only if there exists a finite union \(M\) of open intervals with disjoint closures such that \(f_i(M)\subset M\) for all \(i\). The paper introduces two further parameter loci: the elliptic locus \(\mathcal{E}\), where the generated semigroup contains the identity or an elliptic element, and the semidiscrete and inverse-free locus \(\mathcal{S}\), characterized by the existence of a nontrivial closed subset \(X\subset \mathbb{H}\) mapped strictly inside itself by each generator. It proves that \(\overline{\mathcal{H}\setminus H_p}\subset \mathcal{S}\), that boundary points of non-principal hyperbolic components have finite rank greater than one, and that for \(N\ge 3\) one has \(\mathcal{H}\neq \mathcal{E}^c\), answering negatively the question of whether the hyperbolic locus is the complement of the elliptic locus in general [2009.14511].

## 3. Moduli, exceptional, and bounded-regularity loci in algebraic geometry

In the Hilbert scheme \(\mathrm{Hilb}_{p(t)}^n\), the locus of points corresponding to subschemes of \(\mathbb{P}^n\) with Castelnuovo–Mumford regularity at most \(r'\) is the open subscheme \(\mathrm{Hilb}_{p(t)}^{n,[r']}\). For every \(s\ge r'\), it is realized as a locally closed subscheme of the Grassmannian \(\mathrm{Gr}_{p(s)}^{N(s)}\), where \(N(s)=\binom{n+s}{s}\), and the paper describes it by equations of degree \(\le \deg(p(t))+2\) together with linear inequalities in Plücker coordinates. The construction uses Borel-fixed ideals, marked bases, and \(\mathrm{PGL}(n+1)\)-equivariance, and recovers the full Hilbert scheme when \(r'\) equals the Gotzmann number [1111.2007].

For plane quartics and moduli of curves, “locus” typically refers to a divisor or codimension-two cycle defined by a special ramification or spin condition. Hu determines an explicit Siegel modular form \(F_{77}\) of weight \(11\) for \(\Gamma_3(2)\) whose zero locus on \(A_3(2)\) is the locus where the bitangent corresponding to the odd characteristic \(m_0=(7,7)\) becomes a hyperflex. This yields the divisor class
\[
[HF]=308\lambda-32\delta_0-76\delta_1
\]
in \(\mathrm{Pic}(\overline{\mathcal{M}}_3)\), and the paper shows that the locus of banana curves is contained in the closure of the hyperflex locus [1509.07689]. In low genus, the loci
\[
\mathcal{H}yp_{3,1}\subset M_{3,1},\qquad \mathcal{F}_{3,1}\subset M_{3,1},
\]
of hyperelliptic genus-3 curves with a marked Weierstrass point and non-hyperelliptic genus-3 curves with a marked hyperflex are both codimension two, and their closure classes in \(A^2(\overline{M}_{3,1})\) are computed explicitly. The same paper also computes the class of the genus-4 locus \(\overline{\mathcal{H}_4^+}\) of curves admitting a point \(x\) with \(\mathcal{O}(3x)\) an even theta characteristic [1501.02235].

In the Tannakian theory of Nori motivic local systems, the exceptional locus is defined by
\[
EL(M):=\{\,x\in X(\mathbb{C})\mid G(x^{\dagger}M)\hookrightarrow G^{\odot}(M,x)\text{ is not an isomorphism}\,\}.
\]
Here \(G(M,x)\) is the motivic Galois group of the Tannakian subcategory generated by \(M\), \(G^{Art}(M,x)\) is its Artin quotient, and \(G^{\odot}(M,x)\) is the kernel of \(G(M,x)\twoheadrightarrow G^{Art}(M,x)\). The main theorem is a motivic analogue of Cattani–Deligne–Kaplan: if \(M\) is not Artin, then \(EL(M)\) is a countable union of strict closed algebraic subvarieties. The same geometric description holds for the splitting locus \(WL(M)\) of the motivic weight filtration, and the maximal closed subvarieties are defined over any algebraically closed field of definition for \(X\) and \(M\), with Galois stability under descent [2603.22171].

## 4. Genomic loci and locus-specific chromatin mechanics

In the chromatin-mechanics study, a locus is a fixed-size segment of chromatin represented as a node in a coarse-grained polymer network inferred from contact maps. Two primary resolutions are analyzed: \(25\) kb bins in \(10\) Mb windows for GM12878, \(100\) kb bins for a whole chromosome analysis, and \(250\) bp bins in an mESC Region Capture Micro-C region. The structural ensembles come from the HIPPS-DIMES maximum-entropy framework, which infers effective couplings \(k_{ij}\) and 3D coordinates \(\{r_i\}\); the spectrum of the connectivity matrix \(K\) then governs viscoelastic response [2512.22820].

The global viscoelastic response is Rouse-like, with region-averaged moduli satisfying \(G'(\omega)\propto \omega^{1/2}\) and \(G''(\omega)\propto \omega^{1/2}\) in an intermediate-frequency regime. The locus-specific moduli are obtained by weighting modes by the locus participation \(v_{pi}\):
\[
G'_i(\omega)=\sum_{p\ge 1}v_{pi}^2\frac{(\omega\tau_p)^2}{1+(\omega\tau_p)^2},\qquad
G''_i(\omega)=\sum_{p\ge 1}v_{pi}^2\frac{\omega\tau_p}{1+(\omega\tau_p)^2}.
\]
Using \(\log_{10}\tau_{\max}=1\) as a threshold, the paper separates loci into single-timescale and multi-timescale subpopulations. Multi-timescale loci are strongly enriched in active marks, with H3K27ac at least twofold higher than in the short-\(\tau_{\max}\) group, and \(\tau_{\max}\) follows an approximately inverse trend \(\tau_{\max}\sim \kappa_i^{-1}\) with effective local stiffness \(\kappa_i\) [2512.22820].

The same framework predicts loading-rate-dependent susceptibility under pull–release simulations. For sustained forcing, H3K27ac-rich loci deform more; for a brief, strong impulse with \(F=50\) and \(T=1\), the trend reverses, with reported mean displacements \(\langle\Delta x\rangle\approx 6.0\) for high H3K27ac, \(6.2\) for medium, and \(6.4\) for low. At \(250\) bp resolution, promoters, enhancers, and gene bodies aligned with focal interactions emerge as “viscoelastic islands” with three \(\tan_i(\delta)=1\) crossings and elevated \(\omega_{\max}\), notably in the JUNB/PRDX2 region [2512.22820].

## 5. LoCUS: learning multiscale 3D-consistent features from posed images

“LoCUS” in computer vision stands for “Learning Multiscale 3D-consistent Features from Posed Images.” It is a self-supervised method that learns image-patch descriptors by casting training as a 3D-aware patch retrieval problem over posed views. Given a query landmark embedding \(\theta_j\), the model ranks patch embeddings \(\phi(x_i)\) from other views using cosine similarity
\[
s_{ij}=\frac{\phi(x_i)^\top\theta_j}{\|\phi(x_i)\|\,\|\theta_j\|},
\]
with positives defined by \(\|p_i-\ell_j\|\le \tau_s\) in the same environment and considered negatives restricted to the shell \(\tau_s<\|p_i-\ell_j\|\le \kappa\tau_s\). Patches outside \(\kappa\tau_s\) are ignored. This “don’t-care” outer region is the mechanism used to balance retrieval precision/recall near the landmark with descriptor reuse for semantically similar but spatially distant landmarks [2310.01095].

The optimization target is a smooth differentiable surrogate of Average Precision, used both for descriptor learning and for landmark selection. Scale is explicitly regulated by the spatial tolerance \(\tau_s\): small \(\tau_s\) yields point-/texture-scale invariance, medium \(\tau_s\) yields object-scale invariance, and large \(\tau_s\) yields room/place-scale invariance. The implementation reported in the paper uses a frozen DINO ViT backbone with 768-dimensional features, a two-stage transformer head with 128-dimensional internal features, and two linear layers producing 64-dimensional patch descriptors, for a total of 503,232 trainable parameters; training uses Adam with initial learning rate \(10^{-4}\), temperature \(\tau=0.01\), 20 epochs on Matterport3D, and a mini-batch size of 16 images from the same environment [2310.01095].

On Matterport3D validation patch retrieval, DINO yields \(\alpha\approx 0.20\) and \(AP\approx 0.20\), DINOv2 yields \(\alpha\approx 0.17\) and \(AP\approx 0.17\), and LoCUS yields \(\alpha\approx 0.54\) and \(AP\approx 0.55\). With linear probes on frozen 64-dimensional patch descriptors, the reported overall segmentation results are \(mAP=0.54\), \(mIoU=0.39\), and \(Jac=0.59\), compared with DINO overall \(mAP=0.40\), \(mIoU=0.29\), and \(Jac=0.52\). In relative pose estimation on SparsePlanes-generated Matterport3D pairs, LoCUS reports translation median \(0.92\) m, average \(1.69\) m, \(\le 1\) m accuracy \(0.53\), and rotation median \(22.1^\circ\), average \(34.5^\circ\), \(\le 30^\circ\) accuracy \(0.58\) [2310.01095].

## 6. LOCUS and Locus in lidar odometry and place recognition

“LOCUS” in robotics denotes “Lidar Odometry for Consistent operation in Uncertain Settings,” a lidar-centric front end for real-time odometry and 3D mapping in extreme environments. Its architecture combines motion distortion correction, optional multi-lidar merging, point-cloud filtering, a health-aware sensor integration module, scan-to-scan and scan-to-submap GICP, and an octree map with optional Flat Ground Assumption. The baseline health test is rate \(>1\) Hz, with priority-based switching among VIO, WIO/KIO, IMU, and pure lidar. The reported map update thresholds are \(t=1\) m translation or \(r=30^\circ\) rotation, voxel leaf size is typically \(0.1\) m, and the system is designed to sustain \(10\) Hz lidar in field-deployable settings [2012.14447].

The evaluation reports mean absolute pose error \(0.62\) m on Urban Alpha and \(0.88\) m on Urban Beta, improving to \(0.26\) m and \(0.58\) m with Flat Ground Assumption; on Tunnel Safety Research, mean APE is \(1.67\) m. Map-error RMSE after ICP alignment to the ground-truth map is \(0.29\) m for Urban Alpha, \(0.69\) m for Urban Beta, and \(0.63\) m for Tunnel Safety Research. The system was deployed on multiple robotic platforms and formed a key part of the CoSTAR team’s winning Urban Circuit system in the DARPA Subterranean Challenge [2012.14447].

“Locus” in LiDAR place recognition denotes a global descriptor built from structural appearance, topology, and temporal co-occurrence of scene segments. Starting from Euclidean clusters, the method computes 64-dimensional SegMap-CNN segment features, builds a spatial graph using Minimum Translational Distance between convex hulls, constructs temporal correspondences over a \(k_t=3\) frame window, and aggregates complementary features with second-order outer-product pooling and a Power-Euclidean transform with \(\alpha=0.5\). The final descriptor is a 4096-dimensional, permutation-invariant global vector. On KITTI, the reported mean \(F1_{\max}\) is \(0.942\), with sequence-wise values \(0.983\), \(0.762\), \(0.981\), \(0.992\), \(1.000\), and \(0.931\) on sequences \(00,02,05,06,07,08\), respectively; the method is also reported as robust to viewpoint rotations and severe azimuth-sector occlusions [2011.14497].

## 7. LOCUS in multimodal large language models and the term’s cross-disciplinary structure

In multimodal large language models, “LOCUS” stands for “LOcal visual CUe Search.” The paper identifies a failure mode called visual context rot: decisive evidence may exist in a high-resolution image but fail to be selected and used amid redundant visual context. During training, LOCUS supplies a local crop \(c=\mathrm{Crop}(I,b^*)\) and asks the model to recover its spatial support in the full image. The reward is
\[
r(\hat{y},b^*)=(1-\alpha)r_{\mathrm{loc}}+\alpha r_{\mathrm{format}},\qquad \alpha=0.1,
\]
with \(r_{\mathrm{loc}}=\mathrm{IoU}(\hat{b},b^*)\) for a valid predicted box and \(0\) otherwise. Optimization uses Group-Relative Policy Optimization with KL regularization, leaving inference unchanged: test-time input remains the standard image–question interface without crops, zooms, or tools [2606.16586].

On the primary Qwen2.5-VL-7B run, the paper reports V*Bench \(79.6\to 87.4\), HR-8K \(63.8\to 68.4\), HR-4K \(69.9\to 71.6\), CV-Bench \(75.6\to 76.5\), MME-RealWorld \(58.8\to 62.7\), POPE \(84.9\to 87.6\), HallusionBench \(68.0\to 70.3\), MMStar \(61.5\to 63.1\), RealWorldQA \(61.6\to 66.4\), OCRBench \(82.0\to 85.4\), AI2D \(79.6\to 80.5\), and BabyVision \(11.9\to 12.4\). On the proxy cue-search task, mean IoU rises from \(16.4\) to \(43.0\) and \(ACC@0.5\) from \(16.4\) to \(48.8\); attention-in-box analyses show stronger late-layer concentration on ground-truth evidence regions [2606.16586].

A common source of confusion is to treat “locus” as a single invariant notion. In the cited literature it can mean a traced geometric set, a subset of parameter space defined by stability or ramification conditions, a fixed-size chromatin segment, or an acronym for a computational system. This suggests that the shared semantic core is condition-defined localization, but the object being localized varies sharply: hyperplanes in \(\mathbb{C}^2\), triangle centers, conjugate points, closed-loop spectra, moduli points, chromatin bins, image patches, robot poses, point-cloud places, or fine-grained visual evidence.

Source: https://www.emergentmind.com/topics/locus