Cage: Confinement in Physics, Graphics & ML

• Cage is a multifaceted concept describing an enclosing structure that transiently confines particles or agents in dense liquids, vortex dynamics, and active matter.
• In computer graphics, a cage is a coarse control mesh that drives smooth object deformations using mean value coordinates and efficient geometric processing.
• In nanostructures and machine learning, cages range from hollow carbon shells to acronym-based frameworks (e.g., CAGE for quantization-aware training) that improve model performance.

Searching arXiv for the supplied papers and closely related recent work on “cage” across physics, graphics, and machine learning. In the cited literature, cage denotes an enclosing or constraining structure whose role is to restrict motion, define a transient local environment, or provide a low-dimensional control scaffold. In condensed-matter and fluid dynamics, a cage is the local confining environment formed by neighboring particles or surrounding vortices; in geometric processing it is a coarse enclosing mesh that drives deformation or filtering; in nanostructures it is a hollow carbon shell; and in several machine-learning papers CAGE is an acronym for a named method or framework rather than a literal enclosure (Bernini et al., 2016, Tong et al., 17 Apr 2025, Ahmad et al., 2016, Xia et al., 2024).

1. Cage as a confinement concept in dense and glassy matter

In dense liquids, colloids, hard-sphere systems, and supercooled molecular liquids, the cage is the temporary confinement imposed by nearby neighbors. One paper defines the cage in a dense supercooled molecular liquid as the local first-neighbor environment formed by surrounding monomers, emphasizes that it is not a rigid container, and identifies two characteristic timescales, $t_m \approx 0.175$ and $t^* \approx 1.023$, with the mean-square cage rattling amplitude $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$. Another paper describes the cage in a 2D colloidal liquid as the temporary confinement formed by surrounding particles and associates its onset with a plateau in the mean squared displacement, a shoulder in the self-intermediate scattering function, and increasing dynamical heterogeneity. A hard-sphere study defines the cage geometrically as the free volume accessible to a tagged particle while keeping all of its neighbours fixed, so that “the free volume of a particle is the cage volume.” A supercooled-water study identifies the cage with the first hydrogen-bond shell, defining a caged state when a molecule is H-bonded to four neighboring water molecules and a jumping state when all four of those H-bonds are broken (Bernini et al., 2016, Li et al., 2020, Maiti, 2017, Kikutsuji et al., 2019).

These definitions support distinct but related dynamical pictures. In the colloidal experiment, the onset packing fraction is reported as $\phi_{\rm onset} \approx 0.60$, and the strongest nonlinear response also peaks at $\phi_{\max} \approx 0.60$, with excitations extending over 5–7 particle diameters. In supercooled water, the mean jump time remains nearly constant at $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$, while diffusion is described by

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$

and, because $$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$, approximately by

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$

In hard spheres, the structural relaxation time is related to the average cage volume $v_c$ through

$t^* \approx 1.023$0

with $t^* \approx 1.023$1 at $t^* \approx 1.023$2, which the authors interpret as effectively zero. They explicitly state that this relation between free volume and relaxation “questions the existence of the glass transition in hard sphere systems” (Kikutsuji et al., 2019, Maiti, 2017).

A further refinement is that the “cage effect” need not be a single mechanism. In systems of hard spheres, one paper distinguishes local accommodation of biased displacement distributions by neighboring particles from delayed, non-local collective processes caused by global conservation of the displacement distribution. It represents density and current relaxation through

$t^* \approx 1.023$3

and interprets stretched relaxation as the combined effect of local accommodation and delayed non-local compensation (Megen et al., 2017).

2. Rigidity, uncaging, and the cage state

A mathematically precise cage construction appears in the study of a binary mixture of hard disks. There, a cage is identified with an isostatic, locally rigid subnetwork extracted from a collision network, called a locally rigid framework (LRF). A rigidity matrix is built from the bonds of this LRF; because an isostatic LRF has exactly as many constraints as degrees of freedom, the matrix is square. The criterion for uncaging is the vanishing of the determinant of the rigidity matrix together with a change in its sign. The paper defines the caging parameter

$t^* \approx 1.023$4

where $t^* \approx 1.023$5 is the number of LRFs examined and $t^* \approx 1.023$6 counts sign changes. At packing fraction $t^* \approx 1.023$7, the uncaging time defined by $t^* \approx 1.023$8 for $t^* \approx 1.023$9 is about $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$0 times larger than for $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$1, and the paper connects this timescale gap to two-step relaxation (Sirono, 2011).

A different formalism defines the cage state as the average particle positions while rearrangements are forbidden. One paper introduces the local cage state of particle $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$2 as

$\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$3

using Monte Carlo sampling constrained by the initial cage, approximated through a size-weighted Voronoi-cell condition. It reports that the correlation between $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$4 and dynamic propensity exceeds $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$5, that the cage-state description outperforms both the initial state and the inherent state for intermediate and long times, and that a ridge-regression model built from cage-state descriptors can rival or exceed state-of-the-art machine-learning methods beyond the ballistic regime (Alkemade et al., 2023).

A later study reformulates the cage state as a restricted ensemble average,

$\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$6

and studies its sensitivity to frozen boundaries in a binary hard-sphere mixture. It defines

$\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$7

and a boundary-sensitivity measure

$\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$8

For $\langle u^2\rangle \equiv \langle r^2(t^*)\rangle$9, once the cavity radius exceeds about $\phi_{\rm onset} \approx 0.60$0, the correlation between the cavity cage state and the propensity is almost indistinguishable from the unfrozen system; for $\phi_{\rm onset} \approx 0.60$1, the correlation remains strongly suppressed even for $\phi_{\rm onset} \approx 0.60$2. The paper interprets this as evidence that the cage state becomes increasingly influenced by long-range structural effects and suggests that the CS might be associated with some form of an amorphous growing structural length scale (Alkemade et al., 22 Jul 2025).

3. Cage effects in vortical and active fluids

In vortex dynamics, the word denotes a confinement region generated by a rotating point-vortex crystal. In an inviscid 2D fluid with $\phi_{\rm onset} \approx 0.60$3 identical point vortices on a regular polygon of radius $\phi_{\rm onset} \approx 0.60$4,

$\phi_{\rm onset} \approx 0.60$5

the vortices rotate rigidly as a vortex crystal. In the rotating frame, inertial particles satisfy

$\phi_{\rm onset} \approx 0.60$6

For small $\phi_{\rm onset} \approx 0.60$7, there are $\phi_{\rm onset} \approx 0.60$8 stable equilibria outside the ring, the satellite attracting points. When $\phi_{\rm onset} \approx 0.60$9, the ring center is also a degenerate stable equilibrium. For $\phi_{\max} \approx 0.60$0, the higher-order asymptotic calculation yields

$\phi_{\max} \approx 0.60$1

so particles inside the central recirculation cell drift inward. The paper calls this the cage effect: the surrounding vortices form a rotating boundary that keeps heavy inertial particles confined near the center. It further reports that the outer satellite equilibria disappear above a critical Stokes number $\phi_{\max} \approx 0.60$2, with $\phi_{\max} \approx 0.60$3 for $\phi_{\max} \approx 0.60$4, $\phi_{\max} \approx 0.60$5 for $\phi_{\max} \approx 0.60$6, and $\phi_{\max} \approx 0.60$7 for $\phi_{\max} \approx 0.60$8, whereas the central trapping persists even for larger $\phi_{\max} \approx 0.60$9 (Angilella, 2024).

In dense active matter, the cage is the transient neighbor-induced confinement that determines relaxation. One paper introduces the cage length $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$0, defined as the distance a particle can move before it is blocked by its neighbors, and argues that the key control parameter is the ratio of a short-time active length scale to $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$1. For athermal systems the relevant short-time scale is the persistence length $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$2; for thermal systems it is

$$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$3

The central result is that the dynamics is optimal when

$$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$4

with $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$5 for the studied systems. This gives a unifying interpretation of why increasing persistence time can speed up, slow down, or non-monotonically change glassy relaxation dynamics (Debets et al., 2021).

4. Geometric cages in graphics, Gaussian splatting, and point clouds

In computer graphics and geometric processing, a cage is a coarse enclosing mesh that parameterizes smooth deformation. A deformation cage is defined as a coarse control polyhedron enclosing the object, with enclosed points represented by mean value coordinates (MVC): $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$6 In CAGE-GS, the cage is learned from source and target point clouds through encoders $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$7 and decoders $$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$8, and then used to deform Gaussian centers,

$$\langle \tau_\mathrm{J}\rangle \approx 1\ \text{ps}$$9

Because 3D Gaussian Splatting also depends on anisotropic covariance, the method updates covariance using a Jacobian,

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$0

The paper reports CD $$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$1, DINO $$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$2, and User votes $$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$3, and states that Jacobian sampling reduces computation from 169.7 min to about 7–8 min for chair and from 65.2 min to about 7–8 min for car (Tong et al., 17 Apr 2025).

A different use appears in real-time texture transfer, where an auxiliary cage mesh is repurposed as a geometric reference for filtering Non-Cosmetic Zones (NCZs). The method casts rays

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$4

from target vertices in the direction of the normal of the nearest cage triangle, uses self-intersection tests and cage-intersection tests, partitions the target mesh into connected components

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$5

and evaluates each segment through

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$6

The paper states that spatial queries are optimized using KD-Trees, that runtime scales as

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$7

and that memory use is

$$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$8

It reports ~70 ms on mobile devices for a ~4.8k triangle mesh, specifically 70 ms on an Android Samsung Tablet S6 Lite for a 4,782-triangle lizard head, with approximately ~20 MB total memory (Zhou et al., 23 Jun 2026).

Cage-based point-cloud deformation is also used adversarially. CageAttack initializes a cage as a unit sphere $$D \approx \rho_\mathrm{J}\frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{J}\rangle},$$9, refines it by curvature- and density-aware subdivision with

$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$0

optimizes cage vertices, and propagates deformation through barycentric coordinates,

$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$1

The attack objective is posed in cage space as

$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$2

The paper evaluates on ModelNet40, ShapeNet Part, and ScanObjectNN, using 1024 points, and reports examples such as 68.55 for PointNet $$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$3 DGCNN and 84.54 for DGCNN $$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$4 PointNet++, together with a plausibility user-study preference of 64.0\% (Tang et al., 1 Jul 2025).

5. Carbon cages as hollow nanostructures

In nanomaterials, a cage is a hollow carbon shell rather than a transient dynamical environment. A study of low-density onion-like carbon cages on Cu surfaces describes spheroidal cages with variable shell thickness, radii, and curvature, as well as open and closed multi-shelled cages, cage-inside-cage structures, and cages connected to graphitic layers. These structures are formed on Cu edges and crevices under 0.2–0.8 MeV C$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$5 irradiation at room temperature, with cumulative dose up to

$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$6

and irradiation rates around

$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$7

The paper estimates densities of

$$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$8

far below graphite, and reports cage diameters from 5–30 nm to 20–100 nm, with some structures reaching few hundred nm (Ahmad et al., 2016).

The same work studies in situ transformation under 120 keV TEM electron irradiation. Cages can shrink, smooth, collapse, or coalesce; smaller cages can merge into larger spheroidal structures; and new hollow multiwalled nanotube-like objects with 5–7 shell thickness and heights of 10–40 nm can form in situ. The associated nanoelastic model introduces a shell of radius $$\langle \tau_\mathrm{J}\rangle \ll \langle \tau_\mathrm{C}\rangle$$9, thickness $$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$0, and local protrusion radius $$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$1, with outward lifting force

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$2

protrusion magnitude

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$3

and critical stress

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$4

These expressions are used to explain variable curvature, protrusions, and beam-induced instability (Ahmad et al., 2016).

6. CAGE as acronymic nomenclature in machine learning, robotics, and safety

Several papers use CAGE as an acronym rather than as a literal enclosure. In quantization-aware training, CAGE stands for Curvature-Aware Gradient Estimation. The method augments the straight-through estimator with a correction based on the instantaneous quantization error $$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$5, derived from a multi-objective view of quantized optimization. Its core SGD-style update is

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$6

and the paper introduces the Pareto stationarity condition

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$7

In Llama-style pretraining up to 800M-parameters, the paper states that CAGE recovers over 10\% of the quantization-induced loss increase in the W4A4 regime, with 800M validation perplexity 12.049 for CAGE+HT versus 12.203 for QuEST+HT, against 11.541 in BF16 (Tabesh et al., 21 Oct 2025).

In robotics, CAGE stands for Causal Attention Enables Data-Efficient Generalizable robotic manipulation. The policy uses DINOv2-large, LoRA with rank 16, a causal Perceiver that compresses observation tokens to $$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$8 tokens, and a diffusion-based Attn-UNet action head. It maps stacked RGB observations and proprioception to an action sequence

$$D \approx \frac{\langle r_\mathrm{J}^2\rangle}{\langle \tau_\mathrm{C}\rangle}.$$9

and performs DDIM-style denoising through

$v_c$0

The paper reports that, with as few as 50 demonstrations from a single training environment, CAGE offers an average 42\% increase in task completion rate in similar environments and achieves about 43\% completion rate and 51\% success rate on average in unseen environments where all baselines fail (Xia et al., 2024).

In safety evaluation, CAGE stands for Culturally Adaptive GEneration. The framework adapts English red-teaming benchmarks to new cultures through a three-stage pipeline: seed-prompt collection and taxonomy mapping, slot-based Semantic Mold refinement, and localized instantiation. The mold separates adversarial structure from cultural content by using required and optional slots such as [Act] [Target], [Method/Approach], or [Condition/Context]. The paper’s quality rubric uses a weighted slot-completion score with $v_c$1, and its Korean instantiation, KoRSET, is reported to outperform direct translation baselines in both quality and attack success. The paper argues that direct translation produces a culturally naive benchmark and that the gain comes from modeling localized socio-technical vulnerabilities rather than from surface translation alone (Kim et al., 9 Feb 2026).

7. Conceptual unity and disciplinary divergence

Across these literatures, the common function of a cage is to delimit admissible motion or admissible transformation. In glassy matter, the cage is a transient environment that governs rattling, escape, uncaging, and structural relaxation; in vortex crystals it is a recirculation region bounded by rotating vortices; in graphics it is an enclosing mesh whose vertices induce smooth, global deformation; and in carbon nanostructures it is a hollow shell whose morphology is controlled by accretion, curvature, and irradiation (Bernini et al., 2016, Angilella, 2024, Tong et al., 17 Apr 2025, Ahmad et al., 2016).

The divergences are equally important. Some authors define the cage geometrically through free volume or Voronoi constraints, others dynamically through H-bond rearrangement or collision networks, and others algorithmically through MVC weights, ray casting, or slot-based semantic abstractions. This suggests that cage is less a single object than a family of constrained-state constructions: each version identifies a region, scaffold, or abstract template that preserves local coherence while limiting accessible alternatives. In that sense, the term retains a stable structural meaning even when the underlying system ranges from hard disks and active glasses to 3D Gaussian Splatting and culturally adaptive red-teaming (Maiti, 2017, Kikutsuji et al., 2019, Zhou et al., 23 Jun 2026, Kim et al., 9 Feb 2026).