Moving Sink Phenomenon: Dynamics & Applications

• The moving sink phenomenon is defined by the dynamic relocation and selective absorption of mass, energy, or information in physical and computational systems.
• In granular media, experiments and simulations show that sink behavior is driven by shaking amplitude, foundation geometry, and frictional interactions.
• Applications in wireless sensor networks and Transformer models demonstrate that controlling sink mobility enhances system stability, energy efficiency, and overall performance.

The moving sink phenomenon encompasses a class of physical and computational behaviors in which a distinguished entity—the sink—exhibits both dynamic relocation and unique absorption properties within a system. Across granular physics, wireless sensor networks, and Transformer-based neural networks, the moving sink term is defined by the interaction between the sink's ability to absorb, collect, or redirect mass, energy, or information, and the evolution of its spatial or representational position due to system dynamics. This phenomenon manifests as either physical sinking and tilting of intruders in fluidized granular media, dynamic mobility of data sinks in sensor architectures, or representational migration of attention sinks in neural models. Each context leverages the concept to address challenges inherent in stability, efficiency, robustness, or lifetime. The following sections articulate the foundational experiments, governing equations, implications, and computational analogs of the moving sink phenomenon.

1. Physical Sink Phenomena in Fluidized Granular Media

In granular physics, the moving sink phenomenon arises when rigid objects (intruders) are exposed to laterally shaken, dry granular beds. Experiments by Alonso-Llanes et al. and Altshuler et al. utilize monodisperse spheres (e.g., polystyrene, d≈140 μm, ρ≈1.05 g/cm³) to create fluidized beds. Intruders—cylindrical objects with either flat or ring-shaped bases—are ballasted to match the bulk granular density for neutral buoyancy in the static case (Sanchez-Colina et al., 2016, Altshuler et al., 2013, Alonso-Llanes et al., 2022).

Upon lateral shaking with controlled frequencies (e.g., f=5 Hz) and amplitudes, creating dimensionless accelerations (Γ=A(2πf)²/g), the system transitions from jammed to fluidized. Critical thresholds (e.g., Γ*≈0.27 for fluidization onset) demarcate regimes of sink behavior:

• Flat-bottomed cylinders: For Γ<Γ*, negligible penetration occurs. As Γ increases (0.3≲Γ≲0.75), penetration depth grows until it plateaus at the cylinder height (H), with tilt angles remaining subcritical (θ<5°).
• Ring-based cylinders: Above Γ≈0.25, both vertical penetration and tilt occur, with tilt saturating (θ≈30° observed in 3D) and final depths less than H due to increased resistance.

Discrete element method (DEM) simulations confirm that foundation geometry, granular packing, and aspect ratio determine the torque and drag responsible for tilting and sinking (Sanchez-Colina et al., 2016). Notably, frictional locking between ring elements and fluidized grains introduces nonzero net torque, leading to sustained tilt. The coupled feedback—tilt increasing drag, further reducing penetration—produces the saturated, shallower embedment typical of the ring regime.

2. Governing Equations and Theoretical Models

The vertical dynamics of sinking intruders in shaken granular beds are governed by Newtonian force balances accounting for weight, buoyancy (hydrostatic and depth-dependent), and viscous-like drag emanating from granular rearrangements. For mass m and cross-sectional area S, the core evolution equation is (Sanchez-Colina et al., 2016, Alonso-Llanes et al., 2022):

$$m\,\frac{d^2 h}{dt^2} + D\,\gamma\,\frac{dh}{dt} + \left(\frac{\rho_{sl}\,S\,g}{h_f}\right) h = m\,g$$

Where D is a characteristic length scale (e.g., diameter), γ a granular viscosity coefficient, ρ_sl the density of the unfluidized ("solid") layer, h_f the fluidization depth (h_f(Γ)=α(Γ−Γ*)), and g gravity.

Neglecting inertial terms in the fast-sink regime yields:

$\frac{dh}{dt} + a\,h = b$

With $a = \frac{\rho_{sl} S g}{D \gamma h_f}$, $b = \frac{m g}{D \gamma}$. Analytic solution for p=0 (constant density profile):

$h(t) = \frac{b}{a}\left[1 - e^{-a t}\right]$

This model provides accurate fits to experimental sinking curves over standard Γ ranges. In extraterrestrial gravity fields, as shown by Altshuler et al. (Altshuler et al., 2013), the final sink depth (z_sink) is independent of g_eff, attributed to gravity-loading of friction ($\kappa = \alpha g_{eff}$), while sink time scales as $g_{eff}^{-1/2}$.

3. Moving Sink Mobility in Wireless Sensor Networks

In wireless sensor networks (WSNs), the moving sink paradigm replaces static sinks and energy-intensive relay clusters with mobile sinks that traverse the sensor field, pausing (sojourning) at geometrically determined locations to collect data directly (Akbar et al., 2013, Jafri et al., 2013). Geometric Sink Movement (GSM) partitions the field into sub-cells, wherein one or more sinks move along fixed, synchronous trajectories—inner and outer concentric square loops—visiting each cell center at regular intervals. Sensors buffer data until the sink arrives, achieving delay-tolerant, clusterless, and energy-balanced operation.

A mixed-integer linear programming (MILP) approach models sink positions, node–sink associations, link rates, energy budgets, and aggregate flows, with network lifetime $T = \sum_m t_m$ maximized subject to per-epoch sojourn and transmission constraints. Buffering delay is bounded by the sojourn interval between sink visits:

$$D_i = \sup_{s \ge 0} \left\{ \inf \{ \tau \ge 0 : \bar \alpha_i(s) \le \beta_i(s + \tau) \} \right\}$$

Empirically, moving sink frameworks achieve marked improvements in stability period, total lifetime, and throughput relative to classical protocols (LEACH, SEP), due to minimized transmit distances and energy load balancing. Multi-chain and multi-head chain architectures (MIEEPB) further optimize energy use through regional chain formation and opportunistic head selection based on residual energy and proximity to the moving sink (Jafri et al., 2013).

4. Moving Sink Phenomenon in Transformer-Based Neural Networks

In Transformer architectures, especially Diffusion LLMs (DLMs), the moving sink phenomenon refers to the dynamic relocation of attention mass onto minimal-information tokens throughout iterative denoising steps (Zhang et al., 27 Jan 2026, Shin et al., 5 Jul 2025). Unlike autoregressive models, where the first token serves as a fixed sink due to causal masking, DLMs lack a persistent structural anchor. Instead, attention sinks—tokens with lowest value-space norm—wander across steps and layers, introducing representational noise and inference instability.

Characteristically, sink tokens:

• Exhibit very low $L_2$ norms in value space ($\bar{r}_j^{(\ell)}$), causing attention heads to redistribute excess mass onto them.
• Shift location unpredictably as masking patterns evolve with diffusion steps.

To stabilize this, an explicit sink token is introduced with a modified attention mask ($M_{ij}$): the sink only attends to itself, but is globally visible for all other tokens. Empirical ablations show that a single token suffices; its effectiveness is robust to both positional and representational content (even forced zeroing of its value vectors retains the benefit) (Zhang et al., 27 Jan 2026). Downstream metrics including zero-shot accuracy and perplexity show consistent improvements across scales and tasks.

Autoregressive transformers also manifest moving sink dynamics in normalized hidden-state space. After the initial sink layer, all non-sink tokens progressively drift toward the fixed sink direction, quantified by cosine similarity. The OrthoRank method exploits this by ranking tokens per layer by their orthogonality to the sink, updating only the most orthogonal subset for improved inference efficiency (Shin et al., 5 Jul 2025).

5. Comparative Dynamics, Parameter Regimes, and Engineering Implications

Across domains, the moving sink phenomenon reflects a competition between driving forces (mass, gravity, attention allocation) and medium or representational resistance (granular friction, buffering delay, information bottleneck):

Domain	Sink Dynamics	Critical Parameter(s)
Granular Physics	Vertical sink, tilt	Γ, base geometry, AR
WSNs	Sink mobility, sojourn	Field partition, trajectory
DLMs/Transformers	Attention sinks, orthogonality	Masking, norm, attention

In granular media, foundation geometry and oscillatory forcing determine regimes of pure sinking versus sinking with tilt, with repercussions for soil-structure interaction during seismic events—flat bases resist tilt, while ring foundations enhance shear-locking and torque. In WSNs, systematic sink mobility maximizes network lifetime and alleviates classical multi-hop bottlenecks. In neural models, stabilizing moving sinks directly enhances parallel generation fidelity and computational sparsity.

Common misconceptions, such as the presumed dependence of sink depth on gravity in granular-media analogs, are dispelled by experiments and DEM simulations showing final depth invariance under varying g (Altshuler et al., 2013). In computational models, sink position and norm, not token semantics, govern the redistribution of attention mass (Zhang et al., 27 Jan 2026, Shin et al., 5 Jul 2025).

6. Extensions and Real-World Applications

The moving sink principle provides actionable insight for structural engineering, planetary robotics, sensor network design, and neural computation:

• Earthquake engineering: Foundation shape must be selected to minimize tilt under lateral soil fluidization (Sanchez-Colina et al., 2016, Alonso-Llanes et al., 2022).
• Planetary exploration: Sink predictions for rovers can be made using gravity-invariant penetration models; "lightening" vehicles for low-g simulations is unnecessary (Altshuler et al., 2013).
• Sensor networks: Regular geometric partitioning and delay-tolerant sink mobility yield practical, nearly optimal trade-offs for environmental and industrial monitoring (Akbar et al., 2013, Jafri et al., 2013).
• LLMs and DLMs: Stable attention sinks facilitate robust parallel inference, with simple remedies (extra sink token, OrthoRank) promoting throughput and performance at high sparsity (Zhang et al., 27 Jan 2026, Shin et al., 5 Jul 2025).

A plausible implication is that future work should further exploit the convergence properties and stable anchors provided by moving sinks across disciplines, as both analytical models and system designs benefit from explicit control of sink mobility and absorption characteristics.