Time-Consistent Vertex Tracking

• Time-consistent vertex tracking is a framework that maintains smooth evolution of vertex-indexed signals or positions using methods like DGW, NOA, and incremental AVT.
• Dynamic graph wavelets use joint time-graph frames with sparsity constraints to enable stable localization of propagating events across time and network structures.
• Techniques such as normal-only advection and anchored vertex tracking mitigate numerical artifacts by reducing remeshing and preserving network engagement, ensuring physical fidelity.

Time-consistent vertex tracking refers to the set of methodologies designed to follow the evolution of vertex-indexed quantities—positions, signals, or structural roles—across time and/or dynamic domains (such as graphs or triangulated fronts), ensuring that the inferred temporal trajectories exhibit minimal artificial discontinuities and that transitions reflect genuine dynamical or physical processes. Three primary families of methods are foundational: (1) time-vertex tracking via joint time-graph frames and sparsity, (2) geometric interface tracking with tangential drift suppression, and (3) anchored vertex tracking in evolving graphs. Each class employs distinct mathematical and algorithmic mechanisms to enforce time consistency under application-specific constraints.

1. Time-Vertex Signal Tracking and Dynamic Graph Wavelets

In time-vertex tracking for signals over graphs, the objective is to localize and follow events or sources that propagate across a network, where observables are indexed by vertex $n\in\mathcal{V}$ and time $t\in\{0,\ldots,T-1\}$. The signal $X\in\mathbb{R}^{N\times T}$ is modeled as a time-vertex function, and the combined space is treated as the Cartesian product graph $\mathcal{G}\times G_T$.

The joint Laplacian is defined as: $$\mathbf L_{T}\times\mathbf L_G = \mathbf L_T\otimes\mathbf I_N + \mathbf I_T\otimes\mathbf L_G,$$ allowing decomposition into joint graph and time eigenmodes. The joint Fourier transform (JFT) provides spectral coefficients $\widehat X(\ell,k)$ and underpins time-vertex analysis.

Dynamic Graph Wavelets (DGW) instantiate time-consistent atoms by simulating physical PDEs (e.g., heat or wave propagation) from spatiotemporal impulses (e.g., $\Phi_{m,\tau}(n,t) = \delta_{n,m}\delta_{t,\tau}$). The time-vertex atom $W_{m,\tau,s}(n,t)$ evolves according to physical propagation mechanisms parameterized by $s$ (velocity or scale) and damping.

A stable frame $\{W_{m,\tau,s}\}$ is assured when frame bounds

$$A = \min_{\ell,k}\sum_s|\widehat W_s(\lambda_\ell,\omega_k)|^2 > 0,\quad B = \max_{\ell,k}\sum_s|\widehat W_s(\lambda_\ell,\omega_k)|^2 < \infty$$

hold. Analysis and synthesis operators enable bidirectional mapping between the coefficient space $C$ and signal space $X$.

Time consistency is enforced by seeking sparse expansions in the DGW frame: $\min_C \; \|S_W(C) - Y\|_2^2 + \gamma \|C\|_1,$ where each nonzero coefficient encodes a spatiotemporally coherent wave, ensuring that reconstructed events are temporally smooth and physically grounded, thus avoiding spurious, isolated fittings.

2. Geometric Interface Tracking: Surface-Normal-Only Advection

In three-dimensional front-tracking for interfacial flow or material boundaries, time-consistent vertex tracking addresses the drift and clustering that arise from tangential advection of interface vertices. Classical methods advect triangulated front vertices ${\mathbf x}_i(t)$ by the full local fluid velocity ${\mathbf u}$, decomposed as: $$\mathbf u = (\mathbf u \cdot \mathbf n) \mathbf n + (\mathbf I - \mathbf n\otimes\mathbf n)\mathbf u,$$ with $\mathbf n$ the local surface normal.

The Normal-Only Advection (NOA) scheme projects out tangential components to prevent mesh compaction and clustering, yielding: $$\frac{d\mathbf x_i}{dt} = \mathbf u_{\text{ref}} + [(\mathbf u(\mathbf x_i, t) - \mathbf u_{\text{ref}}) \cdot \mathbf n_i ] \mathbf n_i,$$ where $\mathbf u_{\text{ref}}$ is the volume-averaged enclosed fluid velocity, further suppressing drift. Only normal motion is retained, ensuring vertices remain quasi-stationary in their surface parameterization.

Time consistency is quantified by metrics such as remeshing frequency $R_f$, vertex clustering index $\sigma_\ell/\langle\ell\rangle$, and volume conservation error $\epsilon_V$. NOA yields 80–100% fewer remeshing operations, reduces $\epsilon_V$ by an order of magnitude, and maintains smoother curvature statistics compared to classical schemes.

3. Incremental Tracking of Anchored Vertices in Evolving Graphs

Anchored Vertex Tracking (AVT) in dynamic social or interaction networks seeks to follow a small set of critical (anchored) vertices $S_t$ over time, maximizing network engagement metrics measured via the size of the anchored $k$-core at each snapshot $G_t = (V, E_t)$: $$S_t = \arg\max_{S\subseteq V, |S|\le l} |\mathcal C_k(S, G_t)|,$$ where anchors remain irrespective of degree, and followers are iteratively pruned by degree.

Since the AVT problem is NP-hard for $k\geq 3$, scalable algorithms rely on a greedy per-snapshot anchor selection or, crucially, on incremental algorithms (IncAVT) that exploit the smoothness of network evolution. IncAVT updates the anchor set by (a) maintaining a bounded K-order under edge set changes, (b) identifying “affected” vertices whose core numbers change, and (c) swapping anchors only among affected or neighboring vertices to maximize followers.

Empirical evaluations indicate that IncAVT achieves $10$–$100\times$ speedup over static greedy baselines, with solutions exhibiting $\lvert S_t \triangle S_{t-1}\rvert$ typically small, providing a temporally stable and interpretable anchor sequence. No degradation in solution quality was observed.

4. Algorithmic Structures and Theoretical Guarantees

In time-vertex tracking with wavelet-based sparsity, optimization is performed via accelerated proximal splitting schemes, e.g., FISTA, employing Lipschitz-adapted step sizes and soft-thresholding. The theoretical stability of recovery derives from frame bounds and standard $\ell_1$ reconstruction results: reconstruction error is bounded proportionally to the measurement noise, and vertex localization error is tightly concentrated when events are well-separated in the parameter grid.

For NOA-based geometric tracking, implementation requires only a modification of the advection step in standard front-tracking codes. The suppression of $u_t$ tangential velocity fundamentally alters the evolution of the mesh, directly translating into empirical and quantitative reductions in mesh artifacts and remeshing frequency.

In the AVT problem, static greedy algorithms yield correct solutions with no proven constant-factor approximation guarantees due to non-submodularity, whereas IncAVT incremental update procedures provide identical solution quality with order-of-magnitude reductions in running time and candidate evaluations.

5. Applications and Experimental Evaluation

Time-consistent vertex tracking frameworks have been applied across diverse domains:

• In seismological networks, DGW-based time-vertex tracking localizes earthquake epicenters with errors $\leq 40$ km in noise-free settings, robust up to SNR = $0$ dB, outperforming direct waveform peak-detection (Grassi et al., 2016).
• For multiphase flows, NOA drastically reduces remeshing cost (by 80–100%), improves volume conservation by an order of magnitude, and achieves smoother geometrical properties for rising bubbles and droplets under shear (Gorges et al., 2023).
• In dynamic social graphs (e.g., Enron, Gnutella, Deezer), IncAVT tracks critical user sets stably and efficiently, achieving a 25$\times$ speedup over static methods, with minimal anchor set changes between snapshots and identical downstream utility (Cai et al., 2021).

Framework	Application Domain	Key Impact
Dynamic Graph Wavelets (DGW)	Time-vertex signal propagation	Accurate, stable event localization
Normal-Only Advection (NOA)	Interface tracking (front-tracking)	Reduced remeshing, volume & shape error
Incremental AVT	Temporal (social) graph tracking	Efficient, stable anchor trajectory

6. Limitations and Directions for Extension

The precision of time-consistent vertex tracking depends on algorithmic parameterization:

• In DGW tracking, localization accuracy is sensitive to velocity grid $s$, damping $\beta$, and regularization $\gamma$; coarse $s$-sampling induces bias, and low $\beta$ may violate frame bounds. Real data may violate model assumptions due to scattering or reflections.
• NOA can induce undulations in high-curvature regions due to instability in normal estimation; hybrid strategies or curvature-based switching may be necessary. Smoothing (e.g., TSUR3D) must be balanced to avoid over-smoothing physical features. Extension to moving contact lines requires careful handling.
• IncAVT's efficiency can come at the cost of missing a global optimum if anchor replacements are restricted to a narrow candidate set. Nevertheless, empirical evidence indicates no loss in follower count for benchmarked datasets.

Ongoing research explores adaptive multiscale decompositions, anisotropic diffusion processes, Bayesian inference for uncertainty quantification in event trajectories, and online algorithms tailored for data streams or fast-changing graph topologies.

7. Synthesis

Time-consistent vertex tracking is a common concern across dynamic signal processing, interface geometry evolution, and dynamic network analysis. Frameworks that directly penalize non-physical, instantaneous changes—by imposing sparsity over spatiotemporally extended propagating atoms, projecting out uninformative tangential mesh motions, or restricting incremental changes to structurally affected vertices—achieve robust, physically interpretable, and computationally efficient solutions. The selection and adaptation of these frameworks to specific domains must account for physical modeling fidelity, computational requirements, sensitivity to hyperparameters, and the structural properties of the underlying dynamic system.