Discrete Time Temporal Graphs (DTTGs)

• Discrete Time Temporal Graphs (DTTGs) are combinatorial structures that model network dynamics through nodes and edges annotated with discrete time labels.
• They offer multiple representations—including snapshot, edge-labeled, and mixed-edge models—that facilitate algorithmic, statistical, and logical analysis.
• Applications span dynamic communication, epidemiology, transportation, and machine learning, while posing complex computational and parameterized challenges.

A discrete time temporal graph (DTTG) is a combinatorial structure comprising a finite set of nodes and edges, where each edge is associated with a discrete set of time-labels indicating the specific times at which the edge is available. DTTGs encode dynamic network connectivity with full temporal resolution and serve as a canonical model for evolving graphs in algorithmic, statistical, and dynamic systems studies. They admit several equivalent formalizations and a wide spectrum of algorithmic, statistical, and logical questions.

1. Formal Definitions and Model Variants

The standard definition is a triple $G=(V,E,\lambda)$ where:

• $V$ is a finite vertex set.
• $E \subseteq \binom{V}{2}$ is the underlying static edge set.
• $\lambda: E \to 2^{\mathbb{Z}}$ assigns to each edge $e$ a finite set of time-labels $\lambda(e)$, denoting the (discrete) moments at which $e$ is present (Spirakis et al., 2013).

Snapshot view: Equivalently, a DTTG can be given as a tuple

$G = (V, E_1, E_2, \dots, E_T)$

where each $E_t \subseteq \binom{V}{2}$ specifies the edge set at time $t$ (Michail, 2015, Bui-Xuan et al., 12 Feb 2026).

Edge-labelled view: Another form uses $(V, E, \lambda: E \to 2^{[T]})$, where $\lambda(e)$ is the set of (integer) time-steps when $e$ exists (Bui-Xuan et al., 12 Feb 2026).

Unifying models: Generalizations further permit directed edges and mixed temporal edges $(u, t_a, v, t_b)$, thus supporting both instantaneous edges and transitions across multiple time steps (Wehmuth et al., 2014).

Special cases:

• Uniform random temporal graphs (UNI-CASE): Each edge is assigned exactly one label, chosen independently and uniformly at random from $[1, a]$; 'normalized' designates $a = n$, the number of vertices (Spirakis et al., 2013).
• RWIG (Random Walkers Induced Graphs): A DTTG generator wherein edge activation is induced by the co-location of independent Markovian random walkers on an underlying graph (Almasan et al., 2024).

2. Temporal Paths, Journeys, and Distance

Temporal path (journey): A temporal path from $u$ to $v$ is a time-ordered sequence

$$(u_0, u_1, t_1), (u_1, u_2, t_2), \dots, (u_{k-1}, u_k, t_k)$$

with $u_0 = u$, $u_k = v$, $\{u_{i-1}, u_i\} \in E$, $t_i \in \lambda(\{u_{i-1}, u_i\})$, and $t_1 < t_2 < \dots < t_k$ (Spirakis et al., 2013, Michail, 2015, Bui-Xuan et al., 12 Feb 2026).

Arrival-time: Defined as $t_k$, the time-label of the final edge in the path.

Temporal distance: For any $(u,v)$, the minimal arrival time over all temporal paths from $u$ to $v$: $$\delta'(u, v) = \min\,\{ \text{arrival-time} \mid \text{path from } u \text{ to } v \}$$ To avoid undefined values, a slow-path cost $n'$ may be imposed: $\delta(u, v) := \min\{\delta'(u, v), n'\}$ (Spirakis et al., 2013, Bok et al., 2024).

Reachability: $u$ can reach $v$ temporally if $\delta(u, v) < \infty$.

3. Key Algorithmic and Combinatorial Results

3.1 Counting Temporal Paths in Random DTTGs

For the uniform random clique $K_n$, where each static edge receives a label uniformly at random from $[1, a]$, the expected number of temporal paths of exact length $k$ is

$$E[\#\{\text{paths of length }k\}] = \frac{n!\,a!}{(n-k-1)!\;k!\;a^k(a-k)!}$$

This formula accounts for the combinatorics of simple paths and the distinct, strictly increasing label assignments (Spirakis et al., 2013).

3.2 Maximum Expected Temporal Distance and Temporal Diameter

• Maximum expected temporal distance:

$$\Delta_{\max}(G) := \max_{u,v \in V} \mathbb{E}[\delta(u, v)]$$

• Temporal diameter:

$$\mathscr{D}(G) := \mathbb{E}[\max_{u,v \in V} \delta(u, v)]$$

Always, $\mathscr{D}(G) \ge \Delta_{\max}(G)$ (Spirakis et al., 2013).

Closed formulas are derived for specific DTTG classes:

• For a star with i.i.d. labels in $[1, a]$,

$$\Delta_{\max}(G_{\rm star}) = \frac{(a-1)(a+1)(a+2)}{6 a^2} + n' \cdot \frac{a+1}{2a}$$

• For the normalized clique,

$$\Delta_{\max}(K_n), \mathscr D(K_n) = O(\sqrt{n} (\log n)^2)$$

(Spirakis et al., 2013).

3.3 Algorithmic Problems on DTTGs

Several algorithmic paradigms are central in DTTG analysis:

• Shortest/foremost journey computation: Given a starting node and time, find earliest arrival at other nodes. Classical BFS/Dijkstra methods can be used on the (possibly weighted) time-unfolded digraph (0807.2357, Michail, 2015).
• Temporal reachability queries: Reduced to reachability in the static time-expanded graph (Michail, 2015).
• Temporal separation: Given temporal start/end vertices $(s, z)$, find a minimal node set deleting which blocks all temporal paths from $s$ to $z$; complexity is NP-hard on both split and permutation class layers, but fixed-parameter algorithms exist using measures of temporal "change" (e.g., number of layer switches, total Kendall–tau distance) (Maack et al., 2021).
• Matching and covering: Temporal maximum matching requires that no vertex is matched twice in any $\Delta$-window; even for underlying paths, the problem is NP-complete (Mertzios et al., 2019).
• Diameter shortcut sets: By static expansion, DTTG diameter can be decreased using shortcut edge sets derived from the corresponding static graph, with construction algorithms and complexity closely paralleling their static counterparts, modulo expansion scaling (Quantmeyer, 2 Apr 2025).

4. Metric, Logical, and Statistical Properties

4.1 Temporal Centrality and Availability Metrics

Temporal centrality and related notions are computed using the time-unfolded (static) representation $G_T = (V_T, E_T, w)$ mapping each time-labeled node to a vertex and adding waiting and instantaneous interaction arcs:

• Temporal proximity $p(X,Y;t_a,t_b)$: Minimum total time-weight of any path starting from $(X, t \geq t_a)$ to $(Y, t' \leq t_b)$.
• Average temporal/geodesic proximity $P(X,Y), G(X,Y)$: Averaged over all start times.
• Temporal availability $V(X,Y)$: The fraction of times from which $Y$ is temporally reachable from $X$ (0807.2357).

4.2 Logical Meta-Theorems on DTTGs

• MSO and FO meta-theorems: By static expansion or sliding-window derivative, DTTGs admit meta-algorithms for monadic second-order and local first-order properties, parameterized by tree-width or twin-width of the expansion/derivative graphs:
  • MSO properties are fixed-parameter tractable with respect to the expansion’s tree-width ($\operatorname{tw}(G_{\to})$) and formula size (Bui-Xuan et al., 12 Feb 2026).
  • Local FO properties are FPT parameterized by locally bounded sliding-window twin-width ($\operatorname{dtww}_\Delta(G)$), independent of the full DTTG lifetime (Bui-Xuan et al., 12 Feb 2026).

4.3 Random Models and DTTG Generators

• The RWIG model produces contact sequences via independent random walks; all edge probability and higher-moment statistics have closed forms in terms of the stationary distribution and transition matrices. This enables analytical comparisons to classical static models (e.g., Erdős-Rényi, Barabási–Albert) (Almasan et al., 2024).

5. Applications, Data Structures, and Model Representation

5.1 Representation Frameworks

Multiple explicit frameworks capture DTTGs:

• Edge-labeled lists: Each edge-time pair $(u,v,t)$ specifies availability; this is most common for algorithmic and data-mining studies (Michail, 2015, Spirakis et al., 2013).
• Snapshot sequences: Storing each $G_t$ for $t = 1, \ldots, T$ directly (Michail, 2015, Wehmuth et al., 2014).
• Quadruple-record (mixed-edge) model: Dynamic edges $(u, t_a, v, t_b)$ accommodate both instantaneous and transit interactions, as well as periodic/cyclic temporal behavior (Wehmuth et al., 2014).
• Succinct compression: Compressed Suffix Array (CSA)-based data structures support fast interval searches, neighbor queries, and more, with space proportional to the number of temporal contacts via contact encoding and compressed indexing (Brisaboa et al., 2018).

5.2 Learning and Statistical Models

• Machine learning architectures: State-space models, GNN+RNN hybrids, and Transformer-based techniques have all been formulated for DTTGs. Recent methods incorporate explicitly the temporal and structural evolution, using sequence models regularized by the graph Laplacian or multi-scale transformers for learning long-range dependencies (Li et al., 2024, Longa et al., 2023, Chen et al., 2024).
• Feature representation: DTTGs provide a canonical temporal backbone for node/edge-level and higher-order statistical tasks, including motif discovery, centrality, epidemic spread modeling, and task-specific prediction (Longa et al., 2023, Almasan et al., 2024).

5.3 Algorithmic Applications

DTTGs are foundational in areas such as:

• Dynamic communication and distributed systems: dynamic connectivity, resilience analysis, temporal message dissemination.
• Computational epidemiology: temporal reachability and contact tracing.
• Transportation and logistics: scheduling, dynamic shortest path, and earliest-arrival routing.
• Model checking and logic: analysis of properties via MSO/FO meta-theorems, temporal variants of classical graph theorems (Bui-Xuan et al., 12 Feb 2026, Austin et al., 2023).

6. Complexity, Limitations, and Theoretical Barriers

• Computational complexity: Many problems that are tractable in static graphs (such as matching, separation, metric dimension) become NP-hard or even APX-hard when extended to DTTGs, even on highly restricted classes (paths, stars, or cliques with limited labeling) (Mertzios et al., 2019, Bok et al., 2024, Maack et al., 2021).
• Parameterization and FPT: Notable fixed-parameter tractable islands exist when parameterizing by temporal "change over time" (number of switching vertices, sum of Kendall–tau distances) or by solution size; kernelization and further reductions are intensely studied (Maack et al., 2021, Mertzios et al., 2019).
• Limits of expressiveness: The addition of time increases the expressive power but also the representational cost, especially for dense or long-lived networks. For sufficiently sparse real-world networks, optimal memory is linear in the number of temporal edges (Wehmuth et al., 2014, Brisaboa et al., 2018).
• Random model limitations: For models such as RWIG, full exact pmfs are computable in closed form but scale combinatorially (Bell number) in the number of walkers, making large-scale enumeration infeasible (Almasan et al., 2024).

7. Open Directions and Research Challenges

Significant challenges and future directions in DTTGs include:

• Standardized benchmarks and DTTG-specific datasets to facilitate evaluation and comparison of algorithms (Longa et al., 2023).
• Developing a temporal Weisfeiler–Lehman theory to characterize expressiveness and distinguishability of temporal structural patterns (Longa et al., 2023).
• Deep DTTG models and long-range dependency learning, overcoming issues of over-smoothing and vanishing gradients in spatio-temporal architectures (Li et al., 2024, Chen et al., 2024).
• Approximation and parameterized algorithms for hard combinatorial problems, especially as the number of time-steps or labels increases (Maack et al., 2021, Mertzios et al., 2019).
• Integration of domain-knowledge and temporal logic, to handle application-specific constraints in areas such as network science, epidemiology, and transport (Bui-Xuan et al., 12 Feb 2026).

---

For comprehensive technical details and specific theorems, see (Spirakis et al., 2013) for statistical properties and random models, (Bui-Xuan et al., 12 Feb 2026) for meta-theoretic and logic results, (Maack et al., 2021) and (Mertzios et al., 2019) for complexity, separation, and matching, (Almasan et al., 2024) for analytic models of temporal contact structures, and (Li et al., 2024, Chen et al., 2024, Longa et al., 2023) for recent developments in representation learning and deep temporal graph methods.