---
title: Temporal Graph Realization
url: https://www.emergentmind.com/topics/temporal-graph-realization
type: topic
---

# Temporal Graph Realization

Temporal graph realization denotes a family of constructions, decision problems, and learning procedures in which temporal structure is encoded into a graph so that a specified dynamic property is obtained. In one classical direction, the goal is to build a temporal graph that realizes prescribed pairwise values such as fastest-path durations, earliest-arrival times, shortest temporal lengths, reachability, temporal connectivity, or bounded stretch relative to an underlying static graph [2302.08860]. In another direction, temporal data are transformed into graph objects that retain temporal information, such as time-expanded event graphs, temporal reachability graphs, hierarchical temporal knowledge graphs, or latent operators that reconstruct future graph snapshots [0807.2357].

## 1. Conceptual scope

A common formalization starts from a static graph \(G=(V,E)\) and a time-labeling that determines when edges are available. Realization then asks whether one can choose these labels so that the resulting temporal graph satisfies a prescribed property \(P\), such as an exact distance matrix, a reachability relation, or a global approximation guarantee. In the formulation of simple periodic temporal graph realization, one is given an integer matrix \(D\) and a period \(\Delta\), and asks whether there exists a \(\Delta\)-periodic temporal graph whose pairwise fastest-path durations equal \(D\) [2302.08860].

A broader network-design interpretation appears in periodic transportation settings. There, a fixed infrastructure graph is “realized” as a periodic temporal graph by assigning one time label to each edge so that induced temporal distances remain close to static shortest-path distances. The resulting optimization problem, called **Stretched Temporal Graph Realization (STGR)**, asks for the minimum multiplicative stretch between temporal distance and static distance over all vertex pairs [2504.14258].

A different but compatible usage occurs in data-driven temporal modeling. The 2008 “Temporal Graphs” framework treats realization as a lossless conversion of timestamped events into a graph on which ordinary shortest-path algorithms can be executed while fully retaining temporal information [0807.2357]. More recent learning-based work uses “realization” for latent or hierarchical constructions that encode how a graph evolves, or that make future graph states predictable from present ones [1811.08366].

## 2. Formal models and representations

The dominant combinatorial model is the edge-labeled temporal graph. In the periodic setting, one fixes a period length \(\Delta\in\mathbb{N}\) and a labeling
\[
\lambda : E \to \{1,\dots,\Delta\},
\]
which induces a \(\Delta\)-periodic temporal graph by
\[
L(e) = \{\lambda(e) + i\Delta \mid i\ge 0\}.
\]
A temporal \((s,z)\)-path is a sequence of edge traversals with strictly increasing times, and its duration is
\[
d(P) = t_k - t_0 + 1.
\]
The temporal distance under \(\lambda\) is
\[
\operatorname{dist}_\lambda(u,v) := \min_{P \in \mathcal{P}(u,v)} d(P),
\]
to be compared with the static distance \(\operatorname{dist}_G(u,v)\) [2504.14258].

On trees and tree-like graphs, a useful local quantity is the waiting time or travel delay at an intermediate vertex. For incident edges \(e_1=\{u,v\}\) and \(e_2=\{v,w\}\), the periodic transportation-tree model defines
\[
\tau_v^{u,w} =
\begin{cases}
\lambda(e_2)-\lambda(e_1), & \text{if } \lambda(e_2)>\lambda(e_1),\\
\lambda(e_2)-\lambda(e_1)+\Delta, & \text{otherwise}.
\end{cases}
\]
For a fastest temporal path \(P=(v_0,\dots,v_k)\), the duration decomposes as
\[
d(P)=1+\sum_{i=1}^{k-1} \tau_{v_i}^{v_{i-1},v_{i+1}}.
\]
This decomposition is central in MILP and parameterized algorithms for temporal tree realization [2403.18513].

A different exact representation is the time-expanded graph. Given timestamped events \((u,v,t)\), each entity \(X\) is replaced by time-stamped instances \(X_t\), consecutive appearances of the same entity are linked by directed waiting edges weighted by time differences, and each event becomes a zero-weight edge between the corresponding time-stamped vertices. The resulting graph
\[
G_T = (V_T,E_T,w)
\]
preserves exact event times, exact ordering, and exact waiting times, making temporal shortest-path queries reducible to ordinary graph algorithms [0807.2357].

Reachability-centered models replace metric values by a directed reachability graph. For a temporal graph \(\mathcal{G}=(G,\lambda)\), the reachability graph \(\mathcal{R}(\mathcal{G})=(V,A)\) contains \((u,v)\) exactly when there exists a temporal path from \(u\) to \(v\). Variants depend on whether temporal paths are strict or non-strict, and on whether the temporal graph is undirected or directed, simple, proper, or happy [2503.15771]. A related continuous-time construction is the \((\tau,\delta)\)-reachability graph \(R_\delta\), where \((u,v)\in R_\delta(t)\) iff there exists a journey leaving after \(t\) and arriving within \(t+\delta\), assuming edge traversal time \(\tau\) [1207.7103].

## 3. Objectives and problem families

Temporal graph realization is not a single problem but a class of specifications.

| Objective family | Specification | Representative formulation |
|---|---|---|
| Exact pairwise metrics | Realize a matrix of optimal temporal path values | Fastest, foremost, and shortest temporal path matrices [2510.01702] |
| Upper bounds or distortion | Respect pairwise travel-time bounds or multiplicative stretch | TTR and STGR [2403.18513] |
| Reachability and connectivity | Realize a directed reachability graph or a temporally connected graph | RGR and degree-sequence TC realization [2503.15771] |

Exact metric realization asks for equality \(M(G)=D\), where \(M\) may denote earliest-arrival time, duration of a fastest path, or length of a shortest temporal path. The 2025 metric study states that shortest and earliest-arrival temporal path realization are considered there for the first time, as far as the authors know, and analyzes strict and non-strict, periodic and non-periodic, and label-bounded variants [2510.01702].

Upper-bounded formulations relax equality to inequalities. In periodic upper-bounded temporal tree realization (TTR), the input is a tree \(G\), a period \(\Delta\), and an integer matrix \(D\), and the task is to decide whether there exists a \(\Delta\)-periodic labeling such that the fastest temporal path from every \(v_i\) to \(v_j\) has duration at most \(D_{i,j}\) [2403.18513]. STGR imposes the homogeneous condition
\[
\operatorname{dist}_\lambda(u,v) \le \alpha \cdot \operatorname{dist}_G(u,v)
\]
for all vertex pairs, turning temporal realization into a stretch-bounded scheduling problem [2504.14258].

Connectivity and reachability give rise to additional variants. Reachability Graph Realizability (RGR) asks whether a given directed graph is the reachability graph of some temporal graph, with versions for strict or non-strict temporal paths and for simple, proper, or happy labelings [2503.15771]. A degree-sequence variant asks whether a graphical or multigraphical degree sequence can be realized by a temporally connected graph under simple and proper labeling; the paper gives complete characterizations and linear- or near-linear-time constructive algorithms [2504.17743].

A more local, path-structural realization problem is the **temporal simple path graph** \(tspG_{[\tau_b,\tau_e]}(s,t)\), defined as the subgraph consisting of all temporal simple paths from a source \(s\) to a target \(t\) within a time interval. This realizes not merely reachability but the exact union of all time-respecting simple \(s\)-\(t\) paths [2507.10017].

## 4. Complexity landscape and algorithmic techniques

The algorithmic picture is sharply heterogeneous.

| Problem family | Positive results | Negative results |
|---|---|---|
| Exact fastest durations | Polynomial on trees; FPT by feedback edge number [2302.08860] | NP-hard in general; W[1]-hard by feedback vertex number [2302.08860] |
| Earliest-arrival realization | \(O(n^3\log n)\) for strict, non-strict, periodic, and prescribed-graph variants; ranged version FPT in \(k\) [2510.01702] | NP-hard with one label per edge; ranged/set-valued versions NP-hard [2510.01702] |
| Bounded stretch and bounded trees | Radius algorithm; tree guarantees and MSO-based FPT results [2504.14258] | Strong inapproximability; NP-hard for all fixed \(\Delta\ge 3\) and all fixed \(\alpha\ge 1\) [2504.14258] |
| Transportation trees | TTR is FPT by number of leaves [2403.18513] | TTR is NP-hard even on stars or constant-degree trees [2403.18513] |
| Directed tree realization with slack | Always realizable for odd \(\Delta\le k+1\) or even \(\Delta\le k+2\) [2504.07920] | NP-complete for all remaining \((\Delta,k)\) on bidirected trees [2504.07920] |
| Reachability and temporal connectivity | Tree solid graph polynomial; FPT by feedback edge set; degree-sequence recognition in \(O(n)\) or \(O(n+m)\) [2503.15771] | All undirected RGR variants NP-complete; most directed variants NP-complete [2503.15771] |

For exact fastest-path realization, the foundational 2023 study shows that the problem is NP-hard for all \(\Delta\ge 3\), polynomial-time solvable when the underlying graph is a tree, W[1]-hard when parameterized by feedback vertex number, and FPT when parameterized by feedback edge number [2302.08860]. The FPT algorithm combines a decomposition into segments, guesses of fastest paths between important vertices, and an ILP with \(O(k^2)\) variables, solved via Lenstra-type methods.

For bounded-stretch realization, the principal constructive method is the **radius algorithm**. Given a center \(v_x\), it labels edges between layer \(i-1\) and layer \(i\) by \(\lceil \Delta/2\rceil\) for odd \(i\) and by \(\Delta\) for even \(i\), yielding
\[
\text{stretch}(\lambda) \le \Delta - \frac{\Delta-1}{\min(\operatorname{rad}(G)+1,\operatorname{diam}(G))}.
\]
On trees, the same paper proves a bound of \((\Delta+1)/2\), and for trees of maximum degree at least \(\Delta+1\) this is optimal [2504.14258].

For tree-based transportation realizations with upper bounds, the central positive result is FPT with respect to the number of leaves. The method uses a mixed-integer linear program over travel-delay variables, together with the fact that the fractional part of the constraint matrix is a network matrix and hence totally unimodular [2403.18513]. In the directed case, a root-based labeling gives zero waiting on paths that move only toward or away from the root and bounded waiting on paths that change direction once; this yields a complete threshold classification in terms of the period \(\Delta\) and a minimum slack parameter \(k\) [2504.07920].

Reachability realizability is almost always hard. The 2025 classification proves that all undirected variants are NP-complete, while for directed temporal graphs the only trivial variants are the strict models with arbitrary or simple labeling, where every directed graph is realizable by assigning label \(1\) to each arc [2503.15771]. On the positive side, if the solid graph is a tree, the problem is polynomial-time solvable; if the solid graph has bounded feedback edge set number, the problem is FPT; and this parameter “presumably cannot be replaced by smaller parameters like feedback vertex set or treedepth,” since the problem is W[2]-hard for those [2503.15771].

A complementary positive result arises for temporal connectivity under degree sequences. The 2025 degree-sequence paper gives complete characterizations of the feasible cases and recognition algorithms that run in \(O(n)\) time for graphical degree sequences and \(O(n+m)\) time for multigraphical degree sequences, with constructive output “at essentially no cost” [2504.17743].

## 5. Learning-based and data-driven realizations

In representation learning, temporal graph realization is often reinterpreted as constructing a graph or latent state that encodes temporal evolution rather than merely deciding feasibility.

“Temporal Graph Offset Reconstruction” uses a discrete-time snapshot model
\[
G'=\{G_0,G_1,\ldots,G_t\},
\]
and trains a GCN encoder on snapshot \(G_i\) so that the decoder reconstructs a *future* adjacency \(A_{i+\delta}\) rather than the current adjacency \(A_i\). The deterministic model is
\[
A_{i+\delta} \approx \sigma(Z_i Z_i^\mathsf{T}),
\]
and the variational model replaces \(Z_i\) by a Gaussian latent variable. The paper interprets this as learning a discrete-time temporal evolution operator \(G_i \mapsto G_{i+\delta}\), reports superior performance for future link prediction against non-temporal GAE/GVAE baselines, and states that the method can outperform non-temporal baselines by 38% on a real-world dataset [1811.08366].

“TimeGraphs” realizes temporal dynamics as a hierarchical temporal graph. Starting from frame-wise scene graphs
\[
\{G_t\}_{t=1}^T,\qquad G_t=(\mathcal{V}_t,\mathcal{E}_t),
\]
it builds a level-0 spatio-temporal graph with spatial edges and temporal edges between the same entity across consecutive frames, then constructs higher-level event supernodes using Vertex Infomax Pooling and Graph Cross Networks. Reasoning is performed by an RGCN over the resulting temporal knowledge graph. The method is incremental, supports streaming data, and the paper reports state-of-the-art performance with improvements of up to 12.2% on event prediction and recognition tasks [2401.03134].

In temporal knowledge graph forecasting, “TGL-LLM” treats realization as the construction of historical graph embeddings that, together with an LLM, generate or forecast plausible future edges. The temporal graph learner combines RGCN, GRU, and ConvTransE; hybrid graph tokenization maps historical entity and relation embeddings into the LLM token space; and a two-stage training scheme uses influence-based data pruning plus diversity sampling. The paper reports that the resulting model outperforms a range of state-of-the-art methods on three real-world datasets [2501.11911].

A more explicit subgraph-construction task appears in “Efficient Temporal Simple Path Graph Generation.” There, the realized object is the exact subgraph \(tspG_{[\tau_b,\tau_e]}(s,t)\) containing all temporal simple \(s\)-\(t\) paths in a time window. The method first computes a tight upper-bound graph in polynomial time using polarity times and time-stream common vertices, then applies Escape Edges Verification to construct the exact \(tspG\) without exhaustively enumerating all temporal simple paths [2507.10017].

## 6. Applications, limitations, and open directions

Transportation is one of the main application domains. Periodic temporal graphs model bus, rail, and airline systems in which each link appears once per period and reappears every \(\Delta\) time steps. In this setting, labels encode timetable phases, travel delays encode transfer waiting times, and matrices of upper bounds or stretch factors express service-quality requirements [2504.07920]. Tree-based transportation networks are especially prominent because they model hub-and-spoke or feeder structures, yet even there the upper-bounded realization problem can be NP-hard [2403.18513].

Communication and contact systems form another domain. Time-expanded temporal graphs make it possible to compute temporal proximity, geodesic proximity, and temporal availability exactly from timestamped communication events [0807.2357]. Temporal reachability graphs \(R_\delta\) summarize which ordered pairs can be connected by a journey within maximum delay \(\delta\), and the paper shows that these graphs quantify communication capability, asymmetry, and offloading potential on real and synthetic mobility datasets [1207.7103].

A recurring misconception is to equate temporal graph realization with classical static distance realization. The literature repeatedly shows that the temporal case behaves very differently: exact fastest-path realization is NP-hard in general although static shortest-path realization is polynomial-time solvable, and upper-bounded temporal tree realization can be harder than exact temporal realization on trees [2302.08860]. By contrast, periodic shortest temporal path realization can collapse to the static setting, which suggests that the chosen path metric is itself a primary source of complexity [2510.01702].

Current limitations are model-dependent. Periodic scheduling papers usually assume discrete time, one activation per edge per period, and no node birth or death [2504.14258]. Learning-based future-reconstruction models often assume fixed node sets, no explicit recurrent or continuous-time dynamics, and simple inner-product decoders [1811.08366]. Hierarchical reasoning systems depend on reliable scene graphs or tracking identities across frames, and their higher-level event nodes encode temporal extent only implicitly through constituent nodes [2401.03134]. In LLM-based forecasting, time is typically discretized into snapshot sequences and prediction is evaluated in a constrained candidate-selection setting rather than fully open-ended generation [2501.11911].

Open questions remain broad. The 2025 path-metric study explicitly asks for the smallest per-edge label bound under which Foremost-path TGR remains polynomial-time solvable, whether Shortest-path TGR becomes easier with at most one label per edge or with proper labelings, and whether shortest-path realization is FPT for parameters such as treewidth or vertex cover number [2510.01702]. Earlier representation-learning work suggests extensions to multiple temporal offsets, recurrent architectures, temporal attention, and continuous-time dynamics [1811.08366]. A plausible implication is that temporal graph realization will continue to bifurcate into two interacting strands: exact combinatorial realization under explicit constraints, and learned realizations that encode dynamic structure for prediction, reasoning, or control.

Source: https://www.emergentmind.com/topics/temporal-graph-realization