Temporally-Biased Random Walks

• Temporally-biased random walks are stochastic processes defined on dynamic networks, where time-dependent link availability induces nontrivial biases.
• They quantify key metrics like coverage, mean first-passage time, and steady-state occupation, revealing distinct deviations from static network behaviors.
• Temporal correlations and competing timescales lead to anomalous diffusion, impacting spreading dynamics, search efficiency, and network control strategies.

A temporally-biased random walk is a stochastic process on a network whose topology or link availability evolves over time, such that temporal structure—not merely topology—induces nontrivial biases in the walk's long-term statistical behavior. These biases emerge from the interplay between the time-dependent availability of edges/nodes and the random walker's movement protocol, often resulting in strong deviations from the predictions of classical random walks on static graphs. Temporal bias is especially pronounced when link activation exhibits strong correlations or burstiness, or when multiple timescales interplay (e.g., walker waiting-time, edge up-/down-times). Understanding temporally-biased random walks is essential for modeling spreading, search, and exploration in real-world systems with time-varying interaction networks.

1. Mathematical Formalism and Model Classes

The mathematical structure of temporally-biased random walks depends on the network's temporal rules and the walker's update protocol. The canonical model considers a discrete-time, node-based walker on a temporal network specified by a characteristic function

$\chi(i,j,t)= \begin{cases} 1 &\text{if %%%%0%%%% and %%%%1%%%% are connected at time %%%%2%%%%},\ 0 &\text{otherwise}. \end{cases}$

At each time step, the walker at node $i$ selects uniformly among its instantaneous neighbors $\mathcal V_i(t)=\{j:\chi(i,j,t)=1\}$ if any exist; otherwise, it remains at $i$ (Starnini et al., 2012).

Temporal bias arises in more general settings when:

• Edges have up-times and down-times driven by renewal (often non-Poissonian) processes, yielding time-varying adjacency matrices $[A_{ij}(t)]$;
• The walker's own waiting-time before attempting a step is stochastic and possibly node-dependent;
• Multiple timescales (walker waiting, edge up, edge down) compete, and their ratios control the limiting behavior.

Variants include: (i) activity-driven networks where node activations are heavy-tailed; (ii) “trap models” in which site visitation entails a random holding time, possibly with infinite expectation; (iii) protocols coupling walker state and edge state, leading to non-renewal dynamics in cyclic graphs (Petit et al., 2018, Petit et al., 2019, Betz et al., 2022).

2. Key Observables: Coverage, First-Passage, and Steady State

Performance and statistical bias of temporally-biased walks are characterized through:

• Coverage $C(t)$: the expected number of distinct nodes visited by time $t$;
• Mean First-Passage Time (MFPT): the expected time to reach a given node for the first time;
• Mixing Time: the time to approach stationarity in total-variation distance;
• Steady-State Occupation Probability $\pi_i$: the asymptotic fraction of time spent at each node.

In static networks, coverage and occupation probabilities can often be computed via mean-field (generating function) techniques, yielding e.g., $\pi_i\propto k_i$ for degree $k_i$ or $i$0 for weighted strength $i$1. For temporal networks:

• Direct simulations measure $i$2 and MFPT under realistic contact sequences;
• Analytic treatment is possible for specific renewal-driven temporal models, with the steady-state occupation determined by node-dependent waiting time distributions, e.g.,

$i$3

for power-law inter-event times of exponent $i$4 in activity-driven models (Moinet et al., 2019).

Coverage and MFPT generally exhibit slowdowns relative to their static counterparts and nontrivial scaling laws reflecting burstiness and waiting-time distributions.

3. Temporal Correlations and Emergence of Bias

Temporal correlations—manifested as burstiness (heavy-tailed distributions of contact durations and gap times) and long-range memory—drastically alter random walk dynamics:

• Empirical human contact networks reveal $i$5 and $i$6 with significant consequences for exploration rates (Starnini et al., 2012).
• Sequence replication null models (SRep), which preserve such correlations, cause substantial slowdowns in coverage and increase MFPTs, in contrast to randomized null models (SRan, SStat) which destroy correlations and restore static mean-field scaling.
• In stochastic renewal-edge or activity-driven models, heavy-tailed waiting times (e.g., $i$7 with $i$8) generate anomalous residence-time distributions and can invert classical occupation bias, e.g., suppressing time spent at “hubs” (Moinet et al., 2019, Speidel et al., 2014).

The interplay of walk protocol and memory leads to rich behaviors: on cyclic graphs, even memoryless (exponential) edge and walker clocks can produce non-Markovian path statistics, as the walker's past trajectory retains influence via edge memory (Petit et al., 2018, Petit et al., 2019).

4. Randomization Strategies and Timescale Competition

Careful construction of null models enables isolation of the specific role of temporal patterns in observed bias:

• Sequence Replication (SRep): Preserves all empirical statistics, including correlation structure, leading to maximal temporal bias and exploration slowdown.
• Sequence Randomization (SRan): Destroys all temporal correlations by randomizing time stamps, retaining only static degree/weight structure, thus eliminating temporal bias.
• Statistical Extensions (SStat): Retain marginal distributions (contact weights and durations) but regularize inter-contact gaps, producing intermediate behaviors.

A table comparing preserved features: | Model | Preserved Quantities | |---------|---------------------------------------------| | SRep | $i$9 | | SRan | $\mathcal V_i(t)=\{j:\chi(i,j,t)=1\}$0 | | SStat | $\mathcal V_i(t)=\{j:\chi(i,j,t)=1\}$1 | (Starnini et al., 2012)

Timescale competition governs the bias landscape: when edge up/down and walker clocks are well-separated, familiar static or “fluid” behaviors emerge; otherwise, hybrid regimes with nontrivial bias and memory arise (Petit et al., 2019). Lasting-edge models ultimately interpolate between node-centric (active) and edge-centric (passive) dynamics, with stationary occupation and MFPTs determined by the specifics of the rate hierarchy.

5. Classes of Temporally-Biased Walks and Regimes

The taxonomy of temporally-biased random walks is governed by:

• Reset type (active vs. passive): If the walker resets link-timers upon arrival (active), temporal bias is sensitive to waiting-time distributions; if timers evolve independently (passive), bias can be suppressed, but path correlations arise (Speidel et al., 2014, Petit et al., 2019).
• Inter-event time statistics: For active walks, heavy-tailed timers yield “anti-hub” bias (walkers avoid high-degree nodes), while sub-exponential (Weibull) timers can induce “hub-seeking.” For passive walks, steady-state occupation is uniform for any timer statistics, but trapping effects slow diffusion.
• Edge persistence and cycles: Models with non-instantaneous edge up-times foster a stochastic bias towards long-lived links; in the presence of short cycles, history-dependence (“footprints”) creates non-renewal, non-Markovian trajectories (Petit et al., 2018).
• External controls: Time-dependent biased walks (E-TBRW) enable “boosting” of occupation probabilities and event likelihoods, though optimal control can be computationally hard in directed settings (Haslegrave et al., 2020).

6. Practical Implications and Strategies

Temporal bias is critical for practical algorithms and applications:

• Search and Spreading: Static projections overestimate spreading speed and reachability. Burstiness and temporal inhomogeneity can severely degrade the efficiency of epidemic, information, and search processes.
• Systematic design: For robust search or sampling, protocols should incorporate temporal buffering to mitigate isolation effects, adaptive waiting policies to exploit short bursts, and active rewiring to navigate around persistent inactive gaps (Starnini et al., 2012).
• Biased walks with traps: The ballistic speed in environments with holding/trap times is simply governed by $\mathcal V_i(t)=\{j:\chi(i,j,t)=1\}$2, demonstrating that bias and slowdown are determined by the arithmetic mean of local residence times, and can vanish in heavy-tailed environments (where $\mathcal V_i(t)=\{j:\chi(i,j,t)=1\}$3) (Betz et al., 2022).
• Modeling and inference: Empirical analysis must disentangle the contributions of network topology, temporal scheduling, and protocol to observed bias—neglecting temporal correlations can lead to misleading conclusions regarding accessibility and exploration rates.

7. Anomalous Diffusion and Theoretical Extensions

Temporally-biased random walks frequently exhibit anomalous diffusion:

• Sublinear escape (localization) or superdiffusion (ballistic scaling) depending on the tail of waiting time distributions and system aging (Michelitsch et al., 2022).
• In the limit of diverging mean inter-event times ($\mathcal V_i(t)=\{j:\chi(i,j,t)=1\}$4), stationary occupation becomes uniform—network heterogeneity is effectively erased, even if the activation rates are strongly heterogeneous (Moinet et al., 2019).
• Time-change (subordination) techniques yield new classes of biased walks with rich diffusion regimes (from subdiffusive to superdiffusive), controlled by composing inner and outer renewal processes (Michelitsch et al., 2022).

These extensions highlight the importance of non-Markovian statistics and memory in temporally-biased random walks. Analytical and computational frameworks developed across the cited works serve as the foundation for probing diffusion, sampling, and control in temporally-evolving networked systems.