Relative Random Walk Probabilities (RRWP)

• Relative Random Walk Probabilities (RRWP) are probabilistic descriptors that measure the likelihood of path outcomes in graphs, Markov chains, and time series.
• They leverage methods such as Monte Carlo sampling, combinatorial enumeration, and matrix-based estimators to achieve optimal complexity bounds for endpoint and transition estimations.
• RRWP enhances graph neural network performance by providing multi-scale, permutation-equivariant features that capture structural nuances beyond traditional encodings.

Relative Random Walk Probabilities (RRWP) are a family of probabilistic and structural descriptors that quantify the likelihood of specific outcomes or patterns in random walks on graphs, discrete Markov processes, or time series. The RRWP concept encompasses endpoint probabilities for Markov chains and random walks, ordinal probabilities for path patterns, and multi-scale representations for graph-structured data. RRWPs serve as the foundation of recent algorithmic lower and upper bounds for estimating random walk probabilities in large graphs, as well as for structural encodings in graph neural networks and for closed-form combinatorial formulas for permutation patterns in sampled stochastic processes.

1. Markov Random Walk Probabilities and Endpoint Estimation

RRWP originally arises in the context of estimating $\pi(s, t)$, the probability that a discounted or terminated random walk starting at node $s$ in a (possibly directed) graph $G = (V, E)$ halts at target node $t$ under a geometric stopping rule or equivalently as Personalized PageRank (PPR) with teleport probability $\alpha \in (0,1)$. The random walk proceeds from $s$; at each step, at node $u$, it terminates with probability $\alpha$, or proceeds to a uniform random out-neighbor with probability $1-\alpha$ (Bertram et al., 23 Apr 2025).

The formal personalized endpoint probability is

$$\pi(s, t) = \Pr[\text{walk started at } s \text{ halts first at } t].$$

The RRWP problem is then: for given nodes $s$, $t$, parameter $\delta > 0$, and error tolerance $\varepsilon$, efficiently compute an estimate $\widehat{\pi}(s,t)$ such that

$$|\widehat{\pi}(s,t) - \pi(s,t)| \leq \varepsilon \max\{\pi(s,t), \delta\},$$

with high probability.

Both algorithmic approaches (global push, Monte Carlo sampling, and bidirectional estimators) and information-theoretic lower bounds are developed for this task. Complexity results, depending on access to adjacency primitives and thresholds $\delta$, show that the optimal complexity is

$\tilde{\Theta}(\min\{m, 1/\delta\})$

for the worst case ($m$ is the number of edges), and

$$\tilde{\Theta}(\min\{m, \sqrt{d/\delta}, 1/\delta\})$$

for the average case over targets ($d$ is the average out-degree), with further improvement possible only under enhanced query models (Bertram et al., 23 Apr 2025).

2. RRWP in Markov Chain Inference and Reverse-Ergodicity

RRWP also refers to the empirical estimation of transition probabilities in discrete-time Markov chains from observed sample paths. Given a sequence $x(1), x(2), ..., x(T)$, the RRWP matrix $\hat{P}$ is formed by counting transitions: $\hat P_{ij} = \frac{N_{ij}}{\sum_k N_{ik}},$ where $N_{ij}$ is the number of observed $i \to j$ transitions, subject to optional irreducibility correction (e.g., artificially adding a wrap-around transition) (Schulman, 2016).

This empirical RRWP estimator converges to the true transition matrix $P$ as the sample size increases, under standard Markov ergodicity conditions. It reproduces the relative likelihoods for transition pathways and encodes both metastable structure and possible violations of detailed balance (nonzero probability currents), without requiring prior model structure.

3. RRWP for Ordinal Patterns and Symmetric Random Walks

In time series and permutation pattern analysis, RRWP refers to the probability that a given permutation $\pi \in S_{n+1}$ appears as the ordinal pattern of a random walk of length $n$ with i.i.d. symmetric steps. Formally, for steps $X_1, ..., X_n$, constructing the walk $Z_1 = 0$, $Z_{i+1} = Z_i + X_i$, the RRWP is given by

$$\mathrm{RRWP}(\pi; f) = P_f\{\pi \text{ occurs}\} = \int_{D_{\pi}} \prod_{i=1}^n f(x_i) dx_1 ... dx_n,$$

where $D_\pi$ is the domain where the walk yields permutation $\pi$ on $\{Z_1, ..., Z_{n+1}\}$ (Denoncourt, 2019).

Closed-form RRWP formulas exist for key cases:

• Uniform steps: $$\mathrm{RRWP}(\pi;\text{Uniform}) = |[1,w]| / (2^n n!)$$, where $|[1,w]|$ counts the alcoves in an interval of the affine Weyl group $\widetilde{A}_n$.
• Laplace steps: $$\mathrm{RRWP}(\pi;\text{Laplace}) = 1 / (2^n \prod_{j=1}^n \mathrm{lev}(\pi)_j)$$, with $\mathrm{lev}(\pi)_j$ a level-count determined by adjacency in $\pi$.
• Normal steps: $\mathrm{RRWP}(\pi;\text{Gaussian}) = \det M(\pi)$, with $M$ constructed using two-index level-counts based on $\pi$.

A universality result shows that for “almost consecutive” permutations (consecutive positions differ by at most two), the Laplace-type formula holds for all symmetric densities.

4. RRWP as Structural Encodings in Graph Neural Networks

RRWP functions as a structural encoding for nodes and edges in graph neural networks (GNNs), particularly in the GRASS (Graph Attention with Stochastic Structures) architecture (Liao et al., 2024). In this context:

• RRWP captures, for each node pair $(i, j)$ and walk length $h = 1..k$, the probability that a random walker starting at $i$ reaches $j$ in exactly $h$ steps, using the powers of the transition matrix $T = D^{-1}A$: $P_{h,i,j} = (T^h)_{i,j}$ for sparse graphs, leveraging precomputed tensors of these probabilities.
• RRWP-based encodings are injected as feature vectors for both nodes (“self-return” probabilities) and edges (probabilities of multi-step walks) after batch normalization and learned linear transformation.
• Empirical ablation in GRASS demonstrates that omitting RRWP leads to significant degradation in downstream predictive performance; RRWP features outperform Laplacian and degree-based positional encodings, especially for long-range or higher-order dependencies.

The RRWP encoding is permutation-equivariant and robust to minor structural perturbations, making it a powerful tool for encoding multi-scale affinities beyond shortest path or spectral approaches.

5. RRWP in Information-Theoretic and Maximum Entropy Random Walks

Within the framework of the Ruelle–Bowen random walk, RRWP characterizes the uniformity of conditional path probabilities in continuous-time Markov processes on strongly connected graphs (Chen et al., 2018). The Ruelle–Bowen process on $G$ maximizes entropy rate subject to the graph adjacency, with the generator

$$Q^* = \mathrm{diag}(\varphi)^{-1}A\,\mathrm{diag}(\varphi) - \lambda_A I,$$

where $\lambda_A$ is the Perron–Frobenius eigenvalue and $\varphi$ the right eigenvector of $A$.

The embedded discrete skeleton of the process has the property that, for any fixed number of jumps $n$ between endpoints $i_0,i_n$,

$$\Pr(\text{path }i_0 \to \cdots \to i_n \mid i_0, i_n, n) = \frac{1}{\#\{\text{paths of length }n \text{ from } i_0 \text{ to } i_n\}},$$

meaning all such paths are equiprobable. This unconditional uniformity is a canonical instance of RRWP, reflecting a maximum-entropy principle over stochastic trajectories.

6. Computational Methods and Complexity

RRWP computation depends on the chosen domain:

• For Markov chain or terminating walk models, estimation combines sparse matrix operations (matrix powers, pushes), Monte Carlo sampling, bidirectional strategies, and possibly the use of specialized data structures for adjacency queries (Bertram et al., 23 Apr 2025).
• For empirical estimation from time series, simple count-based algorithms on observed transitions suffice for RRWP matrix inference (Schulman, 2016).
• For path pattern probabilities, explicit combinatorial enumeration, polytope volume computation, and algebraic formulas tied to Coxeter group structure arise (Denoncourt, 2019).
• For GNN encodings, RRWP is precomputed by stacking $k$ powers of the normalized adjacency and aggregating the relevant entries into feature tensors, typically at linear computational and storage cost in $k$ and the number of edges (Liao et al., 2024).

Matching upper and lower bounds for RRWP estimation—$\tilde{\Theta}(\min\{m,1/\delta\})$ for the worst case and $\tilde{\Theta}(\min\{m,\sqrt{d/\delta},1/\delta\})$ for the average case—demonstrate the optimality of existing estimators under standard query models. Enhanced model-specific primitives (EDGE-EXIST, SORTED-OUT) further reduce average-case complexity to $$\tilde{\Theta}(\min\{m,\sqrt{d/\delta},(1/\delta)^{2/3}\})$$, with tight information-theoretic lower bounds established by combinatorial swap and injection-packing arguments (Bertram et al., 23 Apr 2025).

7. Applications, Impact, and Open Directions

RRWP underpins a range of applications:

• Personalized PageRank and centrality scores in large graphs,
• Structural feature design for deep architectures (e.g., GNNs),
• Inference of transition structure from trajectory data in the absence of model knowledge,
• Formal enumeration of ordinal and combinatorial patterns in random walks.

No further asymptotic improvement in worst- or average-case estimation complexity is possible under these query models; significant further progress would require shifting to alternative paradigms such as dynamic, streaming, parallel, or distributed processing (Bertram et al., 23 Apr 2025).

Open questions include the detailed dependence of RRWP estimation on the teleport parameter $\alpha$, the role of graph topology (expanders, power-law networks), fine-grained constant factors, and extensions to non-geometric termination distributions (e.g., heat-kernel PageRank). The RRWP concept continues to serve as a bridge between probabilistic inference, combinatorics, and the efficient representation of high-order graph structure in machine learning.