Sourcewise Approximate Distance Oracle

• The paper introduces a sourcewise approximate distance oracle that efficiently computes fault-tolerant distances from a given source set using landmark sampling and precomputed shortest-path trees.
• It leverages data structures like LCA and SPTs to ensure constant time queries while maintaining a favorable balance between space usage and approximation stretch factors.
• Experimental oracle schemes demonstrate trade-offs, offering stretch factors of 5 and 13, which optimize subquadratic space requirements even under single edge failures.

A sourcewise approximate distance oracle is a compact data structure designed for undirected, positively weighted graphs $G = (V, E)$ with a specified source set $S \subseteq V$, enabling rapid, approximate computation of fault-tolerant distances from sources $s \in S$ to arbitrary targets $v \in V$ following the failure of an edge $f \in E$. The query $Qu(s, v, f)$ requests an approximate distance from $s$ to $v$ in $G-f$. Recent advances have produced oracles with strong trade-offs between query time, space usage, and approximation guarantees, notably achieving $O(1)$ query time and subquadratic space for moderate stretch factors.

1. Formal Problem Definition

Given an undirected, positively weighted graph $G = (V, E)$ and source set $S \subseteq V$, the sourcewise approximate distance oracle problem concerns preprocessing $G$ to answer queries of the form $Qu(s,v,f)$, where $s \in S$, $v \in V$, and $f \in E$. The exact post-fault shortest path distance is denoted $\| s\,v\diamond f\|$. The oracle $\mathcal{D}$ outputs an estimate $d_\mathcal{D}(s,v,f)$ such that

$$\| s\,v\diamond f\| \leq d_\mathcal{D}(s,v,f) \leq \alpha\,\| s\,v\diamond f\|$$

where $\alpha \geq 1$ is the multiplicative stretch. The relevant complexity measures are oracle size (total space) and query time (worst-case time per query). Polylogarithmic factors are typically suppressed in $\widetilde{O}(\cdot)$ notation.

2. Algorithmic Constructs and Data Structures

Modern constructions for sourcewise oracles leverage several universal primitives:

• Unique Shortest Paths: Enforced via random perturbation to avoid path degeneracy and facilitate efficient data structure construction.
• Landmark Sampling: Vertices are sampled independently with probability $p$, producing a landmark set $L$ so that, with high probability, all sufficiently long shortest paths contain a landmark. For parameter $p$, every shortest path of hop-length at least $(3\ln n)/p$ hits $L$.
• Shortest-Path Trees (SPTs): For each root $u \in S \cup L$, an SPT $\mathcal{T}(u)$ is computed, storing $\|uv\|$ and hop-count $|uv|$ for all $v$.
• Lowest Common Ancestor (LCA) Data Structures: Implemented on SPTs for constant-time detection of whether a query edge $f$ lies on the path from root to a given target.

3. Oracle Architectures and Space/stretch Trade-offs

Two primary oracle schemes have been established (Dey et al., 3 Nov 2025):

Oracle	Space	Stretch	Query Time
Oracle 1	$\widetilde{O}(|S|n + n^{3/2})$	$5$	$O(1)$
Oracle 2	$\widetilde{O}(|S|n + n^{4/3})$	$13$	$O(1)$

• Oracle 1 (Stretch 5): Uses one level of landmarks with $p = \frac{3\ln n}{\sqrt{n}}$, constructing exact fault-tolerant $S\times T$ oracles on $G$, SPTs for each $u \in S \cup L$, and storing replacement path distances for edges on $v t_v$ for $t_v \in T$ chosen to minimize $\|vt\|$. Each query involves testing edge membership via LCA and aggregating at most two sub-oracle distances.
• Oracle 2 (Stretch 13): Employs two-level landmarking with landmarks $L_1$ sampled at $3\ln n/n^{1/3}$ and $L_2 \subseteq L_1$ at $1/n^{1/3}$. Targets $t_v$ and $t'_u$ are selected from $T_1 = S \cup L_1$ and $T_2 = S \cup L_2$, leveraging truncated SPTs and multiple replacement-path arrays. Query answering aggregates up to three independently fault-tolerant sub-distances.

4. Query Answering Mechanisms

In both oracles, queries proceed via the following steps:

1. Check if $f$ lies on $sv$ in $\mathcal{T}(s)$ using LCA. If not, return $\|sv\|$.
2. Locate the nearest landmarks or targets ($t_v, t'_u$), determine membership and indices in the corresponding replacement-path arrays, and fetch precomputed distances.
3. For Oracle 1 (two-level), sum at most two component distances ($d_1 + d_2$); for Oracle 2 (three-level), sum three ($d_1 + d_2 + d_3$).
4. Each lookup (array or LCA) executes in $O(1)$ time.

The stretch is established via triangle-type path stitching and path decomposition arguments, ensuring that all computed distances satisfy $d_1 + d_2 \leq 5\,\|s\,v\diamond f\|$ or $d_1 + d_2 + d_3 \leq 13\,\|s\,v\diamond f\|$, respectively.

5. Prior Art and Contextual Comparison

The architecture of these sourcewise oracles is built atop the fault-tolerant $S \times T$ oracle of Bilò et al. (STACS 2018), which constructs an exact distance oracle for all pairs in $S \times T$ under any single edge fault, requiring $\widetilde{O}((|S|+|T|)n)$ space and constant query time. Preprocessing is polynomial, and exact replacement-paths are stored. The sourcewise oracles utilize this as a subroutine alongside additional landmark-driven structures to attain significant space reductions by trading off for small, constant stretches.

Earlier works (e.g., Chechik et al. SODA 2012, Baswana–Khanna 2013, Gupta–Singh ICALP 2018) focus on all-pairs or single-source oracles, often restricting to undirected or unweighted graphs, and do not achieve similar space/stretch combinations for the $S\times V$ scenario.

6. Space/Accuracy Trade-Offs and Query Complexity

Oracle 1 is preferable when $|S| \gg n^{1/2}$ due to its $n^{3/2}$ scaling, whereas Oracle 2 is asymptotically superior for $|S| \ll n^{1/3}$. In both constructions, the use of landmarks serves to bound the maximum hop-length in considered subpaths and minimize space. Both guarantee $O(1)$ query time, enabled by precomputed SPTs, LCA structures, and direct-array access.

A plausible implication is that intermediate stretch and space trade-offs may be achievable by further stratifying the landmark sample or tightening the path decomposition analysis. It is conjectured that, for any integer $k \geq 3$, similar oracles with size $\widetilde{O}(|S|n + n^{1+1/k})$ and stretch $8k-3$ can be realized.

7. Extensions, Open Questions, and Research Directions

Open problems include:

• Extension to dual or multiple-fault sourcewise oracles: current constructions address only single edge failures.
• Generalization from $S \times V$ to arbitrary pair-sets $\mathcal{P} \subseteq V \times V$.
• Application to directed graphs or adaptation to handle vertex faults.
• Further space reduction to nearly linear $|S|n$ for arbitrary stretch constants.
• Path-reporting capabilities, i.e., augmenting oracles to return explicit approximate paths rather than distances alone.

This research direction is closely connected to the study of subquadratic all-pairs fault-tolerant oracles, replacement paths, and graph spanners in fault models. The prevailing methodology—landmark sampling, hierarchical decomposition, and local replacement-path storage—serves as a template for further advancements in fault-tolerant compact routing and sensitivity oracles.

• (Dey et al., 3 Nov 2025) Fault-Tolerant Approximate Distance Oracles with a Source Set
• Bilò et al. STACS 2018 (exact $S \times T$ fault-tolerant oracles)
• Chechik et al. SODA 2012
• Baswana–Khanna 2013
• Gupta–Singh ICALP 2018